Hot-Rolled, Heavy-Rail Image Recognition Based on Deep-Learning Network

Abstract

A new method for image-defect recognition is proposed that is based on a convolution network with repeated stacking of small convolution kernels and a maximum pooling layer. By improving the speed and accuracy of image-defect recognition, this new method can be applied to image recognition such as heavy-rail images with high noise and many types of defects. The experimental results showed that the new algorithm effectively improved the accuracy of heavy-rail image-defect recognition. As evidenced by the simulation study, the proposed method has a lower error rate in heavy-rail image recognition than traditional algorithms, and the method may also be applied to defect recognition of nonlinear images under strong noise conditions. Its robustness and nonlinear processing ability are impressive, and the method is featured with high theoretical depth and important application value.