Gradient adaptive data augmentation and locally compensated feature fusion network for strip steel surface defect classification based on long tailed distribution

Abstract Defect classification is a crucial aspect of strip production. The existing methods for strip steel surface defect classification still require improvement due to the multi-category nature and complexity of defects, as well as the long-tailed distribution characteristics of the defect dataset. In this paper, a gradient-adaptive data augmentation and locally compensated feature fusion network for strip steel surface defect classification is proposed. Firstly, we put forward a gradient-adaptive data augmentation method based on the number of category samples to solve the long-tail distribution problem, which uses gradient-adaptive sample expansion to avoid increasing inter-class variation and preserves the probabilistic information contained in the category distribution through an average-based undersampling operation of the category samples. Secondly, a locally compensated feature fusion network is proposed, which is the first method to increase the multi-scale perception capability of the network using local compensation, mimicking the human behaviour of observing complex image patterns, which is compensating the feature information of the global image with the help of the feature information from the local image. The method not only improves the model’s ability to perceive defects, but also mitigates the negative effects of image compression. The experimental results on one public dataset and three private datasets demonstrate that the proposed method achieves significant classification performance improvement. Thus, it is founded that our method provides a promising solution for the strip surface defect classification problem.