A mask guided cross data augmentation method for industrial defect detection

Expert in pattern recognition, deep learning is widely used for defect detection. Data augmentation is in principle capable of improving the accuracy and robustness of such data-driven models by supplementing patterns. However, the requirements for realistic images, precise annotations, and diverse defect patterns often cannot be addressed simultaneously in current augmentation methods. In this work, a Mask Guided Cross Data Augmentation method dubbed MGCDA using diffusion model is proposed to boost defect detection. Firstly, a generation pipeline in latent diffusion space utilizing autoencoder is formulated to improve the fidelity and resource effort. Based on this, we propose to adopt conditional mechanism to enable samples being synthesized under the guidance of specific masks. To further enhance the information gain, a cross-learning strategy is proposed to empower MGCDA learning and generalizing diverse defect patterns from different categories, making detection more robust. Finally, two strategies are proposed to tackle the demand for data augmentation in different situations. Experiments on eight common industrial datasets show that MGCDA has high applicability to different scenarios and detection models, it can generate high-fidelity samples aligned to guidance and effectively improve the performance of baselines at both image- and pixel-level.