Learning High-fidelity Light Field Images From Hybrid Inputs
share
This paper explores the reconstruction of high-fidelity LF images (i.e., LF images with both high spatial and angular resolution) from hybrid inputs, including a high resolution RGB image and a low spatial but high angular resolution LF image. To tackle this challenging problem, we propose a novel end-to-end learning-based approach, which can comprehensively utilize the specific characteristics of the input from two complementary and parallel perspectives. Specifically, one module efficiently learns a deep multi-dimensional and cross-domain feature representation to regress an intermediate estimation; the other one propagates the information of the input, which is challenging to predict, to construct another intermediate estimation. We finally leverage the advantages of the two intermediate estimations via the learned confidence maps, leading to the final high-fidelity LF image. Extensive experiments demonstrate the significant superiority of our approach over the state-of-the-art ones. That is, our method not only improves the PSNR more than 2 dB, but also preserves the LF structure much better. To the best of our knowledge, this is the first end-to-end deep learning method for reconstructing a high-fidelity LF image with a hybrid input. We believe our framework could potentially decrease the cost of high-fidelity LF data acquisition and also be beneficial to LF data storage.
READ FULL TEXT