Search

The technology for classifying and recognizing underwater targets is crucial for supporting underwater acoustic information countermeasures. The research focus is on the extraction and classification of features of underwater targets. Researchers have conducted an in-depth study from various perspectives. Due to the influence of ambient noise and various operating conditions of different targets, the signal-to-noise ratio of underwater acoustic signals is generally meager. Additionally, the components of these signals are complex, often requiring specific signal pre-processing techniques such as signal enhancement and decomposition. In current methods, there is a primary focus on extracting and classifying features of underwater acoustic signals after multi-step preprocessing. However, these methods do not effectively integrate feature extraction and classification. To address these limitations, we propose a new model called Depthwise Separable Convolutional Attention (DSCA) and use multiple instances of DSCA to construct a neural network, which we call DSCANet. The DSCANet integrates feature extraction and target classification for underwater acoustic targets. The 'target' in our work should be mentioned as it refers to underwater sources of sound. The structure of DSCANet is unified and simple, and no specific pre-processing of the underwater acoustic signal is necessary. The DSCANet is trained and validated on ShipsEars, an open dataset. It achieves a classification accuracy of 93%, which is the highest in the contrast experiment. [ABSTRACT FROM AUTHOR]

Hyperspectral image classification is crucial for a wide range of applications, including environmental monitoring, precision agriculture, and mining, due to its ability to capture detailed spectral information across numerous wavelengths. However, the high dimensionality and complex spatial-spectral relationships in hyperspectral data pose significant challenges. Deep learning, particularly Convolutional Neural Networks (CNNs), has shown remarkable success in automatically extracting relevant features from high-dimensional data, making them well-suited for handling the intricate spatial-spectral relationships in hyperspectral images.This study presents a hybrid approach for hyperspectral image classification, combining 3D Depthwise Separable Convolution (3D DSC) and Depthwise Squeeze-and-Excitation Network (DSENet). The 3D DSC efficiently captures spatial-spectral features, reducing computational complexity while preserving essential information. The DSENet further refines these features by applying channel-wise attention, enhancing the model's ability to focus on the most informative features. To assess the performance of the proposed hybrid model, extensive experimental studies were carried out on four commonly utilized HSI datasets, namely HyRANK-Loukia and WHU-Hi (including HongHu, HanChuan, and LongKou). As a result of the experimental studies, the HyRANK-Loukia achieved an accuracy of 90.9%, marking an 8.86% increase compared to its previous highest accuracy. Similarly, for the WHU-Hi datasets, HongHu achieved an accuracy of 97.49%, reflecting a 2.11% improvement over its previous highest accuracy; HanChuan achieved an accuracy of 97.49%, showing a 2.4% improvement; and LongKou achieved an accuracy of 99.79%, providing a 0.15% improvement compared to its previous highest accuracy. Comparative analysis highlights the superiority of the proposed model, emphasizing improved classification accuracy with lower computational costs. [ABSTRACT FROM AUTHOR]

The typical procedure for Earth Observation Nanosatellites involves the sequential steps of image capture, onboard storage, and subsequent transmission to the ground station. This approach places significant demands on onboard resources and encounters bandwidth limitations; moreover, the captured images may be obstructed by cloud cover. Many current deep-learning methods have achieved reasonable accuracy in cloud detection. However, the constraints posed by nanosatellites specifically in terms of memory and energy present challenges for effective onboard Artificial Intelligence implementation. Hence, we propose an optimized tiny Machine learning model based on the U-Net architecture, implemented on STM32H7 microcontroller for real-time cloud coverage prediction. The optimized U-Net architecture on the embedded device introduces Depthwise Separable Convolution for efficient feature extraction, reducing computational complexity. By utilizing this method, coupled with encoder and decoder blocks, the model optimizes cloud detection for nanosatellites, showcasing a significant advancement in resource-efficient onboard processing. This approach aims to enhance the university nanosatellite mission, equipped with an RGB Gecko imager camera. The model is trained on Sentinel 2 satellite images due to the unavailability of a large dataset for the payload imager and is subsequently evaluated on gecko images, demonstrating the generalizability of our approach. The outcome of our optimization approach is a 21% reduction in network parameters compared to the original configuration and maintaining an accuracy of 89%. This reduction enables the system to allocate only 61.89 KB in flash memory effectively, resulting in improvements in memory usage and computational efficiency. [ABSTRACT FROM AUTHOR]

Copyright of Journal of Test & Measurement Technology is the property of Publishing Center of North University of China and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

SUMMARY: The use of deep learning techniques to identify grape leaf diseases relies on large, high‐quality datasets. However, a large number of images occupy more computing resources and are prone to pattern collapse during training. In this paper, a depth‐separable multifeature generative adversarial network (DMFGAN) was proposed to enhance grape leaf disease data. First, a multifeature extraction block (MFEB) based on the four‐channel feature fusion strategy is designed to improve the quality of the generated image and avoid the problem of poor feature learning ability of the adversarial generation network caused by the single‐channel feature extraction method. Second, a depth‐based D‐discriminator is designed to improve the discriminator capability and reduce the number of model parameters. Third, SeLU activation function was substituted for DCGAN activation function to overcome the problem that DCGAN activation function was not enough to fit grape leaf disease image data. Finally, an MFLoss function with a gradient penalty term is proposed to reduce the mode collapse during the training of generative adversarial networks. By comparing the visual indicators and evaluation indicators of the images generated by different models, and using the recognition network to verify the enhanced grape disease data, the results show that the method is effective in enhancing grape leaf disease data. Under the same experimental conditions, DMFGAN generates higher quality and more diverse images with fewer parameters than other generative adversarial networks. The mode breakdown times of generative adversarial networks in training process are reduced, which is more effective in practical application. [ABSTRACT FROM AUTHOR]

The rise of mobile communication, low-power chips, and the Internet of Things has made smartwatches increasingly popular. Equipped with inertial measurement units (IMUs), these devices can recognize user activities through artificial intelligence (AI) analysis of sensor data. However, most existing AI-based activity recognition algorithms require significant computational power and storage, making them unsuitable for low-power devices like smartwatches. Additionally, discrepancies between training data and real-world data often hinder model generalization and performance. To address these challenges, we propose LIMUNet and its smaller variant LIMUNet-Tiny—lightweight neural networks designed for human activity recognition on smartwatches. LIMUNet utilizes depthwise separable convolutions and residual blocks to reduce computational complexity and parameter count. It also incorporates a dual attention mechanism specifically tailored to smartwatch sensor data, improving feature extraction without sacrificing efficiency. Experiments on the PAMAP2 and LIMU datasets show that the LIMUNet improves recognition accuracy by 2.9% over leading lightweight models while reducing parameters by 88.3% and computational load by 58.4%. Compared to other state-of-the-art models, LIMUNet achieves a 9.6% increase in accuracy, with a 60% reduction in parameters and a 57.8% reduction in computational cost. LIMUNet-Tiny further reduces parameters by 75% and computational load by 80%, making it even more suitable for resource-constrained devices. [ABSTRACT FROM AUTHOR]

Deep learning has greatly promoted intelligent fault diagnosis for bearings and has led to the development of many intelligent fault recognition methods based on deep neural networks. However, bearing fault signals in industrial applications are susceptible to noise interference, and intelligent recognition methods for bearing faults under strong noise interference are lacking. An intelligent fault pattern recognition model termed IFPR-MDL, which is based on a Transformer assisted by depthwise separable convolutional (DSC) and bidirectional long short-term memory (BiLSTM) networks, is designed to identify bearing faults under strong noise conditions. In this model, multiscale features are first extracted by a multiscale feature extraction module designed on the basis of DSC; then, contextual information is further extracted by a BiLSTM, and information from different scales is integrated by Transformer encoders. The DSC network facilitates model parameter reduction, the BiLSTM network has expertise in mining temporal features, and the Transformer network excels at capturing long-range feature correlations. This model fully utilises the advantages of these models and complements them to improve the anti-interference ability of bearing fault pattern recognition. Tests on several public datasets show that the proposed model has higher recognition accuracy, stronger generalisation ability, and noise resistance in strong noise situations. [ABSTRACT FROM AUTHOR]

Deep learning has significantly advanced the field of medical image segmentation. However, the complexity of network structures often leads to high computational demands, limiting their practical efficiency. To enhance the efficiency of image segmentation, this paper introduces an innovative, concise, and lightweight deep learning network. First, to reduce model complexity, we replaced the attention mechanism in the traditional vision transformer (ViT) structure with a shift operation, creating the ShiftViT architecture. This substitution significantly decreased computation and the number of parameters while preserving model performance. Second, to retain and enhance fine-grained features and facilitate more precise information transfer across different layers, we employed a full-scale progressive skip connection strategy. This approach effectively integrates multi-scale feature information, further enhancing model performance. Additionally, to further reduce network complexity, inspired by the independence of probabilities, we opted for depth-wise separable convolution over traditional convolution. This enhances the relative independence between layers. Together, these modifications achieved superior segmentation results on both the Synapse and Automated Cardiac Diagnostic Challenge (ACDC) datasets compared to mainstream models, with substantial advantages in terms of computational efficiency and parameter count. The proposed approach represents an effective solution for medical image applications with limited computational resources and holds great promise for clinical practice. [ABSTRACT FROM AUTHOR]

Emotion recognition plays an important role in human life and healthcare. The EEG has been extensively researched as an objective indicator of intense emotions. However, current existing methods lack sufficient analysis of shallow and deep EEG features. In addition, human emotions are complex and variable, making it difficult to comprehensively represent emotions using a single-modal signal. As a signal associated with gaze tracking and eye movement detection, Eye-related signals provide various forms of supplementary information for multimodal emotion analysis. Therefore, we propose a Pseudo-Siamese Pyramid Network (PSPN) for multimodal emotion analysis. The PSPN model employs a Depthwise Separable Convolutional Pyramid (DSCP) to extract and integrate intrinsic emotional features at various levels and scales from EEG signals. Simultaneously, we utilize a fully connected subnetwork to extract the external emotional features from eye-related signals. Finally, we introduce a Pseudo-Siamese network that integrates a flexible cross-modal dual-branch subnetwork to collaboratively utilize EEG emotional features and eye-related behavioral features, achieving consistency and complementarity in multimodal emotion recognition. For evaluation, we conducted experiments on the DEAP and SEED-IV public datasets. The experimental results demonstrate that multimodal fusion significantly improves the accuracy of emotion recognition compared to single-modal approaches. Our PSPN model achieved the best accuracy of 96.02% and 96.45% on the valence and arousal dimensions of the DEAP dataset, and 77.81% on the SEED-IV dataset, respectively. Our code link is: https://github.com/Yinyanyan003/PSPN.git. [ABSTRACT FROM AUTHOR]

Image correction and deblurring have always posed significant challenges in the field of image processing, and the results of license plate image recognition with low structural similarity index measure (SSIM) and peak signal-to-noise ratio (PSNR) will significantly affect the use of vehicles. To address these problems, we propose a real-time algorithm based on deep learning, which can simultaneously perform image spatial correction and deblurring in the same network, known as 4 × STN + DSC 1 5 + DSK. First, a 15-layer depthwise separable convolutional neural network is designed as the basic network for high-speed image deblurring, and four spatial transformer network (STN) modules are integrated into the basic network to achieve image spatial correction capability. Second, the network is divided into three dense blocks, in which dense residual connections are used to improve the PSNR. Finally, a bottleneck structure is created to improve the real-time performance of image reconstruction that can simultaneously perform image space correction and deblurring. The test results on the CCPD dataset demonstrate that the 4 × STN + DSC 1 5 + DSK algorithm is advanced, with PSNR reaching 30.14 dB, SSIM reaching 0.911, and image reconstruction only 0.12 s. The correction and deblurring modules can promote each other, jointly improving the image reconstruction ability of the network, that is, the network achieves more advanced image reconstruction performance while maintaining high real-time performance. [ABSTRACT FROM AUTHOR]

Accurate segmentation of skin lesions is essential for physicians to screen in dermoscopy images. However, they commonly face three main limitations: difficulty in accurately processing targets with coarse edges; frequent challenges in recovering detailed feature data; and a lack of adequate capability for the effective amalgamation of multi-scale features. To overcome these problems, we propose a skin lesion segmentation network (SFCC Net) that combines an attention mechanism and a redundancy reduction strategy. The initial step involved the design of a downsampling encoder and an encoder composed of Receptive Field (REFC) Blocks, aimed at supplementing lost details and extracting latent features. Subsequently, the Spatial-Frequency-Channel (SF) Block was employed to minimize feature redundancy and restore fine-grained information. To fully leverage previously learned features, an Up-sampling Convolution (UpC) Block was designed for information integration. The network’s performance was compared with state-of-the-art models on four public datasets. Experimental results demonstrate significant improvements in the network’s performance. On the ISIC datasets, the proposed network outperformed D-LKA Net by 4.19%, 0.19%, and 7.75% in F1, and by 2.14%, 0.51%, and 12.20% in IoU. The frame rate (FPS) of the proposed network when processing skin lesion images underscores its suitability for real-time image analysis. Additionally, the network’s generalization capability was validated on a lung dataset. [ABSTRACT FROM AUTHOR]

Aiming at the problems of poor quality of restored images and large number of model parameters in the current occluded face image restoration, a lightweight face image restoration model based on multi-scale feature fusion with improved U-Net, LM-UNET, was proposed. Firstly, the original convolution was replaced by a depthwise separable convolution to enhance the feature expression ability of the model for different channels and contextual information. Secondly, a multi-scale feature attention fusion module was designed in the jump connection to fully fuse the information of different scale features, and the embedded residual block reduced the semantic gap between features to improve the repair accuracy of the model. Finally, a positional attention module was introduced to enhance the salient information of the face image, and improve the model’s effective extraction ability of facial positional pixel information of the model. The algorithm was trained, validated and tested on the occluded face dataset MFD generated based on the CK+ dataset, and the PSNR of the repaired image reached 30.49 dB and SSIM reached 96.85%. The experimental results of comparing the model with the other models show that the model has better image quality and visual effect for restoration of the face in the presence of occlusion.

