Search

Scene text detection is an important research field in computer vision, playing a crucial role in various application scenarios. However, existing scene text detection methods often fail to achieve satisfactory results when faced with text instances of different sizes, shapes, and complex backgrounds. To address the challenge of detecting diverse texts in natural scenes, this paper proposes a multi-scale natural scene text detection method based on attention feature extraction and cascaded feature fusion. This method combines global and local attention through an improved attention feature fusion module (DSAF) to capture text features of different scales, enhancing the network's perception of text regions and improving its feature extraction capabilities. Simultaneously, an improved cascaded feature fusion module (PFFM) is used to fully integrate the extracted feature maps, expanding the receptive field of features and enriching the expressive ability of the feature maps. Finally, to address the cascaded feature maps, a lightweight subspace attention module (SAM) is introduced to partition the concatenated feature maps into several sub-space feature maps, facilitating spatial information interaction among features of different scales. In this paper, comparative experiments are conducted on the ICDAR2015, Total-Text, and MSRA-TD500 datasets, and comparisons are made with some existing scene text detection methods. The results show that the proposed method achieves good performance in terms of accuracy, recall, and F-score, thus verifying its effectiveness and practicality. [ABSTRACT FROM AUTHOR]

This paper proposes a multimodal fusion 3D target detection algorithm based on the attention mechanism to improve the performance of 3D target detection. The algorithm utilizes point cloud data and information from the camera. For image feature extraction, the ResNet50 + FPN architecture extracts features at four levels. Point cloud feature extraction employs the voxel method and FCN to extract point and voxel features. The fusion of image and point cloud features is achieved through regional point fusion and voxel fusion methods. After information fusion, the Coordinate and SimAM attention mechanisms extract fusion features at a deep level. The algorithm's performance is evaluated using the DAIR-V2X dataset. The results show that compared to the Part-A2 algorithm; the proposed algorithm improves the mAP value by 7.9% in the BEV view and 7.8% in the 3D view at IOU = 0.5 (cars) and IOU = 0.25 (pedestrians and cyclists). At IOU = 0.7 (cars) and IOU = 0.5 (pedestrians and cyclists), the mAP value of the SECOND algorithm is improved by 5.4% in the BEV view and 4.3% in the 3D view, compared to other comparison algorithms. [ABSTRACT FROM AUTHOR]

Text summarization is an information compression technology to extract important information from long text, which has become a challenging research direction in the field of natural language processing. At present, the text summary model based on deep learning has shown good results, but how to more effectively model the relationship between words, more accurately extract feature information and eliminate redundant information is still a problem of concern. This paper proposes a graph neural network model GA-GNN based on gated attention, which effectively improves the accuracy and readability of text summarization. First, the words are encoded using a concatenated sentence encoder to generate a deeper vector containing local and global semantic information. Secondly, the ability to extract key information features is improved by using gated attention units to eliminate local irrelevant information. Finally, the loss function is optimized from the three aspects of contrastive learning, confidence calculation of important sentences, and graph feature extraction to improve the robustness of the model. Experimental validation was conducted on a CNN/Daily Mail dataset and MR dataset, and the results showed that the model in this paper outperformed existing methods. [ABSTRACT FROM AUTHOR]

Rails play a vital role in the bearing and guidance of high-speed trains, and the normal condition of rail components is the guarantee of the operation and maintenance safety. Fasteners are critical components for fixing the rails, so it is particularly important to detect whether they are in a normal state or not. The current rail-fastener detection models have some drawbacks, including poor generalization ability, large model volume and low detection efficiency. In view of this, an improved YoLoX-Nano rail-fastener-defect-detection method is proposed in this paper. The CA attention mechanism is added to the three output feature maps of CSPDarknet and the enhanced feature extraction part of the Path Aggregation Feature Pyramid Network (PAFPN); the Adaptively Spatial Feature Fusion (ASFF) is added after the PAFPN output feature map, which enables the semantic information of the high-level features and the fine-grained features of the bottom layer to be further enhanced. The improved YoLoX-Nano model has improved the AP value by 27.42% on fractured fasteners, 15.88% on displacement fasteners and 12.96% on normal fasteners. Moreover, the mAP value is improved by 18.75%, and it is 14.75% higher than the two-stage model Faster-RCNN on mAP. In addition, compared with YoLov7-tiny, the improved YoLoX-Nano model achieves 13.56% improvement on mAP. Although the improved model increases a certain amount of calculation, the detection speed of the improved model has been increased by 30.54 fps and by 32.33 fps when compared with that of the Single-Shot Multi-Box Detector (SSD) model and the You Only Look Once v3 (YoLov3) model, reaching 54.35 fps. The improved YoLoX-Nano model enables accurate and rapid identification of the defects of rail fasteners, which can meet the needs of real-time detection. Furthermore, it has advantages in lightweight deployment of terminals for rail-fastener detection, thus providing some reference for image recognition and detection in other fields. [ABSTRACT FROM AUTHOR]

