A Lightweight Image Segmentation Algorithm Based on YOLOv8

In recent years, deep learning-based image segmentation technology has obtained outstanding achievements in typical jobs, and an increasing number of application fields require efficient and real-time image segmentation. Therefore, research in this area has focused on enhancing the accuracy of image segmentation while maintaining a lightweight model. Addressing these issues, this article takes YOLOv8 as a baseline model and introduces ODConv, an omni-dimensional dynamic convolution, to improve the network structure of the backbone, enhancing its local feature extraction and contextual awareness capabilities. Additionally, a combined loss function is constructed that incorporates Tversky loss, with Tversky loss calculated separately for each detection head, focusing on optimizing the segmentation of small objects and the edges of adjacent objects. According to the results, while in the public COCO dataset, the modified YOLOv8 model has achieved an increase of 4.8% and 0.9% at the mask levels of mAP50 and mAP50-95, the box mAP has also improved by 4% and 0.5%, respectively. Compared to the YOLOv8 baseline model, the proposed lightweight segmentation method enhances both the precision and accuracy of segmentation, enabling rapid and accurate image segmentation, thereby effectively accomplishing the task of image segmentation.