ObjectiveThe monitoring of livestock grazing in natural pastures is a key aspect of the transformation and upgrading of large-scale breeding farms. In order to meet the demand for large-scale farms to achieve accurate real-time detection of a large number of sheep, a high-precision and easy-to-deploy small-target detection model: CSD-YOLOv8s was proposed to realize the real-time detection of small-targeted individual sheep under the high-altitude view of the unmanned aerial vehicle (UAV).MethodsFirstly, a UAV was used to acquire video data of sheep in natural grassland pastures with different backgrounds and lighting conditions, and together with some public datasets downloaded formed the original image data. The sheep detection dataset was generated through data cleaning and labeling. Secondly, in order to solve the difficult problem of sheep detection caused by dense flocks and mutual occlusion, the SPPFCSPC module was constructed with cross-stage local connection based on the you only look once (YOLO)v8 model, which combined the original features with the output features of the fast spatial pyramid pooling network, fully retained the feature information at different stages of the model, and effectively solved the problem of small targets and serious occlusion of the sheep, and improved the detection performance of the model for small sheep targets. In the Neck part of the model, the convolutional block attention module (CBAM) convolutional attention module was introduced to enhance the feature information capture based on both spatial and channel aspects, suppressing the background information spatially and focusing on the sheep target in the channel, enhancing the network's anti-jamming ability from both channel and spatial dimensions, and improving the model's detection performance of multi-scale sheep under complex backgrounds and different illumination conditions. Finally, in order to improve the real-time and deploy ability of the model, the standard convolution of the Neck network was changed to a lightweight convolutional C2f_DS module with a changeable kernel, which was able to adaptively select the corresponding convolutional kernel for feature extraction according to the input features, and solved the problem of input scale change in the process of sheep detection in a more flexible way, and at the same time, the number of parameters of the model was reduced and the speed of the model was improved.Results and DiscussionsThe improved CSD-YOLOv8s model exhibited excellent performance in the sheep detection task. Compared with YOLO, Faster R-CNN and other classical network models, the improved CSD-YOLOv8s model had higher detection accuracy and frames per second (FPS) of 87 f/s in the flock detection task with comparable detection speed and model size. Compared with the YOLOv8s model, Precision was improved from 93.0% to 95.2%, mAP was improved from 91.2% to 93.1%, and it had strong robustness to sheep targets with different degree of occlusion and different scales, which effectively solved the serious problems of missed and misdetection of sheep in the grassland pasture UAV-on-ground sheep detection task due to the small sheep targets, large background noise, and high degree of densification. misdetection serious problems. Validated by the PASCAL VOC 2007 open dataset, the CSD-YOLOv8s model proposed in this study improved the detection accuracy of 20 different objects, including transportation vehicles, animals, etc., especially in sheep detection, the detection accuracy was improved by 9.7%.ConclusionsThis study establishes a sheep dataset based on drone images and proposes a model called CSD-YOLOv8s for detecting grazing sheep in natural grasslands. The model addresses the serious issues of missed detections and false alarms in sheep detection under complex backgrounds and lighting conditions, enabling more accurate detection of grazing livestock in drone images. It achieves precise detection of targets with varying degrees of clustering and occlusion and possesses good real-time performance. This model provides an effective detection method for detecting sheep herds from the perspective of drones in natural pastures and offers technical support for large-scale livestock detection in breeding farms, with wide-ranging potential applications.

With the rapid development of intelligent manufacturing technology, the structure of industrial equipment has become more sophisticated, resulting in frequent equipment failures. However, traditional anomaly diagnosis methods suffer the issues of insufficient accuracy and are always unable to identify anomalies in time. To solve the above problems, a deep learning-based equipment anomaly diagnosis method in industrial production scenarios is proposed in this paper. Specifically, a combination model based on Selective Kernel (SK) convolution blocks is designed to improve the accuracy of anomaly identification; a lightweight model based on Depthwise Separable Convolution (DSC) and Attention Mechanism (AM) is proposed to improve the timeliness of anomaly identification; an anomaly analysis and intelligent diagnosis framework is established and implemented to automatically complete anomaly identification tasks for different equipment and practical problems. Finally, we conduct extensive experiments to validate the effectiveness of our methods. The experimental results show that the accuracy of our anomaly identification method is as high as 99.07%, and the calculation amount of the lightweight model is reduced by 84.95% compared to the baseline model. [ABSTRACT FROM AUTHOR]

Existing deep convolutional neural networks that rely on large datasets typically require images with high resolution and deep neural network models trained and called upon to improve accuracy of image recognition and classification. It is needed to use lightweight model to adapt to such low-power devices. However, lightweight small models are limited in their ability to classify and recognize small-sized images with low-resolution and are constrained by the number of parameters in the model and unable to perform deep-level feature extraction, since the low-resolution indicates small sample information. In the intelligent interaction in digital media, capturing, storing, transmitting, and computing high-resolution, high-precision images incur high power consumption and operating costs. When deploying an image recognition system on the client-side of IoT devices, it is difficult to meet the hardware requirements of high storage space and fast computation speed. It is also challenging to directly use high-resolution image data for model fine-tuning and training, and the size and parameter updates of the model are also limited by the storage and operating capacity of the hardware facilities. We proposed a low-power image recognition framework consists data pre-processing part and lightweight modeling architecture part. The data pre-processing method for image data based on an Auto-Encoder that filters R, G, B color channel data using a resolution filter to realize data compression, that is Downscaling large input data to a smaller size, thus to address the limitations of low-power deep learning model deployment and training. Based on the resolution filter, a channel normalization method is proposed to perform batch normalization on each channel dimension to encode the original image data at the same size and improve the mean squared error discrimination of the image data. And the lightweight model uses a depth-separable convolutional neural network and two kinds of blocks: one with batch normalization and the other without, EdgeNet. The architecture makes it possible to deploy more suitable for IoT device. The proposed framework achieves only a small precision loss within permission, but improves the forward inference speed of the model, and reduce the memory storage to 8.7 MB. [ABSTRACT FROM AUTHOR]

The finishing mill is a critical link in the hot rolling process, influencing the final product's quality, and even economic efficiency. The distribution box of the finishing mill plays a vital role in power transmission and distribution. However, harsh operating conditions can frequently lead to distribution box damage and even failure. To diagnose faults in the distribution box promptly, a fault diagnosis network model is constructed in this paper. This model combines depthwise separable convolution and Bi-LSTM. Depthwise separable convolution and Bi-LSTM can extract both spatial and temporal features from signals. This structure enables comprehensive feature extraction and fully utilizes signal information. To verify the diagnostic capability of the model, five types of data are collected and used: the pitting of tooth flank, flat-headed sleeve tooth crack, gear surface crack, gear tooth surface spalling, and normal conditions. The model achieves an accuracy of 97.46% and incorporates a lightweight design, which enhances computational efficiency. Furthermore, the model maintains approximately 90% accuracy under three noise conditions. Based on these results, the proposed model can effectively diagnose faults in the distribution box, and reduce downtime in engineering. [ABSTRACT FROM AUTHOR]