In view of the existence of remote sensing images with large variations in spatial resolution, small and dense objects, and the inability to determine the direction of motion, all these components make object detection from remote sensing images very challenging. In this paper, we propose a single-stage detection network based on YOLOv5. This method introduces the MS Transformer module at the end of the feature extraction network of the original network to enhance the feature extraction capability of the network model and integrates the Convolutional Block Attention Model (CBAM) to find the attention area in dense scenes. In addition, the YOLOv5 target detection network is improved by incorporating a rotation angle approach from the a priori frame design and the bounding box regression formulation to make it suitable for rotating frame-based detection scenarios. Finally, the weighted combination of the two difficult sample mining methods is used to improve the focal loss function, so as to improve the detection accuracy. The average accuracy of the test results of the improved algorithm on the DOTA data set is 77.01%, which is higher than the previous detection algorithm. Compared with the average detection accuracy of YOLOv5, the average detection accuracy is improved by 8.83%. The experimental results show that the algorithm has higher detection accuracy than other algorithms in remote sensing scenes. [ABSTRACT FROM AUTHOR]

Accurate diagnosis of pear tree nutrient deficiency symptoms is vital for the timely adoption of fertilization and treatment. This study proposes a novel method on the fused feature multi-head attention recording network with image depth and shallow feature fusion for diagnosing nutrient deficiency symptoms in pear leaves. First, the shallow features of nutrient-deficient pear leaf images are extracted using manual feature extraction methods, and the depth features are extracted by the deep network model. Second, the shallow features are fused with the depth features using serial fusion. In addition, the fused features are trained using three classification algorithms, F-Net, FC-Net, and FA-Net, proposed in this paper. Finally, we compare the performance of single feature-based and fusion feature-based identification algorithms in the nutrient-deficient pear leaf diagnostic task. The best classification performance is achieved by fusing the depth features output from the ConvNeXt-Base deep network model with shallow features using the proposed FA-Net network, which improved the average accuracy by 15.34 and 10.19 percentage points, respectively, compared with the original ConvNeXt-Base model and the shallow feature-based recognition model. The result can accurately recognize pear leaf deficiency images by providing a theoretical foundation for identifying plant nutrient-deficient leaves. [ABSTRACT FROM AUTHOR]

To solve the problem of inadequate feature extraction by the model due to factors such as occlusion and illumination in person re-identification tasks, this paper proposed a model with a joint cross-consistency learning and multi-feature fusion person re-identification. The attention mechanism and the mixed pooling module were first embedded in the residual network so that the model adaptively focuses on the more valid information in the person images. Secondly, the dataset was randomly divided into two categories according to the camera perspective, and a feature classifier was trained for the two types of datasets respectively. Then, two classifiers with specific knowledge were used to guide the model to extract features unrelated to the camera perspective for the two types of datasets so that the obtained image features were endowed with domain invariance by the model, and the differences in the perspective, attitude, background, and other related information of different images were alleviated. Then, the multi-level features were fused through the feature pyramid to concern the more critical information of the image. Finally, a combination of Cosine Softmax loss, triplet loss, and cluster center loss was proposed to train the model to address the differences of multiple losses in the optimization space. The first accuracy of the proposed model reached 95.9% and 89.7% on the datasets Market-1501 and DukeMTMC-reID, respectively. The results indicated that the proposed model has good feature extraction capability. [ABSTRACT FROM AUTHOR]

