Search

The simple act of viewing and grasping an object involves complex sensorimotor control mechanisms that have been shown to vary as a function of multiple object and other task features such as object size, shape, weight, and wrist orientation. However, these features have been mostly studied in isolation. In contrast, given the nonlinearity of motor control, its computations require multiple features to be incorporated concurrently. Therefore, the present study tested the hypothesis that grasp computations integrate multiple task features superadditively in particular when these features are relevant for the same action phase. We asked male and female human participants to reach-to-grasp objects of different shapes and sizes with different wrist orientations. Also, we delayed the movement onset using auditory signals to specify which effector to use. Using electroencephalography and representative dissimilarity analysis to map the time course of cortical activity, we found that grasp computations formed superadditive integrated representations of grasp features during different planning phases of grasping. Shape-by-size representations and size-by-orientation representations occurred before and after effector specification, respectively, and could not be explained by single-feature models. These observations are consistent with the brain performing different preparatory, phase-specific computations; visual object analysis to identify grasp points at abstract visual levels; and downstream sensorimotor preparatory computations for reach-to-grasp trajectories. Our results suggest the brain adheres to the needs of nonlinear motor control for integration. Furthermore, they show that examining the superadditive influence of integrated representations can serve as a novel lens to map the computations underlying sensorimotor control. [ABSTRACT FROM AUTHOR]

Visual Attention Prediction (VAP) is widely applied in GIS research, such as navigation task identification and driver assistance systems. Previous studies commonly took color information to detect the visual saliency of natural scene images. However, these studies rarely considered adaptively feature integration to different geospatial scenes in specific tasks. To better predict visual attention while driving tasks, in this paper, we firstly propose an Adaptive Feature Integration Fully Convolutional Network (AdaFI-FCN) using Scene-Adaptive Weights (SAW) to integrate RGB-D, motion and semantic features. The quantitative comparison results on the DR(eye)VE dataset show that the proposed framework achieved the best accuracy and robustness performance compared with state-of-the-art models (AUC-Judd = 0.971, CC = 0.767, KL = 1.046, SIM = 0.579). In addition, the experimental results of the ablation study demonstrated the positive effect of the SAW method on the prediction robustness in response to scene changes. The proposed model has the potential to benefit adaptive VAP research in universal geospatial scenes, such as AR-aided navigation, indoor navigation, and street-view image reading. [ABSTRACT FROM AUTHOR]

Understanding the structural similarity between genomes is pivotal in classification and phylogenetic analysis. As the number of known genomes rockets, alignment-free methods have gained considerable attention. Among these methods, the natural vector method stands out as it represents sequences as vectors using statistical moments, enabling effective clustering based on families in biological taxonomy. However, determining an optimal metric that combines different elements in natural vectors remains challenging due to the absence of a rigorous theoretical framework for weighting different k-mers and orders. In this study, we address this challenge by transforming the determination of optimal weights into an optimization problem and resolving it through gradient-based techniques. Our experimental results underscore the substantial improvement in classification accuracy achieved by employing these optimal weights, reaching an impressive 92.73% on the testing set, surpassing other alignment-free methods. On one hand, our method offers an outstanding metric for virus classification, and on the other hand, it provides valuable insights into feature integration within alignment-free methods.

Visual Attention Prediction (VAP) is widely applied in GIS research, such as navigation task identification and driver assistance systems. Previous studies commonly took color information to detect the visual saliency of natural scene images. However, these studies rarely considered adaptively feature integration to different geospatial scenes in specific tasks. To better predict visual attention while driving tasks, in this paper, we firstly propose an Adaptive Feature Integration Fully Convolutional Network (AdaFI-FCN) using Scene-Adaptive Weights (SAW) to integrate RGB-D, motion and semantic features. The quantitative comparison results on the DR(eye)VE dataset show that the proposed framework achieved the best accuracy and robustness performance compared with state-of-the-art models (AUC-Judd = 0.971, CC = 0.767, KL = 1.046, SIM = 0.579). In addition, the experimental results of the ablation study demonstrated the positive effect of the SAW method on the prediction robustness in response to scene changes. The proposed model has the potential to benefit adaptive VAP research in universal geospatial scenes, such as AR-aided navigation, indoor navigation, and street-view image reading.

The intention to act influences the computations of various task-relevant features. However, little is known about the time course of these computations. Furthermore, it is commonly held that these computations are governed by conjunctive neural representations of the features. But, support for this view comes from paradigms arbitrarily combining task features and affordances, thus requiring representations in working memory. Therefore, the present study used electroencephalography and a well-rehearsed task with features that afford minimal working memory representations to investigate the temporal evolution of feature representations and their potential integration in the brain. Female and male human participants grasped objects or touched them with a knuckle. Objects had different shapes and were made of heavy or light materials with shape and weight being relevant for grasping, not for “knuckling.” Using multivariate analysis showed that representations of object shape were similar for grasping and knuckling. However, only for grasping did early shape representations reactivate at later phases of grasp planning, suggesting that sensorimotor control signals feed back to the early visual cortex. Grasp-specific representations of material/weight only arose during grasp execution after object contact during the load phase. A trend for integrated representations of shape and material also became graspspecific but only briefly during the movement onset. These results suggest that the brain generates action-specific representations of relevant features as required for the different subcomponents of its action computations. Our results argue against the view that goaldirected actions inevitably join all features of a task into a sustained and unified neural representation. [ABSTRACT FROM AUTHOR]

Many studies have shown that the human visual system has two major functionally distinct cortical visual pathways: a ventral pathway, thought to be important for object recognition, and a dorsal pathway, thought to be important for spatial cognition. According to our and others previous studies, artificial neural networks with two segregated pathways can determine objects' identities and locations more accurately and efficiently than one-pathway artificial neural networks. In addition, we showed that these two segregated artificial cortical visual pathways can each process identity and spatial information of visual objects independently and differently. However, when using such networks to process multiple objects' identities and locations, a binding problem arises because the networks may not associate each object's identity with its location correctly. In a previous study, we constrained the binding problem by training the artificial identity pathway to retain relative location information of objects. This design uses a location map to constrain the binding problem. One limitation of that study was that we only considered two attributes of our objects (identity and location) and only one possible map (location) for binding. However, typically the brain needs to process and bind many attributes of an object, and any of these attributes could be used to constrain the binding problem. In our current study, using visual objects with multiple attributes (identity, luminance, orientation, and location) that need to be recognized, we tried to find the best map (among an identity map, a luminance map, an orientation map, or a location map) to constrain the binding problem. We found that in our experimental simulations, when visual attributes are independent of each other, a location map is always a better choice than the other kinds of maps examined for constraining the binding problem. Our findings agree with previous neurophysiological findings that show that the organization or map in many visual cortical areas is primarily retinotopic or spatial. [ABSTRACT FROM AUTHOR]

