Few-Shot Defect Detection of Catheter Products via Enlarged Scale Feature Pyramid and Contrastive Proposal Memory Bank

The automatic detection of defects in manufacturing catheter production assumes a crucial role in ensuring safety within the downstream healthcare industry. However, existing deep-learning-based industrial defect detection methods commonly rely on large-scale datasets, making them unsuitable for catheter products with limited defective samples. Aiming at filling this gap, this work proposes a few-shot learning-based detection method to address three key challenges encountered in the quality inspection of catheter products: limited data, scale variance, and intraclass variation. First, a fine-tuning-based few-shot learning scheme is introduced to gain knowledge from the abundant base dataset in advance. Subsequently, an enlarged scale feature pyramid network is designed to cover the variant size of catheter defects. Finally, a contrastive proposal memory bank is put forward to alleviate the intraclass variation problem caused by different viewpoints and efficiently utilize similar features. Experimental results on the collected catheter defect dataset demonstrate the superior performance of our proposed method compared to other existing prevalent methods.