Domestic trash detection is an essential technology toward achieving a smart city. Due to the complexity and variability of urban trash scenarios, the existing trash detection algorithms suffer from low detection rates and high false positives, as well as the general problem of slow speed in industrial applications. This paper proposes an i-YOLOX model for domestic trash detection based on deep learning algorithms. First, a large number of real-life trash images are collected into a new trash image dataset. Second, the lightweight operator involution is incorporated into the feature extraction structure of the algorithm, which allows the feature extraction layer to establish long-distance feature relationships and adaptively extract channel features. In addition, the ability of the model to distinguish similar trash features is strengthened by adding the convolutional block attention module (CBAM) to the enhanced feature extraction network. Finally, the design of the involution residual head structure in the detection head reduces the gradient disappearance and accelerates the convergence of the model loss values allowing the model to perform better classification and regression of the acquired feature layers. In this study, YOLOX-S is chosen as the baseline for each enhancement experiment. The experimental results show that compared with the baseline algorithm, the mean average precision (mAP) of i-YOLOX is improved by 1.47%, the number of parameters is reduced by 23.3%, and the FPS is improved by 40.4%. In practical applications, this improved model achieves accurate recognition of trash in natural scenes, which further validates the generalization performance of i-YOLOX and provides a reference for future domestic trash detection research. [ABSTRACT FROM AUTHOR]

To address the problem of low efficiency for manual detection in the defect detection field for metal shafts, we propose a deep learning defect detection method based on the improved YOLOv5 algorithm. First, we add a Convolutional Block Attention Module (CBAM) mechanism layer to the last layer of the backbone network to improve the feature extraction capability. Second, the neck network introduces the Bi-directional Feature Pyramid Network (BiFPN) module to replace the original Path-Aggregation Network (PAN) structure and enhance the multi-scale feature fusion. Finally, we use transfer learning to pre-train the model and improve the generalization ability of the model. The experimental results show that the method achieves an average accuracy of 93.6% mAP and a detection speed of 16.7 FPS for defect detection on the dataset, which can identify metal shaft surface defects quickly and accurately, and is of reference significance for practical industrial applications. [ABSTRACT FROM AUTHOR]

Nowadays, faces in videos can be easily replaced with the development of deep learning, and these manipulated videos are realistic and cannot be distinguished by human eyes. Some people maliciously use the technology to attack others, especially celebrities and politicians, causing destructive social impacts. Therefore, it is imperative to design an accurate method for detecting face manipulation. However, most of the existing methods adopt single convolutional neural network as the feature extraction module, causing the extracted features to be inconsistent with the human visual mechanism. Moreover, the rich details and semantic information cannot be reflected with single feature, limiting the detection performance. Therefore, this paper tackles the above problems by proposing a novel face manipulation detection method based on a supervised multi-feature fusion attention network (SMFAN). Specifically, the capsule network is used for face manipulation detection, and the SMFAN is added to the original capsule network to extract details of the fake face image. Further, the focal loss is used to realize hard example mining. Finally, the experimental results on the public dataset FaceForensics++ show that the proposed method has better performance. [ABSTRACT FROM AUTHOR]

Human activity recognition (HAR) based on wearable sensors has emerged as a low-cost key-enabling technology for applications such as human–computer interaction and healthcare. In wearable sensor-based HAR, deep learning is desired for extracting human active features. Due to the spatiotemporal dynamic of human activity, a special deep learning network for recognizing the temporal continuous activities of humans is required to improve the recognition accuracy for supporting advanced HAR applications. To this end, a residual multifeature fusion shrinkage network (RMFSN) is proposed. The RMFSN is an improved residual network which consists of a multi-branch framework, a channel attention shrinkage block (CASB), and a classifier network. The special multi-branch framework utilizes a 1D-CNN, a lightweight temporal attention mechanism, and a multi-scale feature extraction method to capture diverse activity features via multiple branches. The CASB is proposed to automatically select key features from the diverse features for each activity, and the classifier network outputs the final recognition results. Experimental results have shown that the accuracy of the proposed RMFSN for the public datasets UCI-HAR, WISDM, and OPPORTUNITY are 98.13%, 98.35%, and 93.89%, respectively. In comparison with existing advanced methods, the proposed RMFSN could achieve higher accuracy while requiring fewer model parameters. [ABSTRACT FROM AUTHOR]

Sonar imaging technology is widely used in the field of marine and underwater monitoring because sound waves can be transmitted in elastic media, such as the atmosphere and seawater, without much interference. In underwater object detection, due to the unique characteristics of the monitored sonar image, and since the target in an image is often accompanied by its own shadow, we can use the relative relationship between the shadow and the target for detection. To make use of shadow-information-aided detection and realize accurate real-time detection in sonar images, we put forward a network based on a lightweight module. By using the attention mechanism with a global receptive field, the network can make the target pay attention to the shadow information in the global environment, and because of its exquisite design, the computational time of the network is greatly reduced. Specifically, we design a ShuffleBlock model adapted to Hourglass to make the backbone network lighter. The concept of CNN dimension reduction is applied to MHSA to make it more efficient while paying attention to global features. Finally, CenterNet's unreasonable distribution method of positive and negative samples is improved. Simulation experiments were carried out using the proposed sonar object detection dataset. The experimental results further verify that our improved model has obvious advantages over many existing conventional deep learning models. Moreover, the real-time monitoring performance of our proposed model is more conducive to the implementation in the field of ocean monitoring. [ABSTRACT FROM AUTHOR]

