Research On Semantic Segmentation Algorithm Based On Feature Fusion And Deep Learning

At present,image semantic segmentation is widely used in many different fields such as vehicle automatic driving,robot navigation,remote sensing image recognition,and medical imaging tumor detection,and it has a subtle impact on people's living habits.Image semantic segmentation is a pixel-level vision task.The model learned by the computer through training is used to predict each pixel classification of the input image,and each classification in the image is segmented by its corresponding annotation color.The huge application value makes image semantic segmentation get widespread attention,and it is a popular direction of machine vision research.However,in actual application scenarios,problems such as inaccurate segmentation of object edge contours,loss of segmentation details,and loss of information in connection parts will occur.These situations have gradually become a difficulty in image understanding.For these existing difficulties,an improved network structure based on the Pyramid Scene Parsing Network(PSPNet)is proposed.The main research work is as follows:(1)Propose an improved PSPNet+ algorithm based on PSPNet.First of all,in the feature learning module,the input image is based on the original residual network(Residual Network,Res Net)by adding convolution and pooling operations inside the network to further learn the characteristics of each level,and combine the learned multiple low-level Features and high-level features are added to obtain a new feature map with more spatial location information;to obtain rich context information,the PSPNet pyramid pooling structure is used to compare the global context information in the feature map with the local context information of different scales.Combine and perform convolution to obtain the final prediction map.The simulation experiment results show that the improved method has a mean intersection over union(MIo U)ratio of 78.5% in the PASCAL VOC 2012 test set,which is an increase of 1.7% compared with the benchmark algorithm.Perform ablation experiments on the proposed PSPNet+ algorithm and evaluate the original network algorithm of PSPNet and the three algorithms of PSPNet+y3,PSPNet+y2+y3,and PSPNet+ with different low-level feature maps.The results show that: Mio U of PSPNet It is 76.8%,the MIo U of PSPNet+y3 is 77.1%,the MIo U of PSPNet+y2+y3 is 78.3%,and the Mio U of PSPNet+ is 78.5%.(2)Visualized statistical analysis of the ablation experiment results of the PSPNet+algorithm using histograms.The analysis results show that the overall MIo U value is on the rise,and it is found that the pyramid pooling structure is used to predict the results of each classification is relatively stable,but from the accuracy results of each classification,it is not that the more low-level feature map information is introduced,the better the result will be.It will also bring some noise and affect the accuracy performance.The accuracy improvement of PSPNet+y2+y3 to PSPNet+ is only 0.2%.To change this branch,learn from the encoding-decoding idea,apply the CONCAT skip-connection,and directly increase the convolution depth at the decoder side,which helps retention of low-level feature information.The algorithm analyzed and improved on the proposed PSPNet+algorithm is named PSPNet++.Uploading the prediction map obtained through the PSPNet++ algorithm to the PASCAL VOC 2012 test set evaluation platform shows an accuracy result of 79.1%.Compared with PSPNet,the accuracy is improved by 2.3% and compared with PSPNet+ by 0.6%.The algorithm proposed in this paper is an improved end-to-end convolutional neural network algorithm based on the PSPNet,without any post-processing steps.Experiments have shown that it can effectively improve the accuracy of image semantic segmentation.To effectively improve.With the diversity and advancement of algorithms and underlying technologies,further experimental investigations are needed to obtain more excellent experimental results.