MCL-GFAA: Multi-Scale Continual Learning with Gated Feature Alignment for Industrial Surface Defect Segmentation

Accurately segmenting defects on the surface images of industrial product is crucial for ensuring product quality. However, existing defect segmentation models usually suffer from missing detections and false alarms due to the subtlety and diversity of defects, and the lack of continual learning capabilities for adaptively recognizing new-category defects and noises in real-world applications. To solve these issues, this paper proposes a multi-scale continual learning framework with gated feature alignment aggregation (MCL-GFAA) for industrial product surface defect segmentation. Firstly, to improve segmentation accuracy, we design a gated feature alignment aggregation module (GFAAM) to selectively fuse features using a gating mechanism in a fully connected manner and learn the transform offsets of pixels through a learnable interpolation strategy to effectively aggregate features and recognize small defects. Secondly, to achieve continual defect learning, we propose a multi-scale continual learning method (MCLM) to continually learn new-category discriminative defect features from multiple scales in different network layers, greatly enhance the model’s ability to generalize and cope with noise and new categories of defects. The experimental results show that our method outperforms other methods on various incremental defect segmentation tasks.