Surface Defect Segmentation Based on Multi-Scale Hybrid Encoder and Mixture of Experts

In industrial production processes, the application of deep learning techniques for surface defect identification is crucial for current product quality. However, many existing deep learning-based detection methods are primarily intended for general object detection. The small size, irregular shapes, and low distinguishability from the background of defects present significant challenges to these methods. To solve this issue, this paper proposes an efficient, plug-and-play Multi-Scale Semantic Enhancement module. The module adopts a CNN-Transformer hybrid encoding approach, effectively combining the strengths of both, thereby providing the network model with richer global semantic information and higher resolution detail. Additionally, the module incorporates the Mixture of Experts mechanism, which dynamically selects different expert sub-networks to process various types of defects and their associated semantic information, significantly improving the network’s flexibility and efficiency when handling complex defects. Extensive experiments performed on the SD-Saliency-900 and Magnetic Tile Defects datasets validate the effectiveness of this module. With an approximate 5% increase in model parameters, the module achieves improvements of 3.16% and 4.09% in mIoU, respectively.