Research On Semantic Segmentation Method Of RGB-D Image Based On Deep Learning

Semantic segmentation of images is the cornerstone of image understanding.In recent years,with the rapid development of deep learning,image semantic segmentation using deep learning has become a topic of concern in the field of computer vision.The convolutional neural network is an important method of deep learning.In recent years,it has performed well in image semantic segmentation tasks.With the popularity of RGB-D image capture devices,more and more image analysis tasks incorporate depth information.Therefore,this paper introduces a semantic segmentation method of RGB-D image based on convolutional neural network and improves it based on this.The main work is as follows:(1)Analyze the image semantic segmentation method based on the color and depth information of the convolutional neural network.At the same time,the fusion principle of color information and depth information and two fusion methods is studied.The symmetric coded network model combines the advantages of the two kinds of information and solves the problem that the original semantic segmentation network model cannot fully utilize image color and depth information to some extent.(2)For convolutional and downsampling in convolutional neural networks,the problems such as unclear edges and inaccurate segmentation are calculated.This paper proposes an improved RGB-D semantic segmentation network model using full join conditional random fields.Fully connected conditional random field function is calculated using the output of the RGB-D semantic segmentation network model.The binary potential function of the full-connected conditional random field is calculated using the original input RGB image,and then the two are combined to obtain a fine semantic segmentation.As a result,problems such as the inability to make full use of the relationship between pixels and the unclear boundary between the models can be further solved.(3)For the above network model,experiments were conducted in the NYU Depth dataset in the SUN RGB-D dataset to complete the data preprocessing and network model training.Then,the trained model was used for image prediction experiments.In the end,multiple groups of comparisons are performed.Experiments show that compared with other network models,the semantic segmentation accuracy of the network model is obviously improved;the improved method proposed in this paper can improve semantic segmentation compared to the basic RGB-D semantic segmentation method.The accuracy rate proves the feasibility of the improved algorithm in this paper.