In response to the need of establishing an intelligent platform for wheel non-destructive testing, this paper proposed an automatic wheel hub defect segmentation method based on deep learning algorithms. This method used the structure of convolutional neural network and the nonlinear characteristics of radial basis function neural networks to construct a deep learning network structure to simulate human visual perception. This paper made use of the U-Net network to train a wheel hub defect segmentation model based on X-ray images of the car wheel. And based on the region of interest, the human brain hierarchical perception system was simulated. The hierarchical perception system can identify the gray pixels in the region of interest. By means of deep learning layered network and convolutional neural network, it extracted the intrinsic features of the defect area layer by layer, so as to realize the automatic segmentation of wheel defects. The experimental results showed that the recognition rate of this method for complex wheel defects reached over 90%, and the recognition time overhead was about 5ms / sheet, which was better than the traditional method. It can be seen that this method could meet the need for automatic segmentation of wheel hub defects and has potential application prospects.