CCANet: Contextual Collaborative Aggregation Network for Lightweight Surface Defect Detection

Surface defect detection (SDD) is essential for quality assurance in consumer electronics manufacturing. Although existing SDD methods have achieved high accuracy in complex scenarios, their real-world deployment still faces two major challenges. First, high-accuracy models typically involve large model sizes, high computational costs, and slow inference speeds, which hinder real-time applications. Second, although lightweight architectures help to mitigate these issues, they often suffer from weakened feature representations, leading to performance degradation. To address these challenges, we propose a lightweight yet effective Contextual Collaborative Aggregation Network (CCANet) for SDD. CCANet employs a hybrid feature extraction backbone that integrates the first four stages of EfficientNet-B0 with a Context-Aware Vision Transformer (CAViT), enhancing both local and global feature representations while reducing model complexity. We further introduce a Context Collaborative Module (CCM) to leverage both low-level texture details and high-level semantic features. Additionally, we design an Asymmetric Enhancement Module (AEM) that uses asymmetric convolution kernels to explore orthogonal features and gradually fuse multi-level representations. Extensive experiments on four public SDD datasets validate the robustness and generalization of CCANet. For a 352× 352 input image, it requires only 1.85M parameters and 1.69G FLOPs, achieving real-time inference at 33.61 FPS.