Surface Defect Recognition of NdFeb Raw Materials Based on Improved YOLOX-Tiny Model

With the development of science and technology, the detection of material surface defects through machine vision has become a new method. When the NdFeB (neodymium iron boron) raw materials manufacturing and transportation will have broken corners. The damaged raw materials need to be discovered in time. This paper proposes an improved lightweight YOLOX-tiny model, which can ensure high recognition accuracy. Also it will have a high FPS in real-time detection. First, in order to solve the problem of small data sets. We use the data set enhancement method to expand the number of images. In addition, the size and shape of defects are totally different. A ECAnet attention module could effectively solve the problem of low accuracy of detection. The attention module is placed in the back of the Back-bone of the model. Last, we introduce the sgd optimizer in the YOLOX-tiny model which can accelerate the training and recognition speed. In my computer, the detection result reaches 25 FPS and the mAP (mean average precision) reaches to 88.06%. This model has the potential and value to be applied in practical projects.