FS-MSDANet: A Multi-Scale Dynamic Attention Network for Surface Defect Detection in Wear-Resistant Steel

Abstract

Wear-resistish steel is widely applied in industry for its outstanding performance under harsh conditions. And the surface defect plays an increasing role to guarantee quality in this domain. However, the state-of-the-art deep learning methods for surface defect detection is susceptible to overfitting and exhibit poor precision due to limited datasets. To solve these troubles, a few-shot surface defect detection approach, termed Few-Shot Multi-Scale Dynamic Attention Network (FS-MSDANet), is presented in this paper. First, an Improved Cycle Generative Adversarial Networks (ICycleGAN) is constructed by introducing shape consistency constraints into Cycle Generative Adversarial Networks (CycleGAN). By integrating this model with diverse data augmentation techniques, a multi-level data augmentation strategy is developed to expand the dataset and achieve diversified data representation. Second, within the multi-scale convolutional attention framework, we optimize the convolutional structure and the channel-spatial attention mechanism, which not only significantly reduces computational overhead but also enables dynamic perception and detection of multi-scale defects. Lastly, the feature space constraint module, which allows the model to maintain accurate and efficient detection performance in small-sample tasks with fuzzy class borders by introducing metric learning constraints. Ablation studies confirm that FS-MSDANet attains detection accuracy improvements of 2.72%, 2.13%, 2.26%, 3.55%, 3.15%, and 1.55% over the YOLOv12 model for varying sample sizes ( $k =1$ , 2, 3, 5, 10, 20), respectively. Furthermore, subsequent comparisons reveal that FS-MSDANet performs well in accuracy and adeptly addresses challenges such as limited training samples, overfitting, and inadequate generalization.