Search

Leukocytes are mainly composed of neutrophils, basophils, eosinophils, monocytes, and lymphocytes. The number and proportion of different types of leukocytes correspond to different diseases, so an accurate segmentation of each type of leukocyte is important for the diagnosis of disease. However, the acquisition of blood cell images can be affected by external environmental factors, which can lead to variable light and darkness, complex backgrounds, and poorly characterized leukocytes. To address the problem of complex blood cell images collected under different environments and the lack of obvious leukocyte features, a leukocyte segmentation method based on improved U-net is proposed. First, adaptive histogram equalization-retinex correction was introduced for data enhancement to make the leukocyte features in the blood cell images clearer. Then, to address the problem of similarity between different types of leukocytes, convolutional block attention module is added to the four skip connections of U-net to focus the features from spatial and channel aspects, so that the network can quickly locate the high-value information of features in different channels and spaces. It avoids the problem of large amount of repeated computation of low-value information, prevents overfitting, and improves the training efficiency and generalization ability of the network. Finally, to solve the problem of class imbalance in blood cell images and to better segment the cytoplasm of leukocytes, a loss function combining focal loss and Dice loss is proposed. We use the BCISC public dataset to verify the effectiveness of the proposed method. The segmentation of multiple leukocytes using the method of this paper can achieve 99.53% accuracy and 91.89% mIoU. The experimental results show that the method achieves good segmentation results for lymphocytes, basophils, neutrophils, eosinophils, and monocytes. [ABSTRACT FROM AUTHOR]

In computer vision, image recognition is one of the classic tasks. Currently, with the foundation of big data and advanced hardware, deep learning has achieved high accuracy. However, deep learning often fails to perform well when faced with a small number of samples. Therefore, few-shot learning has become a key technology to solve this problem. The learning paradigm of few-shot learning is different from that of deep learning. It aims to learn a universal representation from multiple training categories, used for recognition in new categories. Each few-shot learning training instance consists of a group of images and an unlabeled sample. The goal is to enable the model to perform well in recognizing new categories. To achieve this, the model needs to extract representative and highly generalizable features that enable the correct recognition of new category samples. To address the problem of small sample space being unable to describe enough dataset's semantic features, we propose the attention mechanism and earth mover's distance for few-shot learning (AMEMD-FSL) method. First, we fuse the attention mechanism (AM) to deep learning to help the model extract more semantically rich features. Then we use the earth mover's distance (EMD) metric method to calculate the distance between samples, enabling better classification. Finally, we combine the deep-learning residual network and AMEMD to perform few-shot learning. We validate our algorithm on the Caltech-UCSD Birds-200-2011 dataset and the few-shot public dataset mini-ImageNet, which comes from the DeepMind team. The experimental results demonstrate that we have proposed an end-to-end and effective method in the field of few-shot image classification. [ABSTRACT FROM AUTHOR]

In the context of facial expression recognition (FER), a lot of methods focus on feature extraction of the whole facial image, but sometimes it is equally important to concentrate on local area features that embody stronger emotions. This paper proposes an FER method based on the facial part attention mechanism (FPA). We adopt the attention mechanism to extract emotional rich local area features, which are complementary to the whole facial features for better FER. First, this paper proposes a cluster-based facial landmarks selection method to select facial landmarks that have commonality and best reflect local area emotions. Then, we design an FPA convolutional neural network, which consists of two parts. The first part is the object network, which is used to extract the whole facial features; the second part is the part attention network, and the attention component is used to extract the local part features. Finally, the two parts are merged to train the final classifier. The method was tested on the Real-World Affective Faces database and the recognition accuracy of 87.26% was obtained, which proved the effectiveness of the method. [ABSTRACT FROM AUTHOR]

The two-branch real-time semantic segmentation network can quickly acquire low-level details and high-level semantics. However, the large contextual gap between them results in adverse impact on their fusion, and limits the further improvement of real-time segmentation accuracy. This paper proposes a tripartite real-time semantic segmentation network with scene commonality (TriSCNet) to address this problem. First, we add a parallel scene commonality branch based on the current two-branch architecture to learn intrinsic common features in similar street scene images, such as the spatial location distribution of various objects and the internal connections between them at the semantic level. Further, with the guidance of commonality, we propose an external branch attention module to enrich and enhance the feature information of traditional two branches. Finally, we utilize an alignment and selective fusion module to correct the misaligned context in the semantic branch and highlight the essential spatial information in the detailed branch. Our proposed TriSCNet achieves an excellent trade-off between accuracy and speed, yielding 77.9% mIOU at 67.2 FPS on Cityscapes test set and 75.8% mIOU at 127.4 FPS on CamVid test set, respectively. [ABSTRACT FROM AUTHOR]

This paper proposes a network model based on a three-stream network and improved attention mechanism for blind image quality assessment (TSAIQA). The inputs of the three streams are the distorted image, the pseudoreference image obtained by the improved generative adversarial network (GAN), and the gradient map of the distorted image. The distorted image stream focuses on the holistic quality-related features, the pseudoreference image stream is used to supplement the lost features due to distortion, and the gradient stream explicitly extracts the quality-related structural features. In addition, spatial and channel attention mechanisms combining first- and second-order information are proposed, and the improved attention mechanisms are applied to the three-stream network to optimize spatial and channel-level features effectively. Finally, the three-stream fusion features are input to the quality regression network to predict the image quality. To demonstrate the effectiveness of the proposed model, experiments are conducted on four classical IQA databases and two new large-scale databases. The experimental results show that the results of our TSAIQA model outperform the most advanced IQA methods and confirm the effectiveness of the proposed network structure and attention mechanisms. [ABSTRACT FROM AUTHOR]

Aiming at the detection difficulties in camouflage target detection, such as the high similarity between the target and its background, serious damage to the edge, and strong concealment of the target, a camouflage target detection algorithm YOLO of camouflage object detection based on strong semantic information and feature fusion is proposed. First, the attention mechanism convolutional block attention module (CBAM) is constructed to highlight the important channel features and target spatial locations to further aggregate rich semantic information from the high-level feature map. Then the atrous spatial pyramid pooling module is constructed to repeatedly sample the multiscale feature maps to expand the receptive field of the neural network, reduce feature sparsity in the process of convolution, and ensure dense features and multiscale contextual semantic information enter the feature fusion module. Finally, the attention skip-connections are constructed based on the CBAM module for fusing the original feature maps extracted by the backbone network to the corresponding detection outputs so as to eliminate the redundant features as well as enrich the target information of the network outputs. In order to fully verify the performance of the proposed algorithm, a camouflage target detection dataset named strong camouflage efficiency target dataset (SCETD) is constructed. Experimental results on SCETD show that the precision and recall of the proposed algorithm achieve 96.1% and 87.1%, respectively. The AP0.5 and AP0.5 : 0.95 achieve 92.3% and 54.4%, respectively. The experimental results prove the effectiveness of the proposed method in detecting camouflage targets. [ABSTRACT FROM AUTHOR]

