Research On Depth Reconstruction Algorithm Of Multi-view Images

With the development of naked-eye three-dimensional display technology,the expectation of three-dimensional display content is also very high.Depth reconstruction is a key step in 3D video generation.The quality of depth map directly affects the quality of virtual viewpoint generation.In addition,depth maps are also used in the fields of three-dimensional reconstruction,measurement system and automatic driving.However,the existing depth reconstruction algorithms are difficult to achieve real-time processing speed,and the quality of the depth reconstruction in the occlusion area is unsatisfactory.Therefore,this thesis studies the efficient and high accuracy depth reconstruction technology and proposes a new solution.The contribution of this thesis are as follows:(1)Through the research of depth reconstruction technology,a depth reconstruction algorithm of multi-view images is proposed.This depth reconstruction algorithm of three-view images is implemented on GPU platform.Experiments show that the depth reconstruction algorithm has fast calculation speed and high quality in the occluded area.(2)A three-view sparse viewpoint acquisition and depth reconstruction system is designed.Three industrial cameras are used to collect sparse viewpoints.Camera calibration,image correction and depth reconstruction are realized in GPU platform.The processed sparse viewpoint images and reconstructed depth map are transmitted to the virtual viewpoint rendering module to calculate the dense viewpoint.The synthesized viewpoint can achieve better display effect on the three-dimensional display device and has good robustness in different environments.The processing speed is higher than 30 fps.The system lays a technical foundation for acquiring true three-dimensional video.(3)Due to the relatively simple end-to-end structures of CNNs,the performance for poor and repetitive texture is barely satisfactory.Two main structures are proposed to optimize the depth map.A structure of multi-scale convolution kernels is added to the network.This structure can be processed and aggregated on different scales,so that features can be extracted from different scales at the higher level at the same time.Two sizes of rectangular convolution kernels are introducted to the network structure in order to strengthen the relationship between the poor texture area and the surrounding area.Experimental results demonstrate that our structures of the CNN reduce the error rate from 19.24%to 14.08%.The boundary of reconstructed depth map is clear and the depth of poor texture area is correct.