In response to the issue of short-term fluctuations in photovoltaic (PV) output due to Cloud movement, this paper proposes a method for forecasting short-term PV output based on a Depthwise Separable Convolution Visual Geometry Group (DSCVGG) and a Deep Gate Recurrent Neural Network (DGN). Initially, a cloud motion prediction model is constructed using a DSCVGG, which achieves edge recognition and motion prediction of clouds by replacing the previous convolution layer of the pooling layer in VGG with a depthwise separable convolution. Subsequently, the output results of the DSCVGG network, along with historical PV output data, are introduced into a Deep Gate Recurrent Unit Network (DGN) to establish a PV output prediction model, thereby achieving precise prediction of PV output. Through experiments on actual data, the Mean Absolute Error (MAE) and Mean Squared Error (MSE) of our model are only 2.18% and 5.32 x 10-5, respectively, which validates the effectiveness, accuracy, and superiority of the proposed method. This provides new insights and methods for improving the stability of PV power generation. [ABSTRACT FROM AUTHOR]

Deep learning research has gained significant popularity recently, finding applications in various domains such as image preprocessing, segmentation, object recognition, and semantic analysis. Deep learning has gradually replaced traditional algorithms such as color-based methods, contour-based methods, and motion-based methods. In the context of hand gesture recognition, traditional algorithms heavily rely on depth information for accuracy, but their performance is often subpar. This paper introduces a novel approach using a deep neural network for hand gesture recognition, requiring only a single complementary metal oxide semiconductor (CMOS) camera to operate amidst complex backgrounds. The neural network design incorporates depthwise separable convolutional layers, dividing the model into segmentation and recognition components. As our proposed single-stage model, we avoid the use of the whole model and thus reduce the number of weights and calculations. Additionally, in the training phase, the data augmentation and iterative training strategy further increase recognition accuracy. The results show that the proposed work uses little parameter usage while still having a higher gesture recognition rate than the other works. [ABSTRACT FROM AUTHOR]

Copyright of Telecommunications Science is the property of Beijing Xintong Media Co., Ltd. and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

In this paper, we propose a simple and general lightweight approach named AL-RE2 for text matching models, and conduct experiments on three well-studied benchmark datasets across tasks of natural language inference and paraphrase identification. Firstly, we explore the feasibility of dimensional compression of word embedding vectors using principal component analysis, and then analyze the impact of the information retained in different dimensions on model accuracy. Considering the balance between compression efficiency and information loss, we choose 128 dimensions to represent each word and make the model params 1.6M. Finally, the feasibility of applying depthwise separable convolution instead of standard convolution in the field of text matching is analyzed in detail. The experimental results show that our model’s inference speed is at least 1.5 times faster and it has 42.76% fewer parameters compared to similarly performing models, while its accuracy on the SciTail dataset of is state-of-the-art among all lightweight models. [ABSTRACT FROM AUTHOR]

[Objective] The monitoring of livestock grazing in natural pastures is a key aspect of the transformation and upgrading of large-scale breeding farms. In order to meet the demand for large-scale farms to achieve accurate real-time detection of a large number of sheep, a high-precision and easy-to-deploy small-target detection model: CSD-YOLOv8s was proposed to realize the real-time detection of small-targeted individual sheep under the high-altitude view of the unmanned aerial vehicle (UAV). [Methods] Firstly, a UAV was used to acquire video data of sheep in natural grassland pastures with different backgrounds and lighting conditions, and together with some public datasets downloaded formed the original image data. The sheep detection dataset was gener‐ ated through data cleaning and labeling. Secondly, in order to solve the difficult problem of sheep detection caused by dense flocks and mutual occlusion, the SPPFCSPC module was constructed with cross-stage local connection based on the you only look once (YOLO)v8 model, which combined the original features with the output features of the fast spatial pyramid pooling network, fully re‐ tained the feature information at different stages of the model, and effectively solved the problem of small targets and serious occlu‐ sion of the sheep, and improved the detection performance of the model for small sheep targets. In the Neck part of the model, the con‐ volutional block attention module (CBAM) convolutional attention module was introduced to enhance the feature information capture based on both spatial and channel aspects, suppressing the background information spatially and focusing on the sheep target in the channel, enhancing the network's anti-jamming ability from both channel and spatial dimensions, and improving the model's detection performance of multi-scale sheep under complex backgrounds and different illumination conditions. Finally, in order to improve the real-time and deploy ability of the model, the standard convolution of the Neck network was changed to a lightweight convolutional C2f_DS module with a changeable kernel, which was able to adaptively select the corresponding convolutional kernel for feature ex‐ traction according to the input features, and solved the problem of input scale change in the process of sheep detection in a more flexi‐ ble way, and at the same time, the number of parameters of the model was reduced and the speed of the model was improved. [Results and Discussions] The improved CSD-YOLOv8s model exhibited excellent performance in the sheep detection task. Compared with YOLO, Faster R-CNN and other classical network models, the improved CSD-YOLOv8s model had higher detection accuracy and frames per second (FPS) of 87 f/s in the flock detection task with comparable detection speed and model size. Compared with the YOLOv8s model, Precision was improved from 93.0% to 95.2%, mAP was improved from 91.2% to 93.1%, and it had strong robust‐ ness to sheep targets with different degree of occlusion and different scales, which effectively solved the serious problems of missed and misdetection of sheep in the grassland pasture UAV-on-ground sheep detection task due to the small sheep targets, large back‐ ground noise, and high degree of densification. misdetection serious problems. Validated by the PASCAL VOC 2007 open dataset, the CSD-YOLOv8s model proposed in this study improved the detection accuracy of 20 different objects, including transportation vehi‐ cles, animals, etc., especially in sheep detection, the detection accuracy was improved by 9.7%. [Conclusions] This study establishes a sheep dataset based on drone images and proposes a model called CSD-YOLOv8s for detecting grazing sheep in natural grasslands. The model addresses the serious issues of missed detections and false alarms in sheep detection under complex backgrounds and lighting conditions, enabling more accurate detection of grazing livestock in drone images. It achieves precise detection of targets with varying degrees of clustering and occlusion and possesses good real-time performance. This model provides an effective detection method for detecting sheep herds from the perspective of drones in natural pastures and offers technical support for large-scale livestock detection in breeding farms, with wide-ranging potential applications. [ABSTRACT FROM AUTHOR]

Concrete materials and structures are extensively used in transformation infrastructure and they usually bear cracks during their long-term operation. Detecting cracks using deep-learning algorithms like YOLOv3 (You Only Look Once version 3) is a new trend to pursue intelligent detection of concrete surface cracks. YOLOv3 is a typical deep-learning algorithm used for object detection. Owing to its generality, YOLOv3 lacks specific efficiency and accuracy in identifying concrete surface cracks. An improved algorithm based on YOLOv3, specialized in the rapid and accurate identification of concrete surface cracks is worthy of investigation. This study proposes a tailored deep-learning algorithm, termed MDN-YOLOv3 (MDN: multi-dilated network), of which the MDN is formulated based on three retrofit techniques, and it provides a new backbone network for YOLOv3. The three specific retrofit techniques are briefed: (i) Depthwise separable convolution is utilized to reduce the size of the backbone network; (ii) The dilated-down sampling structure is proposed and used in the backbone network to achieve multi-scale feature fusion; and (iii) The convolutional block attention module is introduced to enhance feature extraction ability. Results show that the proposed MDN-YOLOv3 is 97.2% smaller and 41.5% faster than YOLOv3 in identifying concrete surface cracks, forming a lightweight & efficient YOLOv3 algorithm for intelligently identifying concrete surface cracks. [ABSTRACT FROM AUTHOR]

ABSTRACTAddressing accuracy and computational complexity challenges in hyperspectral image classification for small sample and multi-species scenarios, we developed DSC-DC, a lightweight convolutional neural network. This model is based on Depthwise Separable Convolution and Dilated Convolution and was trained using the Teakettle Experimental Forest dataset (USA). In this study, DSC-DC achieved an overall accuracy (OA) of 99.83%, average accuracy (AA) of 99.64%, and Kappa coefficient of 0.9996. Compared to Support Vector Machine and K-Nearest Neighbors, it demonstrated markedly higher OA (3.88% to 7.55%) and AA (30.71% to 34.09%). Compared to Inception-V3, ResNet50, and MSR-3DCNN, DSC-DC marginally outperformed in accuracy (OA: 0.06% to 0.31%; AA: 0.32% to 3.64%) while reducing the training time by 3.5, 5, and 35 times, and the prediction time by 2, 3, and 17 times, respectively. Moreover, DSC-DC exhibits slightly superior accuracy and efficiency compared to a 5-layer optimal structure of the 3D-CNN model. The application of the DSC-DC model to the hyperspectral dataset from the Jiepai branch of the Gaofeng State Owned Forest Farm in the Guangxi province, China, further demonstrated the reliability, versatility, and practical potential of this model. This study provides a reliable and efficient reference solution for small-sample and multi-tree classification tasks.

In unstructured environments, dense grape fruit growth and the presence of occlusion cause difficult recognition problems, which will seriously affect the performance of grape picking robots. To address these problems, this study improves the YOLOX-Tiny model and proposes a new grape detection model, YOLOX-RA, which can quickly and accurately identify densely growing and occluded grape bunches. The proposed YOLOX-RA model uses a 3 × 3 convolutional layer with a step size of 2 to replace the focal layer to reduce the computational burden. The CBS layer in the ResBlock_Body module of the second, third, and fourth layers of the backbone layer is removed, and the CSPLayer module is replaced by the ResBlock-M module to speed up the detection. An auxiliary network (AlNet) with the remaining network blocks was added after the ResBlock-M module to improve the detection accuracy. Two depth-separable convolutions (DSC) are used in the neck module layer to replace the normal convolution to reduce the computational cost. We evaluated the detection performance of SSD, YOLOv4 SSD, YOLOv4-Tiny, YOLO-Grape, YOLOv5-X, YOLOX-Tiny, and YOLOX-RA on a grape test set. The results show that the YOLOX-RA model has the best detection performance, achieving 88.75 % mAP, a recognition speed of 84.88 FPS, and model size of 17.53 MB. It can accurately detect densely grown and shaded grape bunches, which can effectively improve the performance of the grape picking robot.

Abstract Background Oral mucosal diseases are similar to the surrounding normal tissues, i.e., their many non-salient features, which poses a challenge for accurate segmentation lesions. Additionally, high-precision large models generate too many parameters, which puts pressure on storage and makes it difficult to deploy on portable devices. Methods To address these issues, we design a non-salient target segmentation model (NTSM) to improve segmentation performance while reducing the number of parameters. The NTSM includes a difference association (DA) module and multiple feature hierarchy pyramid attention (FHPA) modules. The DA module enhances feature differences at different levels to learn local context information and extend the segmentation mask to potentially similar areas. It also learns logical semantic relationship information through different receptive fields to determine the actual lesions and further elevates the segmentation performance of non-salient lesions. The FHPA module extracts pathological information from different views by performing the hadamard product attention (HPA) operation on input features, which reduces the number of parameters. Results The experimental results on the oral mucosal diseases (OMD) dataset and international skin imaging collaboration (ISIC) dataset demonstrate that our model outperforms existing state-of-the-art methods. Compared with the nnU-Net backbone, our model has 43.20% fewer parameters while still achieving a 3.14% increase in the Dice score. Conclusions Our model has high segmentation accuracy on non-salient areas of oral mucosal diseases and can effectively reduce resource consumption.

A Facial Expression Recognition (FER) method based on Conv3D-ConvLSTM-SEnet in an online education environment is proposed to address the issue of low accuracy in current classroom expression recognition (FER) methods. Firstly, ConvLSTM is used to integrate the local feature extraction ability of CNN and the temporal modeling ability of LSTM, and based on this, the rich spatial features of the image are characterized. Then, by introducing Depth Separated Convolution (DSC) to change the number of output channels and separate each channel, the Feature Maps (F-M) are sequentially concatenated to obtain a multi-channel output F-M. Finally, based on ConvLSTM and SEnet modules, a Conv3D ConvLSTM-based FER model was proposed by redistributing the extracted abstract features and basic texture features, achieving high accuracy classroom FER. The proposed classroom FER method and the other five methods were compared and analyzed through simulation experiments using the CK + , FER2013 and JAFFE datasets, respectively. The results indicate that the proposed method achieves the highest accuracy, precision, recall, and F1 score, with improvements of at least 1.85%, 2.41%, 1.18% and 2.05%, respectively, compared to the other five methods on the FER2013 dataset. The proposed method can perform FER on students in online classroom teaching, detect their emotions in real-time, and help teachers make better course teaching adjustments based on this. [ABSTRACT FROM AUTHOR]

Introduction: Precise fault diagnosis is crucial for enhancing the reliability and lifespan of the flexible converter valve equipment. To address this issue, depthwise separable convolution, bidirectional gate recurrent unit, and multi- head attention module (DSC-BiGRU-MAM) based fault diagnosis approach is proposed. Methods: By DSC and BiGRU operation, the model can capture the correlation between local features and temporal information when processing sequence data, thereby enhancing the representation ability and predictive performance of the model for complex sequential data. In addition, by incorporating a multi-head attention module, the proposed method dynamically learns important information from different time intervals and channels. The proposed MAM continuously stimulates fault features in both time and channel dimensions during training, while suppressing fault independent expressions. As a result, it has made an important contribution to improving the performance of the fault diagnosis model. Results and Discussion: Experimental results demonstrate that the proposed method achieves higher accuracy compared to existing methods, with an average accuracy of 95.45%, average precision of 88.67%, and average recall of 89.03%. Additionally, the proposed method has a moderate number of model parameters (17,626) and training time (935 s). Results indicate that the proposed method accurately diagnoses faults in flexible converter valve equipment, especially in real-world situations with noise overlapping signals. [ABSTRACT FROM AUTHOR]

