Research On Image Semantic Segmentation Based On Encoder-decoder Structure

Image semantic segmentation is a key part of image understanding,and it is also a popular and challenging research direction.Semantic segmentation is based on pixel-level,high-precision image segmentation,dense prediction of each pixel in the image,so that each pixel is marked with the category of the corresponding object or region.Image semantic segmentation is different from simple tasks such as target recognition and classification.In addition to judging the category of each object in the image,it also needs to accurately describe the outline of the object,accurately locate the position information of the object,and make the image change.It has to be truly meaningful.It provides ground object classification information for urban planning,meteorology,oceanography,military reconnaissance,and other fields to assist scientific decision-making and deployment.Image semantic segmentation algorithms are developing rapidly.This article conducts in-depth research on current domestic and foreign related algorithms and results,and analyzes many challenges that still exist.For example,segmentation accuracy is easily affected by image quality in actual application scenarios,and large segmentation training sets cause storage Computational burden,lack of scale information,and spatial information affect the segmentation effect.Therefore,in response to the existing problems,this thesis proposes two image semantic segmentation algorithms based on encoder-decoder structure,namely,multi-scale image semantic segmentation based on convolutional neural network and image semantic segmentation based on global context information.The main research work of this thesis includes:(1)Aiming at the problem of poor segmentation effect due to lack of scale information in existing image semantic segmentation algorithms,a multi-scale image semantic segmentation model based on convolutional neural network is designed.In the encoder module,the improved Res Net network and the hollow space pyramid pooling structure are used to extract the high-level semantic information of the image,and the decoder module combines multiple outputs to fuse the low-level information of the image at various scales to solve the problem of loss of target details.The experimental verification on the PASCAL VOC data set shows that the algorithm has a good performance in object detail processing.(2)Aiming at the problems that existing semantic segmentation algorithms do not fully understand the context and the results of semantic segmentation are not refined enough,a model of image semantic segmentation that integrates global context information is designed.The network adopts an encoder-decoder structure,and the encoder part uses a double-branch structure.One branch is used to extract high-level abstract information of the picture,and the other branch uses the twin neural network to obtain global context information,which is fused to generate the final feature map.The decoder part selects low-level feature maps of the same resolution from the encoder for fusion,and finally obtains the segmentation result through upsampling.The experiment demonstrates that the segmentation result has been effectively improved.(3)In this thesis,an image semantic segmentation system is designed for the algorithm in this thesis,and the algorithm model proposed in this thesis is embedded in the system to verify the effectiveness of the algorithm in this thesis.