Batch-normalized Convolutional Neural Networks for Defect Detection of the Steel Strip

Surface defect detection uses advanced machine vision inspection technology to detect defects such as spots, pits, scratches and chromatic aberrations on the surface of the workpiece. The traditional machine vision detection method requires manual selection of defect features as the basis of defect identification, which is time-consuming and laborious and has low accuracy in defect detection. To overcome the aforementioned deficiencies, the convolutional neural network (CNN) is proposed as a deep learning model to extract the defect features autonomously in an elegant way. In this paper, two smaller convolution kernels form a parallel channel in two layers of the convolutional neural network architecture, and then the results of the operation are fused to extract multi-scale information, which increases the adaptability of the network to scale. Besides, the batch normalization (BN) is introduced into convolutional neural network to standardize the data distribution, offering an easy starting condition for training and improving the generalization characteristics of the network. A steel strip defect data sets are adopted to conform the effectiveness of the proposed method. The experimental results show that the proposed method accelerate the training process through reducing the training epoch number, the accuracy and detection consistency on the steel strip defect data sets achieve a superior performance to the existing methods.