Convolutional Neural Networks (CNNs) have achieved remarkable results in many application fields. However, these CNNs have a large number of network parameters, thereby consuming a lot of computation and storage resources. This makes CNNs unable to be effectively applied to these platforms with limited storage and computation resources. To address this issue, this paper proposes a new compact convolution module called DSC-Ghost-Conv, which combines the advantages of both depthwise separable convolution (DSC) and Ghost convolution module (Ghost-Conv). DSC-Ghost-Conv replaces the standard convolution used in the Ghost convolution module with depthwise separable convolution so as to reduce resource costs of the Ghost convolution module. DSC-Ghost-Conv can be used as a plug-and-play component to implement ordinary convolutional layers in typical CNNs such as VGG-16, ResNet-50 and GoogleNet. Experimental results on the MNIST and CIFAR-10 datasets show that implementing the ordinary convolutional layers of CNNs with DSC-Ghost-Conv not only obtains the competitive performance to typical CNNs, but also greatly reduces the number of network parameters and floating point operations (FLOPs) of CNNs. This demonstrates that the proposed DSC-Ghost-Conv can effectively reduce the resource costs of CNNs. [ABSTRACT FROM AUTHOR]

Copyright of Journal of Computer Engineering & Applications is the property of Beijing Journal of Computer Engineering & Applications Journal Co Ltd. and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

Single image super‐resolution (SISR) has gained significant attention in image processing and computer vision, driven by deep learning‐based models like convolutional neural networks (CNN). Yet, the resource‐intensive nature of these models poses challenges when deploying them on edge devices. To address this issue, resource‐constrained models need to be developed. While recent models like the information distillation network (IDN), the information multi‐distillation network (IMDN), the residual feature distillation network (RFDN), and so on, have attempted to reduce parameters and computational complexity, further optimization remains vital. Therefore, this paper presents an approach to enhancing the efficiency and lightweight nature of the SISR. We introduce a novel lightweight SR model by building upon the RFDN architecture, the winner of the AIM2020 and NTIRE2022 SR challenges. The proposed depthwise channel attention network (DWCAN) model makes some key changes to RFDN. First, it replaces the main residual feature distillation block (RFDB) with a depthwise channel attention block (DWCAB). Additionally, DWCAN includes a shallow residual block (SRB) with depthwise separable convolution (DW) and a channel attention (CA) block. The primary goal of our work is to significantly reduce model parameters, computational operations, inference time, and memory size while maintaining or improving a peak signal‐to‐noise ratio (PSNR) of 29 dB. The experimental results demonstrate the effectiveness of the proposed model. By applying our modifications, we achieve a notable reduction in model complexity, leading to an improved PSNR of 29.07 dB, up from RFDN's 29.04 dB on a diverse 2 K resolution (DIV2K) dataset. This underscores the potential of our lightweight model to balance computational efficiency and SR quality. Additionally, the proposed work is essential for the metaverse for two key reasons: (1) Enhancing visual quality by adding complex details to textures and objects, making the digital world feel more like reality. (2) Ensuring device compatibility across a range of gadgets, from smartphones to VR headsets, optimizing the metaverse experience for all users. In short, a lightweight single image super‐resolution model for image reconstruction is proposed in this paper. [ABSTRACT FROM AUTHOR]

In human-robotcollaboration tasks, hand key point detection provides target point coordinates for the robotic arm.A2J(Anchor-to-Joint) is a representative method of key point detection using anchor points.A2J can achieve better detection effect with depth map input, but it has insufficient ability to acquire global features. In this study, a GLF(Global-Local Feature Fusion) module is designed to fuse the shallow and deep features of the backbone network. In order to improve the detection speed, the backbone network of A2J is replaced with ShuffleNetv2 and reformed, and 3×3 depth separable convolution is replaced with 5×5 depth separable convolution to increase the sensitivity field and effectively improve the backbone network's ability to extract global features. ECA(Efficient Channel Attention) is introduced into the anchor weight estimation branch to improve the network's attention to important anchor points. The results of training and testing on the mainstream data sets ICVL and NYU show that the average error of the proposed method is reduced by 0.09 mm and 0.15 mm, respectively, compared with A2J.The detection rate of 151 frame·s-1 is realized on GTX1080Ti graphics card, which fully meets the real-time requirement of man-machine collaboration task. [ABSTRACT FROM AUTHOR]

The classification of vegetation types worldwide plays a significant role in studies involving remote sensing. This method, used notably in agriculture, aids producers in devising more efficient agricultural management models. It relies on satellite and aircraft technologies to analyze agricultural lands. Nevertheless, the recent emergence of unmanned aerial vehicles (UAVs) has introduced faster and more cost-effective alternatives to traditional satellite and aircraft systems. These UAVs provide higher resolution images, leading to a shift in remote sensing practices. For deep learning in UAV-based image classification, convolutional neural network (CNN) techniques are commonly employed due to their advantageous features and exceptional extraction capabilities. This study proposes a hybrid approach based on CNN, combining 2D depthwise separable convolution (DSC) with a conventional 2D CNN and a Squeeze-and-Excitation network (SENet). The inclusion of SENet aims to boost classification performance without significantly increasing the overall parameter count. By integrating 2D DSC, computational costs and the number of trainable parameters are notably reduced. The multipath network structure's core purpose is to amplify the extracted features from UAV-derived images. The effectiveness of this multipath hybrid approach was evaluated using an orthophoto from Harran University's campus captured by a UAV. The primary goal was to distinguish between mature and immature lavender plants. The results indicate a high accuracy, with immature lavender plants classified at 99.77% accuracy and mature lavender plants at 95.15% accuracy. These findings from experimental studies demonstrate the high effectiveness of our hybrid method in identifying immature lavender plants. [ABSTRACT FROM AUTHOR]

This research introduces a novel approach to Alzheimer's disease detection by combining Xception's efficiency with machine learning classifiers, notably XGBoost. The hybrid model strategically uses Xception for feature extraction and integrates machine learning algorithms to enhance early detection accuracy, leveraging depthwise separable convolution for reduced computational complexity. Addressing imbalanced data, the study incorporates SMOTE, showcasing the hybrid model's effectiveness. Before SMOTE, the model achieved 72.89% accuracy and a 74.35% F1 score, outperforming the non-hybrid Xception model. Post-SMOTE, accuracy increases to 86.75%, and the F1 score to 86.84%, demonstrating substantial improvement without excessive computational demands. In comparison, the non-hybrid Xception model exhibits 78.71% accuracy and a 78.27% F1 score after SMOTE, emphasizing the pronounced enhancement achieved by the hybrid model. The Kaggle-derived dataset, totaling 6400 images, undergoes meticulous preprocessing, acknowledging dataset-specific constraints on generalizability. Emphasizing the importance of addressing data imbalance for robust classification, the hybrid model offers a promising solution for accurate and efficient Alzheimer's disease detection. This study contributes valuable insights to the field, showcasing the potential of innovative hybrid models to address complex healthcare challenges while balancing accuracy and computational efficiency. [ABSTRACT FROM AUTHOR]

The computing and storage resources of the industrial Internet of things (IIoT) equipment in the coal mine are limited, making it vulnerable to illegal network intrusion, causing sensitive data leakage or malicious tampering, and threatening the safety of coal mine production. Precise recognition of coal mine IIoT equipment can achieve effective management and maintenance of equipment operation, improve equipment safety and protection capabilities. However, existing equipment recognition algorithms suffer from complex feature construction, high memory and computing requirements, making it difficult to deploy in resource limited coal mine IIoT equipment. In order to solve the above problems, a coal mine IIoT equipment recognition model is proposed. Firstly, the model performs traffic segmentation, irrelevant field removal, deduplication, and fixed length field truncation operations on traffic data that supports TCP/IP protocol transmission. The model then converts it to IDX format for storage. Secondly, the model uses convolutional block attention module (CBAM) to optimize depthwise separable convolu-tion(DSC). A lightweight DSC-CBAM model is constructed to filter Non-IIoT equipment. Thirdly, the Wasserstein generative adversarial network with gradient penalty (WGAN-GP) is used to expand the data of coal mine IIoT equipment with less traffic, achieving the goal of balancing offset traffic data. Finally, multi-scale feature fusion (MFF) technology is introduced on the basis of DSC-CBAM to capture shallow global feature information, and Mish activation function is added to improve model training stability. The MFF-DSC-CBAM-Mish (MDCM) model is established to achieve precise recognition of coal mine IIoT equipment. The experimental results show that the model has a fast convergence speed, with accuracy, recall, precision, and F1 score all reaching 99.98%. The model has a small number of parameters, which can accurately and efficiently recognize IIoT equipment in coal mines.