Research on the identification of urban functional regions is of great significance for the understanding of urban structure, spatial planning, resource allocation, and promoting sustainable urban development. However, achieving high-precision urban functional region recognition has always been a research challenge in this field. For this purpose, this paper proposes an urban functional region identification method called ASOE (activity–scene–object–economy), which integrates the features from multi-source data to perceive the spatial differentiation of urban human and geographic elements. First, we utilize VGG16 (Visual Geometry Group 16) to extract high-level semantic features from the remote sensing images with 1.2 m spatial resolution. Then, using scraped building footprints, we extract building object features such as area, perimeter, and structural ratios. Socioeconomic features and population activity features are extracted from Point of Interest (POI) and Weibo data, respectively. Finally, integrating the aforementioned features and using the Random Forest method for classification, the identification results of urban functional regions in the main urban area of Xi'an are obtained. After comparing with the actual land use map, our method achieves an identification accuracy of 91.74%, which is higher than other comparative methods, making it effectively identify four typical urban functional regions in the main urban area of Xi'an (e.g., residential regions, industrial regions, commercial regions, and public regions). The research indicates that the method of fusing multi-source data can fully leverage the advantages of big data, achieving high-precision identification of urban functional regions. [ABSTRACT FROM AUTHOR]

Abstract Due to the ignoring of rich spatio-temporal and global contextual information with convolutional neural networks in features extraction, the traditional method is prone to tracking drift or even failure in complex scenario, especially for the tiny targets in aerial photography scenario. In this work, it proposes a transformer feature integration network (TFITrack) to obtain diverse and comprehensive target feature for the robust object tracking. Based on the typical transformer architecture, it optimizes encoder and decoder structure for aggregating discriminative spatio-temporal information and global context-awareness feature. Furthermore, the encoder introduces the similarity calculation layer and dual-attention module; the aim is to deepen the similarity between features and make corrections for channel and spatial dimensions, and feature representation is improved. Finally, with the introduction of the temporal context filtering layer, unimportant feature information is ignored adaptively, obtaining a balance between the parameters number reduction and stable performance. Experimental results show that the proposed tracking algorithm exhibits excellent tracking performance on seven benchmark datasets, especially on the aerial dataset UAV123, UAV20L, and UAV123@10fps, which presents the advantages of the novel method in dealing with fast motion and external interference.

The field of person re-identification (re-ID) encounters challenges such as inter-person similarity and occlusion. In this paper, we introduce an occluded pedestrian re-ID method based on Multiple Fusion and Semantic feature Mining (MFSM) to mitigate the adverse effects of occlusion on feature representation. Specifically, we devise a Multi-branch Strategic Enhancement Module (SEM) to bolster the robustness of data augmentation in addressing the issues posed by data memorization more consistently. This module emulates various real-world disturbances by implementing distinct and independent augmentation strategies. Furthermore, we propose the Triplet Cross Unit (TCU) to comprehensively exploit and consolidate both regional and relational visual cues, as well as multi-level features. The TCU facilitates transformers in acquiring an early understanding of translation invariance in images by transferring local patterns such as edges, textures, and colors from the lower CNN layers to the shallower transformer layers. Simultaneously, deeper transformer features offer more abstracted semantic visual representations that complement CNN's high-level semantic features. Lastly, we introduce the Global Squeeze-Excitation Fusion (GSEF) module to address the challenge of global feature segregation across different models in final outputs. The GSEF selectively merges features based on attention mechanisms, prioritizing the comprehensive utilization of valid global features. Extensive experiments demonstrate the state-of-the-art performance of our model in handling pedestrian occlusion, thus validating the efficacy of our method.

The field of person re-identification (re-ID) encounters challenges such as inter-person similarity and occlusion. In this paper, we introduce an occluded pedestrian re-ID method based on Multiple Fusion and Semantic feature Mining (MFSM) to mitigate the adverse effects of occlusion on feature representation. Specifically, we devise a Multi-branch Strategic Enhancement Module (SEM) to bolster the robustness of data augmentation in addressing the issues posed by data memorization more consistently. This module emulates various real-world disturbances by implementing distinct and independent augmentation strategies. Furthermore, we propose the Triplet Cross Unit (TCU) to comprehensively exploit and consolidate both regional and relational visual cues, as well as multi-level features. The TCU facilitates transformers in acquiring an early understanding of translation invariance in images by transferring local patterns such as edges, textures, and colors from the lower CNN layers to the shallower transformer layers. Simultaneously, deeper transformer features offer more abstracted semantic visual representations that complement CNN's high-level semantic features. Lastly, we introduce the Global Squeeze-Excitation Fusion (GSEF) module to address the challenge of global feature segregation across different models in final outputs. The GSEF selectively merges features based on attention mechanisms, prioritizing the comprehensive utilization of valid global features. Extensive experiments demonstrate the state-of-the-art performance of our model in handling pedestrian occlusion, thus validating the efficacy of our method. • MFSM framework enables CNN-Transformer interaction, improving person re-identification accuracy, especially under occlusion. • Our strategic enhancement module expands data richness, suppressing noise and improving occlusion handling. • TCU combines CNN and Transformer features, improving re-ID performance without overemphasizing local features. • GSEF module consolidates diverse semantic features, enriching re-ID representation and improving accuracy. [ABSTRACT FROM AUTHOR]

Plaids consist of two orthogonally oriented gratings. However, their structure may be perceived as a checkerboard with an oblique orientation or as two overlaid gratings. We used a visual search paradigm to investigate how plaids with different perceived structures influence search efficiency for determining whether a preattentive plaid-specific or grating-specific mechanism is involved in representing the plaid. Participants searched for a grating among plaids (Experiment 1) or a plaid among gratings (Experiment 2). Contrast (low vs. high), composed spatial frequency (SF) of the plaid (coherent vs. mixed), and target--distractor orientation relationship (component orientation-- aligned vs. structure orientation--aligned) were manipulated to vary the perceived structure of the plaid. Parallel search was observed when the target was a coherent checkerboard plaid, and orientation manipulation did not influence search efficiency. However, search performance was less efficient when the target plaid was perceived as overlaid gratings. When searching for a grating among plaids, low-SF gratings were located more quickly than high-SF gratings, regardless of the plaid distractor type. The results provide evidence for the existence of a preattentive plaid-specific mechanism that is distinct from the grating-sensitive mechanism. [ABSTRACT FROM AUTHOR]

view of the large amount of garbage and the fact that a picture contains multiple garbage objects, this study proposes a garbage detection and classification algorithm based on the improved MobileNet network, which integrates the MobileNet network into YOLOv5(You Only Look Oncev5) target detection algorithm. At the same time, the CBAM(Convolutional Block Attention Modul) module is introduced in the backbone to filter meaningful information, and the vision transformer is used to aggregate and form image features. In addition, the weighted bidirectional feature pyramid network is used to distinguish the contribution of different features. At the same time, the ECA(Efficient Channel Attention) module is introduced to combine the image features and transmit them to the prediction layer. Finally, in order to obtain better performance when there is occlusion between garbage targets, soft-NMS(soft-Non Maximum Suppression) method and Alpha-IoU(Alpha-Intersection over Union) loss function is used to predict the extracted features. The experimental results show that the method proposed in this study can realize the location and recognition of multi-target and multi-category garbage., and the mAP(mean Average Percision) value reaches 90.31%, which is 4.95% higher than that of YOLOv5 network, and the processing speed is shortened by about 2.4 seconds. Compared with the Faster R-CNN(Region-based Convolutional Neural Network) algorithm which integrates ResNet(Residual Network) network, the algorithm proposed in this study improves the processing efficiency on the premise of ensuring the accuracy. [ABSTRACT FROM AUTHOR]

