Original Feature Map Research Articles

Segmentation of the left atrial appendage based on fusion attention.

In clinical practice, the morphology of the left atrial appendage (LAA) plays an important role in the selection of LAA closure devices for LAA closure procedures. The morphology determination is influenced by the segmentation results. The LAA occupies only a small part of the entire 3D medical image, and the segmentation results are more likely to be biased towards the background region, making the segmentation of the LAA challenging. In this paper, we propose a lightweight attention mechanism called fusion attention, which imitates human visual behavior. We process the 3D image of the LAA using a method that involves overview observation followed by detailed observation. In the overview observation stage, the image features are pooled along the three dimensions of length, width, and height. The obtained features from the three dimensions are then separately input into the spatial attention and channel attention modules to learn the regions of interest. In the detailed observation stage, the attention results from the previous stage are fused using element-wise multiplication and combined with the original feature map to enhance feature learning. The fusion attention mechanism was evaluated on a left atrial appendage dataset provided by Liaoning Provincial People's Hospital, resulting in an average Dice coefficient of 0.8855. The results indicate that the fusion attention mechanism achieves better segmentation results on 3D images compared to existing lightweight attention mechanisms.

Improved skeleton-based activity recognition using convolutional block attention module

Inferring human activities from the skeletons extracted from activity photos or videos is a fundamental yet important issue in the research community of computer vision. Current skeleton-based activity recognition methods have accomplished this task by extracting features in either spatial or temporal perspective, or combining both views; however, the extracted features lack further after-processing like feature augmentation. Aiming at the above shortcomings, this article presents a new human activity recognition approach based on human skeleton data, named Convolutional Block Attention Module-based Spatial Temporal Convolutional Network (CBAM-STGCN). To enhance the discriminative validity of the generated features, we leverage the attention mechanism to weight feature units within each dimension by introducing two attention modules designed in terms of different aspects into the model architecture. Specifically, the original feature map of an activity sample is adaptively refined by the convolutional block attention module, where two separate attention maps obtained with regard to spatial and channel dimensions are adopted to generate the corresponding attended feature maps. Evaluation results on the Kinetics benchmark dataset showed that the proposed model has improved 1.76 % and 2.43 % according to top-1 and top-5 metrics, respectively, compared to the conventional baseline model. Further ablation experiments proved that both channel and spatial attention modules contribute to the model performance.

Mitigating the Concurrent Interference of Electrode Shift and Loosening in Myoelectric Pattern Recognition Using Siamese Autoencoder Network.

The objective of this work is to develop a novel myoelectric pattern recognition (MPR) method to mitigate the concurrent interference of electrode shift and loosening, thereby improving the practicality of MPR-based gestural interfaces towards intelligent control. A Siamese auto-encoder network (SAEN) was established to learn robust feature representations against random occurrences of both electrode shift and loosening. The SAEN model was trained with a variety of shifted-view and masked-view feature maps, which were simulated through feature transformation operated on the original feature maps. Specifically, three mean square error (MSE) losses were devised to warrant the trained model's capability in adaptive recovery of any given interfered data. The SAEN was deployed as an independent feature extractor followed by a common support vector machine acting as the classifier. To evaluate the effectiveness of the proposed method, an eight-channel armband was adopted to collect surface electromyography (EMG) signals from nine subjects performing six gestures. Under the condition of concurrent interference, the proposed method achieved the highest classification accuracy in both offline and online testing compared to five common methods, with statistical significance (p <0.05). The proposed method was demonstrated to be effective in mitigating the electrode shift and loosening interferences. Our work offers a valuable solution for enhancing the robustness of myoelectric control systems.

TECO: A Unified Feature Map Compression Framework Based on Transform and Entropy.

The massive memory accesses of feature maps (FMs) in deep neural network (DNN) processors lead to huge power consumption, which becomes a major energy bottleneck of DNN accelerators. In this article, we propose a unified framework named Transform and Entropy-based COmpression (TECO) scheme to efficiently compress FMs with various attributes in DNN inference. We explore, for the first time, the intrinsic unimodal distribution characteristic that widely exists in the frequency domain of various FMs. In addition, a well-optimized hardware-friendly coding scheme is designed, which fully utilizes this remarkable data distribution characteristic to encode and compress the frequency spectrum of different FMs. Furthermore, the information entropy theory is leveraged to develop a novel loss function for improving the compression ratio and to make a fast comparison among different compressors. Extensive experiments are performed on multiple tasks and demonstrate that the proposed TECO achieves compression ratios of 2.31 × in ResNet-50 on image classification, 3.47 × in UNet on dark image enhancement, and 3.18 × in Yolo-v4 on object detection while keeping the accuracy of these models. Compared with the upper limit of the compression ratio for original FMs, the proposed framework achieves the compression ratio improvement of 21%, 157%, and 152% on the above models.

