MCINet: a deep fusion network based on multi-scale context integration for industrial inspection

Abstract Automatic industrial inspection of products has become a key task for improving quality and efficiency in industrial production. Researchers have conducted in-depth research on this task using semantic segmentation technology. However, due to the characteristics of industrial defects, such as diversity, low contrast, blurred boundaries and anisotropic background texture, there are still challenges in small target and fuzzy boundary defect detection. This paper proposes a deep fusion network based on multi-scale context integration (MCINet) for pixel-level industrial inspection. A contextual aggregation pooling module is proposed as an encoder to extract multi-scale defect features effectively. To enhance context information in the decoder, a feature deep fusion module is introduced to carry out information exchange between features at different levels. Moreover, a spatial information perception module is proposed to enable the decoder to focus dynamically for accurate defect location and segmentation. Based on the above modules and the backbone, an efficient coding-decoding industrial defect detection network model is proposed. Extensive experiments on three challenging industrial defect datasets demonstrate that MCINet achieves competitive performance compared to state-of-the-art methods, while balancing accuracy and efficiency.