Accurate understanding of maritime traffic flow is crucial for optimizing waterway and port resources. However, the complex spatiotemporal characteristics and non-stationary sequences in maritime traffic pose challenges for feature extraction. This study proposes a deep learning approach that integrates Convolutional Neural Network (CNN) and Bidirectional Gated Recurrent Unit Network (BIGRU) for predicting maritime traffic flow. CNN is employed to extract feature information, while BIGRU captures temporal dependencies, enhancing the understanding of intrinsic features’ temporal changes. To further improve predictive performance, this study introduce the Whale Optimization Algorithm (WOA) for multiple objective optimization, which fine-tunes the CNN-BIGRU model parameters. By using eight years of monthly traffic flow data collected from the Dayaowan Port area in Dalian, the validity of this optimized model was rigorously verified. The results demonstrate that the WOA-optimized CNN-BIGRU provides an efficient forecast of maritime traffic flow with the Mean Absolute Error (MAE) of 4.084, Root Mean Square Error (RMSE) of 7.9719 and Mean Absolute Percentage Error (MAPE) of 1.262%. Compared to the original CNN-BiGRU models, this optimized approach reduced the prediction errors by 59.4%, 33.6%, and 54.2%, respectively. This further highlights its effectiveness in addressing the complexities of maritime traffic flow and provides valuable references for further research in this region.

Research on the identification of urban functional regions is of great significance for the understanding of urban structure, spatial planning, resource allocation, and promoting sustainable urban development. However, achieving high-precision urban functional region recognition has always been a research challenge in this field. For this purpose, this paper proposes an urban functional region identification method called ASOE (activity–scene–object–economy), which integrates the features from multi-source data to perceive the spatial differentiation of urban human and geographic elements. First, we utilize VGG16 (Visual Geometry Group 16) to extract high-level semantic features from the remote sensing images with 1.2 m spatial resolution. Then, using scraped building footprints, we extract building object features such as area, perimeter, and structural ratios. Socioeconomic features and population activity features are extracted from Point of Interest (POI) and Weibo data, respectively. Finally, integrating the aforementioned features and using the Random Forest method for classification, the identification results of urban functional regions in the main urban area of Xi’an are obtained. After comparing with the actual land use map, our method achieves an identification accuracy of 91.74%, which is higher than other comparative methods, making it effectively identify four typical urban functional regions in the main urban area of Xi’an (e.g., residential regions, industrial regions, commercial regions, and public regions). The research indicates that the method of fusing multi-source data can fully leverage the advantages of big data, achieving high-precision identification of urban functional regions.

This study explores fine-grained feature-based face attribute recognition techniques to enhance the accuracy of face recognition in low-resolution and complex environments. The article proposes a global feature extraction method and a local texture feature extraction method to extract global and regional features by enhancing feature reuse and information flow through dense connectivity and ShuffleNet V2 framework. Then, a multiscale feature exchange method is used to fuse different scale features to enhance the capture of detail information. Finally, efficient feature integration is achieved by the multiscale feature fusion method. Experimental results on the CK+ and FER2013 datasets show that the accuracy of this method on face expression recognition reaches 97.24% and 95.93%, respectively, and the average recognition accuracy in the face attribute recognition experiments on the CelebA dataset is 97.11%, which is significantly better than the comparison algorithm. In addition, the analysis of the recognition effect on low-resolution faces shows that this paper’s method achieves a recognition accuracy of 54.03% at a resolution of 15 × 15 and a high accuracy of over 99% at resolutions of 70 × 70 and above. These results show that the face attribute recognition technique based on fine-grained features proposed in this paper significantly improves recognition accuracy.

Remote sensing scene objective recognition (RSSOR) plays a serious application value in both military and civilian fields. Convolutional neural networks (CNNs) have greatly enhanced the improvement of intelligent objective recognition technology for remote sensing scenes, but most of the methods using CNN for high-resolution RSSOR either use only the feature map of the last layer or directly fuse the feature maps from various layers in the "summation" way, which not only ignores the favorable relationship information between adjacent layers but also leads to redundancy and loss of feature map, which hinders the improvement of recognition accuracy. In this study, a contextual, relational attention-based recognition network (CRABR-Net) was presented, which extracts different convolutional feature maps from CNN, focuses important feature content by using a simple, parameter-free attention module (SimAM), fuses the adjacent feature maps by using the complementary relationship feature map calculation, improves the feature learning ability by using the enhanced relationship feature map calculation, and finally uses the concatenated feature maps from different layers for RSSOR. Experimental results show that CRABR-Net exploits the relationship between the different CNN layers to improve recognition performance, achieves better results compared to several state-of-the-art algorithms, and the average accuracy on AID, UC-Merced, and RSSCN7 can be up to 96.46%, 99.20%, and 95.43% with generic training ratios. [ABSTRACT FROM AUTHOR]

Real-time and accurate awareness of the grain situation proves beneficial for making targeted and dynamic adjustments to cleaning parameters and strategies, leading to efficient and effective removal of impurities with minimal losses. In this study, harvested maize was employed as the raw material, and a specialized object detection network focused on impurity-containing maize images was developed to determine the types and distribution of impurities during the cleaning operations. On the basis of the classic contribution Faster Region Convolutional Neural Network, EfficientNetB7 was introduced as the backbone of the feature learning network and a cross-stage feature integration mechanism was embedded to obtain the global features that contained multi-scale mappings. The spatial information and semantic descriptions of feature matrices from different hierarchies could be fused through continuous convolution and upsampling operations. At the same time, taking into account the geometric properties of the objects to be detected and combining the images' resolution, the adaptive region proposal network (ARPN) was designed and utilized to generate candidate boxes with appropriate sizes for the detectors, which was beneficial to the capture and localization of tiny objects. The effectiveness of the proposed tiny object detection model and each improved component were validated through ablation experiments on the constructed RGB impurity-containing image datasets. [ABSTRACT FROM AUTHOR]

Visual discomfort prediction (VDP) is used to evaluate a user's visual discomfort when viewing stereoscopic images. The focus of this study is to explore more representative and distinctive features to describe complex visual discomfort perception. We consider integrating hand-crafted features and high-level semantic features based on deep learning to describe discomfort perception, and propose a multi-task VDP model (M2VDP) based on multi-view feature fusion. In particular, the model reinforces the positive effects of scene structure and depth difference on visual discomfort, and dynamically learns high-level semantic features at different scales by constructing a multi-scale neural network structure. In addition, using the multi-task learning, the prediction task of discomfort score can be guided by correctly classifying the discomfort level of stereoscopic images. The proposed method is verified on the public IEEE-SA database, and the results show that the proposed method achieves optimal performance. The ablation experimental results show that the multi-view feature fusion structure and multi-task learning are effective. [ABSTRACT FROM AUTHOR]

