Research on Few-Shot Defect Classification Based on Improved MobileNetV3

To address the limited generalization capability of detection models caused by emerging rare defect types in steel surface inspection, this study proposes a meta-learning framework integrating data augmentation and transfer initialization. The framework first acquires generic feature representations through large-scale pre-training, then enhances sample diversity via self-supervised data augmentation strategies. The MAML (Model-Agnostic Meta-Learning) algorithm enables rapid adaptation to new classification tasks, while the SE (Squeeze-and-Excitation) attention mechanism in Bottleneck modules is replaced with self-attention to optimize feature extraction efficiency. Experimental results demonstrate that the improved model achieves significant performance gains on the few-shot image classification benchmark FSC-20, with 34.5% higher accuracy and 29.68% improved F1-score under the 10-shot setting. This method provides an efficient and practical solution for industrial defect classification under few-shot learning conditions.