Enhancing multi-scale learning with dual-path adaptive feature fusion for mixed supervised-industrial defect detection

Abstract In the surface defect detection of industrial products, existing deep learning methods heavily rely on a large amount of precisely labeled data, leading to high labeling costs. To address this issue, we investigate the use of hybrid supervision networks and explore the impact of different amounts of labeled data on detection performance. This paper focuses on key issues in the mixed supervision scenario, such as the high rate of missed detection for small defects and insufficient cross-scale feature fusion, and proposes a multi-scale feature fusion network architecture based on cross-layer feature interaction. This architecture is capable of extracting multi-scale feature information and effectively integrating it with segmentation masks to complete the surface defect detection task. The core contributions of this method include: (1) Designing a Dual-Path Adaptive Feature Fusion Module (DAFFM), which achieves effective fusion of low-level detail features and deep-level semantic features, significantly enhancing feature representation capability; (2) Introducing a low-to-high detail compensation residual branch that effectively captures the fine defect features lost in shallow networks. Experiments on the KolektorSDD2 and Severstal Steel datasets demonstrate that by adding less than ten percent of pixel-level labeled data, an average precision (AP) of 88.4% is achieved, representing a relative improvement of 6.25 percentage points over the existing state-of-the-art hybrid supervision methods. This work provides a practical solution for the industrial inspection field with lower labeling costs and better detection performance.