This work proposes a new approach based on YOLOX model enhancement for the helmet-wearing real-time detection task, which is plagued by low detection accuracy, incorrect detection, and missing detection. First, in the backbone network, recursive gated convolution (gnConv) is utilized instead of traditional convolution, hence addressing the issue of extracting many worthless features due to excessive redundancy in the process of feature extraction using conventional convolution. Replace the original FPN layer in the Neck network with the EfficientNet-BiFPN layer. Realize top–down and bottom–up bidirectional fusion of deep and shallow features to improve the flow of feature data between network layers. Lastly, the SIOU cross-entropy loss function is implemented to address the issue of missed detections in crowded environments and further increase the model's detection precision. Experiments and data comparisons indicate that the modified model's average detection accuracy is 95.5%, which is 5.4% higher than that of the original network model, and that the detection speed has been dramatically increased to suit actual production requirements. [ABSTRACT FROM AUTHOR]

Change detection (CD) aims to detect change objects of interest from bi-temporal images and is a hot research direction due to its value in human civilization. Existing CD methods usually employ convolution or transformer structures to extract image features. However, they neglect the shortcomings and complementary advantages of different structures. In addition, the potential relationship between bi-temporal input features is not adequately correlated, resulting in the difficulty of learning changed region features. To this end, we propose a novel Semantic-aware Transformer with Feature Integration (STFI) method to adequately correlate bi-temporal features by exploring complementary advantages from convolution or transformer structures for remote sensing change detection. Specifically, the residual module is employed to extract local information in the shallow layer, while the transformer module is adopted to model long-distance dependencies in the deep layer. We skillfully link convolution and transformer operations to learn complementary local and global features. In addition, we propose to integrate high-level bi-temporal semantic information in two aspects. On the one hand, the feature generation module is built to generate fused features in the local view from the convolutional operation. On the other hand, the feature integration module is designed to fuse global features from transformer operation. Furthermore, we develop an objective function containing two items, i.e., the cross-entropy loss and the dice loss, to optimize the whole CD model. Abundant experiments demonstrate that the proposed STFI surpasses other superior methods by 1.18%, 1.99%, and 3.05%, 5.15% in terms of F1 and IoU on LEVIR-CD and WHU-CD, respectively. [Display omitted] • We fuse convolution and transformer modules to model the local and global contexts. • We design feature generation/integration modules to correlate bi-temporal features. • We develop a function with the cross-entropy and dice losses to optimize the model. • We provide an in-depth analysis and prove that STFI achieves superior performance. [ABSTRACT FROM AUTHOR]

In the dynamic landscape of saliency detection, convolutional neural networks have emerged as catalysts for innovation, but remain largely tailored for RGB imagery, falling short in the context of infrared images, particularly in memory-restricted environments. These existing approaches tend to overlook the wealth of contour information vital for a nuanced analysis of infrared images. Addressing this notable gap, we introduce the novel Two-branch Edge Guided Lightweight Network (TBENet), designed explicitly for the robust analysis of infrared image saliency detection. The main contributions of this paper are as follows. First, we formulate the saliency detection task as two subtasks, contour enhancement and foreground segmentation. Therefore, the TBENet is divided into two specialized branches: a contour prediction branch for extracting target contour and a saliency map generation branch for separating the foreground from the background. The first branch employs an encoder–decoder architecture to meticulously delineate object contours, serving as a guiding blueprint for the second branch. This latter segment adeptly integrates spatial and semantic data, creating a precise saliency map that is refined further by an innovative edge-weighted contour loss function. Second, to enhance feature integration capabilities, we propose depthwise multi-scale and multi-cue modules, facilitating sophisticated feature aggregation. Third, a high-level linear bottleneck module is devised to ensure the extraction of rich semantic information, and by replacing the standard convolution with the depthwise convolution, it is beneficial to reduce model complexity. Additional, we reduce the number of channels of the feature maps from each stage of the decoder to further enhance the lightweight of the model. Last, we construct a novel infrared ship dataset Small-IRShip to train and evaluate our proposed model. Experimental results on the homemade dataset Small-IRShip and two publicly available datasets, namely RGB-T and IRSTD-1k, demonstrate TBENet's superior performance over state-of-the-art methods, affirming its effectiveness in harnessing edge information and incorporating advanced feature integration strategies. [ABSTRACT FROM AUTHOR]

Remote sensing scene objective recognition (RSSOR) plays a serious application value in both military and civilian fields. Convolutional neural networks (CNNs) have greatly enhanced the improvement of intelligent objective recognition technology for remote sensing scenes, but most of the methods using CNN for high-resolution RSSOR either use only the feature map of the last layer or directly fuse the feature maps from various layers in the “summation” way, which not only ignores the favorable relationship information between adjacent layers but also leads to redundancy and loss of feature map, which hinders the improvement of recognition accuracy. In this study, a contextual, relational attention-based recognition network (CRABR-Net) was presented, which extracts different convolutional feature maps from CNN, focuses important feature content by using a simple, parameter-free attention module (SimAM), fuses the adjacent feature maps by using the complementary relationship feature map calculation, improves the feature learning ability by using the enhanced relationship feature map calculation, and finally uses the concatenated feature maps from different layers for RSSOR. Experimental results show that CRABR-Net exploits the relationship between the different CNN layers to improve recognition performance, achieves better results compared to several state-of-the-art algorithms, and the average accuracy on AID, UC-Merced, and RSSCN7 can be up to 96.46%, 99.20%, and 95.43% with generic training ratios.

The healthy liver plays more than 500 organic roles in the human body, while a malfunction may be dangerous or even deadly. Early diagnosis and treatment of liver disease can improve the likelihood of survival. Machine learning (ML) is a powerful tool that can assist healthcare professionals during the diagnostic process for a hepatic patient. The standard ML system includes the methods of data pre-processing, feature extraction, and classification. In the feature extraction stage, ML researchers frequently use projection-based feature extraction approaches to remove data redundancy, but this does not produce the desired results. In addition, most statistical projection methods have different purposes when projecting original features. The Indian liver patient dataset (ILPD) from the University of California, Irvin (UCI) repository is used in this study to classify chronic liver disease. The data set has 583 patient disease records; 416 patients have liver disease, and 167 do not. Using several projection methods, we proposed an integrated feature extraction approach to categorize liver patients. In the pipeline, the proposed method first imputes the missing values and outliers for pre-treatment. Then, integrated feature extraction applies the pre-processed data to extract the significant features for classification. A simulation study is also being conducted to strengthen the suggested methodology. The proposed approach incorporates several ML algorithms, including logistic regression (LR), random forest (RF), K-nearest neighbor (KNN), support vector machine (SVM), multilayer perceptron (MLP), and the ensemble voting classifier. The offered system has an accuracy of 88.10%, a precision of 85.33%, a recall of 92.30%, an F1 score of 88.68%, and an AUC score of 88.20% in predicting liver diseases. Our proposed technique yielded 0.10–18.5% better results than the latest existing studies. The findings suggest that the recommended system could be used to supplement a physician's diagnosis of liver disease.