A Small Intestinal Stromal Tumor Detection Method Based on an Attention Balance Feature Pyramid.

Small intestinal stromal tumor (SIST) is a common gastrointestinal tumor. Currently, SIST diagnosis relies on clinical radiologists reviewing CT images from medical imaging sensors. However, this method is inefficient and greatly affected by subjective factors. The automatic detection method for stromal tumors based on computer vision technology can better solve these problems. However, in CT images, SIST have different shapes and sizes, blurred edge texture, and little difference from surrounding normal tissues, which to a large extent challenges the use of computer vision technology for the automatic detection of stromal tumors. Furthermore, there are the following issues in the research on the detection and recognition of SIST. After analyzing mainstream target detection models on SIST data, it was discovered that there is an imbalance in the features at different levels during the feature fusion stage of the network model. Therefore, this paper proposes an algorithm, based on the attention balance feature pyramid (ABFP), for detecting SIST with unbalanced feature fusion in the target detection model. By combining weighted multi-level feature maps from the backbone network, the algorithm creates a balanced semantic feature map. Spatial attention and channel attention modules are then introduced to enhance this map. In the feature fusion stage, the algorithm scales the enhanced balanced semantic feature map to the size of each level feature map and enhances the original feature information with the original feature map, effectively addressing the imbalance between deep and shallow features. Consequently, the SIST detection model's detection performance is significantly improved, and the method is highly versatile. Experimental results show that the ABFP method can enhance traditional target detection methods, and is compatible with various models and feature fusion strategies.

Joint Segmentation and Identification Feature Learning for Occlusion Face Recognition.

The existing occlusion face recognition algorithms almost tend to pay more attention to the visible facial components. However, these models are limited because they heavily rely on existing face segmentation approaches to locate occlusions, which is extremely sensitive to the performance of mask learning. To tackle this issue, we propose a joint segmentation and identification feature learning framework for end-to-end occlusion face recognition. More particularly, unlike employing an external face segmentation model to locate the occlusion, we design an occlusion prediction module supervised by known mask labels to be aware of the mask. It shares underlying convolutional feature maps with the identification network and can be collaboratively optimized with each other. Furthermore, we propose a novel channel refinement network to cast the predicted single-channel occlusion mask into a multi-channel mask matrix with each channel owing a distinct mask map. Occlusion-free feature maps are then generated by projecting multi-channel mask probability maps onto original feature maps. Thus, it can suppress the representation of occlusion elements in both the spatial and channel dimensions under the guidance of the mask matrix. Moreover, in order to avoid misleading aggressively predicted mask maps and meanwhile actively exploit usable occlusion-robust features, we aggregate the original and occlusion-free feature maps to distill the final candidate embeddings by our proposed feature purification module. Lastly, to alleviate the scarcity of real-world occlusion face recognition datasets, we build large-scale synthetic occlusion face datasets, totaling up to 980193 face images of 10574 subjects for the training dataset and 36721 face images of 6817 subjects for the testing dataset, respectively. Extensive experimental results on the synthetic and real-world occlusion face datasets show that our approach significantly outperforms the state-of-the-art in both 1:1 face verification and 1:N face identification.

IGly-IDN: Identifying Lysine Glycation Sites in Proteins Based on Improved DenseNet.

Lysine glycation is one of the most significant protein post-translational modifications, which changes the properties of the proteins and causes them to be dysfunctional. Accurately identifying glycation sites helps to understand the biological function and potential mechanism of glycation in disease treatments. Nonetheless, the experimental methods are ordinarily inefficient and costly, so effective computational methods need to be developed. In this study, we proposed the new model called iGly-IDN based on the improved densely connected convolutional networks (DenseNet). First, one hot encoding was adopted to obtain the original feature maps. Afterward, the improved DenseNet was adopted to capture feature information with the importance degrees during the feature learning. According to the experimental results, Acc reaches 66%, and Mathews correlation coefficient reaches 0.33 on the independent testing data set, which indicates that the iGly-IDN can provide more effective glycation site identification than the current predictors.

