Hybrid compressed light field optimization algorithm based on stochastic gradient descent

Abstract

As we all know, the traditional compressed light field 3D display technology has the problems of limited 3D depth of field and low display brightness. In this paper, a hybrid compressed light field device based on polarization multiplexing is proposed, which combines multiplicative and superimposed compressed light field 3D display to improve the light intensity perceived by human eyes and enlarge the depth of field. In addition, when using high-brightness mini-leds, noise can appear at the edges of the reconstructed image. This is because non-negative tensor matrix (NTF) algorithm adopts hierarchical iteration, which is easy to fall into the local optimal solution, resulting in poor optimization effect of the edge part and noise. Then we introduce the stochastic gradient descent (SGD) algorithm which can better improve the problem of edge noise because all spatial light modulator pixel values are updated at the same time in the iteration process. In terms of perception indicators, NTF uses the mean square error coefficient, which cannot account for many nuances of human perception, resulting in iterative results that sometimes do not conform to the subjective perception of human eyes. In contrast, the loss function of SGD can be self-defined. This paper introduces the Learned Perceptual Image Patch Similarity, which is more in line with human perception. Through simulation and experiments, we verify the advantages of the proposed device and the effectiveness of the corresponding optimization algorithm.