Automatically recognizing and grading spangle on the galvanized steels surface based on convolutional neural network

Currently, quality control of zinc barrier coatings on the galvanized steels, especially the homogeneity of spangle pattern, is becoming increasingly stringent due to the higher coating performance demanded from the end users. However, manual check by naked-eyes dominates the on-line inspection process in the steel factories, which is labor-consuming and reliant on experience. In this paper, the artificial intelligence method is applied to recognize spangles based on our proposed features which are different from the ones recognized by typical texture descriptors. During spangle recognition process, combination of Multi-column Convolutional Neural Network (MCNN) and Gaussian Mixture Model (GMM) is proposed to extract features, nuclei number and average nearest neighbor distance (ANND), for grading by support vector machine (SVM). The final grading results indicate that the accuracy reaches 91.3%, much higher than 41.3%−67.3% based on the typical texture descriptors, such as LBP, Laws, Harris, Laplace and SLIC. The effects of training data scale, network structure and hyperparameters of Gaussian kernel on the errors of counting and localization are discussed.