Research On Image Semantic Segmentation Based On Deep Learning

Image semantic segmentation is one of the most challenging research topics in the field of computer vision.It makes realizing scene understanding more possible.It has a wide range of applications in many tasks,such as target detection,scene annotation and 3D reconstruction.The appearance of deep learning accelerates the development of image semantic segmentation,and many models have been proposed.In this paper,we use convolutional neural network to extract the feature of an image.We also study the combination of RGB feature and depth feature of input images,multi-scale image input and locating the boundary contour of an image using the fully connected conditional random field.The main work of this paper is as follows:(1)In order to solve the problem that the information of the input image may get lost after pooling layer of the convolutional neural network,we propose an encoding-decoding network structure based on the dilation idea.The method can extract the features of the RGB image and the depth image at the same time,and the depth information is added as additional information to the color information of the image in a manner of element summation,and ensure that the size of the output feature image is consistent with the size of the input feature image after the image is processed,and the receptive field of the network is not affected at the same.The network is trained and tested on the NYUv2 dataset,which is superior to the basic model named Fuse Net in the MIo U evaluation standard,and the feasibility of the proposed method is verified.(2)Aiming at the problem that the image semantic segmentation field can't make full use of the context information of the image and can't clearly segment the edge of the image,a network model combining multi-scale feature extraction and fully connected conditional random field is proposed.The input of the network is multi-scale RGB image and depth image.The convolutional neural network is used to extract the features of the RGB image and the features of the depth image respectively,and the depth information is added as supplementary information to the RGB feature map to obtain the semantic coarse segmentation result.Finally,the results of semantic fine segmentation are obtained by using the results of fully connected conditional random field optimization semantic coarse segmentation.The experimental results on the NYUv2 dataset show that the proposed method improves the accuracy of semantic segmentation and optimizes the edge ofsegmentation results,which has practical application value.