Separable Convolution Research Articles

Fusion of transformer attention and CNN features for skin cancer detection

Melanoma is a skin cancer that spreads quickly and has serious risks. Early diagnosis is essential, but since the symptoms of skin lesions in the early stages are vague and similar, they can be difficult for specialists to detect. Therefore, machine learning-based alternative diagnostic methods can be developed in addition to existing ones. This study proposes a new deep learning model, a modified lightweight vision transformer (ViT) architecture, and a hybrid framework developed with an integrated deep learning model and an Ensemble Learning (EL) model for the early-stage diagnosis of skin lesions. The proposed deep learning model was developed based on convolution layers and transformers. The model is called multi-head attention block depthwise separable convolution network (MABSCNET). The proposed hybrid framework was developed by combining modern deep learning and EL models pre-trained with the ImageNet dataset along with the MABSCNET model. In the experimental process, the effectiveness of the proposed methods was evaluated on the ISIC 2020 dataset. Additionally, additional experiments were conducted on ISIC 2018 and a Kaggle dataset to analyze the proposed hybrid framework's classification performance. Image enhancement techniques were used in the datasets. In the ISIC 2020 dataset, the MABSCNET model reached 78.63 % accuracy, the ViT model obtained 76.50 %, and the hybrid framework reached 92.74 % accuracy. Moreover, the proposed hybrid framework achieved 100 % on the ISIC 2018 dataset and 94.24 % on the Kaggle dataset.

A brain tumor diagnosis approach based on deep separable convolutional networks and attention mechanisms

Brain tumors rank among the most lethal types of tumors. Magnetic Resonance Imaging (MRI) technology can clearly display the position, size, and borders of the tumors. Hence, MRI is frequently used in clinical diagnostics to detect brain tumors. Deep learning and related methods have been widely used in computer vision research recently for diagnosis and classification of MRI images of brain tumors. One trend is to achieve better performance by increasing model complexity. However, at the same time, the trainable parameters of the model increases accordingly. Too many parameters will lead to more difficult model training and optimization, and are prone to overfitting. To address this problem, this study constructs a model that incorporates a depthwise separable convolution technique and an attention mechanism to balance model performance and complexity. The model achieves 97.41% accuracy in the brain tumor classification task, which exceeds the 96.72% accuracy of the pre-trained model MobileNetV2, and shows good image classification ability. Future work could test the model's classification results on noisy images and investigate how to optimize the model's generalization ability.

A brain tumor diagnosis approach based on deep separable convolutional networks and attention mechanisms

Brain tumors rank among the most lethal types of tumors. Magnetic Resonance Imaging (MRI) technology can clearly display the position, size, and borders of the tumors. Hence, MRI is frequently used in clinical diagnostics to detect brain tumors. Deep learning and related methods have been widely used in computer vision research recently for diagnosis and classification of MRI images of brain tumors. One trend is to achieve better performance by increasing model complexity. However, at the same time, the trainable parameters of the model increases accordingly. Too many parameters will lead to more difficult model training and optimization, and are prone to overfitting. To address this problem, this study constructs a model that incorporates a depthwise separable convolution technique and an attention mechanism to balance model performance and complexity. The model achieves 97.41% accuracy in the brain tumor classification task, which exceeds the 96.72% accuracy of the pre-trained model MobileNetV2, and shows good image classification ability. Future work could test the model's classification results on noisy images and investigate how to optimize the model's generalization ability.