Violence behavior recognition is an important research scenario in behavior recognition and has broad application prospects in the field of network information review and intelligent security. Inspired by the long-short-term memory network, we estimate that temporal shift module (TSM) may have more room for improvement in the feature extraction ability of long-term information. In order to verify the above conjecture, we explored based on TSM. After many attempts, it was finally proposed to connect the two TSMs in a cascaded manner, which can expand the receptive field of the model. In addition, an efficient channel attention module was introduced at the front end of the network, which strengthened the model's spatial feature extraction capabilities. At the same time due to behavior recognition prone to over-fitting, we extended and processed on the basis of some open-source datasets to form a larger violence dataset and solved the problem of over-fitting. The final experimental results show that the algorithm proposed can improve the model's feature extraction ability of violent behavior in the space and temporal dimension and realize the recognition of violent behavior, which verified the above point of view. [ABSTRACT FROM AUTHOR]

In practical scenarios, detecting and grasping objects accurately can be very challenging due to the uncertainty of their positions and orientations, as well as environmental interference. Especially when the target object is occluded by other objects, traditional machine vision methods have difficulty in accurately recognizing it. To address this problem, we propose the double-attention mechanism-based segmentation grasping detection network (DAM-SGNET). DAM-SGNET is a technique used for detecting and grasping objects accurately in cluttered environments. It utilizes a deep neural network that incorporates two attention mechanisms to predict the optimal grasping posture for RGB images at the pixel level without relying on depth images. The method begins by reannotating datasets, such as the Cornell dataset, cluttered scenes objects dataset, and VMRD dataset, with a new labeling method proposed by previous researchers. These datasets are then used to train an occlusion detection model. DAM-SGNET uses a residual network (SERESNET) with channel attention mechanisms to extract features from the images, and an adaptive decoder including a feature pyramid deformation network and an efficient channel attention module to enhance robustness in cluttered, unstructured open environments. DAM-SGNET ultimately achieves grasp detection accuracy of 99.43%, 99.24%, and 85.38% for the official Cornell grasp dataset, the cluttered scenes grasping dataset, and the VMRD grasping dataset, respectively. Real-world experiments demonstrate the efficacy of DAM-SGNET in self-built robotic arm platforms, achieving a single-target grasping success rate of 99.6%, and an average grasping success rate of 96.46% for cluttered stacked objects. [ABSTRACT FROM AUTHOR]

Aiming at the subtle differences between fabric defects and complex texture background and the problem of small size defects, we propose a fabric surface defect recognition method based on multibranch residual network. First, the convolution kernel is decomposed and the multibranch shared feature extraction module is constructed using the residual connection to mine the feature information at different scales. At the same time, squeeze-and-excitation network (SE-Net) with a reduction ratio of 16 is added to the module to suppress the interference of complex texture background. Second, the channel attention mechanism is embedded in the residual block of the network backbone to realize the compensation for the lost feature information. Finally, the swish activation function is used to enhance the accuracy and robustness of the deep network, and transfer learning is used to further improve the network accuracy. The experimental results show that the proposed model is superior to the existing model in terms of average recognition accuracy. The proposed model has the highest recognition accuracy of 96.17% for eight defect categories, which proves the effectiveness of fabric surface defect class recognition. [ABSTRACT FROM AUTHOR]

Ensuring reliable and steady operation of power equipment is paramount to safeguard the livelihoods and labor of the populace. However, traditional detection techniques face obstacles when adapting to the intricate background of transmission lines, leading to innumerable incorrect and missed detections. To resolve these issues, an improved YOLOv7 transmission line insulator defect recognition algorithm has been introduced. An anchor frame, matching the size of the transmission line fault, has been constructed using K-means++, followed by a convolutional block attention module to enhance the ability to extract features. Next, the wise-IoU loss function has been incorporated, providing a gradient gain allocation strategy, thereby enhancing the positioning performance and speed of the model. Finally, the SiLU activation function has been replaced with meta-ACON to enhance the adaptive feature capability of the YOLOv7 network. Experimental results have shown that the proposed method has an average accuracy (mAP) of 91.8%, and the detection accuracy for transmission lines can be improved to 98.8%. This model resolves the persisting issues of erroneous and missing detections by addressing the traditional technical difficulties of detecting complex backgrounds and defects with insufficient accuracy. [ABSTRACT FROM AUTHOR]

Extracting dense rotating objects accurately from remote sensing images is an emerging task in object detection. To increase the applicability of existing algorithms in the above tasks, an optimized anchor-free network optimized by a dual attention mechanism (DAM) and gate multiscale feature fusion (GMFF) is designed. The DAM module is composed of two attention mechanisms with different functions. This part can enhance the backbone network's ability to extract and model information at different levels and reduce the accuracy loss caused by object density changes in the image. The GMFF module uses the gating structure to realize adaptive transmission and integration of multiscale information. Through this module, the useless information in features will be filtered, and the key information will be retained. Several experiments are designed to verify the feasibility of the algorithm. Compared with the baseline model, adding DAM and GMFF to the dense rotating object extraction task in remote sensing images improves the model accuracy by 3.5% and 2.1%, respectively, while adding two modules simultaneously, and the accuracy increases from 79.1% to 84.3%. In conventional object extraction tasks, such as dataset for object detection in aerial images and HRSC2016, our method has the highest accuracy compared to other similar algorithms, with 76.5% and 90.3%, respectively. [ABSTRACT FROM AUTHOR]

With the increasing global concern regarding public health, accurate diagnosis and treatment of diseases have become critical. In the context of liver computed tomography (CT) image diagnosis, obtaining precise liver segmentation output samples can save consultation time and reduce the risk of misdiagnosis. We propose a full-scale connected attention-aware segmentation network, called Attention UNet3+. To fully leverage semantic information at different scales, we redesign the depth supervised decoder and adopt a full-scale skip connection, which can effectively extract features from different layers thus increasing accuracy. The proposed Attention UNet3+ model uses an attention gate connection instead of the skip connection, which effectively suppresses irrelevant regions and highlights salient features of specific local regions during feature extraction, therefore, improving the segmentation accuracy. Additionally, the classification-guided module enhances the liver boundary and reduces over-segmentation of non-liver regions, obtaining accurate segmentation results. Our experimental evaluation on the medical image computing and computer assisted intervention Liver Tumor Segmentation Challenge 2017 dataset shows that the proposed Attention UNet3+ outperforms other improved UNet algorithms for liver image segmentation by a minimum of 2.9% in intersection over union and a minimum of 1.1% in Dice. [ABSTRACT FROM AUTHOR]

