Research On Semantic Segmentation Method Based On Deep Learning Convolutional Neural Network

In recent years,with the rapid development of the field of deep learning,many emerging fields have been derived from the field of computer vision,among which semantic segmentation is also one of many emerging fields.This paper analyzes and improves the model from the perspectives of solving the problems of semantic segmentation in practical applications,respectively,from the perspectives of solving the accuracy of segmentation and the timeliness of segmentation.From the perspective of segmentation accuracy,the accuracy of the model is improved by introducing an attention mechanism to extract important features.From the perspective of timeliness,multi-size feature maps are used to extract features hierarchically to reduce the time required for the network to extract features.First of all,this article summarizes and analyzes related models of the research and development in the field of semantic segmentation in recent years.Summarizes the traditional semantic segmentation methods before deep learning,and analyzes the theory of related methods briefly.The classic semantic segmentation model and segmentation method using deep learning convolutional neural network are analyzed.Secondly,in view of the inaccurate positioning of the image target image by the Deeplabv3+ model,the inaccurate target edge segmentation,the slow image feature fitting,and the ineffective use of attention information,this paper proposes to use two attention branch modules to analyze the spatial information of the image and The channel information is extracted by feature fusion,and the extracted target features are post-processed with the fully connected conditional random field to optimize the detailed information of the target image.This paper designs the connection between the attention module and the encoding module of Deeplabv3+,and inputs the output features of the Deeplabv3+ encoding module into the attention module for convolution operation to realize the recalibration of the original features,and assign greater feature weights to important features.In this way,the problem of accurate positioning of spatial channel information by the model is solved.And design the decoding module to obtain the spatial feature and channel feature from the two branches respectively,and input the feature information obtained from the decoding structure into the fully connected conditional random field for local detail optimization to solve the problem of not obvious local detail optimization and rough segmentation boundary.The segmentation result of this model on the Voc2012 dataset is 1.95% higher than the original model,and the test result on the cityscape dataset is 1.05% higher than the original model.Finally,for the spatial pyramid pooling model's long operation time for image feature sampling,poor timeliness and high computational cost,this paper uses Mobile Net V2 to replace the network in the spatial pyramid pooling model to solve the problem of high computational cost,and design the corresponding codec structure to The edge feature extraction of the target image is refined,and the input method of images of different sizes is used to solve the problem of long operation time for feature sampling of high-resolution images.The model in this paper uses the Mobile Net V2 network as the basic network,designs the codec structure,and uses image inputs of different sizes for feature extraction of high-resolution target images.The model in this paper reduces the calculation time of semantic segmentation at the cost of slightly lowering the accuracy rate,and achieves the effect of real-time semantic segmentation.The time-consuming test of this model on the Voc2012 dataset is 44% less than that of the original model,and the time-consuming test on the cityscape dataset is 42% less than that of the original model.