Little is known about expertise-related plasticity of neural mechanisms for auditory feature integration. Here, we contrast two diverging hypotheses that musical expertise is associated with more independent or more integrated predictive processing of acoustic features relevant to melody perception. Mismatch negativity (MMNm) was recorded with magnetoencephalography (MEG) from 25 musicians and 25 non-musicians, exposed to interleaved blocks of a complex, melody-like multi-feature paradigm and a simple, oddball control paradigm. In addition to single deviants differing in frequency (F), intensity (I), or perceived location (L), double and triple deviants were included reflecting all possible feature combinations (FI, IL, LF, FIL). Following previous work, early neural processing overlap was approximated in terms of MMNm additivity by comparing empirical MMNms obtained with double and triple deviants to modeled MMNms corresponding to summed constituent single-deviant MMNms. Significantly greater subadditivity was found in musicians compared to non-musicians, specifically for frequencyrelated deviants in complex, melody-like stimuli. Despite using identical sounds, expertise effects were absent from the simple oddball paradigm. This novel finding supports the integrated processing hypothesis whereby musicians recruit overlapping neural resources facilitating more integrative representations of contextually relevant stimuli such as frequency (perceived as pitch) during melody perception. More generally, these specialized refinements in predictive processing may enable experts to optimally capitalize upon complex, domain-relevant, acoustic cues. [ABSTRACT FROM AUTHOR]

In view of the problems of low efficiency of manual garbage sorting, heavy tasks and harsh environment, this study proposes a YOLO-based target detection method to realize garbage detection and classification. The model is adjusted through making a specific dataset, using K-means clustering algorithm and Mish activation function. According to the characteristics of the convolutional neural network, the CBAM attention module is embedded in front of each detection head of the YOLO model, combined with PANet to enhance the feature integration ability to improve the accuracy of small target detection. The experimental results show that the garbage detection and classification method proposed in this study can accurately and quickly identify garbage. Compared with YOLOv4, the map value of the proposed model on the garbage data set has increased by 2.81%. The recognition accuracy of Cans can reach 94.56%, and the accuracy of PlasticBottle has increased by 6.36%. [ABSTRACT FROM AUTHOR]

Recently, correlation filter (CF)-based tracking algorithms have attained extensive interest in the field of unmanned aerial vehicle (UAV) tracking. Nonetheless, existing trackers still struggle with selecting suitable features and alleviating the model drift issue for online UAV tracking. In this paper, a robust CF-based tracker with feature integration and response map enhancement is proposed. Concretely, we develop a novel feature integration method that comprehensively describes the target by leveraging auxiliary gradient information extracted from the binary representation. Subsequently, the integrated features are utilized to learn a background-aware correlation filter (BACF) for generating a response map that implies the target location. To mitigate the risk of model drift, we introduce saliency awareness in the BACF framework and further propose an adaptive response fusion strategy to enhance the discriminating capability of the response map. Moreover, a dynamic model update mechanism is designed to prevent filter contamination and maintain tracking stability. Experiments on three public benchmarks verify that the proposed tracker outperforms several state-of-the-art algorithms and achieves a real-time tracking speed, which can be applied in UAV tracking scenarios efficiently. [ABSTRACT FROM AUTHOR]

Little is known about expertise-related plasticity of neural mechanisms for auditory feature integration. Here, we contrast two diverging hypotheses that musical expertise is associated with more independent or more integrated predictive processing of acoustic features relevant to melody perception. Mismatch negativity (MMNm) was recorded with magnetoencephalography (MEG) from 25 musicians and 25 non-musicians, exposed to interleaved blocks of a complex, melody-like multi-feature paradigm and a simple, oddball control paradigm. In addition to single deviants differing in frequency (F), intensity (I), or perceived location (L), double and triple deviants were included reflecting all possible feature combinations (FI, IL, LF, FIL). Following previous work, early neural processing overlap was approximated in terms of MMNm additivity by comparing empirical MMNms obtained with double and triple deviants to modeled MMNms corresponding to summed constituent single-deviant MMNms. Significantly greater subadditivity was found in musicians compared to non-musicians, specifically for frequency-related deviants in complex, melody-like stimuli. Despite using identical sounds, expertise effects were absent from the simple oddball paradigm. This novel finding supports the integrated processing hypothesis whereby musicians recruit overlapping neural resources facilitating more integrative representations of contextually relevant stimuli such as frequency (perceived as pitch) during melody perception. More generally, these specialized refinements in predictive processing may enable experts to optimally capitalize upon complex, domain-relevant, acoustic cues.

Real-time and accurate awareness of the grain situation proves beneficial for making targeted and dynamic adjustments to cleaning parameters and strategies, leading to efficient and effective removal of impurities with minimal losses. In this study, harvested maize was employed as the raw material, and a specialized object detection network focused on impurity-containing maize images was developed to determine the types and distribution of impurities during the cleaning operations. On the basis of the classic contribution Faster Region Convolutional Neural Network, EfficientNetB7 was introduced as the backbone of the feature learning network and a cross-stage feature integration mechanism was embedded to obtain the global features that contained multi-scale mappings. The spatial information and semantic descriptions of feature matrices from different hierarchies could be fused through continuous convolution and upsampling operations. At the same time, taking into account the geometric properties of the objects to be detected and combining the images’ resolution, the adaptive region proposal network (ARPN) was designed and utilized to generate candidate boxes with appropriate sizes for the detectors, which was beneficial to the capture and localization of tiny objects. The effectiveness of the proposed tiny object detection model and each improved component were validated through ablation experiments on the constructed RGB impurity-containing image datasets.

Considering the high requirements of current kiwifruit picking recognition systems for mobile devices, including the small number of available features for image targets and small-scale aggregation, an enhanced YOLOX-S target detection algorithm for kiwifruit picking robots is proposed in this study. This involved designing a new multi-scale feature integration structure in which, with the aim of providing a small and lightweight model, the feature maps used for detecting large targets in the YOLOX model are eliminated, the feature map of small targets is sampled through the nearest neighbor values, the superficial features are spliced with the final features, the gradient of the SiLU activation function is perturbed, and the loss function at the output is optimized. The experimental results show that, compared with the original YOLOX-S, the enhanced model improved the detection average precision (AP) of kiwifruit images by 6.52%, reduced the number of model parameters by 44.8%, and improved the model detection speed by 63.9%. Hence, with its outstanding effectiveness and relatively light weight, the proposed model is capable of effectively providing data support for the 3D positioning and automated picking of kiwifruit. It may also successfully provide solutions in similar fields related to small target detection. [ABSTRACT FROM AUTHOR]

