Surface defects inspection of cylindrical metal workpieces based on weakly supervised learning

Weakly supervised learning applies image tag labels to train convolutional neural networks locate defect. In industrial vision system, metal surface is anisotropic under light in all directions and it inevitable cause local overexposure due the natural reflection of active strong light, especially on cylindrical surface. this paper, injector valve taken as representative workpieces. Since variety complexity workpiece defects which pixel-level annotation require expensive manual work. This problem hinders application network industries. order solve these above challenges, paper proposed an end-to-end weakly framework named Integrated Residual Attention Convolutional Neural Network (IRA-CNN). IRA-CNN only uses for training performs defect classification segmentation simultaneously. achieved by extracting category-related spatial features from scores. composed multiple Block (IRA-Block) feature extractor improves accuracy achieves real-time performance. IRA-Block adds Module (IAM) includes channel attention submodule submodule. The adaptively extracts map improve its bilateral nonlinearity robustness. IAM can be well integrated into makes suppress interference useless background area highlight area. Satisfied performance method our own dataset could meet requirements process. Experimental results show that has good generalization ability. reaches 97.84% significantly improved compared with benchmark method.