Due to excellent mechanical and optical properties, cover glasses have become essential components in display modules for mobile phones and other electronic accessories. With the popularity of electronic displays and the development of smart devices, the demand for cover glasses has also increased year by year. However, the inspection of cover glasses can no longer meet the increasing market demand constrained by the inefficiency of artificial. Therefore, the develop of automatic defect detection equipment for surface defects of cover glasses is meaningful in practical applications.
Based on the above background, this paper aims to solve the common problems in the surface defect detection of cover glasses. This paper has conducted in-depth research on the bottleneck problems and established an on-line defect detection system. To deal with the difficulty of data acquisition, a data generation method based on a small sample set is proposed. For high-resolution image processing, a fast preprocessing algorithm is designed, and various detection methods are designed for corresponding defects. The main works accomplished in this paper are listed below:
（1）A compact and fast on-line inspection system is designed for common defects on the surface of printing cover glasses. The system which can acquire high-resolution images consists of a lighting imaging subsystem, a transmission subsystem, an image acquisition subsystem and an image processing subsystem. The cover glasses illuminated by the system pass through the imaging subsystem driven by transmission subsystem. Finally, the images captured by acquisition subsystem are transfered to the processing subsystem for detection.
（2）For the semi-transparent printing area such as the infrared radiation hole, an image segmentation algorithm based on deep learning is designed. This paper also proposes a performance improvement method based on adversarial training, which improves the robustness and noise immunities of the original model.
（3）Aiming at the difficulties of data acquisition and labeling the in the industrial vision inspection area, a data generation method based on a small unlabeled sets is proposed. The method uses a generative model to realistically render the results of the image synthesis algorithm, and the adversarial training method enables the generated model to learn the characteristics of the unlabeled data. The generative algorithm solves the problem of data acquisition to some extent in the field of visual inspection.
（4）This paper designs variety defect extraction methods. To improve the real-time performance in time-consuming tasks of high-resolution image processing, an adaptive binarization algorithm based on integral image is employed. Besides, a multidimensional feature based on statistical density information is designed to identify and classify weak defects.
In summary, this paper realizes key technologies in the on-line inspection system for printing defects. On the premise of guaranteeing the accuracy of micro-level detection, the detection and processing of typical defects can be completed within 3 seconds, which meets the efficiency of industrial production.