Self-Training Semi-Supervised Semantic Segmentation Combined with Image Search

Self-training semi-supervised semantic segmentation technology has exerted a crucial role in semantic segmentation tasks. It not only significantly reduces the reliance on a large quantity of manually annotated data but also effectively enhances segmentation performance. Consequently, this technology has garnered widespread attention in various practical applications. However, the pseudo labels' quality problem remains a major challenge for self-training semi-supervised semantic segmentation, seriously impacting the overall performance of the model. This article presents a series of targeted improvement measures to address this problem. We have designed an advanced image search module based on self-trained semi-supervised semantic segmentation, aiming to fully utilize annotated data to improve the quality of pseudo label generation. This module enhances the model's access to high-quality annotation information by precisely matching and filtering relevant samples. In the experiment, our method was extensively tested on the PASCAL VOC2012 dataset. The experimental results show that under annotation ratios of 1/16, 1/8, and 1/4, the mIoU accuracy of our method is improved by 0.41%, 0.59%, and 0.31 % respectively compared to traditional methods. These results not only demonstrate the outstanding performance of our method in segmentation tasks but also confirm the effectiveness of our method. Compared with existing self-training semi-supervised semantic segmentation techniques, the approach proposed by this paper demonstrates higher accuracy in segmentation performance, laying a solid foundation for future research and applications.