Enhanced Multiview attention network with random interpolation resize for few-shot surface defect detection

Few-shot surface defect detection aims to detect or identify potential defects with limited data and has significant application value in improving the quality of industrial products. Challenges in few-shot surface defect detection include sparse samples and the diversity of defects. Existing methods typically employ transfer learning and meta-learning techniques, utilizing a single interpolation method for image resizing and single-view method for feature extraction. However, single-view feature extraction may lead to discriminative defect features being drowned out, and using a single interpolation method hinders the model's generalization capacity. To address these issues, we propose the ERNet (Enhanced Multiview Attention Model with Random Interpolation Resize). Firstly, we employ the Multiview Attention Module (MAM) by utilizing three parallel channel attention mechanisms to learn more discriminative defect features and enhance the diversity of the extracted defect characteristics. Secondly, we employ the Random Interpolation Resize (RIR) data augmentation method to enhance the diversity of training data and improve the model's generalization. Experimental results on the GC10-DET, NEU-DET, and TCAL datasets demonstrate that our method achieves outstanding performance across different experimental settings. The code is available at: https://github.com/lph656/ERNet.