Research On Image Semantics Segmentation Algorithm Based On Deep Learning

Image semantic segmentation is an important part of computer vision research and plays a key role in image understanding.Early traditional image segmentation is to divide the image into several areas with the same meaning through the features of color,texture,brightness and shape.However,the feature recognition of this segmentation technique is low,and only low-level features can be extracted when the image is segmented,resulting in poor segmentation performance.With the upgrading of computer technology and the successful application of deep learning in computer vision,image segmentation has gradually entered the stage of image semantic segmentation which can extract features through deep network.In recent years,the increasing strength of deep learning has greatly promoted the development of image semantic segmentation technology,so that many researchers at home and abroad have carried out in-depth research on image semantic segmentation algorithms based on deep learning,and achieved excellent results.However,due to the limitations of deep learning and the complexity of image semantic segmentation task,the image semantic segmentation algorithm based on deep learning still has many problems.First,for the small categories,the outline is too small to accurately locate the outline,resulting in the problem of blurred target edge and inaccurate segmentation.Secondly,in the research of segmentation algorithm,more and more algorithms use deep network to better extract features,but this has a greater dependence on hardware equipment,and the slow segmentation speed limits the application of algorithm;Moreover,the feature relations obtained in many existing algorithm structures are local,and they fail to extract semantic context information effectively and make full use of feature information among different levels,resulting in insufficient feature expression ability.Therefore,in view of the above three problems,this paper mainly does the following three aspects:(1)A semantic segmentation algorithm combined with edge detection is proposed.Firstly,the edge detection network and the semantic segmentation network are constructed in parallel.The edge detection network is used to extract the edge features of the image,and the semantic segmentation network is used to extract the preliminary semantic segmentation features.Then,edge features and semantic segmentation features are fused by feature fusion module to obtain the final semantic segmentation result.Finally,the improved algorithm is experimentally analyzed on the data set.Experimental results show that the proposed method can effectively solve the problems of fuzzy target edges and inaccurate segmentation in semantic segmentation,and effectively improve the accuracy of semantic segmentation.(2)A real-time semantic segmentation algorithm based on void separable convolution module and attention mechanism is proposed.Firstly,by combining the deep separable convolution with the void convolution of different void rates,a void separable module for feature extraction is designed.This module can extract features more efficiently while reducing the calculation amount of the model.Secondly,channel attention module and spatial attention module are added to the output end of the network,and they are integrated with the original features to enhance the expression ability of the features.Finally,the fused features are supersampled to the size of the original image to obtain the segmentation results.Experimental results show that the proposed method can not only reduce the size of the model,but also improve the semantic segmentation accuracy while ensuring real-time segmentation.(3)A semantic segmentation algorithm based on feature aggregation and bidirectional feature fusion is proposed.The method is based on encoder and decoder.At the encoder stage,the feature extraction is carried out through the backbone network of Resnet101,and the multi-scale context information is obtained by Atrous Spatial Pyramid Pooling(ASPP).Then,the multi-scale context information is fused with the features obtained from the backbone network.Finally,the fused features are weighted by features,and the relative importance of different scale information is modeled and selected to effectively aggregate the features.In the decoder stage,two-way feature fusion is used to fuse the high level features and the low level features.First,the low level features are subsampled and then fused with the high level features output by the encoder.Then,the fused features are upsampled,and then fused with the underlying features.Finally,the feature is output by up sampling and the segmentation result is obtained.Experimental results show that this method can effectively improve the accuracy of image semantic segmentation.