Improving surface defect detection with context-guided asymmetric modulation networks and confidence-boosting loss

Abstract

Segmentation networks based on deep learning are widely used in the field of surface defect detection to ensure product quality. However, due to the complexity of defects and limited datasets, it is difficult for segmentation networks to achieve good performance in single-shot predictions. Therefore, we propose a novel context-guided asymmetric modulation network (called CAMNet) to improve the segmentation performance of existing methods. Inspired by the idea of cascading, CAMNet employs a coarse-to-fine segmentation framework that treats coarse predictions as priors to refine the extracted deep features. For this purpose, two asymmetric feature modulation modules, APM and ACM, are constructed in the spatial and channel dimensions, respectively. They are specially designed to leverage the rich context to aggregate and update features for fine prediction, and they are also lightweight and efficient, allowing for a significant reduction in computational complexity without sacrificing performance. In addition, a confidence-boosting loss is proposed to further widen the performance gap between the two prediction stages of CAMNet. Extensive experiments on three industrial datasets (PCB, Magnetic-tile, and DAGM2007) confirm the effectiveness of our method, showing that CAMNet can consistently achieve performance gains across different baselines and input scales. For example, CAMNet improves the IoU scores of FCN and GCNet on DAGM2007 by 5.92% and 4.98%, respectively.