Research On Image Semantic Segmentation Algorithm Based On Deep Learning

Image semantic segmentation technology becomes much more popular in the computer vision field during recent years due to its significance in military and civilian appliance.With higher requirements of military intelligence,the semantic information plays an important role,which further promotes the research of semantic segmentation technology.In this paper,the image semantic segmentation algorithm is studied by using the scene and objects from large-scale data set as research objects.The research content mainly includes the following three parts:(1)The current development status of semantic segmentation technology is introduced,and the current mainstream semantic segmentation algorithms are also summarized.Meanwhile,this paper describes the two most widely used two-dimensional semantic segmentation datasets PASCAL VOC 2012 and Cityscapes datasets in detail,and elaborates the essential evaluation indexes of semantic segmentation.(2)Aiming at the deficiency of the Full Convolutional Neural Network(FCN)algorithm,an image semantic segmentation algorithm based on super pixel is proposed.The algorithm front-end neural network model adopts pyramid pooling structure to fuse multi-scale features,and introduces a new supervised loss strategy in Resnet.The back-end introduces super-pixel segmentation SLIC algorithm and fuses super-pixel information to design a super-pixel semantic annotation strategy.This algorithm effectively combines the advanced features of neural networks with the low-level features of superpixel segmentation,and then achieves a good semantic segmentation effect.(3)Based on the idea that the deep learning framework is used as the front end and the probability graph model as the back end,the semantic segmentation algorithm based on conditional random field is deeply studied for the shortcomings of the conditional random field model.The deep learning front-end framework introduces Atrous convolution technology and designs the Atrous convolutional space pyramid model.The back-end is optimized with an improved conditional random field and compared with existing conditional random field correlation algorithms.The algorithm realizes the effective fusion of the deep learning framework and the improved conditional random field.It greatly improves the algorithm time while ensuring the accuracy is basically unchanged,and achieves a good semantic segmentation effect on the standard dataset.Based on the full investigation of the existing semantic segmentation algorithm,this paper studies the algorithms under two different deep learning frameworks based on the idea of combining context information,using pyramid pooling,Atrous convolution and other techniques,using superpixel segmentation and conditions.With the method of random field and other methods,the algorithm studied in this paper has achieved good results and has certain research prospects.