TSNet: Token Sparsification for Efficient Video Transformer

In the domain of video recognition, video transformers have demonstrated remarkable performance, albeit at significant computational cost. This paper introduces TSNet, an innovative approach for dynamically selecting informative tokens from given video samples. The proposed method involves a lightweight prediction module that assigns importance scores to each token in the video. Tokens with top scores are then utilized for self-attention computation. We apply the Gumbel-softmax technique to sample from the output of the prediction module, enabling end-to-end optimization of the prediction module. We aim to extend our method on hierarchical vision transformers rather than single-scale vision transformers. We use a simple linear module to project the pruned tokens, and the projected result is then concatenated with the output of the self-attention network to maintain the same number of tokens while capturing interactions with the selected tokens. Since feedforward networks (FFNs) contribute significant computation, we also propose linear projection for the pruned tokens to accelerate the model, and the existing FFN layer progresses the selected tokens. Finally, in order to ensure that the structure of the output remains unchanged, the two groups of tokens are reassembled based on their spatial positions in the original feature map. The experiments conducted primarily focus on the Kinetics-400 dataset using UniFormer, a hierarchical video transformer backbone that incorporates convolution in its self-attention block. Our model demonstrates comparable results to the original model while reducing computation by over 13%. Notably, by hierarchically pruning 70% of input tokens, our approach significantly decreases 55.5% of the FLOPs, while the decline in accuracy is confined to 2%. Additional testing of wide applicability and adaptability with other transformers such as the Video Swin Transformer was also performed and indicated its progressive potentials in video recognition benchmarks. By implementing our token sparsification framework, video vision transformers can achieve a remarkable balance between enhanced computational speed and a slight reduction in accuracy.

Substation rotational object detection based on multi-scale feature fusion and refinement

In modern energy systems, substations are the core of electricity transmission and distribution. However, similar appearance and small size pose significant challenges for automatic identification of electrical devices. To address these issues, we collect and annotate the substation rotated device dataset (SRDD). Further, feature fusion and feature refinement network (F3RNet) are constructed based on the classic structure pattern of backbone-neck-head. Considering the similar appearance of electrical devices, the deconvolution fusion module (DFM) is designed to enhance the expression of feature information. The balanced feature pyramid (BFP) is embedded to aggregate the global feature. The feature refinement is constructed to adjust the original feature maps by considering the feature alignment between the anchors and devices. It can generate more accurate feature vectors. To address the problem of sample imbalance between electrical devices, the gradient harmonized mechanism (GHM) loss is utilized to adjust the weight of each sample. The ablation experiments are conducted on the SRDD dataset. F3RNet achieves the best detection performance compared with classical object detection networks. Also, it is verified that the features from global feature maps can effectively recognize the similar and small devices.

Self-supervised dual-head attentional bootstrap learning network for prostate cancer screening in transrectal ultrasound images

Current convolutional neural network-based ultrasound automatic classification models for prostate cancer often rely on extensive manual labeling. Although Self-supervised Learning (SSL) have shown promise in addressing this problem, those data that from medical scenarios contains intra-class similarity conflicts, so using loss calculations directly that include positive and negative sample pairs can mislead training. SSL method tends to focus on global consistency at the image level and does not consider the internal informative relationships of the feature map. To improve the efficiency of prostate cancer diagnosis, using SSL method to learn key diagnostic information in ultrasound images, we proposed a self-supervised dual-head attentional bootstrap learning network (SDABL), including Online-Net and Target-Net. Self-Position Attention Module (SPAM) and adaptive maximum channel attention module (CAAM) are inserted in both paths simultaneously. They captures position and inter-channel attention and of the original feature map with a small number of parameters, solve the information optimization problem of feature maps in SSL. In loss calculations, we discard the construction of negative sample pairs, and instead guide the network to learn the consistency of the location space and channel space by drawing closer to the embedding representation of positive samples continuously. We conducted numerous experiments on the prostate Transrectal ultrasound (TRUS) dataset, experiments show that our SDABL pre-training method has significant advantages over both mainstream contrast learning methods and other attention-based methods. Specifically, the SDABL pre-trained backbone achieves 80.46% accuracy on our TRUS dataset after fine-tuning.

Attention-Guided Instance Segmentation for Group-Raised Pigs.

