Motion-Blur Robust MobileNet v3 with Clustered Data Purification for Aluminum Alloy Defect Classification

The automatic detection of surface defects in aluminum alloys is crucial for industrial quality control. However, current methods face three major challenges: (1) label noise in clustered datasets due to overlapping feature distributions, (2) performance degradation caused by motion blur in high-speed production lines, and (3) class imbalance resulting from the rarity of defects. To address these issues, we propose a dual-enhanced MobileNet v3 framework.First, we design a semi-automated data purification process that combines expert-guided minority class sample augmentation with confidence-based filtering to simultaneously optimize label purity and class distribution. Second, we develop a motion blur-robust training scheme, synthesizing directional blur kernels that mimic the motion of physical conveyor belts and integrating an attention module to capture global spatial dependencies.Experiments on industrial datasets demonstrate that our method improves baseline accuracy on purified data and achieves a Top-1 accuracy of 62.1% under strong motion blur conditions (blur kernel size of 8). This outperforms ResNet-50 by 54.5% under such conditions. The lightweight design, which has 55.2% fewer parameters than the Deep Multi-Patch Hierarchical Network based on ResNet-50, enables real-time deployment, highlighting its practical value in automated inspection systems.