Research On Semantic Segmentation Based On Convolutional Neural Networks

Semantic segmentation plays an important role in image understanding,wearable devices,and smart car driving.Its purpose is to classify every pixel in the image,thereby segmenting the specific outline of the target,and has a wide range of research prospects and application values.The traditional segmentation method is limited by the accuracy and speed,and does not meet the requirements of complex scene understanding.With the development of deep learning technology,more and more researchers have used convolutional neural networks to solve basic visual problems and emerged a variety of different convolutional neural network models.The application of classified convolutional networks to the semantic segmentation problem has been pioneered since the fully convolutional network,which has promoted the rapid development of semantic segmentation,and the accuracy and speed performance have surpassed the traditional segmentation methods.Semantic segmentation generates enough spatial semantic information for each pixel to comprehend the content of the complex scene.Semantic segmentation of each pixel to generate enough information to understand the semantic space for complex scene content,segmentation models are usually designed around lower semantic information loss and enhanced details to expand and improve the accuracy of segmentation is still the main problem at this stage.On the basis of this problem,this paper develops semantic segmentation based on the convolutional neural networks.From the perspective of multi-scale feature integration,two different semantic segmentation models are implemented.The main work of this article is as follows:(1)In-depth study of the key technologies of semantic segmentation,including transpose convolution,atrous convolution and conditional random field,summarized and summed up the general framework of semantic segmentation.Based on the idea of fully convolutional networks,this paper implements a semantic segmentation model of DeepLab based on Aligned-Inception-ResNet network.(2)A segmentation model MsNet based on multi-scale feature integration is proposed.It includes two stages of semantic feature extraction and semantic feature integration.It can effectively eliminate the differences between the low-level features and high-level features and integrate the useful information contained in different convolution stages to improve the pixel classification accuracy and reduce the spatial position blur.MsNet as a technique to optimize the segmentation effect can smooth the edge contour of the target and bring more context information.Using the DeepLab segmentation algorithm as the reference network,MsNet-4 achieved a 5.4%accuracy improvement on the PASCAL VOC2012 dataset.(3)A scene semantic segmentation algorithm AugNet based on encoder-decoder structure is proposed.In the encoder stage,the DRN(Dilated Residual Networks)network is used to extract high-resolution semantic features and effectively maintain the receptive field of the pre-trained network.Based on the idea of multi-scale feature integration proposed in Chapter 3,the decoder stage is different from the simple linear interpolation method.It effectively combines the results of multi-branch prediction with the strong spatial information extracted by the PSPNet pyramid pooling module to enhance the spatial cues of scene understanding and local details.AugNet uses the scene datasets CityScapes and ADE20K to complete the verification.Compared with the PSPNet network,the segmentation effect is improved significantly.