Disparity optimization using improved denoising diffusion probabilistic model

Disparity estimation is crucial for light field applications. How to improve the accuracy of disparity estimation in occluded areas, textureless regions, and non-Lambertian surfaces remains a pressing issue. Disparity estimation methods generally consist of two stages: initial disparity estimation and disparity optimization. Disparity optimization methods can be divided into guided filtering-based methods and deep learning-based methods. The first type of method is based on the boundary information of the RGB image to locate the disparity boundaries, while other regions are constrained by piecewise smooth priors. The second type of method relies on deep learning to learn the mapping relationship between the light field and the real disparity from labeled datasets, and can obtain more accurate disparity images. The prior assumptions of the first kind of methods for disparity images are not accurate enough and are greatly affected by RGB images. The second type of method depends on labeled datasets and lacks generalization ability for data collected from different systems. In recent years, the ability of generative models to mine and represent prior information in data has become increasingly prominent. This paper proposes a disparity optimization method based on a conditional diffusion model. This method learns prior information about disparity images from existing public datasets and uses a conditional diffusion model to generate a disparity image with higher accuracy, conditioned on the initial disparity image estimated from the light field. Experimental results show that the prior information about scene disparity learned by this method is more comprehensive than the piecewise smooth property, is not affected by the RGB image, and has stronger generalization ability.