This paper reviews the current status and trends of fully supervised learning methods in image semantic segmentation tasks. First, the currently popular fully supervised learning image semantic segmentation methods are classified and outlined, including methods based on codecs, DeepLab series, recurrent neural networks, image pyramids, attention mechanisms, probabilistic graph models, optimized convolutions, and generative adversarial networks. Then, the characteristics and limitations of existing methods are comprehensively evaluated, and their performance in terms of datasets, computational complexity, accuracy, and real-time performance is analyzed. Finally, this paper discusses the problems of existing methods and proposes future research directions and development trends. The research in this paper helps to provide researchers in related fields with a comprehensive understanding of the current status and trends of fully supervised learning methods in image semantic segmentation tasks, promote the development of image semantic segmentation technology, and improve its application effects and performance in various fields.

Keywords: fully supervised, semantic segmentation, current status

1. Introduction

1.1 Background of Topic Selection

With the rapid development of deep learning, many tasks in the field of computer vision have been greatly improved. Among them, image semantic segmentation, as an important computer vision task, has received widespread attention and has made great progress. Image semantic segmentation refers to dividing an image into several different regions, each corresponding to a different semantic category. This task has important application value in many application fields, and has been widely used in fields such as autonomous driving, medical image analysis, and security monitoring. In the task of image semantic segmentation, supervised learning methods are one of the most commonly used methods. It uses known labeled data as training data for the model and trains the model to achieve semantic segmentation of new data.

Fully supervised learning is a method that can use all pixels in an image and their labeled information to train the model. Compared with semi-supervised and unsupervised learning methods, fully supervised learning methods can achieve better performance in image semantic segmentation tasks. With the development of deep learning technology, more and more fully supervised learning methods have been applied to image semantic segmentation tasks.

1.2 Research Purpose

Fully supervised learning methods have been widely studied and applied in image semantic segmentation tasks. This paper aims to review the current application of fully supervised learning methods in image semantic segmentation tasks. First, the development history of fully supervised learning image semantic segmentation networks is reviewed, and the currently popular fully supervised learning image semantic segmentation methods are classified and outlined. Secondly, the characteristics and limitations of existing methods are summarized, and existing methods are comprehensively evaluated, analyzing their performance in terms of data sets, computational complexity, accuracy, real-time performance, etc., and their advantages and disadvantages are discussed from theoretical and practical perspectives. Finally, in response to the problems of existing methods, future research directions and development trends are proposed.

1.3 Research significance

The review research in this paper, through the systematic collection, organization, analysis and evaluation of fully supervised learning image semantic segmentation methods, helps to provide researchers in related fields with a comprehensive understanding of the application status and trends of fully supervised learning methods in image semantic segmentation tasks, and helps to further promote the development of image semantic segmentation technology and improve its application effects and performance in various fields. At the same time, the research in this paper can also provide guidance and reference for developers in related fields, helping them to develop efficient and accurate image semantic segmentation applications more quickly.

2. Comprehensive Analysis

2.1 Horizontal Comparison

In general, fully supervised learning image semantic segmentation methods have their own advantages and disadvantages, and the most appropriate method needs to be selected according to specific tasks and data sets. The codec-based method has the advantages of simplicity and efficiency, but lacks details and location information; the DeepLab series of methods uses dilated convolution and ASPP modules to improve the receptive field while retaining location information, but the amount of calculation is large; the recurrent neural network method considers sequence information by establishing temporal correlation, but it is prone to gradient disappearance and explosion; the image pyramid method performs multi-scale segmentation at different scales, but it is prone to insufficient context information; the attention mechanism method improves segmentation performance by introducing spatial attention mechanism and channel attention mechanism, but the model complexity is high; the probabilistic graph model method improves segmentation accuracy by modeling the interaction relationship between pixels, but it needs to manually annotate the interaction relationship between pixels; the optimized convolution method improves segmentation performance by optimizing the convolution kernel, but it takes a long time to train; the generative adversarial network method improves segmentation accuracy by introducing adversarial loss, but it is prone to training instability.