Surface Defect Recognition of CSP Typical Slab Based on Improved ResNet

The CSP process is compact, and pseudo defects such as water strain, seam, and rolling scales cause the surface of slab to be very complicated, so it is necessary to develop a fast and accurate defect recognition model. In this paper, an improved ResNet is proposed to identity the surface defects of typical CSP slabs. The original ResNet model struc-ture is adjusted to halve the channel of convolution kernels, which can reduce the amount of calculation and enable the network to run under low conditions. Meanwhile, the location of the activation function is changed so that it is not used for shortcut connections, and we use exponential linear units (ELU) instead of rectified linear units (ReLU), which can speed up convergence and enhance model generalization ability. The results show that the defect detection accuracy of improved ResNet is 84%, which is 4% higher than that of original network, and the pseudo defect detection accuracy of improved ResNet is 97%, which is 1% higher than that of original network. The proposed network parameters are only 6 million, and the test time is 31 milliseconds.