Research On Occlusion Boundary Detection Approaches Of Depth Image Based On Neural Network

The occlusion problem has always been a classic topic in the field of computer vision.In the visual task of target tracking,3D reconstruction,object recognition,etc.the processing result of occlusion problem occupies a pivotal position.Therefore,Among many computer vision processing problems,how to accurately and efficiently detect occlusion has became a technical point which to be solved.The depth image has more powerful information in the three-dimensional real scene.At the same time,the depth image as the research object at home and abroad is still less.Based on these factors,this paper takes the depth image as the research object and combines with the neural network model in the current machine learning theory to solve image occlusion boundary detection task.Firstly,with the depth image as the starting point,the related concepts about occlusion,occlusion boundary and occlusion related features are introduced.At the same time,the relevant knowledge content of all neural networks used in this paper is described and summarized.Secondly,an image occlusion boundary detection method based on BP neural network is proposed.In this method,based on the pixel information in the depth image,a novel occlusion related feature is proposed,and the importance is evaluated by random forest algorithm compared with existing features.Meanwhile,combined with the distribution of pixel feature value vector in the depth image design a nonlinear normalization method to mapped the feature value vector.and according to the BP neural network,to train and detect the relevant samples,and then divided all the pixels in the image into occlusion boundary points or non-occlusion boundary points.At last complete the occlusion detection task.Thirdly,an occlusion detection method based on convolutional neural network model is proposed.Based on VGG16 network,the fully connected layers are replaced with convolutional layer,then modify the network structure.The next use these images which are tailored to train the model.Then use the cross entropy loss function to count the final error result of model.At the end,use the model to test whole image to finsh the coolusion task.Finally,the two detection methods designed in this paper are experimentally verified,and compared with the existing occlusion detection approaches of depth image,proving the feasibility of the proposed approaches.