In the pig farming environment, complex factors such as pig adhesion, occlusion, and changes in body posture pose significant challenges for segmenting multiple target pigs. To address these challenges, this study collected video data using a horizontal angle of view and a non-fixed lens. Specifically, a total of 45 pigs aged 20-105 days in 8 pens were selected as research subjects, resulting in 1917 labeled images. These images were divided into 959 for training, 192 for validation, and 766 for testing. The grouped attention module was employed in the feature pyramid network to fuse the feature maps from deep and shallow layers. The grouped attention module consists of a channel attention branch and a spatial attention branch. The channel attention branch effectively models dependencies between channels to enhance feature mapping between related channels and improve semantic feature representation. The spatial attention branch establishes pixel-level dependencies by applying the response values of all pixels in a single-channel feature map to the target pixel. It further guides the original feature map to filter spatial location information and generate context-related outputs. The grouped attention, along with data augmentation strategies, was incorporated into the Mask R-CNN and Cascade Mask R-CNN task networks to explore their impact on pig segmentation. The experiments showed that introducing data augmentation strategies improved the segmentation performance of the model to a certain extent. Taking Mask-RCNN as an example, under the same experimental conditions, the introduction of data augmentation strategies resulted in improvements of 1.5%, 0.7%, 0.4%, and 0.5% in metrics AP50, AP75, APL, and AP, respectively. Furthermore, our grouped attention module achieved the best performance. For example, compared to the existing attention module CBAM, taking Mask R-CNN as an example, in terms of the metric AP50, AP75, APL, and AP, the grouped attention outperformed 1.0%, 0.3%, 1.1%, and 1.2%, respectively. We further studied the impact of the number of groups in the grouped attention on the final segmentation results. Additionally, visualizations of predictions on third-party data collected using a top-down data acquisition method, which was not involved in the model training, demonstrated that the proposed model in this paper still achieved good segmentation results, proving the transferability and robustness of the grouped attention. Through comprehensive analysis, we found that grouped attention is beneficial for achieving high-precision segmentation of individual pigs in different scenes, ages, and time periods. The research results can provide references for subsequent applications such as pig identification and behavior analysis in mobile settings.

Fine-Grained Image Retrieval via Object Localization

In this paper, a network consisting of an object localization module and a discriminative feature extraction module is designed for fine-grained image retrieval (FGIR). In order to reduce the interference of complex backgrounds, the object localization module is introduced into the network before feature extraction. By selecting the convolutional feature descriptors, the main object is separated from the background, and thus, most of the interference is filtered out. Further, in order to improve the overall performance of the network, a discriminative filter bank is introduced into the network as the local feature detector. Hence, the local discriminative features can be extracted directly from the original feature map. The experimental results based on the CUB-200-2011 and Cars-196 datasets demonstrate that the proposed method can improve the performance of FGIR.

Global and local multi-feature fusion-based active contour model for infrared image segmentation

Infrared (IR) image segmentation technology plays a pivotal role in many urgent fields, such as traffic surveillance, nondestructive detection and autonomous driving. In recent years, active contour model (ACM) has been one of the most commonly used tools for image segmentation, but the precision sharply decreases when dealing with IR images with intensity inhomogeneity. To solve this problem, a new ACM based on global and local multi-feature fusion (GLMF) is proposed in this paper. First of all, the multi-feature fitting maps inside and outside the contour are calculated using the strategy of global and local information fusion. Then, a hybrid signed pressure function (SPF) is designed by combining multiple fitting error maps, which are measured by the similarity between the multi-feature fitting map and the original feature map. Next, a level set formulation (LSF) is constructed using the proposed hybrid SPF and the level set function is thus evolved. Finally, the contour of IR foreground object with visual saliency can be extracted using the zero level set of the converged level set function. Both qualitative and quantitative experiments based on IR datasets verify that the presented ACM has remarkable advantages in terms of accuracy and robustness when compared with other typical ACMs. Our codes are available at https://github.com/MinjieWan/Global-and-Local-Multi-Feature-Fusion-Based-Active-Contour-Model-for-Infrared-Image-Segmentation.

