Detection of Polarizer Surface Defects Based on an Improved Lightweight YOLOv3 Model

Polarizer is an important part of Liquid Crystal Display (LCD). Due to the selection of raw materials, the production process of polarize attachment, the transportation and storage of products, defects often occur, which not only seriously affect the display quality of LCD, but also lead to the scrapping of the whole LCD. Some previous methods mainly consider manual detection and traditional machine vision methods, but these methods show low accuracy and efficiency. In this paper, a lightweight polarizer surface defects detection method based on improved YOLOv3 is proposed. More specifically, firstly, MobileNet is introduced to replace DarkNet53 as backbone, which reduces the amount of network parameters and improves the detection speed. Secondly, the Mixed Convolution Efficient Attention (MECA) module is proposed to further improve the detection accuracy. Finally, the polarizer surface images are collected through the acquisition system, and the dataset of four defect classes are established. Experiments show that based on the above dataset, the mAP of the improved network is 89.8%, slightly higher than YOLOv3, and the average detection speed is 121 frames per second, increased by 44%.