The methodology to generate new views for an object from provided input object view is called Novel View Synthesis (NVS). Humans imagine novel views through prior knowledge gathered through their lifetime. NVS-GAN predicts the novel views through computation. Literature survey reveals that there are limited NVS models with low Trainable Parameter Count (TPC) and low model size. Also, a study on the effect of different loss functions on NVS models was lacking. Lowering the TPC indicates less computational steps for the model to predict the output, therefore desirable. Combined with a low model size, the proposed model will become more suitable for deployment in diverse devices having limited resources for computation. Application of right combination of loss functions yield better accuracy. To address these research gaps, NVS-GAN is proposed. NVS-GAN is a Generative Adversarial Network (GAN) approach which yields NVS-Generator which performs NVS. NVS-Generator incorporates identity skip connections, bilinear sampling module, Depthwise Separable Convolution (DSC) as design features and results in low TPC, model size. In addition to discriminator loss, NVS-GAN is trained with different combinations of loss functions i.e. Mean Absolute Error (MAE) loss, Structural Similarity Index Measure (SSIM) loss, Huber loss on chair and car objects of ShapeNet dataset. The performance of NVS-Generator on test set measured in terms of MAE and SSIM is tabulated and analysed. The performance is compared with existing NVS models. The proposed NVS-GAN experiment recorded reduction in NVS-Generator TPC in 37 %–54.6 % range and reduction in model size between 37.2 % and 47.6 % range. NVS-Generator reduced MAE upto 55 % and improved SSIM upto 4 % than existing models. Summarily, NVS-GAN increased model performance and made the model “lightweight”.

Low-light images are susceptible to poor visibility, low contrast, and additive noise. Although various techniques can enhance brightness to a reasonable extent, they inevitably amplify noise and cause color distortion, halo artifacts, and other forms of deterioration. Most low-light image enhancement (LLIE) methods utilize multiple algorithms or models to remove the corresponding deterioration types while brightening the image. However, an optimization strategy can be difficult, primarily because each sub-model has its own set of hyperparameters that must be tuned. To address this problem, we propose an end-to-end network, called CBMD-Net, which concurrently brightens and mitigates the deterioration in low-light images. Specifically, our network is developed on a U-shaped backbone with several stacked pyramid depthwise separable convolution (PDC) modules to accurately extract meaningful spatial features from the network. Additionally, a new upsampling module is presented to efficiently restore the spatial resolution of the feature maps. Finally, we discovered that the latent representations at the network bottleneck contain relevant information regarding the input image, including deteriorated features that can have repercussions on image quality. Therefore, we introduce an intermediate loss function at the bottleneck to mitigate the additive noise and undesirable artifacts. Comprehensive experiments demonstrate that the developed network can obtain satisfactory results from both subjective and objective comparisons against state-of-the-art methods.

Purposes Aiming at the problems of large computation and high reference quantity in existing traffic sign detection algorithms, in this paper we proposed a lightweight traffic sign detection algorithm based on improved YOLOv7. Methods The algorithm is divided into four parts: input, backbone network of feature extraction, neck network of feature fusion, and head network of target prediction. Large kernel convolution was introduced into the backbone network, which increases the effective receptive field and improves the ability of feature extraction. The detection of neck fusion coordinate attention, random pooling, and other methods can not only build channel attention and capture accurate position, but also improve the generalization ability of the network. In addition, a comprehensive depth-separable convolution module was proposed to extract image features by reducing the number of parameters and the radical sign. Results Experimental results show that the detection accuracy of the proposed algorithm on the CCTSDB2021 data set reaches 93.13%, and mAP also reaches 87.59%, which is a great improvement with respect to other methods of the same type. The network achieves a high accuracy rate under the condition of low parameter number and calculation amount, which can not only accurately capture the location information of traffic signs, but also achieve a high accuracy rate. At the same time, it can accurately predict the traffic signs.

In this paper, we present our result of research in video deepfake detection. We built a deepfake detection system to detect whether a video is a deepfake or real. The deepfake detection algorithm still struggle in providing a sufficient accuracy values, especially in challenging deepfake dataset. Our deepfake detection system utilized spatiotemporal feature that extracted using Video Vision Transformer (ViViT). The main contribution of our research is providing a deepfake detection system that based on ViViT architecture and using landmark area images for the input of the system. Our system extracted the feature from a number of spatial features. The spatial feature was extracted using Depthwise Separable Convolution (DSC) block combined with Convolution Block Attention Module (CBAM) from tubelet. The tubelet was a representation of facial landmark area that was extracted from the input video. In our system, we used 25 facial landmark area for an input video. In our experiment we used Celeb-DF version 2 dataset because it is considered to be a challenging deepfake dataset. We conducted augmentation to the dataset, so we obtained 8335 videos for training set, 390 videos for validation set, and 1123 videos for testing set. We trained our deepfake detection system using Adam optimizer, with learning rate of 10–4 and 100 epoch. From the experiment, we obtained the accuracy score of 87.18% and F1 score of 92.52%. We also conducted the ablation study to display the effect of each part of our model to the overall system performance. From this research, we obtained that by using landmark area images, our ViViT based deepfake detection system had a good performance in detecting deepfake videos.

The fast and accurate series arc fault (SAF) identification method and its hardware implementation are the key to the development of arc fault circuit interrupter (AFCI) or arc fault detection device (AFDD). The SAF experiments under household multi-branch circuit conditions were conducted. And a novel SAF identification model based on lightweight one-dimensional (1-D) convolutional neural network was proposed. First, the main-circuit current signal was used as the input of the model. The 1-D convolutional layers and 1-D maximum pooling layers of the model were used to extract the features of the current signal. The fully connected neural network (FCNN) was used to identify whether or not there is a SAF in the circuit and determine the branch-circuit where the fault is located. Second, the second to fourth standard convolutional layers of the model were improved by using depthwise separable convolution, and the batch normalization layers were added to the model, so as to realize the optimal design of the model. Finally, the model was deployed to an embedded device and its performance was tested. When the sampling frequency is higher than 5 kHz, the accuracy of fault identification and fault line selection of the model in the embedded device is higher than 98.05% and 99.11%, respectively. The average runtime of single identification is 5.26 ms. It meets the technical requirements of household AFCI or AFDD.

