NeRF多视角去模糊的三维光场内容生成

o address the problem of poor reconstruction of Neural Radiation Field (NeRF) due to blurred and noisy input images, this paper presents a multi-view images deblurring for 3D light-field display based on neural radiance field. In this paper, the image blur process is modeled as sparse convolution kernel, the positions and weights of convolution kernels are learned by using Multi-Layer Perceptron, optimized rays is generated to simulate blurred light, and finally the optimized rays are sent to the NeRF network, and the poses of the camera and convolution kernels are jointly optimized in the training process. The experimental results show that compared with the classical algorithm, this paper has stronger robustness and higher view synthesis quality, improves the objective index and get better display performance.