Relational Enhancement Network for Industrial Defect Detection

As industrial manufacturing quality standards rise, demand for advanced defect detection models has surged. Compared to generic objects, industrial defects exhibit more diverse and complex shapes and sizes. Traditional detection models typically process each instance in isolation, leading to incomplete detections (e.g. fragmented or redundant bounding boxes) when facing such complex defect patterns. To address these challenges, we propose Relational Enhancement Network for defect detection, which enhances defect features by exploring implicit spatial and semantic relations. Our model introduces a position embedding module to map geometric features into a high-dimensional space. A relational enhancement module is proposed to integrate geometric and semantic features, capturing complex interactions among defects to enhance the original features. This process is dynamically adjusted through a relational refining mechanism. The proposed position-sensitive loss further aligns classification task with localization task using spatial metrics. Experiments on three industrial defect benchmark datasets (metals, bearings, engines, and LEDs) show our method outperforms state-of-the-art approaches in detection precision and addresses incomplete defect detection. Additionally, our method exhibits strong transferability, theoretically offering clear improvements to any similar-structured methods. The code is available at https://github.com/lhht/Relational-Enhancement-Network