LED Chip Defect Detection Method Based on a Hybrid Algorithm

LED is an extremely important energy-saving lighting products, which has greatly facilitated human life. Meanwhile, it also makes a positive contribution to global carbon neutrality and carbon peaking. Defect detection is a vital part of the production process to control the quality of LED chips. The traditional methods use a microscope for manual visual inspection, which is time-consuming and has inconsistent testing standards, low efficiency, and other deficiencies. To solve these problems, a hybrid algorithm based on geometric computation and a convolutional neural network is proposed for LED chip defect detection. The method takes advantage of the dimensionality reduction of geometric computation to perform coarse detection of defects on preprocessed chip lithography graphs in the form of grid segmentation, which realizes fast coarse screening of large-scale chip samples and reduces postcomputational costs. The convolutional neural network model is used for the secondary fine detection of “suspected defective” chips after geometric coarse screening, and the SPP (spatial pyramid pooling) network model is improved by directly introducing the original feature map into the SPP pooling layer for summation to enhance the global and local feature information of the output feature map. Furthermore, we construct an LED chip image acquisition platform using a high-frequency multimagnification zoom lens, collect training samples of defective chips, and increase the number of samples through image processing techniques. The research introduces the R-CNN, SDD, and YOLO methods to evaluate the superiority of our method in a number of experiments. The experimental results show that our algorithm proposed in this paper has an average precision (AP) of 96.7% for large-scale chip detection with a low defect rate. Compared with other methods, the testing mean average precision (mAP) is 10.39% higher than traditional YOLO v2. The testing mIoU is also 3.63% higher than traditional YOLO v5, the detection speed is also significantly improved, and it has good robustness.

V-BANet: Land cover change detection using effective deep learning technique

Urban development alters landscapes, frequently degrading environmental services and quality of life. High-resolution remote sensing images provide a chance to detect subtle changes in land cover and can capture the features of a ground object. However, traditional approaches usually experience difficulties when processing large and quickly expanding datasets, low levels of automation, limited computational efficiency, and inconsistent identification accuracies and standards brought on by inconsistent operators. Conducting change detection in a more accurate, automated, and standardized manner has become crucial and increasingly difficult due to the quick collection of remote sensing data. Therefore, in this paper, V-Net and Bilateral Attention Network (V-BANet) based deep learning is implemented to segment the landscapes and extract the features from the images. Initially, the bi-temporal images are segmented using V-Net to independently identify the objects in each image. Then spatial and channel attention blocks are employed in Bilateral Attention Network to learn more discriminative features from the images. Finally, the features' relationships are discovered by contrasting the original feature map in one image with the updated feature map in the other. Objective and subjective experiments are performed on a public bi-temporal high-resolution ONERA Satellite Change Detection (OSCD) dataset and the LEVIR-CD dataset. Moreover, the proposed approach reached an accuracy of 99.29% and IoU of 98.31% with the OSCD Dataset and 99.42% accuracy and 98.83% IoU with the LEVIR-CD Dataset. The experimental outcomes with each specified dataset demonstrated that the suggested methodology outperformed several state-of-the-art techniques and produced superior results.

Safety Helmet Wearing Detection Model Based on Improved YOLO-M

In recent years, construction accidents have occurred frequently. The safety guarantee measures for construction personnel are thrown into sharp focus. Wearing helmets is one of the most important requirements to protect the safety of construction personnel, and the detection of wearing safety helmets has become necessary. For the problems of the existing helmet wearing detection algorithm such as too many parameters, substantial detection interferences, and low detection accuracy, in this paper a helmet wearing detection model YOLO-M is proposed. Firstly, MobileNetv3 is adopted as the backbone network of YOLOv5s for the feature extraction, which can reduce the number of model parameters and model size. Secondly, a residual edge is introduced in the feature fusion. The original feature map information is fused during feature fusion, and the detection ability of small targets is enhanced. At last, by changing the connection between CAM and SAM, a new attention module BiCAM is designed. The comparison experiments show that the detection accuracy of YOLO-M is 2.22% higher than YOLOv5s, and the model parameter quantity is reduced to 3/4 of YOLOv5s. Under the same detection conditions, the detection speed of YOLO-M is better than the other models, which meets the accuracy requirements of helmet detection in the construction scene.

