Ambiguous surface defect detection and classification of display module

Abstract

In precision industries, the defect detection and classification based on machine vision approaches are the endless problems, which are directly related to the quality control and the productivity. In recent years, the mobile display industry has grown rapidly. As new types of display panel modules and production methods are being deployed, surface defects have become harder to inspect using conventional methods. In these cases, mostly the final inspection relies on the human visual inspection. Due to the ambiguous boundaries, uneven illumination on surfaces, and different light source effects of many vision systems, these defects cannot be handled simply by applying the traditional inspection algorithm. To be more generalizable and deployable, the inspection algorithm has to be able to manage diversities that come from different vision systems including various background patterns and local illumination differences in high resolution images. In this study, we present a novel framework to detect and classify surface defects on a display module that is applicable in practical terms. First, to solve the faint and ambiguous shaped surface defects detection problem, we propose a novel threshold-free defect detection method which is robust against variations in local illumination, structured patterns and noises for both color and gray cameras. We built a Gaussian mixture model (GMM) for modeling defect denoted as foreground and non-defect regions denoted as background separately so that no tuning threshold is necessary for detection. We exploit additional saliency information in the GMM learning scheme to enhance the discriminative power. This model initializes automatically the background components with credible pixel points that are chosen by the saliency filter value. The probability for the foreground is guided by the saliency filter for each iteration and only pixels in the foreground components are allowed to move to the background components to maintain reliable the background components. Compared to traditional inspection methods which need various preset parameters for morphological image processing depending on the surface regions and light source configuration, the proposed method is a threshold free detection method. Second, to solve the problem of ambiguous surface defect classification, we introduce a novel filtering method called the neighboring difference filter (NDF) that effectively separates the foreground defective regions from the background, which has structured patterns, local illumination variation, and different light conditions for each of several cameras in an inspection system. The NDF is designed to strengthen the information given by dissimilar patches and avoid that of similar ones using the properties inherent in textured patterns that distinctive information is more likely to be found further away from an image patch which contains a pixel of interest. Applying the NDF to defect images, we propose an optimum feature composition that composed of geometrical, statistical, intensity, and texture features by adopting a wrapper based feature selection method using a random forest as a learning algorithm. Successful results of our proposed approach are demonstrated using a real world dataset provided by an industrial production plant that show the possibility of the fully automated detection and classification on an industrial line, for which the inspection is currently performed by cumbersome human visual inspection.