Human activity recognition (HAR) performs a vital function in various fields, including healthcare, rehabilitation, elder care, and monitoring. Researchers are using mobile sensor data (i.e., accelerometer, gyroscope) by adapting various machine learning (ML) or deep learning (DL) networks. The advent of DL has enabled automatic high-level feature extraction, which has been effectively leveraged to optimize the performance of HAR systems. In addition, the application of deep-learning techniques has demonstrated success in sensor-based HAR across diverse domains. In this study, a novel methodology for HAR was introduced, which utilizes convolutional neural networks (CNNs). The proposed approach combines features from multiple convolutional stages to generate a more comprehensive feature representation, and an attention mechanism was incorporated to extract more refined features, further enhancing the accuracy of the model. The novelty of this study lies in the integration of feature combinations from multiple stages as well as in proposing a generalized model structure with CBAM modules. This leads to a more informative and effective feature extraction technique by feeding the model with more information in every block operation. This research used spectrograms of the raw signals instead of extracting hand-crafted features through intricate signal processing techniques. The developed model has been assessed on three datasets, including KU-HAR, UCI-HAR, and WISDM datasets. The experimental findings showed that the classification accuracies of the suggested technique on the KU-HAR, UCI-HAR, and WISDM datasets were 96.86%, 93.48%, and 93.89%, respectively. The other evaluation criteria also demonstrate that the proposed methodology is comprehensive and competent compared to previous works. [ABSTRACT FROM AUTHOR]

Network traffic anomaly detection is a key step in identifying and preventing network security threats. This study aims to construct a new deep-learning-based traffic anomaly detection model through in-depth research on new feature-engineering methods, significantly improving the efficiency and accuracy of network traffic anomaly detection. The specific research work mainly includes the following two aspects: 1. In order to construct a more comprehensive dataset, this article first starts from the raw data of the classic traffic anomaly detection dataset UNSW-NB15 and combines the feature extraction standards and feature calculation methods of other classic detection datasets to re-extract and design a feature description set for the original traffic data in order to accurately and completely describe the network traffic status. We reconstructed the dataset DNTAD using the feature-processing method designed in this article and conducted evaluation experiments on it. Experiments have shown that by verifying classic machine learning algorithms, such as XGBoost, this method not only does not reduce the training performance of the algorithm but also improves its operational efficiency. 2. This article proposes a detection algorithm model based on LSTM and the recurrent neural network self-attention mechanism for important time-series information contained in the abnormal traffic datasets. With this model, through the memory mechanism of the LSTM, the time dependence of traffic features can be learned. On the basis of LSTM, a self-attention mechanism is introduced, which can weight the features at different positions in the sequence, enabling the model to better learn the direct relationship between traffic features. A series of ablation experiments were also used to demonstrate the effectiveness of each component of the model. The experimental results show that, compared to other comparative models, the model proposed in this article achieves better experimental results on the constructed dataset. [ABSTRACT FROM AUTHOR]

Facial expression methods play a vital role in human–computer interaction and other fields, but there are factors such as occlusion, illumination, and pose changes in wild facial recognition, as well as category imbalances between different datasets, that result in large variations in recognition rates and low accuracy rates for different categories of facial expression datasets. This study introduces RCL-Net, a method of recognizing wild facial expressions that is based on an attention mechanism and LBP feature fusion. The structure consists of two main branches, namely the ResNet-CBAM residual attention branch and the local binary feature (LBP) extraction branch (RCL-Net). First, by merging the residual network and hybrid attention mechanism, the residual attention network is presented to emphasize the local detail feature information of facial expressions; the significant characteristics of facial expressions are retrieved from both channel and spatial dimensions to build the residual attention classification model. Second, we present a locally improved residual network attention model. LBP features are introduced into the facial expression feature extraction stage in order to extract texture information on expression photographs in order to emphasize facial feature information and enhance the recognition accuracy of the model. Lastly, experimental validation is performed using the FER2013, FERPLUS, CK+, and RAF-DB datasets, and the experimental results demonstrate that the proposed method has superior generalization capability and robustness in the laboratory-controlled environment and field environment compared to the most recent experimental methods. [ABSTRACT FROM AUTHOR]