Due to the small size, high density, and background noise associated with strip surface defects, the current object detection model commonly faces limitations in performance. To address this issue, we propose a spatial-to-depth feature-enhanced detection method called STD-Detector. The method consists of two types STD-Conv-A and STD-Conv-B. First, the STD-Conv-A module is used in the backbone feature extraction network to expand the field of perception and enable the model to learn a wider range of background information. Then, the STD-Conv-B module is used for feature fusion networks to improve the expression of output features. In addition, we incorporate the convolutional block attention module to mitigate background interference and enhance the performance of the model. Finally, experimental results on the NEU-DET dataset show that our method achieves a mean average precision of 82.9%, which represents a 3.9% improvement over the baseline. Compared with the state-of-the-art model, our method exhibits greater competitiveness in detecting strip surface defects. Moreover, experimental results on the road surface defect dataset show that our method has good robustness. [ABSTRACT FROM AUTHOR]

Detecting pavement cracks from images is a complex computer vision task due to their varying shapes, backgrounds, and sizes. We propose CrackF-Net, an end-to-end convolutional neural network for automatic crack detection in road images. We construct the CrackF-Net network using an encoder–decoder architecture to extract image features in convolutional blocks with residuals and fuse the multiscale convolutional features produced by the decoder. Convolutional blocks with residuals are used to capture the strong semantic features of cracks, and an adaptive filter fusion module is proposed to assist the network make a selection of filter fusion features on the channels. CrackF-Net fuses the multiscale features in decoder to improve crack detection performance. The proposed CrackF-Net is compared to other advanced crack detection methods using three public datasets. The experimental results show that CrackF-Net achieves state-of-the-art performance, which obtains F-measures of 0.866, 0.737, and 0.852 on the three datasets. [ABSTRACT FROM AUTHOR]

We propose a method to change the shadows cast by vehicles based on unpaired data, which is used to solve the problem of insufficient datasets for target detection in autonomous driving scenarios. This method can increase the training samples of road vehicles in different time periods and reduce the time needed to make complex annotations of datasets. Direct use of generative adversarial networks with cycle-consistency constraints can achieve mapping transformation of unpaired domains, but it is difficult to completely learn the mapping between different shadow domains. To solve this problem, V-ShadowGAN is proposed. First, to train the model, we construct an unpaired dataset of vehicle shadows in different time periods. Second, the shadow mask extractor and the shadow mask discriminator are constructed to extract shadow masks and guide the generation of shadows in the network. Finally, the spatial attention mechanism is added to the generator, which improves the overall quality of the generated image by changing the channel weight of the network and deepening the network. Various experiments show that V-ShadowGAN can effectively change the vehicle shadow. [ABSTRACT FROM AUTHOR]

In practical applications, to improve the real-time performance of end-to-end stereo matching networks, the existing methods build cost volume at low resolution. However, with detailed information missing in low-resolution features, it is difficult to get accurate disparity estimation results in weak texture regions. Besides, smooth L1 loss supervision also results in a loss of accuracy in disparity discontinuity areas. To solve these problems, we propose an efficient stereo-matching network based on multiple attention mechanisms and edge optimization, which can achieve high accuracy in a short time. The multi-scale attention module is applied to enhance the feature expression in detail regions. For weak texture areas, we construct a concatenation cost volume and a multi-level patch matching volume, which can be combined to improve the network's attention to weak texture regions. In terms of edge optimization, we perform bimodal Laplace modeling of the sampled edge points' disparity distribution and optimize the edge region of the initial disparity map using likelihood loss to obtain sharp edges. The experimental results show that, on the SceneFlow and KITTI datasets, the proposed network improves by 32% and 27% in accuracy compared with BGNet+. [ABSTRACT FROM AUTHOR]

Insulators that connect high-voltage transmission lines may experience various faults due to long-term exposure to the natural environment, leading to safety degradation and reliability issues in power grids. Therefore, detecting defective insulators through daily maintenance and long-term overhaul is crucial. We propose an insulator defect detection method to achieve this objective, using a novel neural network named more information-you only look once (MI-YOLO). MI-YOLO includes several modifications compared to YOLOv5s: a skip connection module with down-sampling in the backbone, a spatial pyramid dilated convolution module in the neck, and a novel serial-parallel spatial-channel attention module in between. These changes enhance the feature abstraction process by providing more information (MI) through the YOLO-like structure, hence the name MI-YOLO. To evaluate the superiority of MI-YOLO, multiple experiments are performed. First, an ablation study demonstrates the effectiveness of novel substructures and their combinations compared to YOLOv5s. Then, MI-YOLO is compared to baseline structures, including faster R-CNN, SSD, and five YOLOs. Furthermore, the performances of MI-YOLO and 10 state-of-the-art object detection methods are evaluated. Lastly, a second dataset is adopted to evaluate the generalization property and heat maps are presented. The experimental results show that the new designs substantially improve the accuracy of insulator fault detection by abstracting more information from the features. [ABSTRACT FROM AUTHOR]

Arbitrary style transfer has flexible applicability and aims to learn the painting styles of different artists through one model training and re-render to everyday photos. Existing domain-to-domain style transfer methods based on generative adversarial networks achieve good results in image generation quality but need more flexibility in various style transfer tasks. In the image-to-image style transfer method, focusing on the fusion of content image and style image can learn local detail texture very well. Still, there are some defects in the preservation of the content structure. To alleviate the above two problems, we leverage both advantages, using attention improves style transfer via contrastive learning. Specifically, the attention mechanism is used to fuse the multi-scale features of the content and style images. The previous shallow features are fully utilized to supplement the texture information lost by the deep features. Then the currently learned style image is distinguished from other style images by contrastive learning modules, thereby enhancing the ability of the network to learn style images. We also derive a local loss based on the transferred blocks to improve the visual quality of the generated images. We conduct extensive qualitative and quantitative analyses, the results of which demonstrate the effectiveness of our approach. [ABSTRACT FROM AUTHOR]

