Abstract
In industrial production, visual defect inspection is crucial for ensuring product quality. However, due to difficulties in sample collection during defect inspection, there is often a scarcity of diverse defective samples, leading most data generation methods to fail in generating high-diversity data. Therefore, this paper proposes the Automatic Iterative Diversity Improvement training strategy to improve the diversity of generated data with only a few training samples, which can improve the performance and generalization of downstream tasks. During the training process, given the characteristics of visual defects, we design the Defect Diversity-based Picking and Evaluation to pick up diverse samples and assess the diversity regarding defect location and shape. Experiments show that the method can generate high-diversity defect samples with high quality for industrial defect scenarios even in zero-shot learning tasks. The effectiveness of the generated data is also verified by the down-streaming defect detection task.

This publication has 12 references indexed in Scilit: