Research Advanced in the Image Segmentation

Due to the progress of computer technology, image processing-based applications become more and more important not only in advanced research fields but also in our daily life. For example, biomedical image processing, autonomous vehicles, video surveillance, computer vision, augmented reality, and many more are some hotspots in research as well as the applications. The technology of image segmentation is the basis of all these applications. In this review, we summarize the traditional image segmentation methods and algorithms including segmentation based on thresholding, segmentation based on graph theory, and segmentation based on region. Also, due to the exceptional performance of the CNN model, a series of CNN-based segmentation algorithms were proposed. Along with the invention of these new approaches, segmentation task has been pushed to a higher accuracy and efficiency level. Thus, we also summarize the major CNN-based model, such as the FCN model, the Encoder-Decoder designed architectures, the Multi-scale and Pyramid-based architectures, and Deeplab Family. Moreover, we provide some information on some popular datasets and metrics to rigorously evaluate a model, and we provide tables that contain the quantitative information on each model's performance. Even though the outcomes of current models have reached a significant level, there are still many problems that need to be tackled. At last, we provide our opinion regarding the current situation in image segmentation field.