CAFS-Net: Context-Aware Feature Selection Network for Accurate and Efficient Tiny Surface Defect Segmentation

Although deep learning-based methods have achieved remarkable performance in the industrial defect segmentation area, tiny defect segmentation in ultrahigh-resolution image scenarios still remains unexplored. Most of the existing methods utilize attention mechanisms and a sliding window strategy for tiny defect segmentation. However, this approach is not only computationally demanding but also prone to texture noise interference, likely stemming from a lack of global contextual understanding. To alleviate this challenge, we propose a context-aware feature selection network (CAFS-Net) which consists of a context-aware encoder network (CANet), a novel feature selection module (FSM), and a lightweight feature-fusing decoder. The CANet is constructed by low-level convolutional blocks and high-level transformer blocks to capture both local and global context information, thereby enhancing the discrimination ability between tiny defects and texture noise. The FSM includes a multilayer perceptron (MLP) classifier and a selector for selecting defective features of image blocks from the feature pyramid of image blocks based on the classification outcomes. Then, the selected defective features are fed to the lightweight feature-fusing decoder to perform multiscale feature fusion and obtain the segmentation masks of the defective image patches. Additionally, we propose a concentrated random cropping data augmentation strategy to address the class imbalance problem during training. We conducted extensive experiments on two defect segmentation datasets, including the compact camera module (CCM) defect segmentation dataset and the printed circuit board (PCB) defect segmentation dataset, to demonstrate the superiority and generalization performance of our proposed model. The results show that our CAFS-Net outperforms other state-of-the-art (SOTA) methods in both accuracy and efficiency.