The worldwide spread of the coronavirus disease 2019 (COVID-19) has had a significant impact on healthcare systems globally. Timely and accurate diagnosis is crucial for controlling and preventing outbreaks. Computed tomography (CT) has been demonstrated as an effective diagnostic tool for COVID-19. Deep learning techniques have been applied in related fields to solve relevant problems. Normally, they achieve good performance on large datasets directly. However, due to the need to protect patient privacy and the sudden outbreak of the epidemic, the amount of publicly available CT data is insufficient for utilizing these methods. In addition, some natural images, which are used for model pre-training, are quite different from chest CT images, causing some inevitable errors. To address these issues, a targeted self-supervised attention network based on an improved inception ResNet called SIRANet is proposed. SIRANet incorporates a self-supervised learning method during the pre-training process and introduces a attention module to improve feature selection and focus on key informative features. Moreover, a transfer learning-based strategy for COVID-19 detection is presented and experiments are conducted on three widely used datasets, including ImageNet, an unlabeled chest CT dataset, and a COVID-19 CT dataset. Experimental results show that SIRANet outperforms other state-of-the-art techniques in terms of accuracy, precision, recall, and f1-score, demonstrating its effectiveness in capturing key informative features from unlabeled and labeled images. The implementation of SIRANet is available at https://github.com/Xiiin6/COVID-19. [ABSTRACT FROM AUTHOR]

Object detection is an enabling technology of computer vision for locating instances and categorizing classes of objects in images or videos. It has made enormous strides with significant growing of deep learning research over the last decade. However, there are severe challenges on small object detection in complex scenes as they appear with low resolution and have not enough contrast from the background information. This disturbance may cause missed detection and detection accuracy decline for small objects. We propose a small object detection network in the dark light scene based on improved YOLOv5. The network takes YOLOv5 as the baseline and incorporates a channel and spatial dual-branch backbone module, which enhances the details by fusing the features of the two branches. We also introduce a densely linked feature fusion network before detection layers with receptive field block. The fusion network integrates deep features with shallow ones across different scales. A data augmentation module is used to enhance the brightness and limit the contrast for small object detection in the dark light scene. Experiments are carried out on the Dark Face and darkened Vis Drone dataset. The results show that the evaluation index of the proposed method is better than that of the comparison methods. From the detected images, it is obvious that the undetected frame of the small object in the dark light scene decreases and the detection accuracy improves. All of the results show that our model has better performance than some existing methods for small object detection in the dark light scene. [ABSTRACT FROM AUTHOR]

Extracting vehicle information is highly significant to the development of intelligent transportation systems. Vehicle logo detection (VLD) plays an important role in the extraction of vehicle information. Traditional VLD methods rely on the detection of the license plate and have difficulty in accurately locating the logo. VLD methods based on deep learning have greatly improved detection accuracy and speed. To further increase the detection accuracy and speed of VLD, a VLD method, VLDetNet, is proposed. This method first introduces a combined multibranch block for VLD (denoted as vBlock), a new network block that can be combined with other network structures to greatly improve their recognition accuracy. By combining vBlock and MobileNetV1, we then construct a new backbone network structure, vMNet, which replaces the original backbone (DarkNet53) in YOLOv3. Finally, VLDetNet extracts spatial attention information from lower-level feature maps and uses it as spatial attention weights, to mine spatial relationships of pixels and facilitate vehicle logo extraction. To further promote research on VLD, we have labeled a new vehicle logo dataset, named HFUT-VL3, which contains 6000 images and 54 types of vehicle logos obtained from an real-time highway surveillance systems in China. We conducted experiments using both HFUT-VL3 and an open VLD-45 dataset. And the experimental results demonstrate that the proposed method achieves high detection accuracy and speed and outperforms most existing methods. [ABSTRACT FROM AUTHOR]

The purpose of image steganalysis is to detect whether the secret information is hidden in the image. The current advanced adaptive steganography algorithm hides the secret information in the complex area of the image texture. However, it is difficult for the image steganalysis model to capture enough noise residual, which makes the detection ability of the model insufficient. To further improve the detection ability of the spatial image steganalysis model, a U-Net-based auxiliary information generation network is first constructed to increase the size of noise residual in the image so that the model can capture more favorable information. Besides, the spatial and channel attention mechanisms are fused to guide the model to pay more attention to the regions of the image that are globally favorable for steganalysis. To verify the effectiveness of the proposed model, experiments are conducted on the BOSSbase-v1.01 dataset through advanced spatial adaptive steganography algorithms S-UNIWARD and WOW. The experimental results show that the proposed model can improve the detection accuracy by up to 4.5% compared with the current optimal deep learning-based spatial image steganalysis models SR-Net and Siamese-Net. [ABSTRACT FROM AUTHOR]

Semantic scene segmentation has become an important application in computer vision and is an essential part of intelligent transportation systems for complete scene understanding of the surrounding environment. Several methods based on convolutional neural networks have emerged, but they have some problems, including small-scale target loss, inaccurate detailed region segmentation, and boundary category confusion. Using shallow features, we exploit the capabilities of global context information according to the theory of pyramids. A weighted pyramid feature fusion module is constructed to fuse the feature maps of different scales generated by the backbone network, and the proportion of feature fusion is dynamically updated by trainable parameters. After that, a self-attention mechanism is introduced to discover information about spatial channel interdependencies. Finally, the atrous spatial pyramid pooling module of the DeepLabv3+ network is improved by connecting the atrous convolution with different dilation rates at the receptive field. The experimental results show 4.1% mean pixel accuracy and 3.92% mean intersection over union improvements in the proposed method compared with the DeepLabv3+, and the result of semantic segmentation is more accurate. [ABSTRACT FROM AUTHOR]

Single-image deraining is a classical problem in the field of low-level computer tasks. Most of the recent state-of-the-art image rain removal methods are trained on synthetic images, which have the problems of incomplete rain removal on real images and inability to process complex rain conditions. Based on these limitations, we propose a single-image deraining network based on multistage feature fusion (SIDNMFF). The network performs rain removal in four stages, with the first three stages using an improved encoder–decoder subnetwork for feature fusion to extract global and local information from the image, resulting in more detailed texture information of the obtained image. In the last stage, the network fuses the feature information extracted in the first three stages, performs feature extraction on the original resolution image, and finally outputs the final result of multistage deraining. The proposed method conducts a large number of comparative experiments on synthetic and real datasets as well as experiments on real datasets using no-reference metrics and the target detection method as the evaluation basis. Experimental results confirm that the proposed method achieves a satisfactory rain removal effect and outperforms the other methods. [ABSTRACT FROM AUTHOR]

