Military circuit surface defect detection method based on perception autoencoder

This study primarily addresses the task of surface defect detection in military-grade circuits. Given the diversity, small-batch nature, and complex surfaces of these circuits, existing methods demonstrate inadequate performance in image reconstruction. To resolve this, we propose a novel approach based on a Perceptual AutoEncoder(PAE), integrating perceptual loss and self-attention mechanisms into an unsupervised method, thereby enhancing the method's transferability and image reconstruction capabilities. Compared to traditional detection methods, our Perceptual AutoEncoder-based approach eliminates the need for alignment, illumination balance, and color balance, significantly improving transferability across different products and effectively addressing the challenge of detecting various small-batch military circuits. Specifically, we employ feature pyramids and convolutional methods to extract feature vectors of different scales and perform clustering. After clustering, we use the self-attention module to automatically weigh and enhance the features of interest, then reconstruct them into an image, which serves as a template for comparison with the input. Evaluations on our proprietary dataset indicate that the autoencoder, once integrated with perceptual loss, can more accurately detect defects.