Research On Indoor Image Segmentation Method Based On RGBD

Image segmentation technology has always been in a very important position in image processing.In recent years,deep convolution neural network is gradually applied to image semantic segmentation tasks.In many large-scale semantic segmentation competitions,machine learning algorithm based on convolution neural network achieves higher detection accuracy than traditional algorithm.Conditional random field model is a kind of probability model based on undirected graph.It can mark sequence data and has strong probability reasoning ability.It can fully consider the position relationship between different objects in the image and reasonably infer the object category.For indoor scenes,when the existing methods are faced with problems such as too similar illumination,color and so on,it is easy to cause image edge blur and low segmentation accuracy.To this end,we have done research work from two aspects,as follows:In this paper,a semantic segmentation method of indoor image based on depth information is proposed.Firstly,using the depth similarity between pixels,the similar image depth information is seamlessly incorporated into the convolutional neural network(CNN)to generate the one-dimensional potential energy response of each category at the pixel level;secondly,the image depth information is fused into the contrast sensitive dual kernel potential energy conditional random field(CRF),which is combined with the previous one-dimensional potential energy for indoor image segmentation and refinement Finally,on the nyuv2 data set,five existing algorithms and four evaluation indexes are selected to carry out comparative experiments on this method.The results show that the proposed method is better than the existing methods in the accuracy of indoor scene classification and image segmentation.A conditional random field image segmentation model with 3D voxels is proposed.Firstly,the depth information is integrated into the one-dimensional potential function in conditional random field,and the traditional two-dimensional information is replaced by three-dimensional position information to improve the segmentation result;secondly,through the training of texture,color and position information of rgbd image,the two-dimensional potential function with Gaussian kernel pair potential function is obtained to complete the image segmentation with depth information;finally,pascal3d data set is used Compared with the existing methods,the results show that this method has a good performance and can effectively improve the accuracy of indoor image segmentation.