With the increasing demand for authentication using iris biometrics, iris recognition systems are gradually being deployed on low-power devices. Iris segmentation is a critical step for iris recognition systems. However, existing iris segmentation methods often require a more complex network to obtain higher accuracy, which makes the segmentation model have a greater number of parameters and a higher computation cost. Therefore, these algorithms are not suitable for deployment on low-power devices. To address this problem, we propose a lightweight iris segmentation network, referred to as ATTention Linknet. First, we redesign the feature extraction part of the Linknet, which combines traditional convolution based on an attention mechanism with depthwise separable convolution to reduce computation cost while maintaining accuracy. Then, a proposed loss function that better segments the iris images taken under different cross-sensors is used to train the network. Experimental results on the Chinese Academy of Sciences (CASIA-V4), Indian Institute of Technology Delhi (IITD), and the University of Beira Interior (UBIRIS.V2) iris databases show that the proposed method better balances iris segmentation precision and efficiency compared with state-of-the-art iris segmentation methods. Specifically, our model obtains an error rate of 1.47% on the CASIA-V4 database with 1.69 M parameters. [ABSTRACT FROM AUTHOR]

The object detectors based on deep learning, such as Yolov3, Yolov3-tiny, and Yolov4-tiny, have achieved real-time object detection in various vision tasks. However, long-tailed distribution of classes, small objects, and incomplete numbers on the dataset have brought challenges to the automatic recognition of wheel-type water meter digit reading using these object detectors. We propose a Mix suppression data augmentation method based on information deletion and resampling. The method improves the long-tailed distribution of the water meter dataset and the performance of small object detection. Moreover, to further improve the performance of detection model, we propose a block spatial attention module to connect with the output of the backbone of object detection model. This module can introduce priori knowledge into the convolutional neural network, reducing the phenomenon of missing objects. The experiments show that our methods successfully improve the performance of neural networks for the water meter digit reading recognition, and they could also be conveniently embedded into other mainstream general object detection models. In addition, we construct a dataset named waterDataset for noncommercial use, which is available at a Github repository: https://github.com/l1158984238/waterDataset. [ABSTRACT FROM AUTHOR]

We propose a saliency-enhanced two-stream convolutional network (SETNet) for no-reference image quality assessment. The proposed SETNet contains two subnetworks of image stream and saliency stream. The image stream focuses on the whole image content, while the saliency stream explicitly guides the network to learn spatial salient features that are more attractive to humans. In addition, the spatial attention module and dilated convolution-based channel attention module are employed to refine multiple levels features in spatial and channel dimensions. Finally, the image stream and saliency stream features fusion strategy is proposed to integrate features at the corresponding layer, and the final quality scores are predicted by using multiple levels of integrated features and weighting strategy. The experimental results of the proposed method and several representative methods on four synthetic distortion datasets and two real distortion datasets show that our SETNet has higher prediction accuracy and generalization ability. [ABSTRACT FROM AUTHOR]

Attention mechanism originates from the study of human visual behavior, which has been widely used in various fields of artificial intelligence in recent years and has become an important part of neural network structure. Many attention mechanism-based trackers have gained improved performance in both accuracy and robustness. However, these trackers cannot suppress the influence of background information and distractors accurately and do not enhance the target object information, which limits the performance of these trackers. We propose new Siamese point attention (SPA) networks for robust visual tracking. SPA networks learn position attention and channel attention jointly on two branch information. To construct point attention, each point on the template feature is used to calculate the similarity on the search feature. The similarity calculation is based on the local information of the target object, which can reduce the influence of background, deformation, and rotation factors. We can obtain the region of interest by calculating the position attention from point attention. Position attention is integrated into the calculation of channel attention to reduce the influence of irrelevant areas. In addition, we also propose the object attention, and integrate it into the classification and regression module to further enhance the semantic information of the target object and improve the tracking accuracy. Extensive experiments are also conducted on five benchmark datasets. The experiment results show that our method achieves state-of-the-art performance. [ABSTRACT FROM AUTHOR]

You only look once (YOLO) target detection algorithms have garnered substantial attention in engineering inspection tasks due to their remarkable accuracy and swift detection speed. However, adapting the YOLOv7 algorithm to engineering construction presents challenges in balancing lightweight design and high precision. This work introduces a YOLOv7 lightweight detection algorithm, leveraging partial convolution to efficiently reduce network parameters without compromising detection speed. Integrating partial convolution into the network's backbone is a highly effective parameter reduction strategy. Additionally, we propose the SE_Conv module following quantitative analyses through experiments, thereby significantly enhancing the algorithm's detection accuracy. To address the impact of low-quality data on detection effectiveness in engineering tasks, the advanced bounding box loss function WIoU is employed. In a practical application scenario involving helmet detection, the algorithm achieves a notable 23.4% reduction in parameters and a substantial 28.2% decrease in operations through partial convolution. Subsequent experiments with the SE_Conv module yield a noteworthy 1.4% improvement in results. Significantly, in head detection, our algorithm surpasses YOLOv8 algorithm by 2.6% and single shot multibox detector algorithm by 9.4%. These findings underscore the efficacy of our proposed algorithm in engineering applications. [ABSTRACT FROM AUTHOR]

The performance of face detection has achieved great progress with the rapid development of the convolutional neural network. However, detecting faces with a wide range of scales is still a challenging and crucial problem, especially for small faces. We focus on designing a strong network architecture, aiming at generating high-quality feature representation to handle the multiscale problem. Specifically, we conduct an in-depth study of three aspects, including feature fusion, attention mechanism, and receptive field, then propose an effective one-stage face detector. First, we design a feature fusion module, which adopts a more effective manner to sufficiently fuse multiscale features for building discriminative feature pyramid representation. Then, we propose an attention enhancement module for introducing the attention mechanism, which makes the network keep more valid information to emphasize the features from facial regions. Finally, we construct a prediction module, which ensures that each detection layer contains rich receptive fields to bring more useful context information. A large number of experiments on the widely used face detection benchmarks (i.e., WIDER FACE, FDDB, PASCAL FACE, and AFW) indicate that the results of our method outperform overwhelming majority of the existing methods and are comparable to the most recent state-of-the-art face detector, which demonstrate that our method is able to detect faces with a wide range of scales effectively. [ABSTRACT FROM AUTHOR]

