Research On Image Semantic Segmentation Based On Deep Learning

The image plays an important role in human perception.It is one of the media through which humans perceive the world,and the research work on images has never stopped.In recent years,with the improvement of computer computing power,the bottleneck of computing power limitation has been gradually released.Deep learning has become more and more popular in the field of image semantic segmentation.Compared with traditional semantic segmentation methods,deep learning methods have brought significant progress to the research of segmentation problems.Image semantic segmentation tasks have high-standard requirements.Its goal is to accurately identify each pixel category of a given input image,so as to obtain a pixel-level classification result map with the same size as the original image.It can be seen from this definition that semantic segmentation is not a completely independent research field,but is closely related to basic image classification,which can be regarded as a pixel-level classification problem.The classification informa-tion at the pixel level of the image not only includes the category information to which the pixel belongs.After the pixels are marked with spatial coordinates,the position infor-mation of the object can also be obtained.From the perspective of practical application,semantic segmentation technology,as a computer's visual perception organ,provides the possibility of machine automation for scene understanding tasks such as medical image analysis,human body analysis,satellite image analysis,and even automatic driving.In order to better complete computer vision tasks and enable them to be applied to real scenes as soon as possible,the study of image semantic segmentation tasks is an essential part.This article aims to explore the basic and challenging computer vision research tasks of image semantic segmentation based on deep learning methods.In addition to focusing on multi-scale feature information,semantic segmentation tasks also need to find a balance between local and global information.The information contained in the picture is diverse,and depending on the scene or combination of objects,the information presents different complexity.It is very difficult to understand the relationship between pixels based on local information.Although local information is very critical for pixel-level segmenta-tion prediction,the solution to the local ambiguity problem is still inseparable from global context information.The combination of local and global information can complete the segmentation task more accurately.Recent research attempts to integrate feature infor-mation of different depths to improve the performance of segmentation tasks.Some of them strengthen the features before the feature fusion,but focus on the enhancement of the overall feature,ignoring the local differences of the features.In other words,not all positions of every feature need to be enhanced.Therefore,it is still uncertain which areas of the feature should be enhanced and how to enhance.In addition,in real-time semantic segmentation tasks,such as automatic driving tasks with high real-time requirements,the existing high-precision networks are difficult to meet the real-time requirements due to their huge computational complexity.How to achieve a good balance between accuracy and speed has become a new challenge.For the above two problems,this paper proposes the following solutions:(1)We propose an adaptive feature enhancement method that adaptively enhances the key regions of low-level features by filtering high-level features.In addition,we combined this method to design an adaptive feature enhancement network to solve the problem of which areas should be enhanced and how to enhance.(2)For real-time semantic segmentation tasks that need to balance accuracy and speed,we use a lightweight network as the backbone to propose a new aggregation struc-ture design idea and build an aggregation structure network.In addition,we propose a network model training strategy with multiple auxiliary losses,which can improve the segmentation accuracy of the network model without increasing the complexity of the network structure and the amount of calculation in the test phase.A large number of experiments show that our method has good results on common semantic segmentation scene datasets.