An Industrial Defect Detection Network with Fine-Grained Supervision and Adaptive Contrast Enhancement

Object detection approaches based on deep learning have made remarkable results in Automated Defect Inspection (ADI). However, some challenges still remain. Firstly, many defect objects lack semantic information, which causes the convolutional kernels tend to capture simple gray anomalies, thus making it challenging for the network to distinguish between foreground and background interference. Secondly, the poor image quality like low contrast makes it even more difficult for convolutional networks to extract effective features. To address these issues, this paper propose a one-stage defect detection network with additional fine-grained supervision to enable the model to learn richer features aside from the grayscale, as well as an image enhancement module to adaptively adjust image contrast and highlight object areas. Comprehensive experiments demonstrate significant performance improvements of our proposed method compared to the baseline and other defect detection methods, while maintaining high efficiency, which confirm the correctness and effectiveness of our model.