The Research On Super Resolution Algorithm Of Depth Map Based On Deep Learning

Joint image super-resolution refers to methods to enhance the resolution of an image with the guidance of a higher resolution(HR)image.It has become one of the hot topics in the field of deep learning,computer vision,and other related research fields as people's demand for image accuracy increases.Generally speaking,depth image super resolution(SR)algorithms can be roughly divided into three categories: local-based SR algorithms,globalbased SR algorithms and CNN-based(or deep learning-based)SR algorithms.The filter weights of local-based SR algorithms depend solely on the local structure of the guidance image.Therefore,erroneous structures may be transferred to the target image due to the lack of consistency check.However,global-based SR algorithms rely on hand-designed objective functions that may not reflect the complexities in natural images.Instead of using designed features,using features representation learned by algorithms becomes a recent trend.A typical example is deep neural networks,which have achieved great success in single image SR in recent year.This dissertation is based on the deep learning algorithm to research joint image SR.Based on larger receptive fields is helpful for image completion and other applications,and image SR is similar to image completion,larger receptive fields can improve the effect of image SR reconstruction.However,larger receptive fields increase the depths and parameters of the network,which may cause degradation and large memory consumption.In this paper,the convolutional neural residual pyramid(CNRP)is proposed,which is used to fully expand the receptive fields without significantly increasing network parameters and memory space.The structure of the deep learning network model is prone to generate gradient vanishing.Therefore,this paper proposes the joint residual pyramid(JRP)model,which introduces residual module and linear interpolation layer into the convolutional neural pyramid(CNP)to effectively solve the problem of gradient vanishing and improve the performance of the network model.Experimental results show that our JRP model outperforms existing state-of-the-art algorithms not only on data pairs of intensity/depth images,but also on data pairs like intensity/saliency images and color-scribbles/intensity images,without significantly sacrificing computation efficiency and memory space.Meanwhile,the JRP model can also be applied to the application of image completion,and good experimental results are obtained.