Research On Depth Map Reconstruction Based On Sparse Representation

In recent years,with the fast development of 3D camera,such as Microsoft Kinect and time-of-flight(TOF)camera,depth maps have been widely used in computer vision,such as human-computer interaction,augmented reality and scene analysis.However,limited by the external condition and the camera itself,depth map captured by these cameras often suffer from low resolution,large among of noise interference and structure loss around boundaries.Thus,these depth maps can hardly be directly used in depth perception and 3D reconstruction.Therefore,it is of great importance for super resolution reconstruction from an input low resolution depth map.Many methods have been proposed to solve this problem.At the base of these,we propose two depth map reconstruction works in this paper that are based on sparse representation,the contents are as follows:First,a novel multiclass dictionary learning with geometrical directions based depth map super resolution model is proposed based on the texture features of depth map.The multiclass dictionary in the model is realated to the geometric directions of image patches,and it is sparser and more efficient to express an image than the traditional dictionary.This model divides the reconstruction work into the dictionary training stage and the reconstruction stage.In the dictionary training stage,depth map is divided into classified patches according to their geometrical directions and a sparse dictionary is trained within each class.In the reconstruction stage,a dictionary is selected according to the geometrical direction of an input image patch.Besides,texture information in color image is taken into account as prior information.Experimental results demonstrate that our method achieves better subjective and objective results than other state-of-the-art methods under different upsampling factors.Second,a novel depth map super resolution model based on multiclass dictionary learning and autoregressive model is proposed.This model combines the characteristics of sparse representation model and autoregressive model,and uses the advantage that the autoregress model can effectively deal with the edges and boundaries,making the reconstruction results with clearer boundary details.This model divides the reconstruction work into dictionary training stage and the joint constraint reconstruction stage.In the dictionary training stage,we put all the dictionaries that are trained from different geometrical directions together to form a new dictionary.This dictionary is sparser than the dictionary in one class and is more accurate to express the signal.In the joint constraint reconstruction stage,a new model is established based on sparse representation model and autoregressive model.Then we solve it efficiently.Experimental results show that this model achieves richer texture details and higher quality results than main stream methods.