CvT-UNet: A weld pool segmentation method integrating a CNN and a transformer

In the acquisition of weld seam images, the presence of light and impurity interference often leads to weak edge problems in weld seam images, making the identification of weld seam defects challenging. In such scenarios, segmentation processing must be performed specifically at the fusion pool of the weld seam. In recent years, Transformers, which leverage self-attention mechanisms, have demonstrated superior sequence modeling capabilities compared to convolutional neural networks (CNNs). The Transformer excels at weighting different positions within a sequence, enabling effective handling of long-range dependencies. However, Transformers lack the ability of CNNs to extract local information. To address this limitation, this paper introduces a novel CvT-UNet semantic segmentation model that combines the strengths of both the convolutional neural network and the transformer. The model combines the global context information capabilities of the Transformer with the spatial information advantages provided by the convolutional neural network. The proposed model employs an encoder-decoder architecture by building a U-shaped network with designed CvT blocks. Depthwise separable convolutions are utilized to balance parameter quantities, and the skip-connection module is redesigned to enable the model to achieve more precise segmentation of weld seam fusion pools with fewer parameters. Experimental evaluations on diverse weld seam datasets under different environmental conditions demonstrate promising results. The model achieves average intersection over union (IoU) values of 93.75 %, 88.31 %, and 90.86 %, respectively. Specifically, on an automotive weld seam dataset, CvT-UNet outperforms the purely convolutional neural network structure UNet3+, with a 1.63 % increase in the mean IoU (MIOU) and a 1.39 % increase in the mean pixel accuracy (MPA). Compared to the TransUNet fusion model, CvT-UNet results in a 0.33 % increase in the MIOU and a 0.6 % increase in the MPA. Furthermore, compared to the state-of-the-art curve segmentation model LIOT, CvT-UNet achieves a 1.94 % increase in the MIOU and a 0.61 % increase in the MPA, demonstrating favorable segmentation performance.

A two-stream-network based steganalysis network: TSNet

The three-stage framework consisting of preprocessing, feature extraction and classification is the most commonly used structure in traditional deep learning based steganalysis networks (SNs). Preprocessing, as the core part of the three-stage framework, typically adopts fixed high-pass filters to generate filter residuals, and exploits a feature enhancement module following high-pass filters to improve signal-to-noise ratio of the stego signal, thus strengthening the expression of steganographic features. However, fixed high-pass filters can only provide filter residuals through the linear relationships, rather than complex dependencies automatically learned, among local pixels, which suggests that fixed high-pass filters lack powerful ability of extracting comprehensive steganographic features from the beginning of SNs. Inspired by the motivation, a two-stream-network based SN called TSNet is proposed to extract comprehensive steganographic features from two different perspectives, namely automatic learning and fixed filtering. The discrete cosine transform attention mechanism (DCTAM)-based fusion module (DCTAM-FM) with considering the steganalysis properties is proposed to fuse two streams and enhance channel representation using DCTAM, which is conductive to improving the detection accuracy. Two streams aim to explore different types of residuals from two perspectives, but inevitably lead to a considerable number of parameters and high computational complexity. To this end, the feature extraction stage contains lightweight depthwise separable convolution blocks and the linearly increasing style of the number of channels. Combining the two-stream network, DCTAM-FM as well as the lightweight feature extraction stage, TSNet obtains deep composite features, which can fully express the comprehensive steganographic features, and thus brings a high detection accuracy. Experimental results from three benchmark datasets demonstrate that TSNet outperforms current state-of-the-art methods in detection accuracy and has an acceptable computational complexity.

OMC-YOLO: A Lightweight Grading Detection Method for Oyster Mushrooms

In this paper, a lightweight model—OMC-YOLO, improved based on YOLOv8n—is proposed for the automated detection and grading of oyster mushrooms. Aiming at the problems of low efficiency, high costs, and the difficult quality assurance of manual operations in traditional oyster mushroom cultivation, OMC-YOLO was improved based on the YOLOv8n model. Specifically, the model introduces deeply separable convolution (DWConv) into the backbone network, integrates the large separated convolution kernel attention mechanism (LSKA) and Slim-Neck structure into the Neck part, and adopts the DIoU loss function for optimization. The experimental results show that on the oyster mushroom dataset, the OMC-YOLO model had a higher detection effect compared to mainstream target detection models such as Faster R-CNN, SSD, YOLOv3-tiny, YOLOv5n, YOLOv6, YOLOv7-tiny, YOLOv8n, YOLOv9-gelan, YOLOv10n, etc., and that the mAP50 value reached 94.95%, which is an improvement of 2.62%. The number of parameters and the computational amount were also reduced by 26%. The model provides technical support for the automatic detection of oyster mushroom grades, which helps in realizing quality control and reducing labor costs and has positive significance for the construction of smart agriculture.