GeSANet: Geospatial-Awareness Network for VHR Remote Sensing Image Change Detection

The characteristics of very high resolution (VHR) remote sensing images (RSIs) have higher spatial resolution inherently, and are easier to obtain globally compared with hyperspectral images (HSIs), making it possible to detect small-scale land cover changes in multiple applications. RSI change detection (RSI-CD) based on deep learning has been paid attention to and become a frontier research field in recent years, and is currently facing two challenging problems: The first is high dependence on registration between bi-temporal images caused by high spatial resolution; The other is high pseudo-change information response caused by low spectral resolution. In order to address the above-mentioned two problems, a novel RSI-CD framework called Geospatial-Awareness Network (GeSANet) based on the geospatial Position Matching Mechanism (PMM) with multi-level adjustment and the geo-spatial Content Reasoning Mechanism (CRM) with diverse pseudo-change information filtering is proposed. First of all, the PMM assigns independent two-dimensional offset coordinates to each position in the previous temporal image, afterwards, bilinear interpolation is employed to obtain the subpixel feature value after the offset, and the sparse results based on the difference are transmitted to the next level prediction to realize multi-level geospatial correction. The CRM extracts global features from the corrected sparse feature map in terms of dimensions, implementing effective discriminant feature extraction on basis of the original feature map in a stepwise refinement manner through the cross-dimension exchange mechanism, to filter out various pseudo-change information as well as maintain real change information. Comparison experiments with five recent SOTA methods are carried out on two popular datasets with diverse changes, the results show that the proposed method has good robustness and validity for multi-temporal RSI-CD. In particular, it has a strong comparative advantage in detecting small entity changes and edge details. The source code of the proposed framework can be downloaded from https://github.com/zxylnnu/GeSANet.

Pyramid attention object detection network with multi-scale feature fusion

With the development of deep learning, object detection has made substantial progress. However, when the object to be detected in the image is small or partially occluded, the detection network often fails to detect it successfully. We propose a multi-scale feature fusion pyramid attention module, which effectively combines the global average pooling results of multiple scales with the upper features in the residual blocks of the feature extraction network to obtain more spatial context information in the original feature map. We added the multi-scale feature fusion pyramid attention module proposed in this paper based on YoloV3 and conducted experiments on the PASCALL VOC and MS COCO datasets. The experimental results show that the attention module can effectively help the network detect small objects and accurately detect partially occlusion objects.

Remote sensing aircraft object detection algorithm based on memory CenterNet

In order to solve the problem of poor detection accuracy of CenterNet in remote sensing aircraft with dense targets and complex backgrounds, an improved memory CenterNet object detection algorithm is proposed. First, the multiscale receptive field module and residual attention module are introduced to improve ResNet50, which is used as a new backbone feature extraction network. It has a strong feature extraction ability and detection speed at the same time. Different dilated ratios are used in the multiscale receptive field to expand the receptive field and extract multiscale features, and the multiscale target features are highlighted through the attention mechanism. Then the last set of residual networks in ResNet50 is deleted, and the residual attention network is introduced and stacked in the same way. The feature parameters of different channels in the residual attention network are calculated by two-dimensional cosine transform and weighted with the original feature map, and the target features are highlighted to enhance the ability to locate the target at the end of the network. Second, three times deconvolution is replaced by a memory feature fusion module. The shallow detail information and the deep semantic information are fully fused. The key information in the output feature map is further enriched, which is beneficial to the improvement of model detection performance. Finally, the experimental results show that the average accuracy of remote sensing aircraft dataset is 88.46%, which is 18.21% higher than that of the original algorithm and effectively improves the detection accuracy of remote sensing aircraft targets.

Surface Defect Detection of Heat Sink Based on Lightweight Fully Convolutional Network

The surface defect detection plays an important role in industrial production, and directly affects production efficiency and product quality. In this paper, we focus on the surface defect detection of heat sink and propose a mothed based on a Ghost-SE light U-Net (GSLU-Net). The GSLU-Net is a novel combination of the lightweight convolution module, self-attention mechanism, and fully convolutional network. It has a symmetrical architecture inspired by the U-Net. By introducing the Ghost module which can generate feature maps with cheap operations, we reduce the computation cost while maintaining high accuracy. The Ghost module uses an ordinary convolution with small kernels to obtain original feature maps, then a depth-wise convolution to generate more feature maps. The SE block is also introduced to improve the representational power of the network and further improve accuracy. It can adaptively recalibrate channel-wise feature responses at a slight additional computational cost. Fewer down-sampling layers and more skip connections enable this network to retain more location information and have stronger detection capabilities for tiny targets than previous fully convolutional networks. The ablation experiments validate that our improvements have played their due role. Then the comparative experiments demonstrate that our network outperforms the state-of-the-art fully convolutional networks. The GSLU-Net reaches an accuracy of 97.96% at a speed of 14.115ms per image on the dataset of heat sink surface defect, achieving a balance of efficiency and accuracy. The GSLU-Net also outperforms other networks on open datasets, further demonstrating the superiority of the GSLU-Net.