A Lightweight Reconstruction Network for Surface Defect Inspection

At present, the majority of deep learning approaches are unable to address the challenges posed by the limited availability of the industrial product samples that are defect and the substantial variations in their features. The present study introduces an unsupervised defect detection algorithm that utilizes a reconstruction network, requiring solely a large quantity of readily accessible sample data which is defect-free for implementation. The network consists of two parts: reconstructing image and detecting the area of surface defect. Reconstruction module is designed as a lightweight, fully convolutional autoencoder, trained with a restricted quantity of non-defective setsto create a defect-free reconstructed image. To address the challenge of detecting surface defects with irregular textures, we propose a loss function that combines structural and $L 1$ losses for the reconstruction network. Moreover, we leverage the dissimilarity between reconstructed and test images to identify potential defect regions, which can be located using classical image operations. We apply the proposed unsupervised algorithm for identifying defects based on the reconstruction network on various sets of defect images. The results reveal that our algorithm exhibits high accuracy and robustness compared to similar approaches. (Abstract)