Research On Axial Super-resolution Algorithm Of Light Field Based On Adaptive Separable Convolutional Network

With the development of science and technology,the traditional imaging mode can no longer meet the requirements of human development.Because the light field imaging model can record the position and direction of light,so the industrialization of the light field imaging model will also become an inevitable requirement for future development.However,there are still a lot of problems to be solved in the practical application of light field imaging.How to better use the light field imaging data has also become a hot spot of current research,and it still needs a lot of outstanding scientific researchers to make breakthroughs.This paper focuses on the axial super-resolution of light field refocused images.The user does not need to refocus the light field image pixel by pixel.The algorithm doubles the refocused image or more based on the existing light field refocused image.The algorithm makes the light field refocusing between images more smooth,and also saves more computing power for the user.Therefore,the light field axial super-resolution algorithm will have certain application value in the future.(1)This paper proposes a generative adversarial network model for axial super-resolution of the light field.The article uses two generative modules to learn the image sequence features.The third network module integrates the network outputs of the previous two modules.The final network outputs super-resolution refocused images.The feature extraction part of generative adversarial network uses DenseNet's network ideas.Previous network extraction feature is passed directly to the back-end network.At the same time,we construct a light field refocusing dataset to train the model.The experimental results show that our proposed generative adversarial network model has excellent results on the axial super-resolution of the light field.(2)We apply an adaptive separable convolutional neural network model to the axial super-resolution of the light field.The main difference between algorithms and generative adversarial networks is that the model generates an adaptive convolution kernel based on a set refocused images of the light field.Therefore,different image convolution kernels will change adaptively.Then use the generated convolution kernel to convolve with the input image of the network and integrate to obtain the target image.The experimental results show that the proposed algorithm has excellent effect on optical field axial super-resolution,and the output image has obvious advantages in visual and numerical evaluation indexes.(3)We use PyQt to build the UI interface for the two super-resolution algorithms.We also use software development ideas to test the four main functions of the software.Tested and used,the software all functions can be display smoothly.The software greatly reduces the difficulty of using the algorithm.