Supervised semantic segmentation based on deep learning: a survey

Recently, many semantic segmentation methods based on fully supervised learning are leading the way in the computer vision field. In particular, deep neural networks headed by convolutional neural networks can effectively solve many challenging semantic segmentation tasks. To realize more refined semantic image segmentation, this paper studies the semantic segmentation task with a novel perspective, in which three key issues affecting the segmentation effect are considered. Firstly, it is hard to predict the classification results accurately in the high-resolution map from the reduced feature map since the scales are different between them. Secondly, the multi-scale characteristics of the target and the complexity of the background make it difficult to extract semantic features. Thirdly, the problem of intra-class differences and inter-class similarities can lead to incorrect classification of the boundary. To find the solutions to the above issues based on existing methods, the inner connection between past research and ongoing research is explored in this paper. In addition, qualitative and quantitative analyses are made, which can help the researchers to establish an intuitive understanding of various methods. At last, some conclusions about the existing methods are drawn to enhance segmentation performance. Moreover, the deficiencies of existing methods are researched and criticized, and a guide for future directions is provided.