EAAI Journal 2026 Journal Article
A lightweight network for weak texture surface defect detection
- Lingxi Peng
- Binxiong Lv
- xun gao
- Haohuai Liu
- Guangyan Huang
- Zhiwen Yu
With the continuous improvement of industrial processes, massive obvious surface defects are replaced by weak texture surface defect while with fewer defective samples. Currently, most existing defect detection models are designed for detecting the former but not suitable for detecting the latter, since the latter needs more efficient lightweight models that are only trained by a small available sample dataset. Therefore, we propose a novel end-to-end lightweight network using Multi-scale Adversarial Mixing Loss function Network (MAML-Net) for weak texture surface defect detection. Different from other end-to-end network approaches, a multi-scale perception loss is introduced to achieve a min-max game between the generator and discriminator, enabling the generator to achieve good segmentation performance. In addition, we propose a novel data augmentation method to deal with small-sample scenarios and introduce a novel self-distillation scheme to enhance the segmentation performance of the model. The experiments on two public datasets (Kaggle State Farm Distracted Driver Detection 2 and Magnetic Tile), show that the proposed MAML-Net achieves higher segmentation quality, better time efficiency and more accurate detection performance. In particular, compared to several counterpart methods, the proposed model achieved the highest efficiency (single-image processing time of 20 ms) and the highest area under the curve (above 98 %). It also attained the highest values in the Cost-Performance Score and Balanced Performance Score.