LIN Gang, FENG Hao, CAO Ligang , PAN Haipeng, CAO Xuming

(School of Mechanical and Electronic Engineering, Jingdezhen Ceramic Institute, Jingdezhen 333403, Jiangxi, China)

Abstract: At present, the defect of daily-use ceramics is mainly manually detected. In the current visual inspection algorithms, there are various problems, such as limited defect type, high sensitivity to light intensity, poor adaptability and universality. In this paper, a defect classification algorithm based on multi-scale feature fusion was proposed. Firstly, a median filter was used to filter the detail noise. According to the gray scales of the training image and the test image, linear transformations with different modes were carried out to reduce the feature distance of the same defects and increase adaptability to the light intensity to a certain extent meanwhile. Secondly, the statistical data characteristics of multi-scale gray histogram were calculated. Then, the image was processed with the Sobel operator, in order to calculate the gray level co-occurrence matrix of the gradient image, thus resulting energy, correlativity, uniformity, contrast, entropy and isotropy. Finally, the statistical features of the gray histogram and the six groups of texture features in the gray level co-occurrence matrix were fused and then transferred to the KNN classifier for training, after which ceramic disc defect classification model was developed. This method exhibited an average defect recognition rate of 86.86% on the surface of ceramic disk, demonstrating the advantages of real-time and robustness.

Key words: linear transformation; multiscale gray statistics; adaptability; KNN algorithm; gray level co-occurrence matrix