关键词: 缺陷检测与分割, YOLO-seg, 多尺度特征融合, 下采样, 上采样

Abstract: With the continuous updating of cell phone products, the production standards of cell phone screens have become more and more stringent. Given that traditional manual quality inspection methods are struggling to meet the rigorous demands of intelligent manufacturing, a screen appearance defects instance segmentation model — i. e. the double-head multi-purpose high-efficient YOLO (DMH-YOLO) is proposed. Firstly, a dynamic segment (DySegment) segmentation head is designed to realize the instance segmentation functionality of YOLOv7. Secondly, an innovative lobster attention downsampling module is incorporated into the head of the backbone network, which not only reduces the size of the feature map but also enhances feature representation capability while retaining more feature information, thus providing richer information for subsequent feature extraction. Furthermore, the original downsampling approach is improved by adopting Haar wavelet transform downsampling and space-to-depth downsampling techniques. Simultaneously, the upsampling is optimized to dynamic sample (DySample) upsampling to further enhance the accuracy. Lastly, to reduce the parameter count and computational complexity while improving computational efficiency, ELAN-tiny and SimSPPF are introduced into the neck network. Experimental results demonstrate that DMH-YOLO exhibits exceptional performance on an enhanced screen defect dataset subjected to augmentation processes such as flipping, Gaussian blurring, and the addition of salt-and-pepper noise. It achieves an average precision of 98.77% for defect detection and 96.74% for semantic segmentation. Compared to semantic segmentation models, instance segmentation DMH-YOLO offers broader applicability and more flexible in its application.

Key words: defect detection and segmentation, YOLO-seg, multiscale feature fusion, downsampling, upsampling