Research On Light Field Depth Estimation Based On Deep Learning

Depth estimation is an important task in computer vision,which aims to infer the depth information of objects in the scene from images by computers.The multi viewpoint characteristics of a light field image make it naturally carry more accurate depth information.Light field depth estimation plays an important role in virtual reality and augmented reality,autonomous driving,three-dimensional reconstruction,and other fields.Therefore,depth estimation based on light field is also receiving increasing attention.Current light field depth estimation still faces the following challenges:(1)It is difficult to collect light field data sets.Due to the small parallax of the light field data sets,it is difficult to obtain relatively effective depth truth values when using a light field camera for photography,so synthetic data sets are the main focus of research.(2)The number of optical field sub aperture images is too large,with a lot of redundant information.If training is conducted in the network at the same time,it will not only greatly increase the amount of network parameters,but also affect the learning effect of the network due to excessive useless and redundant repetitive information,affecting the final depth estimation result.In response to the above issues,the work and innovation done in this article are as follows:1.The unsupervised method avoids the difficulty of obtaining Ground Truth during supervised learning.Warp operations on optical field images can reconstruct the central sub aperture image,and then perform a loss function calculation with the input central sub aperture image.This allows for the training of unsupervised networks without Ground Truth.Experimental results prove the effectiveness of the unsupervised network proposed in this paper,and even outperform some supervised algorithms on some indicators.2.In supervised learning,it is possible to train only through sub aperture images in four directions: 0 °,45 °,90 °,and 135 °,obtain a finer disparity distribution through sub pixel translation,and then use the Attention mechanism to increase the learning weight of sub aperture images in unobstructed and noiseless areas to achieve better learning results.Experimental results on commonly used datasets show that our algorithm has excellent performance compared to current mainstream light field depth learning methods,and is at the forefront of the mainstream in all indicators.