In the era of We-media, articles are written by independently individuals that are not officially registered with the authorities. Propaganda hidden in the We-media articles have the potential to polarize public opinion and influence the mindset of the target audience. Currently, graph neural networks (GNNs) have been remarkably successful in Natural Language Processing (NLP). However, there are still some challenges to apply existing GNN-based for propaganda detection due to the limitation of extracting diverse word dependencies and capturing non-consecutive and long-range context. In this paper, we have proposed a Graph-based Hierarchical Feature Integration Network (abbreviated as G-HFIN) for Propaganda Detection. Specifically, semantic, syntactic, and sequential features are extracted to construct three heterogeneous graphs. Then, the Residual-connected Dual-layer Coarsening and Refining procedures (abbreviated as RDCR) are proposed to promote information interactions between distant nodes that are not directly connected, preserving both local and global node information during the intra-graph information propagation. Subsequently, an Attention-based Three-channel Feature Integration (abbreviated as ATFI) is proposed to harmonize sequence, semantic and syntactic information from three channels during the inter-graph information enhancement. Intra-graph and Inter-graph Joint Information Propagation is to implement homogeneous and heterogeneous information interaction respectively. Finally, these news representations are pooled and fed into the propaganda detection classifier. The experiments on three public datasets demonstrate that our model has outperformed state-of-the-art methods. [ABSTRACT FROM AUTHOR]

This study presents a novel feature integration method for interest level estimation using a semi-supervised multimodal Gaussian process latent variable model with pseudo-labels (semi-MGPPL). Semi-MGPPL is an extended version of the multimodal Gaussian process latent variable model (mGPLVM). It integrates features calculated from multiple modalities to predict the users’ interest levels in content. It is known that reflecting known interest levels of known users in the latent space effectively improves the accuracy of interest level estimation. However, previous methods have difficulty reflecting the interest levels when the number of samples is insufficient. Semi-MGPPL efficiently reflects interest levels in the latent space by pseudo-labeling of unlabeled samples and increasing the number of available pairs among labeled samples. In addition, obtaining behavior features is difficult for a new test sample. However, requirement of features of all modalities by previous mGPLVM-based methods makes the calculation of latent variables of a test sample challenging. Semi-MGPPL solves this problem by training a projection function from the original feature to the latent space. The experimental results on real data demonstrate the effectiveness and robustness of semi-MGPPL.

Post-translational modification (PTM) refers to the covalent and enzymatic modification of proteins after protein biosynthesis, which orchestrates a variety of biological processes. Detecting PTM sites in proteome scale is one of the key steps to in-depth understanding their regulation mechanisms. In this study, we presented an integrated method based on eXtreme Gradient Boosting (XGBoost), called iRice-MS, to identify 2-hydroxyisobutyrylation, crotonylation, malonylation, ubiquitination, succinylation and acetylation in rice. For each PTM-specific model, we adopted eight feature encoding schemes, including sequence-based features, physicochemical property-based features and spatial mapping information-based features. The optimal feature set was identified from each encoding, and their respective models were established. Extensive experimental results show that iRice-MS always display excellent performance on 5-fold cross-validation and independent dataset test. In addition, our novel approach provides the superiority to other existing tools in terms of AUC value. Based on the proposed model, a web server named iRice-MS was established and is freely accessible at http://lin-group.cn/server/iRice-MS. [ABSTRACT FROM AUTHOR]

Techniques for integrating different types of multiple features effectively have been actively studied in recent years. Multiset canonical correlation analysis (MCCA), which maximizes the sum of pairwise correlations of inter-view (i.e., between different features), is one of the powerful methods for integrating different types of multiple features, and various MCCA-based methods have been proposed. This work focuses on a supervised MCCA variant in order to construct a novel effective feature integration framework. In this paper, we newly propose supervised fractional-order embedding geometrical multi-view CCA (SFGMCCA). This method constructs not only the correlation structure but also two types of geometrical structures of intra-view (i.e., within each feature) and inter-view simultaneously, thereby realizing more precise feature integration. This method also supports the integration of small sample and high-dimensional data by using the fractional-order technique. We conducted experiments using four types of image datasets, i.e., MNIST, COIL-20, ETH-80 and CIFAR-10. Furthermore, we also performed an fMRI dataset containing brain signals to verify the robustness. As a result, it was confirmed that accuracy improvements using SFGMCCA were statistically significant at the significance level of 0.05 compared to those using conventional representative MCCA-based methods.

The use of machine learning has increased over the years, especially in the world of molecular data. Generally, the inference of relationships between features is determined by statistical models. The phenotype (observable clinical characteristics) can result from the expression of the genotype (genetic code) or environmental factors. Molecular datasets have limited information, while supporting clinical data is ambiguous. There are no well-established approaches for combining clinical information with genomic repositories. The genomic tests that are available only use molecular data and give physicians a result which can be integrated clinically. In this article, we present the strategy where clinical data, regardless of its limitations, is combined in one predictive model with molecular features. We predict the risk of malignancy in the thyroid nodules based on the results of fine-needle aspiration biopsy and expression of selected genes. We utilize a Bayesian network (BN) framework to discover relationships between molecular features and assess the impact of added clinical data quality on the performance of the chosen gene set. Bayesian network offering both prognostic and diagnostic perspectives is a perfect non-parametric technique for feature selection, feature extraction, and prediction purposes. We show that certain clinical factors could work as a synthetic feature and provide predictive abilities beyond what genes alone can offer. The experimental results demonstrate a higher performance of predictive models based on molecular and clinical data than when using only molecular data. We also explain why, one should consider the source of clinical data, but be aware of the quality of variables.

Many studies have shown that the human visual system has two major functionally distinct cortical visual pathways: a ventral pathway, thought to be important for object recognition, and a dorsal pathway, thought to be important for spatial cognition. According to our and others previous studies, artificial neural networks with two segregated pathways can determine objects' identities and locations more accurately and efficiently than one-pathway artificial neural networks. In addition, we showed that these two segregated artificial cortical visual pathways can each process identity and spatial information of visual objects independently and differently. However, when using such networks to process multiple objects' identities and locations, a binding problem arises because the networks may not associate each object's identity with its location correctly. In a previous study, we constrained the binding problem by training the artificial identity pathway to retain relative location information of objects. This design uses a location map to constrain the binding problem. One limitation of that study was that we only considered two attributes of our objects (identity and location) and only one possible map (location) for binding. However, typically the brain needs to process and bind many attributes of an object, and any of these attributes could be used to constrain the binding problem. In our current study, using visual objects with multiple attributes (identity, luminance, orientation, and location) that need to be recognized, we tried to find the best map (among an identity map, a luminance map, an orientation map, or a location map) to constrain the binding problem. We found that in our experimental simulations, when visual attributes are independent of each other, a location map is always a better choice than the other kinds of maps examined for constraining the binding problem. Our findings agree with previous neurophysiological findings that show that the organization or map in many visual cortical areas is primarily retinotopic or spatial., Competing Interests: The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. The author(s) declared that they were an editorial board member of Frontiers, at the time of submission. This had no impact on the peer review process and the final decision., (Copyright © 2024 Han and Sereno.)

