Rare Earth Magnetic Materials Surface Defect Recognition via Multi-Scale Residual Neural Network

The work aims to solve the problems of low detection accuracy, and poor reliability of magnetic material surface defects. A lightweight multi-scale feature ResNet (LMSF-ResNet) method was proposed in this paper. The method reduces the calculation and parameters of the neural network by using group convolutions and channel split operator and merging features from different branches of distinct scales by using the multi-scale model's structure. The algorithm employs the strategy of using the channel attention model to improve accuracy at the same time. This paper takes the defect images of rare earth magnetic materials as an example to verify its effectiveness. The results show that the method has higher detection accuracy and stronger stability, and can be widely used to accurately detect surface defects of magnetic materials.