Coarse-to-Fine Few-Shot Defect Recognition With Dynamic Weighting and Joint Metric

Deep learning-based methods have been widely used in the defect recognition and achieved great success. However, deep learning-based methods need a large-scale dataset. While in the real industrial scenarios, the training samples are always insufficient since the defect data acquisition is difficult and time-consuming. Therefore, in this paper, the few-shot learning theory is introduced to address the challenge. We propose to achieve the few-shot defect recognition in a coarse-to-fine manner with dynamic weighting and joint metric. In the coarse-grained phase, following the feature embedding, we propose an affine dynamic weighting module to control the embedding output of all channels according to the global context. By the dynamic weighting, the model can extract discriminative features better with learnable affine parameters. In the fine-grained phase, we propose a joint metric method which contains a K-L divergence based covariance metric module (KLCM) and cosine classifier. In this method, KLCM exploits the covariance matrix from the local descriptors to represent the distribution of a special defect class and then measures the similarity between the support and query defect images. A novel few-shot defect recognition dataset FSDR which contains a variety of defects from four different surfaces is constructed to evaluate our method. The results show the state-of-the-art performance compared to the mainstream methods.