Image Segmentation Techniques in Computer Vision

This review paper provides an in-depth study of the image segmentation algorithms in computer vision, focusing on academic literature from 2020 to 2024. Specifically, this paper aims to analyze the development of image segmentation algorithms, starting from transformer networks and self-supervised learning models to vision-language models and foundation models. A total of 30 articles are selected for systematic literature review, which are highly cited papers in top-tier conferences like CVPR, NeurIPS, ICCV, and ECCV. Based on the results of this study, the years from 2020 to 2024 can be considered a groundbreaking period in the history of image segmentation, when the new era of foundation models has started with the invention of Segment Anything Model (SAM) and its variants, which have enabled promptable and zero-shot image segmentation for the first time. Contemporary algorithms exhibit significantly better performance than their predecessors based on CNNs in terms of accuracy and generalizability, although problems with high computational cost, real-time application, and domain adaptation have yet to be solved.