The rise of mobile communication, low-power chips, and the Internet of Things has made smartwatches increasingly popular. Equipped with inertial measurement units (IMUs), these devices can recognize user activities through artificial intelligence (AI) analysis of sensor data. However, most existing AI-based activity recognition algorithms require significant computational power and storage, making them unsuitable for low-power devices like smartwatches. Additionally, discrepancies between training data and real-world data often hinder model generalization and performance. To address these challenges, we propose LIMUNet and its smaller variant LIMUNet-Tiny—lightweight neural networks designed for human activity recognition on smartwatches. LIMUNet utilizes depthwise separable convolutions and residual blocks to reduce computational complexity and parameter count. It also incorporates a dual attention mechanism specifically tailored to smartwatch sensor data, improving feature extraction without sacrificing efficiency. Experiments on the PAMAP2 and LIMU datasets show that the LIMUNet improves recognition accuracy by 2.9% over leading lightweight models while reducing parameters by 88.3% and computational load by 58.4%. Compared to other state-of-the-art models, LIMUNet achieves a 9.6% increase in accuracy, with a 60% reduction in parameters and a 57.8% reduction in computational cost. LIMUNet-Tiny further reduces parameters by 75% and computational load by 80%, making it even more suitable for resource-constrained devices.

The parameter and computational requirements of neural networks have increased with the rapid development of deep learning technology, making it increasingly difficult to deploy neural networks on hardware platforms with limited resources. This paper proposed a method to address the challenge of deploying deep learning models on small edge computing platforms. The method utilized a depthwise separable network model applied to a custom dataset. This method carried out model compression on the software end by employing steps as transfer learning, sensitivity analysis,and pruning quantization. On the hardware end, it analyzed and designed a pipeline hardware accelerator suitable for FPGA with limited resources. Experimental results demonstrated that after software-based network compression optimization, this quantized deployment model achieved a high accuracy rate of 94.60%, with a lower single-inference fixed-point operation count of 16.64M and a parameter count of 0.079M. Furthermore, after hardware resource optimization, the pipeline deployment on a domestic FPGA development board achieved an inference frame rate of 366FPS and a computational efficiency of 8.57GOPS/W. This research provides a solution for high-performance deployment of deep learning models on small-scale edge computing platforms. [ABSTRACT FROM AUTHOR]

Copyright of Journal of Mine Automation is the property of Industry & Mine Automation Editorial Department and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

Copyright of Journal of South-Central Minzu University (Natural Science Edition) is the property of Journal of South-Central Minzu University (Natural Science Edition) Editorial Office and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

The critical nature of passive ship-radiated noise recognition for military and economic security is well-established, yet its advancement faces significant obstacles due to the complex marine environment. The challenges include natural sound interference and signal distortion, complicating the extraction of key acoustic features and ship type identification. Addressing these issues, this study introduces DWSTr, a novel method combining a depthwise separable convolutional neural network with a Transformer architecture. This approach effectively isolates local acoustic features and captures global dependencies, enhancing robustness against environmental interferences and signal variability. Validated by experimental results on the ShipsEar dataset, DWSTr demonstrated a notable 96.5\% recognition accuracy, underscoring its efficacy in accurate ship classification amidst challenging conditions. The integration of these advanced neural architectures not only surmounts existing barriers in noise recognition but also offers computational efficiency for real-time analysis, marking a significant advancement in passive acoustic monitoring and its application in strategic and economic contexts. [ABSTRACT FROM AUTHOR]

In many scenarios, edge devices perform computations for applications such as target detection and tracking, multimodal sensor fusion, low-light image enhancement, and image segmentation. There is an increasing trend of deploying and running multiple different network models on one hardware platform, but there is a lack of generic acceleration architectures that support standard convolution (CONV), depthwise separable CONV, and deconvolution (DeCONV) layers in such complex scenarios. In response, this paper proposes a more versatile dynamically reconfigurable CNN accelerator with a highly unified computing scheme. The proposed design, which is compatible with standard CNNs, lightweight CNNs, and CNNs with DeCONV layers, further improves the resource utilization and reduces the gap of efficiency when deploying different models. Thus, the hardware balance during the alternating execution of multiple models is enhanced. Compared to a state-of-the-art CNN accelerator, Xilinx DPU B4096, our optimized architecture achieves resource utilization improvements of 1.08× for VGG16 and 1.77× for MobileNetV1 in inference tasks on the Xilinx ZCU102 platform. The resource utilization and efficiency degradation between these two models are reduced to 59.6% and 63.7%, respectively. Furthermore, the proposed architecture can properly run DeCONV layers and demonstrates good performance. [ABSTRACT FROM AUTHOR]

Accurately monitoring the assembly sequence of mechanical parts is significant for reducing assembly process errors and improving assembly accuracy. To monitor the newly assembled parts of mechanical assemblies from multiple views, this paper proposes a multi-view change detection network based on feature fusion and feature refinement with depthwise separable convolution (DWFR Net) for mechanical assembly images. Based on the BIT (Bitemporal Image Transformer) network, the DWFR Net adds a feature fusion structure to the convolutional neural network module to reduce the loss of shallow features due to the increase in network depth. Then, a feature refinement module is designed to refine the features extracted by the convolutional neural network and make the Q of the Transformer decoder more accurate. The depthwise separable convolution with a convolution kernel of 5 × 5 is used in the feature fusion structure and feature refinement module. It reduces the number of parameters added by the large convolution kernel while increasing the receptive field. Finally, the DWFR Net is evaluated on a mechanical assembly image change detection dataset and a public dataset LEVIR-CD. The experimental results show that the F1 of the DWFR Net can reach 95.60% on the mechanical assembly image change detection dataset and 90.12% on the LEVIR-CD dataset, and the latter is 0.81% higher than that of the BIT network. [ABSTRACT FROM AUTHOR]

Remote sensing is one of the major domains witnessing the increasingly significant interest in Hyperspectral image (HSI) classification. One recent approach achieving great success in HSI classification is deep learning. However, most current HSI classification methods are performed on small datasets, using overlapped sub-regions both for training and testing, causing information leakage. Data leaks lead to improper ideal classification results. Thus, more research work is still needed to boost classification performance for non-leaking methods. This paper proposes an Encoder-Decoder deep learning-based method that can further improve the accuracy of HSI classification. The method contains two main parts: an encoder and a decoder. The encoder consists of depthwise separable convolution, squeeze, and excitation blocks based on the popular MobileNetV3 deep learning network, and the decoder is inherited from the U-Net deep learning model. This encoder-decoder model aims to take into account the relationships between neighboring pixels when performing HSI pixel classification. Since the pixel distribution in an HSI dataset is usually imbalanced, the focal loss function is employed during the model's training process to counterbalance the weight difference between the pixel classes and avoid overfitting the classes with few pixels. For performance assessment, we carried out experiments on the AeroRIT dataset, consisting of non-overlapping training, validation, and test sets. We achieved an overall accuracy of 95.54% and a mean intersection over union score of 83.53%, setting state-of-the-art results on this dataset. [ABSTRACT FROM AUTHOR]