UniverDetect: Universal landmark detection method for multidomain X-ray images

Landmark detection using X-ray imaging is vital for disease screening, treatment, and prognosis. It provides a framework for subsequent tasks, including segmentation, classification, and target detection. Existing landmark detection methods exhibit limits with the increasing number of X-ray examiners. Not only they serve specific data types, but also their ability to learn the complete semantic information from this dataset is limited, potentially affecting their widespread adoption. We address these challenges by proposing a universal landmark detection model for multidomain X-ray imaging, UniverDetect. By progressively acquiring local semantic information, global semantic insights, and anatomical structure knowledge at various levels, the model ensured that the detected landmarks were closely aligned with the ground-truth labels. UniverDetect consists of three key components. The landmark detection module (LDM) utilizes a U-net network equipped with our innovative pyramid depthwise separable (PDS) convolution module for initial landmark detection. The landmark refinement module (LRM) integrates a fine-tuning module comprising continuously extended convolution blocks. Finally, the landmark correction module (LCM) incorporates a graph convolutional network (GCN) to rectify offset errors during partial landmark detection. An inherent feature of this model is its domain- generalization capability, which enables continuous learning across diverse domains. This model can concurrently learn from eight domains, covering 118 landmarks within a diverse dataset of 5969 images. Extensive experiments on multiple datasets demonstrate that this method consistently outperforms state-of-the-art approaches. This study presents a versatile and effective solution to reduce doctors’ workload while providing precise quantitative analysis.

A Low-Cost Laser Welding Monitoring Framework Based on Depth-Wise Separable Convolution with Photoelectric Signals

A Low-Cost Laser Welding Monitoring Framework Based on Depth-Wise Separable Convolution with Photoelectric Signals

Night target detection algorithm based on improved YOLOv7

Aiming at the problems of error detection and missing detection in night target detection, this paper proposes a night target detection algorithm based on YOLOv7(You Only Look Once v7). The algorithm proposed in this paper preprocesses images by means of square equalization and Gamma transform. The GSConv(Group Separable Convolution) module is introduced to reduce the number of parameters and the amount of calculation to improve the detection effect. ShuffleNetv2_×1.5 is introduced as the feature extraction Network to reduce the number of Network parameters while maintaining high tracking accuracy. The hard-swish activation function is adopted to greatly reduce the delay cost. At last, Scylla Intersection over Union function is used instead of Efficient Intersection over Union function to optimize the loss function and improve the robustness. Experimental results demonstrate that the average detection accuracy of the proposed improved YOLOv7 model is 88.1%. It can effectively improve the detection accuracy and accuracy of night target detection.

Detection Method of Manipulator Grasp Pose Based on RGB-D Image

In order to better solve the visual detection problem of manipulator grasping non-cooperative targets, we propose a method of grasp pose detection based on pixel point and feature fusion. By using the improved U2net network as the backbone for feature extraction and feature fusion of the input image, and the grasp prediction layer detects the grasp pose on each pixel. In order to adapt the U2net to grasp pose detection and improve its detection performance, we improve detection speed and control sampling depth by simplifying its network structure, while retaining some shallow features in feature fusion to enhance its feature extraction capability. We introduce depthwise separable convolution in the grasp prediction layer, further fusing the features extracted from the backbone to obtain predictive feature maps with stronger feature expressiveness. FocalLoss is selected as the loss function to solve the problem of unbalanced positive and negative samples in network training. We use the Cornell dataset for training and testing, perform pixel-level labeling on the image, and replace the labels that are not conducive to the actual grasping. This adaptation helps the dataset better suit the network training and testing while meeting the real-world grasping requirements of the manipulator. The evaluation results on image-wise and object-wise are 95.65% and 91.20% respectively, and the detection speed is 0.007 s/frame. We also used the method for actual manipulator grasping experiments. The results show that our method has improved accuracy and speed compared with previous methods, and has strong generalization ability and portability.

MC-ShuffleNetV2: A lightweight model for maize disease recognition

Maize has a long history of cultivation and is renowned for its high yield, superior quality, and adaptability. Currently, maize holds a significant position in grain cultivation and occupies a significant place in the agricultural structure. However, maize is susceptible to various diseases during its growth process, which can have a significant impact on the quality and yield. Traditional machine learning is heavily reliant on feature extraction, whereas deep learning has demonstrated notable success in image recognition for computer vision.The use of bloated models and the resulting wastage of computational resources represent significant challenges. The paper proposes a lightweight model, MC-ShuffleNetV2 (Mish + Convolutional Block Attention Module + ShuffleNetV2), to meet the practical needs of convolutional neural networks in maize disease image recognition. The model has designed with a focus on network lightweighting and accurate feature extraction. The model was constructed upon the foundation of the high-performance ShuffleNetV2 1 × network. The Convolutional Block Attention Module was integrated into the network architecture to enhance the model’s adaptive expressiveness. The depthwise separable convolution kernel of the depth-separable module was modified from a 3 × 3 kernel to a 5 × 5 kernel. This modification was implemented with the objective of expanding the image receptive field and extracting more detailed features of the image. It was necessary to modify the activation function in each stage for Mish. The model was compressed through the application of pruning operations. In the maize disease dataset test, the accuracy of the test set recognition accuracy of the network model constructed in this paper reaches 99.86 %, the model parameters are only 873,936, and the FLOPs (Floating-point Operations) are only 1,751,286. Compared with LeNet, AlexNet, MobileNetV2, and EfficientNetV2 models, the MC-ShufflenetV2 model’s recognition ability and size have obvious advantages, and it is more conducive to the actual deployment of the agricultural mobile terminal.