This work proposes a new approach based on YOLOX model enhancement for the helmet-wearing real-time detection task, which is plagued by low detection accuracy, incorrect detection, and missing detection. First, in the backbone network, recursive gated convolution (gnConv) is utilized instead of traditional convolution, hence addressing the issue of extracting many worthless features due to excessive redundancy in the process of feature extraction using conventional convolution. Replace the original FPN layer in the Neck network with the EfficientNet-BiFPN layer. Realize top–down and bottom–up bidirectional fusion of deep and shallow features to improve the flow of feature data between network layers. Lastly, the SIOU cross-entropy loss function is implemented to address the issue of missed detections in crowded environments and further increase the model’s detection precision. Experiments and data comparisons indicate that the modified model’s average detection accuracy is 95.5%, which is 5.4% higher than that of the original network model, and that the detection speed has been dramatically increased to suit actual production requirements.

Understanding the structural similarity between genomes is pivotal in classification and phylogenetic analysis. As the number of known genomes rockets, alignment-free methods have gained considerable attention. Among these methods, the natural vector method stands out as it represents sequences as vectors using statistical moments, enabling effective clustering based on families in biological taxonomy. However, determining an optimal metric that combines different elements in natural vectors remains challenging due to the absence of a rigorous theoretical framework for weighting different k -mers and orders. In this study, we address this challenge by transforming the determination of optimal weights into an optimization problem and resolving it through gradient-based techniques. Our experimental results underscore the substantial improvement in classification accuracy achieved by employing these optimal weights, reaching an impressive 92.73% on the testing set, surpassing other alignment-free methods. On one hand, our method offers an outstanding metric for virus classification, and on the other hand, it provides valuable insights into feature integration within alignment-free methods., Competing Interests: None., (© 2024 Published by Elsevier B.V. on behalf of Research Network of Computational and Structural Biotechnology.)

Recently, correlation filter (CF)-based tracking algorithms have attained extensive interest in the field of unmanned aerial vehicle (UAV) tracking. Nonetheless, existing trackers still struggle with selecting suitable features and alleviating the model drift issue for online UAV tracking. In this paper, a robust CF-based tracker with feature integration and response map enhancement is proposed. Concretely, we develop a novel feature integration method that comprehensively describes the target by leveraging auxiliary gradient information extracted from the binary representation. Subsequently, the integrated features are utilized to learn a background-aware correlation filter (BACF) for generating a response map that implies the target location. To mitigate the risk of model drift, we introduce saliency awareness in the BACF framework and further propose an adaptive response fusion strategy to enhance the discriminating capability of the response map. Moreover, a dynamic model update mechanism is designed to prevent filter contamination and maintain tracking stability. Experiments on three public benchmarks verify that the proposed tracker outperforms several state-of-the-art algorithms and achieves a real-time tracking speed, which can be applied in UAV tracking scenarios efficiently.

Bayesian models of object recognition propose the resolution of ambiguity through probabilistic integration of prior experience with available sensory information. Color, even when task-irrelevant, has been shown to modulate high-level cognitive control tasks. However, it remains unclear how color modulations affect lower-level perceptual processing. We investigated whether color affects feature integration using the flash-jump illusion. This illusion occurs when an apparent motion stimulus, a rectangular bar appearing at different locations along a motion trajectory, changes color at a single position. Observers misperceive this color change as occurring farther along the trajectory of motion. This mislocalization error is proposed to be produced by a Bayesian perceptual framework dependent on responses in area V4. Our results demonstrated that the color of the flash modulated the magnitude of the flash-jump illusion such that participants reported less of a shift, i.e., a more veridical flash location, for both red and blue flashes, as compared to green and yellow. Our findings extend color-dependent modulation effects found in higher-order executive functions into lower-level Bayesian perceptual processes. Our results also support the theory that feature integration is a Bayesian process. In this framework, color modulations play an inherent and automatic role as different colors have different weights in Bayesian perceptual processing.

Bayesian models of object recognition propose the resolution of ambiguity through probabilistic integration of prior experience with available sensory information. Color, even when task-irrelevant, has been shown to modulate high-level cognitive control tasks. However, it remains unclear how color modulations affect lower-level perceptual processing. We investigated whether color affects feature integration using the flash-jump illusion. This illusion occurs when an apparent motion stimulus, a rectangular bar appearing at different locations along a motion trajectory, changes color at a single position. Observers misperceive this color change as occurring farther along the trajectory of motion. This mislocalization error is proposed to be produced by a Bayesian perceptual framework dependent on responses in area V4. Our results demonstrated that the color of the flash modulated the magnitude of the flash-jump illusion such that participants reported less of a shift, i.e., a more veridical flash location, for both red and blue flashes, as compared to green and yellow. Our findings extend color-dependent modulation effects found in higher-order executive functions into lower-level Bayesian perceptual processes. Our results also support the theory that feature integration is a Bayesian process. In this framework, color modulations play an inherent and automatic role as different colors have different weights in Bayesian perceptual processing. [ABSTRACT FROM AUTHOR]

This paper presents an advanced methodology that combines Convolutional Neural Networks (CNNs) and Transformers to tackle the issue of feature loss and elevate segmentation accuracy. While Transformers excel at capturing global features, there is a potential risk of losing crucial global information when CNNs concentrate on local feature extraction. To counteract this, we introduce the FM module, which re-extracts local information from the Encoder's output, reinforcing local feature expression and enhancing segmentation accuracy. Additionally, our approach places a significant emphasis on the judicious fusion of features in the Decoder. By incorporating a technology that aggregates both global and local features, we aim to prevent the loss of feature information. To augment the effective representation of data features and heighten accuracy across diverse segmentation tasks, we introduce the GS Feature Combination method, which adjusts the weights of different features during the aggregation process. Experimental results demonstrate noteworthy performance improvements, achieving an Intersection over Union (IOU) of 71.3%, surpassing existing methods. This innovative approach bears substantial importance in the diagnosis, treatment, and prognosis prediction of osteosarcoma, affording doctors the opportunity to reduce workload and time while upholding diagnostic precision. • Pathological image segmentation, crucial for medical diagnosis. • U-Net and others address challenges but risk losing essential features. • GSFormer integrates a transformer for global features and a CNN for localization. • High accuracy in analyzing osteosarcoma images. • GSFormer addresses medical deficiencies in developing countries. [ABSTRACT FROM AUTHOR]

