Few-Shot Classification of Screen Defects With Class-Agnostic Mask and Context-Based Classifier

In the real mobile screen production line, the yield of normal products is relatively high, so some defects are difficult to collect. Besides, samples of emerging defects are difficult to collect in large numbers in a short time. To address the few-shot problem caused by these, we propose a few-shot classification method to simultaneously recognize both the known defects with adequate training data and novel defects with scarce labeled training samples. First, we design a feature extractor with the class-agnostic mask. This module applies normal images to weaken irrelevant background information, thus extracting defect-highlighted features. Then, to avoid drowning limited important features in noise, the class prototype generator (CPG) is introduced to produce more representative class prototypes. Finally, a context-based classifier (CBC), which applies global context information, is proposed. This module can make better classification decisions among all classes. Experiments demonstrate that our method significantly outperforms other few-shot methods on the mobile screen defect dataset. The proposed method achieves 89.19% accuracy, 96.96% accuracy, and 97.55% accuracy in the one-shot, five-shot, and ten-shot test scenarios, respectively. The code is available at https://github.com/CFZ1/FSL_Cls_Screen.