Deep learning has become an essential technique in image steganography. Most of the current deep-learning-based steganographic methods process digital images in the spatial domain. There are problems such as limited embedding capacity and unsatisfactory visual quality. To improve capacity-distortion performance, we develop a steganographic method from the frequency-domain perspective. We propose a module called the adaptive frequency-domain channel attention network (AFcaNet), which makes full use of the frequency features in each channel by a fine-grained manner of assigning weights. We apply this module to the state-of-the-art SteganoGAN, forming an Adaptive Frequency High-capacity Steganography Generative Adversarial Network (AFHS-GAN). The proposed neural network enhances the ability of high-dimensional feature extraction through overlaying densely connected convolutional blocks. In addition to this, a low-frequency loss function is introduced as an evaluation metric to guide the training of the network and thus reduces the modification of low-frequency regions of the image. Experimental results on the Div2K dataset show that our method has a better generalization capability compared to the SteganoGAN, with substantial improvement in both embedding capacity and stego-image quality. Furthermore, the embedding distribution of our method in the DCT domain is more similar to that of the traditional method, which is consistent with the prior knowledge of image steganography. [ABSTRACT FROM AUTHOR]

Deep learning approaches have been recently applied to traffic prediction because of their ability to extract features of traffic data. While convolutional neural networks may improve the predictive accuracy by transiting traffic data to images and extracting features in the images, the convolutional results can be improved by using the global-level representation that is a direct way to extract features. The time intervals are not considered as aspects of convolutional neural networks for traffic prediction. The attention mechanism may adaptively select a sequence of regions and only process the selected regions to better extract features when aspects are considered. In this paper, we propose the attention mechanism over the convolutional result for traffic prediction. The proposed method is based on multiple links. The time interval is considered as the aspect of attention mechanism. Based on the dataset provided by Highways England, the experimental results show that the proposed method can achieve better accuracy than the baseline methods. [ABSTRACT FROM AUTHOR]

Sawn timber is an important component material in furniture manufacturing, decoration, construction and other industries. The mechanical properties, surface colors, textures, use and other properties of sawn timber possesed by different tree species are different. In order to meet the needs of reasonable timber use and product quality of sawn timber products, sawn timber must be identified according to tree species to ensure the best use of materials. In this study, an optimized convolution neural network was proposed to process sawn timber image data to identify the tree species of the sawn timber. The spatial pyramid pooling and attention mechanism were used to improve the convolution layer of ResNet101 to extract the feature vector of sawn timber images. The optimized ResNet (simply called "AM-SPPResNet") was used to identify the sawn timber image, and the basic recognition model was obtained. Then, the weight parameters of the feature extraction layer of the basic model were frozen, the full connection layer was removed, and using support vector machine (SVM) and XGBoost classifier which were commonly used in machine learning to train and learn the 21 × 1024 dimension feature vectors extracted by feature extraction layer. Through a number of comparative experiments, it is found that the prediction model using linear function as the kernel function of support vector machine learning the feature vectors extracted from the improved convolution layer performed best, and the F1 score and overall accuracy of all kinds of samples were above 99%. Compared with the traditional methods, the accuracy was improved by up to 12%. [ABSTRACT FROM AUTHOR]

Due to the limitation of less information in a single image, it is very difficult to generate a high-precision 3D model based on the image. There are some problems in the generation of 3D voxel models, e.g., the information loss at the upper level of a network. To solve these problems, we design a 3D model generation network based on multi-modal data constraints and multi-level feature fusion, named as 3DMGNet. Moreover, 3DMGNet is trained by self-supervised method to achieve 3D voxel model generation from an image. The image feature extraction network (2DNet) and 3D feature extraction network (3D auxiliary network) are used to extract the features of the image and 3D voxel model. Then, feature fusion is used to integrate the low-level features into the high-level features in the 3D auxiliary network. To extract more effective features, each layer of the feature map in feature extraction network is processed by an attention network. Finally, the extracted features generate 3D models by a 3D deconvolution network. The feature extraction of 3D model and the generation of voxelization play an auxiliary role in the training of the whole network for the 3D model generation based on an image. Additionally, a multi-view contour constraint method is proposed, to enhance the effect of the 3D model generation. In the experiment, the ShapeNet dataset is adapted to prove the effect of the 3DMGNet, which verifies the robust performance of the proposed method. [ABSTRACT FROM AUTHOR]