SegAuxClsNet: Segmentation Task-assisted Classification Neural Network for Surface-defect Detection

Feature sharing is widely employed in neural networks to enhance the generalization capability and robustness of models. Multi-task learning, a paradigm in machine learning, leverages the correlations between tasks by sharing feature representations to simultaneously address multiple related tasks, thereby improving performance. This involves balancing multiple tasks to enhance the performance of all tasks, which is often challenging. This study proposes a classification network that incorporates a segmentation auxiliary task, SegAuxClsNet. Both tasks share a backbone network for feature extraction, and the pixel-level information from segmentation masks can be utilized to enable the model to learn more discriminative features, thereby improving classification performance. Additionally, we introduce the Efficient Multi-Scale Attention Module (EMA) and Omni-Dimensional Convolution (ODConv) to enhance the model's ability to recognize multi-scale objects. The effectiveness of the model is validated on the industrial surface defect datasets KolektorSDD and KolektorSDD2, achieving classification APs of 100% and 95.01%, and classification recalls of 100% and 84.54%, respectively.