Optimizing Mobile Vision Transformers for Land Cover Classification

Image classification in remote sensing and geographic information system (GIS) data containing various land cover classes is essential for efficient and sustainable land use estimation and other tasks like object detection, localization, and segmentation. Deep learning (DL) techniques have shown tremendous potential in the GIS domain. While convolutional neural networks (CNNs) have dominated image analysis, transformers have proven to be a unifying solution for several AI-based processing pipelines. Vision transformers (ViTs) can have comparable and, in some cases, better accuracy than a CNN. However, they suffer from a significant drawback associated with the excessive use of training parameters. Using trainable parameters generously can have multiple advantages ranging from addressing model scalability to explainability. This can have a significant impact on model deployment in edge devices with limited resources, such as drones. In this research, we explore, without using pre-trained weights, how the inherent structure of vision transformers behaves with custom modifications. To verify our proposed approach, these architectures are trained on multiple land cover datasets. Experiments reveal that a combination of lightweight convolutional layers, including ShuffleNet, along with depthwise separable convolutions and average pooling can reduce the trainable parameters by 17.85% and yet achieve higher accuracy than the base mobile vision transformer (MViT). It is also observed that utilizing a combination of convolution layers along with multi-headed self-attention layers in MViT variants provides better performance for capturing local and global features, unlike the standalone ViT architecture, which utilizes almost 95% more parameters than the proposed MViT variant.

Memory-Augmented 3D Point Cloud Semantic Segmentation Network for Intelligent Mining Shovels.

The semantic segmentation of the 3D operating environment represents the key to intelligent mining shovels' autonomous digging and loading operation. However, the complexity of the operating environment of intelligent mining shovels presents challenges, including the variety of scene targets and the uneven number of samples. This results in low accuracy of 3D semantic segmentation and reduces the autonomous operation accuracy of the intelligent mine shovels. To solve these issues, this paper proposes a 3D point cloud semantic segmentation network based on memory enhancement and lightweight attention mechanisms. This model addresses the challenges of an uneven number of sampled scene targets, insufficient extraction of key features to reduce the semantic segmentation accuracy, and an insufficient lightweight level of the model to reduce deployment capability. Firstly, we investigate the memory enhancement learning mechanism, establishing a memory module for key semantic features of the targets. Furthermore, we address the issue of forgetting non-dominant target point cloud features caused by the unbalanced number of samples and enhance the semantic segmentation accuracy. Subsequently, the channel attention mechanism is studied. An attention module based on the statistical characteristics of the channel is established. The adequacy of the expression of the key features is improved by adjusting the weights of the features. This is done in order to improve the accuracy of semantic segmentation further. Finally, the lightweight mechanism is studied by adopting the deep separable convolution instead of conventional convolution to reduce the number of model parameters. Experiments demonstrate that the proposed method can improve the accuracy of semantic segmentation in the 3D scene and reduce the model's complexity. Semantic segmentation accuracy is improved by 7.15% on average compared with the experimental control methods, which contributes to the improvement of autonomous operation accuracy and safety of intelligent mining shovels.

DdeNet: A dual-domain end-to-end network combining Pale-Transformer and Laplacian convolution for sparse view CT reconstruction