Existing methods for fusing infrared and visible images prioritize the fusion effect at the expense of the model size and inevitably tend to be more oriented toward infrared images during fusion, which results in fused images that can lack the texture detail information of visible images. Therefore, a new feature gradient attention block is designed in our model to extract the texture detail of the image, in which the module extracts the gradient information of the original image while extracting features, then uses depthwise separable convolution to optimize and enhance the information with rich edge features. To preserve the original image information, we also use short links to reference previous features. Since the important features are strengthened in the feature extraction stage, we design an adaptive weight energy attention network based on the energy fusion strategy in the fusion stage to further preserve the thermal radiation area of the infrared image and the spatial details of the visible image. The proposed method is experimentally verified on the public visible-infrared paired dataset for low-light vision and the TNO dataset, and six objective evaluation indicators are used to prove that our model is better to the existing fusion algorithms. In addition, we further verify the effectiveness of the proposed method for high-level vision task models by object detection and semantic segmentation models. [ABSTRACT FROM AUTHOR]

Most trackers based on the correlation operation mainly carry out a straightforward fusion of template features and search region features. The correlation operation, however, is a local linear matching process that might cause the loss of semantic information, leading to local optimums. We propose a tracker method that takes advantage of both the attention mechanism and feature fusion. This approach uses an attention-based feature fusion network to integrate the characteristics of the template area and search area, which effectively avoids the semantic information loss and local optimum induced by the correlation operation. Meanwhile, by adopting a modified ResNet50 network for extracting multilayer target features to fully utilize both shallow and deep features, our method takes full advantage of the template information to get better tracking results. Extensive experiments on four large-scale tracking benchmark datasets—VOT2019, OTB100, GOT-10k, and LaSOT—show that our tracker obtains more accurate tracking results and has a better performance. [ABSTRACT FROM AUTHOR]

We propose a solution to the issue of hyperspectral unmixing methods that only consider local spectral–spatial information of the pixel level or pixel block. The proposed method is a two-stage convolutional autoencoder network that takes into account global spatial context information. In stage-I, the parallel dual-branch module is dedicated to extracting multi-scale spatial and spectral features. The extracted spatial–spectral prior information is then propagated from stage-I to stage-II to assist in the extraction of joint spectral–spatial features. The proposed method also uses a spectral–spatial attention residual module to refine spectral–spatial features into distinct deep spectral–spatial features and suppress irrelevant redundant features. We validate the proposed method using synthetic and real datasets. It is found that the proposed method outperforms existing unmixing methods in terms of endmembers extraction and abundance estimation. The source code for the proposed model will be made public in the Github repository available at https://github.com/xzw001212/two-stage-DBMRANet. [ABSTRACT FROM AUTHOR]

Achieving real-time performance while maintaining high accuracy in semantic segmentation can be a challenging task. Many existing methods adopt multi-branch architectures to extract both spatial and semantic information, resulting in increased computational complexity and a lack of communication between branches. We propose a pseudo bilateral segmentation network (PBSNet) that can extract rich spatial and semantic features from a single path, without incurring additional computational cost or time consumption. Our proposed PBSNet utilizes a semantic enhancement module to explore the relationship between high-level semantic features, an interchange module to enhance feature representation through bi-directional vertical propagation and adaptive spatial attention, and an attention fusion module to aggregate multi-scale features to produce the final segmentation prediction. Our results on the Cityscapes dataset demonstrate the superiority of PBSNet over state-of-the-art methods, achieving a balance of accuracy and efficiency with 74.52% mean intersection over union and 82.5 frames per second. [ABSTRACT FROM AUTHOR]

Human emotions are known to always have four phases in the temporal domain: neutral, onset, apex, and offset. This has been demonstrated to be of great benefit for emotion recognition. Therefore, temporal segmentation has attracted considerable research interest. Although state-of-the-art techniques use recurrent neural networks to highly increase the performance, they ignore the relevance of each frame (time step) of a video, and they do not consider the changing contribution of different features when fusing them. We propose a framework called dual-level attention-aware bidirectional grated recurrent unit, which integrates ideas from attention models to discover the most important frames and features for improving temporal segmentation. Specifically, it applies attention mechanisms at two levels: frame and feature. A significant advantage is that the two-level attention weights provide a meaningful value to depict the importance of each frame and feature. The experiments demonstrated that the proposed framework outperforms state-of-the-art methods. © 2018 SPIE and IS&T [DOI: 10.1117/1.JEI.27.3.033012] [ABSTRACT FROM AUTHOR]

Herein, a lightweight "You Only Look Once" (YOLOv5) ultrahigh voltage (UHV) transmission line insulator self-detonation detection algorithm based on the attention mechanism and Ghostnet is proposed to address the multi-scale features of UHV insulator aerial images. Small-scale defective pieces cannot be detected accurately by conventional algorithms in a complex power system inspection environment. To suppress the complex background interference, a YOLOv5-based spatial and channel convolution attention model is used to enhance the saliency of the target to be detected by weighting the feature layers and feature maps. The convolutional structure of the original cross stage partial (CSP_X) structure is replaced by the Ghostnet network to generate more similar feature maps by performing linear operations on the redundant feature maps so that more feature maps can be generated using fewer parameters to achieve a lightweight network and reduce the network parameters. To address the problem of missed and false detection caused by inadequate expression of the target features to be detected, the original neck feature pyramid network (FPN) + path aggregation network (PAN) structure is modified to a bidirectional FPN structure such that the target multi-scale features are effectively fused. Finally, SCYLLA-IoU is used as the loss function to accelerate the model convergence and improve the detection accuracy, and migration learning parameter sharing combined with the model training strategy of freeze and thaw training is performed for non-generalization of the network owing to small sample datasets. The effectiveness of the proposed algorithm is verified by training it with data obtained from UHV patrols. Experimental results show that the proposed algorithm accurately monitors insulator self-detonation targets in complex environments with an average detection accuracy of 97.4% and a significant reduction in the size of models and parameters, which has good practical value. [ABSTRACT FROM AUTHOR]

To balance noise reduction and texture details of low-dose computed tomography (LDCT) images during deep learning training, we propose a network combining edge features and a pyramid structure for LDCT denoising (PLEDNet). PLEDNet contains three states: superficial feature acquisition, deep feature acquisition, and feature confluence. First, the superficial feature acquisition stage uses a simple convolutional block to perform feature extraction on the input LDCT image. Second, in the deep feature extraction phase, one branch uses pyramid pooling to extract multiscale features, and the other uses the trainable Laplacian of the Gaussian module to extract edge detail features. Finally, the feature fusion phase uses a fusion attention module to adaptively fuse multiscale and edge features. The results of the experiments demonstrate that the method not only enhances various relative objective indicators but also effectively represses the noise in LDCT images while fully conserving the image texture details, which has great advantages in terms of visual effects. [ABSTRACT FROM AUTHOR]