• scMVFI integrates linear and non-linear features of scRNA-seq data to address the disadvantage of zero-inflated noise caused by technical factors. • scMVFI performs multi-view similarity estimation using subsets of features with different sampling rates to identify highly similar cell pairs. • Experimental results on five real datasets and three simulated datasets show that the proposed method has certain advantages in terms of clustering performance compared to other methods. Accurate and consistent estimation of cell-to-cell similarity is crucial for clustering single-cell RNA-seq (scRNA-seq) data. However, the high sparsity of scRNA-seq data often leads to suboptimal mining and decreased accuracy in identifying cell types. Moreover, using a larger number of features (genes) does not necessarily improve clustering accuracy due to redundant information. In this paper, we propose a framework, called scMVFI (single-cell Multi-View Feature Integration), which integrates linear and non-linear features of scRNA-seq data to address the disadvantage of zero-inflated noise caused by technical factors. By employing an autoencoder model for data reconstruction, scMVFI performs multi-view similarity estimation using subsets of features with different sampling rates to identify highly similar cell pairs. We evaluate the effectiveness of scMVFI using five real scRNA-seq datasets and three simulated datasets. The results demonstrate that scMVFI can effectively mitigate the impact of data "dropout" events compared to other methods. Moreover, the affinity networks constructed from both linear and non-linear perspectives can accurately capture sample relationships, thereby enhancing the identification of cell types when combined with existing clustering methods. [ABSTRACT FROM AUTHOR]

In this paper, we propose a novel method named Discrimination-Aware Integration (DAI) for person re-identification (re-ID) in camera networks, which not only integrates multiple re-ID models but also adaptively learns integration weights for different feature dimensions. To avoid the tough selection of deep models, we employ different data sources to train the same re-ID model for learning features from different views, and then, we obtain multiple features for each pedestrian image. To effectively integrate these features, the proposed DAI learns integration weights for each feature dimension according to their importance. Finally, the features extracted from different re-ID models are integrated with the learned integration weights to form the final representation for the pedestrian images. We evaluate the performance of the proposed DAI on three public large-scale person re-ID datasets, i.e., Market1501, CUHK03, and DukeMTMC-reID, and the experimental results show that the proposed DAI outperforms the state-of-the-art results.

Deep neural networks (DNNs) have been widely applied to the automatic analysis of medical images for disease diagnosis and to help human experts by efficiently processing immense amounts of images. While the handcrafted feature has been used for eye disease detection or classification since the 1990s, DNN was recently adopted in this area and showed a very promising performance. Since handcrafted and deep feature can extract complementary information, we propose, in this paper, three different integration frameworks to combine handcrafted and deep feature for optical coherence tomography image-based eye disease classification. In addition, to integrate the handcrafted feature at the input and fully connected layers using existing networks, such as VGG, DenseNet, and Xception, a novel ribcage network (RC Net) is also proposed for feature integration at middle layers. For RC Net, two “rib” channels are designed to independently process deep and handcrafted features, and another so-called “spine” channel is designed for the integration. While dense blocks are the main components of the three channels, sum operation is proposed for the feature map integration. Our experimental results showed that the deep networks achieved better classification accuracy after the integration of the handcrafted features, e.g., scale-invariant feature transform and Gabor. The RC Net showed the best performance among all the proposed feature integration methods.

Visual object tracking is an active and challenging research topic in computer vision, as objects often undergo significant appearance changes caused by occlusion, deformation, and background clutter. Although convolutional neural network (CNN)-based trackers have achieved competitive results, there are still some limitations. Most existing CNN-based trackers track the object by leveraging high-level semantic features of the highest convolutional layer, which may lead to low-spatial resolution feature maps and degrade the localization precision of tracking. Furthermore, these trackers hardly benefit from end-to-end training since the extraction of features and the learning of classifier are separated. To deal with the above-mentioned issues, we design an adaptive weighted CNN features-based Siamese network for tracking. To capture spatial and semantic information of the object, we design a feature extraction network that derives feature maps by concatenating features of all convolutional layers. To make the features representation more discriminative, we propose a feature integration network. In the feature integration network, we propose a holistic-part network to capture strong visual cues and learn the semantic relations between the holistic object and its parts and combine the holistic-part network with spatial and channel attention mechanisms to adaptively assign weights to each region and channel of the feature maps. In addition, the designed Siamese network can be trained offline end-to-end. The experimental results on the benchmark datasets OTB50 and OTB100 demonstrate that the proposed tracker achieves favorable performance against several state-of-the-art trackers while running at an average speed of 20.5 frames/s.

A novel method for interest level estimation based on tensor completion via feature integration for partially paired users' behavior and videos is presented in this paper. The proposed method defines a novel canonical correlation analysis (CCA) framework that is suitable for interest level estimation, which is a hybrid version of semi-supervised CCA (SemiCCA) and supervised locality preserving CCA (SLPCCA) called semi-supervised locality preserving CCA (S2LPCCA). For partially paired users' behavior and videos in actual shops and on the Internet, new integrated features that maximize the correlation between partially paired samples by the principal component analysis (PCA)-mixed CCA framework are calculated. Then videos that users have not watched can be used for the estimation of users' interest levels. Furthermore, local structures of partially paired samples in the same class are preserved for accurate estimation of interest levels. Tensor completion, which can be applied to three contexts, videos, users and “canonical features and interest levels,” is used for estimation of interest levels. Consequently, the proposed method realizes accurate estimation of users' interest levels based on S2LPCCA and the tensor completion from partially paired training features of users' behavior and videos. Experimental results obtained by applying the proposed method to actual data show the effectiveness of the proposed method.

With the wide application of isometric 3D models, the representation and recognition of them receive increasing attention. Most existing methods for shape representation and analysis either focus on using the constrained hand-craft models or purely devote efforts on developing complicated deep learning methods. In order to make full use of the advantages of both, this paper presents a novel mixture modeling approach by integrating both the shape function map (SFM) and the deep convolutional neural network (CNN). First, multiple SFM maps are constructed to grasp the rigid and non-rigid information that are usually considered separately for shape representation. Then, to fully characterize the low-level information existing in SFM, diverse sets of deep features are learned on different SFM maps during the training process of classifiers. Finally, both the rigid and non-rigid deep features are integrated for more discriminative feature abstraction. The experimental results on standard shape benchmark datasets have validated the superior performance of our proposed approach on feature extraction, classification and retrieval. Besides, our evaluation results on extensive shape datasets (e.g. noisy, CAD and protein shapes) have again verified the effectiveness of the proposed algorithm.

Considering the high requirements of current kiwifruit picking recognition systems for mobile devices, including the small number of available features for image targets and small-scale aggregation, an enhanced YOLOX-S target detection algorithm for kiwifruit picking robots is proposed in this study. This involved designing a new multi-scale feature integration structure in which, with the aim of providing a small and lightweight model, the feature maps used for detecting large targets in the YOLOX model are eliminated, the feature map of small targets is sampled through the nearest neighbor values, the superficial features are spliced with the final features, the gradient of the SiLU activation function is perturbed, and the loss function at the output is optimized. The experimental results show that, compared with the original YOLOX-S, the enhanced model improved the detection average precision (AP) of kiwifruit images by 6.52%, reduced the number of model parameters by 44.8%, and improved the model detection speed by 63.9%. Hence, with its outstanding effectiveness and relatively light weight, the proposed model is capable of effectively providing data support for the 3D positioning and automated picking of kiwifruit. It may also successfully provide solutions in similar fields related to small target detection.