Sparse view (SV) computed tomography (CT) is a clinical diagnostic technique aimed at reducing radiation dose to the human body from X-rays. However, SVCT reconstructed using conventional filtered back-projection (FBP) algorithms will produce severe streak artifacts. In recent years, convolutional neural networks (CNNs) have demonstrated considerable success in addressing the inverse problem of SVCT reconstruction. However, the local convolutional operations of CNNs ignores long-range dependencies of contextual feature information, thereby limiting their capability in feature extraction. To address this limitation, we propose a dual-domain end-to-end network (DdeNet) combining Pale-Transformer and Laplacian convolution for SVCT reconstruction. Initially, in the projection domain, a UNet is employed to restore the interpolated sinogram. Subsequently, in the image domain, we introduce Pale-Transformer and Laplacian convolution to design a dual-stream feature fusion network for finely restoring the CT images reconstructed by FBP. The dual-stream feature fusion network comprises an image restoration network (IRN) module and an edge enhancement network (EEN) module. The depthwise separable convolution based modified pale-shaped self-attention mechanism in Pale-Transformer captures global feature information of the CT image in the IRN, while the Laplacian convolution and multi-scale convolution block with a squeeze and excitation mechanism in the EEN further extract and restore edge feature information of the CT image. Finally, the feature information output from IRN and EEN is fused along the channel dimension and then convolved to obtain the final reconstructed CT image. Experimental results demonstrate that DdeNet exhibits superior generalization and reconstruction performance compared to the previous CNN or Transformer-based reconstruction methods.

Automatic soft-tissue analysis on orthodontic frontal and lateral facial photographs based on deep learning.

To establish the automatic soft-tissue analysis model based on deep learning that performs landmark detection and measurement calculations on orthodontic facial photographs to achieve a more comprehensive quantitative evaluation of soft tissues. A total of 578 frontal photographs and 450 lateral photographs of orthodontic patients were collected to construct datasets. All images were manually annotated by two orthodontists with 43 frontal-image landmarks and 17 lateral-image landmarks. Automatic landmark detection models were established, which consisted of a high-resolution network, a feature fusion module based on depthwise separable convolution, and a prediction model based on pixel shuffle. Ten measurements for frontal images and eight measurements for lateral images were defined. Test sets were used to evaluate the model performance, respectively. The mean radial error of landmarks and measurement error were calculated and statistically analysed to evaluate their reliability. The mean radial error was 14.44 ± 17.20 pixels for the landmarks in the frontal images and 13.48 ± 17.12 pixels for the landmarks in the lateral images. There was no statistically significant difference between the model prediction and manual annotation measurements except for the mid facial-lower facial height index. A total of 14 measurements had a high consistency. Based on deep learning, we established automatic soft-tissue analysis models for orthodontic facial photographs that can automatically detect 43 frontal-image landmarks and 17 lateral-image landmarks while performing comprehensive soft-tissue measurements. The models can assist orthodontists in efficient and accurate quantitative soft-tissue evaluation for clinical application.

ALdamage-seg: A Lightweight Model for Instance Segmentation of Aluminum Profiles

Aluminum profiles are widely used in various manufacturing sectors due to their flexibility and chemical properties. However, these profiles are susceptible to defects during manufacturing and transportation. Detecting these defects is crucial, but existing object detection models like Mask R-CNN and YOLOv8-seg are not optimized for this task. These models are large and computationally intensive, making them unsuitable for edge devices used in industrial inspections. To address this issue, this study proposes a novel lightweight instance segmentation model called AL-damage-seg, inspired by the YOLOv8n-seg architecture. This model utilizes MobileNetV3 as the backbone. In YOLOv8n-seg, the role of C2f is to enhance the nonlinear representation of the model to capture complex image features more efficiently. We upgraded and improved it to form multilayer feature extraction module (MFEM) and integrates a large separable kernel attention (LSKA) mechanism in the C2f module, resulting in C2f_LSKA, to further optimize the performance of the model. Additionally, depth-wise separable convolutions are employed in the feature fusion process. The ALdamage-seg’s weight on the Alibaba Tian-chi aluminum profile dataset constitutes 43.9% of that of YOLOv8n-seg, with its GFLOPs reduced to 53% relative to YOLOv8-seg, all the while achieving an average precision (mAP) of 99% relative to YOLOv8-seg. With its compact size and lower computational requirements, this model is well-suited for deployment on edge devices with limited processing capabilities.

