Abstract: In order to reduce the impact and improve the detection efficiency of transformer oil leakage inspection images, a transformer oil leakage detection model based on depth-separable void convolution pyramid is proposed. Firstly, the ordinary convolutional blocks in the cavity pyramid are modified to depth-separable convolutional blocks, so as to expand the pyramid sensing field and make the semantic information of the feature map extracted by the feature extraction network richer; then, the fusion process of the low-order semantic features and the high-order semantic features at the feature extraction stage is improved, so as to further enhance the semantic information of the feature map generated by the feature extraction network; finally, in order to avoid the loss of semantic information of the feature map after multiple convolution and pooling operations, a feature map is generated by the feature extraction network in the fusion process of the low-order and high-order semantic features. Finally, in order to avoid the loss of semantic information in the feature map after multiple convolution and pooling operations, spatial attention mechanism and channel attention mechanism are introduced in the fusion process to further enhance the semantic information in the feature map. Compared with UNet (Convolutional Networks for Biomedical Image Segmentation), PSPNet (Pyramid Scene Parseing Network), DeepLabv3+ (Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation), DeepLabv3+ (Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation), and MCNN (Multi-Class Convolutional Neural Network) algorithms are compared and experimented and it is found that the network detection model proposed in this paper is good, with the detection rate 76.85%, the average intersection and merger ratio reaches 64.63%, the recall rate reaches 73.56%, and the FPS (Frames Per Second) reaches 30 frames/second. In order to verify the effectiveness of the proposed method in this paper, ablation experiments are designed, and compared with the basic network model, the detection rate is improved by 9.33%, the average intersection and merger ratio is improved by 7.15%, and the recall rate is improved by 5.66%.
Download 2023.6.8