Low-Shot Defect Detection Method for Metal Multisurface Under Dynamic Uncertain Manufacturing Scenarios via Data Decoupling Augmentation

Industrial manufacturing environments are inherently unpredictable, leading to diverse defect manifestations influenced by factors such as ambient lighting, oil contamination, and varying damage levels. The limited diversity of defect samples constrains the effectiveness of deep learning models in defect detection. To address this challenge, this study introduces a low-shot defect detection method for metal multisurface defects under dynamic uncertain manufacturing conditions through data decoupling augmentation. The proposed framework features a progressive upsampling decoder (Conv_ResUp) that captures fine-grained features in small-scale feature maps using standard convolution, enhancing decoding residual connections and linear interpolation. Additionally, an adaptive discriminator disturbance augmentation (ADDA) module dynamically adjusts disturbance strength based on the discriminator’s overfitting level, ensuring stable adversarial training. A defect decoupling loss module further enables unsupervised attribute separation, improving defect generation quality. Experimental results demonstrate that the proposed method outperforms conventional generative models, producing high-fidelity, diverse defect samples that better align with real manufacturing conditions. When integrated with YOLOv5 and EfficientNet, the augmented training dataset enhances defect detection performance, improving mean average precision (mAP) and F1-score by 5.95% and 5.80% for end-face and tooth-face defect detection, respectively.