Region based level sets for image segmentation: a brief comparative review with a fast model FREEST

• Published in: Multimedia tools and applications 2023-10, Vol.82 (24), p.37065-37095
• Main Authors: Feng, Chaolu, Chen, Shuaizheng, Zhao, Dazhe, Yang, Jinzhu
• Format: Article
• Language: English
• Subjects: Bias; Complexity; Computer Communication Networks; Computer Science; Data Structures and Information Theory; Image segmentation; Inheritances; Internal forces; Multimedia; Multimedia Information Systems; Smoothness; Special Purpose and Application-Based Systems
• ISSN: 1380-7501; 1573-7721
• DOI: 10.1007/s11042-023-15073-x

Region based level sets are one class of popular image segmentation models. Sorting out the inheritance relationship and comparing their performance on same image repositories are of guiding significance. In this paper, we first propose a generalization model to cover external forces of representative region based level sets and give a brief comparative review on them by describing their evolution process and performing a capacity and complexity analysis. We then briefly review regularizations on level sets, known as internal force in the literature. As the models become more and more complicated to perform well on challenging images, it is significant to ensure both segmentation performance and time performance. Thirdly, we propose a fast region based level set (FREEST) model to segment images with intensity inhomogeneities where smoothness of the estimated intensity bias field is ensured by a convolution operation. We then improve FREEST by introducing global intensity variances and rename it as FREEST σ to deal with images with different variances between objects of interest and the background. Experiments on representative 120 images (natural images from BSDS500 and well known synthetic images in the field) with simulated intensity biases show that time performances of the proposed models are close to the simplest but most famous level set model and its time incrementing is only about 1/20 of existing models. Qualitative and quantitative comparison with the representative models on the images in terms of Dice Similarity Coefficient and Jaccard Similarity Coefficient demonstrate advantages of the proposed models. Compared with the second-best model, the evaluation indicators increased by 0.09 and 0.13, respectively. Parameter settings and representative influence are discussed, which indicates robustness of the proposed models. Grand challenges and still open problems such as initialization sensitivity, complex background segmentation, and multi-class object segmentation are finally discussed. Codes will be released if this paper was accepted.