A Lightweight Detection Algorithm for Surface Defects in Small-Sized Bearings

Background: To address issues in current deep learning models for detecting defects on industrial bearing surfaces, such as large parameter sizes and low precision in identifying small defects, we propose a lightweight detection algorithm for small-sized bearing appearance defects. Methods: First, we introduce a large separable convolution attention module on the spatial pyramid pooling fusion module. The deep convolutional layer with large convolutional kernels effectively captures more extensive context information of small-sized bearing defects while reducing the computation burden and learns attention weights to adaptively select the importance of input features. Secondly, we integrate the SimAM (simple attention mechanism) into the model without increasing the original network parameters, thereby augmenting the capacity to extract small-sized features and enhancing the model’s feature fusion capability. Finally, utilizing SIoU (Scylla IoU) as the regression loss and Soft-NMS (soft non-max suppression) for handling redundant boxes strengthens the model’s capacity to identify overlapping areas. Results: Experimental results demonstrate that our improved YOLOv8n model, sized at 6.5 MB, outperforms the baseline in terms of precision, recall, and mAP (mean average precision), with FPS (frames per second) of 146.7 (f/s), significantly enhancing bearing defect recognition for industrial applications.

Dual residual and large receptive field network for lightweight image super-resolution

In recent years, there has been rapid progress in single-image super-resolution (SISR) reconstruction technology based on deep learning, but many methods face challenges in practical application due to their excessive computational requirements. To address this issue, numerous lightweight super-resolution (SR) methods have been proposed. Convolutional neural networks (CNNs) have demonstrated outstanding performance in lightweight SR reconstruction tasks. However, existing CNN-based lightweight SR networks still suffer from redundancy due to the use of small convolutional kernels and deeper network architectures for feature extraction. To solve these problems, we propose a novel network called Dual Residual and Large Receptive Field Network (DRLN), which includes two main designs. Firstly, a Large Receptive Field Feature Extraction Block (LRFEB) is proposed, which concatenates large kernel depth separable convolutions and large kernel depth separable dilated convolutions, effectively capturing spatial contextual information without increasing the number of parameters or computational burden, and enhancing the overall performance of the model. Secondly, in the Efficient Feature Distillation Block (EFDB), we innovatively propose a dual residual structure, which greatly improve the learning ability and feature expression richness of the network, and significantly improve the reconstruction performance. Extensive experiments demonstrate that DRLN surpasses BSRN which is the champion of the model complexity track of the NTIRE 2022 Efficient SR Challenge on all benchmark datasets, with fewer parameters and less computation.

A lightweight spatially-aware classification model for breast cancer pathology images

Breast cancer is a prevalent malignant tumour with high global incidence. Its diagnosis relies primarily on the analysis of pathological breast images. Owing to the complex organisation of the tumour microenvironment, neural network models are essential as efficient classification tools in the field of pathological image analysis. This study introduced spatially-aware attention swift parallel convolution network (SPA-SPCNet), a lightweight and low-latency model for classifying breast pathologies. A novel module for multi-scale feature extraction was constructed using a depthwise separable convolution method. It focuses on the multi-scale features of pathological images to alleviate recognition problems caused by similar local features in breast cancer tissues. The module concatenates the convolutions of different kernels from three branches. Second, a lightweight dynamic spatially-aware attention module was introduced to integrate the visual graph convolutional architecture in a branch. This allowed the model to capture the spatial structure and relationships in image, enabling better handling of the unique spatial distribution relationship between breast cancer tissue structures. The other branch utilises a self-attention mechanism in the transformer. The module can dynamically adjust the attention of the model to different regions in the image, allowing it to focus on the key features of the complex spatial distribution of breast cancer tissue. This feature fusion method enabled the model to capture both global semantics and local details. Compared with existing lightweight models, the proposed model has advantages in terms of tissue structure classification accuracy, parameter quantity, floating-point operations, and real-time inference speed, providing a powerful tool for computer-aided breast pathological image classification.

A Method for Identifying Power Equipment in Foggy Substation Based on Improved YOLOv4

Abstract In foggy weather, inspection robots detect low visibility and unclear targets in substation power equipment, and the commonly used target detection algorithms have too many parameters, making it difficult to achieve real-time requirements when deployed on inspection robots. To this end, a lightweight substation power equipment identification method combining Cycle Dehaze defogging algorithm is proposed. Pre process the collected fog map of the substation to improve the clarity and visibility of the target power equipment; Introducing MobileNetv3 and deep separable convolution in YOLOv4 network to achieve model lightweight; To increase the precision of power equipment identification, incorporate the SE attention mechanism into the target detection network. Based on the testing results, it is possible to reach a recognition mAP of 92.5% by combining Cycle Dehaze and improved YOLOv4, as the improved YOLOv4 has a detection speed that is 15.59 frames/second faster than the YOLOv4 method. The model parameter quantity has been reduced by 80.21%, which is more conducive to deployment on substation inspection robots.