An Enhancement and Detection Method for a Glue Dispensing Image Based on the CycleGAN Model

During the active alignment focusing process of car camera assembly, lenses and holders need to be gummed by creamy white and translucent UV glue. The quality of glue dispensing can directly influence the performance of car cameras. Because of the translucency of UV glue, the glue dispensing image may present a low contrast situation, which increases the difficulty of vision detection. This paper proposes a method based on CycleGAN to enhance the glue dispensing image and effectively overcome the problems of blurred and low contrast edges. First, the glue part of the image is segmented into twenty regions. Second, the VGG16 model is used to divide the abovementioned twenty regions into high-contrast images and low-contrast images. Next, the CycleGAN model is trained to enhance the low-contrast images, and then convert them to high-contrast images. Finally, glue contours are extracted by using thresholding segmentation and edge detection to ensure that the quality of glue dispensing can be detected. The success rates of the VGG16 model and the CycleGAN model are 96% and 58%, respectively. The results show that the proposed method can effectively enhance the low contrast part of the glue region and improve the detection accuracy. Specifically, it can increase the gray value difference between the glue and the background from 20 to 55, while the background is substantially retained. The detailed information of the edges of the images is enriched. The accuracy of glue edge extraction can be increased to 99%, which is an approximately 75% improvement compared to the methods without enhancement.