Research On Image Semantic Segmentation Algorithm Based On Convolutional Neural Network

Image semantic segmentation is a very important research direction in the field of image processing,computer vision and deep learning.Semantic segmentation is to classify the image pixel by pixel,so that the original image can be segmented into a semantic segmentation image with a specific pixel mark,which is the most challenging in image processing.With the rise of automatic driving technology,image semantic segmentation can accurately analyze and locate the object information in the scene,then to conduct navigation and precise guidance;in addition,it can also detect the diseases and insect pests on the surface of the plant through semantic segmentation.Therefore,image semantic segmentation is more and more meaningful.In the actual scene,image semantic segmentation also faces many challenges.First of all,at the object level,the effect of semantic segmentation is different due to the intersection and deformation of objects,and the influence of light and distance when photographing images;second,at the object category level,the differences and similarities between objects within and between classes also affect the semantic segmentation;finally,at the background level,the object background in the actual scene is often very complex,which is not conducive to semantic segmentation.The rapid development of image semantic segmentation benefits from the development of deep convolution neural network.The main purpose of this paper is to improve the effect and real-time performance of semantic segmentation in scene understanding.Using Pascal VOC 2012 data set as the object and Res Net101 as the backbone network.The proposed method first improves the effect of image semantic segmentation,and then improves the real-time performance.The main work are as follows:(1)First,we proposed joint feature pyramid networks(JFP)to improve the effect of semantic segmentation.Combined with the attention mechanism,we design an auxiliary semantic segmentation model to assist the training of the neural network and accelerate the convergence of the network.(2)Second,we proposed the SDC-JFP model by combining separable convolution and dilated convolution(SDC)and JFP model to improve the real-time performance of semantic segmentation.And we analyzed the effect of dilated convolution rate in separable dilated convolution on semantic segmentation.(3)Third we used the SDC-JFP model for reference,and removed ASPP model and auxiliary network to improve the real-time performance of semantic segmentation.We proposed a parallel normalization method(Parallel Batch and Instance Normalization,PBIN)and a cascaded normalization method(Cascaded Batch and Instance Normalization,CBIN)to improve the effect of semantic segmentation.The primary purpose of these three methods is to improve the effect of semantic segmentation,and further improve the real-time performance of semantic segmentation.The results also show that the three methods can improve the real-time performance of semantic segmentation step by step while ensuring the effect of semantic segmentation.