A Novel Glass Defect Detection Method for Mobile Phone Cover Combining Vision Transformer and Faster R-CNN

Due to the high transparency and precision of mobile phone glass covers, there are still challenges in the production process, such as a large range of dimensional variations in defects, difficulty in capturing small defect features, and high computational requirements for high-precision image acquisition. These challenges make it difficult to meet the accuracy and speed requirements of surface quality inspection for production and quality control. Therefore, this paper proposes a novel method (ViT-Faster-RCNN) combining Vision Transformer (ViT) and Faster-RCNN to detect defects in glass covers, achieving rapid global detection of high-resolution image acquisition. By introducing the encoding and decoding structures of the Transformer framework into the backbone network, the model's ability to extract context information and global information is effectively enhanced, building upon its capability to accurately capture key local features. The model is able to ensure precise localization of defects of different sizes under complex background interference. Furthermore, the performance of the encoder is optimized by introducing residual structures and window attention mechanisms, which alleviates the gradient disappearance problem caused by network depth. This enhances the model's convergence speed and reduces computational complexity. The Simple Feature Pyramid Network (SFPN) is used to simplify the multi-scale feature fusion structure, further enhancing detection efficiency and accuracy. Experimental results demonstrate that the proposed method can accurately and automatically detect small defects on the glass cover plates, achieving an accuracy rate reaching 97.9%.