The single shot multibox detector (SSD) is one of the most important algorithms in single-stage target detection, having the characteristics of multiscale detection and rapid detection speed. However, the effective SSD feature layers are independent of one another, which can lead to object detection difficulties. To address this problem, we proposed an improved SSD object detection algorithm. First, the global attention mechanism (GAM)—which can enhance spatial and channel information—was introduced into the multiscale feature layer. The channel attention module of the GAM was improved. Second, a feature fusion module was introduced to strengthen the relationship between feature layers. Finally, the cross stage partial structure was introduced into the feature fusion module, and used to improve the model's learning ability. For model training and detection based on the PASCAL VOC dataset, the mean average precision and frames per second obtained by the improved SSD algorithm were 84.67% and 18.67, respectively, which could effectively detect difficult targets. [ABSTRACT FROM AUTHOR]

Model-based methods utilize atmospheric scattering model to effectively dehaze images but introduce unwanted artifacts. By contrast, recent model-free methods directly restore dehazed images by an end-to-end network and avoid artificial errors. However, their dehazing ability is limited. To address this problem, we combine the advantages of supervised and unsupervised learning and propose a semisupervised knowledge distillation network for single image dehazing named SSKDN. Specially, we respectively build a supervised learning branch and an unsupervised learning branch by four attention-guided feature extraction blocks. In the supervised learning branch, the network is optimized by synthetic images. In the unsupervised learning branch, we dehaze real-world images by dark channel prior and refine dehazing network (RefineDNet) (another dehazing method) and use these dehazed images as fake ground truths to optimize network using prior information and knowledge distillation. Experimental results on synthetic and real-world images demonstrate that the proposed SSKDN performs better than state-of-the-art methods and owns powerful generalization ability. [ABSTRACT FROM AUTHOR]

Human pose estimation is a fundamental yet challenging task in computer vision. Although many methods have achieved significant improvement, they are still insufficient for the fusion of feature maps at different stages, such as the stacked hourglass network (SHNet). The SHNet is a classic human pose estimation network that extracts multiscale features through stacked multistage downsampling and upsampling operations. We propose a multistage attention mechanism to fuse the multistage feature maps. Furthermore, we apply it in the SHNet to propose a multistage attention network (MANet). In the experiments, we demonstrated the effectiveness of MANet in human pose estimation on the common objects in context dataset and the MPII human pose dataset. [ABSTRACT FROM AUTHOR]

The demand for applying the iris segmentation model on mobile devices has been growing rapidly. Most current segmentation networks have an enormous amount of parameters, hence unsuitable for mobile devices, while other small memory footprint models follow the spirit of classification networks and ignore the inherent characteristic of segmentation. To address the challenge, we propose a lightweight segmentation network (LiSeNet) for iris segmentation of noisy images. Unlike previous studies that only focus on improving the accuracy of segmentation masks, LiSeNet can simultaneously obtain segmentation masks, parameterized pupillary and limbic boundaries of the iris, further enabling CNN-based iris segmentation to be applied in any regular iris recognition systems. We first propose a multiscale concatenate (MSC) Block, which connects multiple sizes of convolution kernels in a dense manner, gradually reduces the dimension of feature maps and uses the aggregation of them for image representation. Based on the MSC block, we develop a two-stage refinement encoder to aggregate discriminative features through subnetwork feature reuse and substage feature reassess, thus obtaining a sufficient receptive field and enhancing the model learning ability. To exploit object contextual information more efficiently, we further devise a grouped spatial attention to emphasize the important features and suppress irrelevant noises through a gating mechanism in the decoder. Extensive experiments on three challenging iris datasets show that LiSeNet, without any complicated postprocessing, achieves competitive or state-of-the-art performance with only 2.2M parameters, being 14 × smaller than the previous best method. Code will be publicly available. [ABSTRACT FROM AUTHOR]

As an effective method of information hiding, steganography embeds secret information into images in a way that is not perceived by humans. Recent interest in the combination of image steganography and generative adversarial networks (GANs) has yielded rapid progress. However, existing steganography frameworks still suffer from the low quality of the steganographic images and weak resistance to detection by steganalysis algorithms. To overcome these limitations, we propose an effective GAN-based image steganography framework with multiscale features integration. Specifically, we construct the secret image feature extraction network (SfeNet), which is driven by the spatial attention mechanism to extract multiscale features of secret images. And the encoder combined with the efficient channel attention mechanism is presented to embed multiscale features of secret images into the cover image. Subsequently, a steganalyzer is incorporated as the discriminator with the encoder in GAN to strengthen the ability of the model to resist steganalysis. Besides, a mixed loss function is proposed by combining perceptual loss, MS-SSIM, and L1 loss to preserve the structural similarity of images. Experimental results on ImageNet, Pascal VOC2012, and LFW show that the proposed method achieves better quality steganographic images and better resistance to steganalysis compared to some of the state-of-the-art steganography algorithms. [ABSTRACT FROM AUTHOR]

To solve the problems of poor authenticity and lack of clarity for short-time typhoon prediction, we propose a self-attentional spatiotemporal adversarial network (SASTA-Net). First, we introduce a multispatiotemporal feature fusion method to fully extract and fuse the multichannel spatiotemporal feature information to effectively enhance feature expression. Second, we propose an SATA-LSTM prediction model that incorporates spatial memory cell and attention mechanisms in order to capture spatial features and important details in sequences. Finally, a spatiotemporal 3D discriminator is designed to correctly distinguish the generated predicted cloud image from the real cloud image and generate a more accurate and real typhoon cloud image by adversarial training. The evaluation results on the typhoon cloud image data set show that the proposed SASTA-Net achieves 67.3, 0.878, 31.27, and 56.48 in mean square error, structural similarity, peak signal to noise ratio, and sharpness, respectively, which is superior to the most advanced prediction algorithm. [ABSTRACT FROM AUTHOR]

