Substation defect detection model based on self-supervised learning

To enhance the accuracy of substation equipment defect detection with a small number of labeled samples, an improved defect detection algorithm based on self-supervised model SimSiam was proposed. Different from the original SimSiam, the improved algorithm does not need to use iconic images, but directly uses non-iconic images for contrast learning. The performance of the supervised method is comparable to that of the downstream defect detection task. By using full convolutional network and spatial attention module instead of MLP in projection head and prediction head, the spatial structure and local information of high dimensional features are preserved. At the same time, before calculating the similarity, the eigenvector is obtained by means pooling of the feature graph, and the eigenvector is normalized to calculate the Euclidean distance, thus improving the loss function of self-supervised contrast learning. Experimental results show that the proposed algorithm can make full use of non-iconic images for contrast learning and improve the accuracy of substation equipment defect detection when only a few samples are marked.