Recently, single object tracking has demonstrated great success. However, due to various problems caused by fast motion, occlusion, and deformation, it is still intractable for traditional trackers to adapt to changes in object appearance. In this work, a dynamic memory network is proposed for visual tracking to handle the template matching process globally. More specifically, we first build a memory model, which consists of a memory feature fusion module and a memory bank. By the memory model, the network not only accepts the first frame as initial information but also memorizes the selective frames in the video sequence to provide rich time-domain information. Second, an innovative sampling strategy is adopted in the tracking process. By updating the template and guiding the selection of memory frames, our model can output higher-quality features. In addition, a spatial-channel fused attention module that effectively improves the representational capability and discriminability of the model is introduced. Our proposed method obtains compelling results on six challenging tracking benchmarks, including the OTB100, VOT2019, UAV123, NFS, GOT-10k, and LaSOT datasets. Extensive experiments demonstrate that our approach shows satisfactory robustness and leading application potential in real-time speed. [ABSTRACT FROM AUTHOR]

Solving the problem that features prediction of self-supervised monocular depth estimation is still ambiguous in low-texture regions and boundaries. We proposed an innovative self-supervised monocular depth estimation method, monocular depth self-supervised network, which integrates three effective strategies to construct an innovative self-supervised monocular depth estimation framework: (1) the attention mechanism and feature fusion module are adopted to enhance the semantic and spatial information of feature images, (2) the threshold segmentation mask is utilized to solve object motion and low-texture regions to increase image details, and (3) the residual pose module and deep reconstruction loss are used to enhance the feature extraction capability of the model to improve the accuracy of depth and pose estimation. Comprehensive experiments and visual analysis results demonstrate the effectiveness of each component in isolation. Compared to existing self-supervised methods, our model not only achieves outstanding results on KITTI and NYU Depth V2 datasets but also can be suitable to different environments. [ABSTRACT FROM AUTHOR]

The goal of pedestrian trajectory prediction is to predict the future trajectory according to the historical one of pedestrians. Multimodal information in the historical trajectory is conducive to perception and positioning, especially visual information and position coordinates. However, most of the current algorithms ignore the significance of multimodal information in the historical trajectory. We describe pedestrian trajectory prediction as a multimodal problem, in which historical trajectory is divided into an image and coordinate information. Specifically, we apply fully connected long short-term memory (FC-LSTM) and convolutional LSTM (ConvLSTM) to receive and process location coordinates and visual information respectively, and then fuse the information by a multimodal fusion module. Then, the attention pyramid social interaction module is built based on information fusion, to reason complex spatial and social relations between target and neighbors adaptively. The proposed approach is validated on different experimental verification tasks on which it can get better performance in terms of accuracy than other counterparts. [ABSTRACT FROM AUTHOR]

The purpose of low-light image enhancement is to improve image quality assessment by human visual perception and bolster the performance of subsequent visual tasks. It is necessary to consider not only the complexity of illumination in a real scene but also the issues of color distortion after image enhancement. We propose a two-stage network to solve these problems. In the first stage, we use a two-branch encoder–decoder subnetwork to learn multiscale features of the image, in which the brightness encoder branch is used to improve the network's attention to a low-light region and improve the problem of uneven illumination. In the second stage, we utilize the detail recovery subnetwork to preserve details. In addition, we introduce an attention mechanism and cross-stage connection between the two stages to utilize the features learned by different subnetworks effectively. Extensive experiment results demonstrate that the proposed method outperforms several state-of-the-art methods in terms of quantitative metrics and the visual effect. Thus, our network enhances low-light images with high quality. [ABSTRACT FROM AUTHOR]

To address the problems of error and omission detection in remote sensing image detection caused by the diverse scale changes of remote sensing object scales and the abundant proportion of small-scale objects, as well as the global and dense distribution of remote sensing objects, a remote sensing image detection improvement method based on YOLOv5-S is proposed. First, according to the characteristics of remote sensing objects, the data enhancement strategy is adopted to expand the dataset samples for the characteristics of remote sensing objects to improve the generalization ability of the model. Second, the contextual transformer module is introduced to the backbone feature extraction network and the feature fusion network to ensure the local feature extraction capability while improving the global information acquisition capability of the model, making full use of the input contextual information and guiding the dynamic attention matrix learning to improve the visual representation ability. Third, based on the original model, a shallow detection scale is added, and then a multiscale complex fusion structure is adopted. Meanwhile, the K-means++ algorithm replaces the original K-means algorithm and then clusters 12 anchor box sizes. Fourth, the efficient intersection over union loss is used to improve the accuracy of the remote sensing object recognition prediction. In the experiment on the on two optical remote sensing image datasets, a comparison with several object detection algorithms based on convolutional neural network is made, the results show that the mAP@0.5 tested on the remote sensing datasets is higher than the original YOLOv5-S. Compared with other models, the detection efficiency is higher, and the problems of small-scale object detection in remote sensing image have been significantly improved. [ABSTRACT FROM AUTHOR]

In recent years, various convolutional neural networks have successfully applied to single-image super-resolution task. However, most existing models with deeper or wider networks require heavy computation and memory consumption that restrict them in practice. To solve the above questions, we propose a lightweight multiscale residual attention network, which not merely can extract more detail to improve the quality of the image but also decrease the usage of the parameters. More specifically, a multiscale residual attention block (MRAB) as the basic unit can fully exploit the image features with different sizes of convolutional kernels. Meanwhile, the attention mechanism can be adaptive to recalibrate channel and spatial information of feature mappings. Furthermore, a local information integration module (LFIM) is designed as the network architecture to maximize the use of local information. The LFIM consists of several MRAB and a local skip connection to complement information loss. Our experimental results show that our method is superior to the representative algorithms in performance with fewer parameters and computational overhead. Code is available at https://github.com/xiaotian3/EMRAB. [ABSTRACT FROM AUTHOR]

The autonomous perception of unmanned surface vessels in the natural sea environment has improved significantly; however, it is adversely affected by haze. Improving the U-Net network can retain and restore spatial information effectively and improve image dehazing. However, many dehazing algorithms that use the U-Net network as the main architecture face problems, such as distorted images and blurred contours of small distant objects. To overcome these limitations, we propose a generative adversarial network (GAN)-U-Net++ network that uses the U-Net++ network with a coordinate attention mechanism as the main part of a GAN generator. We used the coordinate attention mechanism to perform horizontal and vertical pixel level pooling to improve the attention on small targets in hazy images. The addition of the edge preservation loss function and image gradient loss function to small targets improved the sharpness of their edges. A Markov discriminator was used to evaluate the authenticity of the images. Several experiments were conducted to compare the proposed method with existing dehazing algorithms. The structural similarity value and peak signal-to-noise ratio increased by 15.74% and 29.14%, respectively. [ABSTRACT FROM AUTHOR]