IEST: Interpolation-Enhanced Shearlet Transform for Light Field Reconstruction Using Adaptive Separable Convolution

Abstract

The performance of a light field reconstruction algorithm is typically affected by the disparity range of the input Sparsely-Sampled Light Field (SSLF). This paper finds that (i) one of the state-of-the-art video frame interpolation methods, i.e. adaptive Separable Convolution (SepConv), is especially effective for the light field reconstruction on a SSLF with a small disparity range (<10 pixels); (ii) one of the state-of-the-art light field reconstruction methods, i.e. Shearlet Transformation (ST), is especially effective in reconstructing a light field from a SSLF with a moderate disparity range (10-20 pixels) or a large disparity range (> 20 pixels). Therefore, to make full use of both methods to solve the challenging light field reconstruction problem on SSLFs with moderate and large disparity ranges, a novel method, referred to as Interpolation-Enhanced Shearlet Transform (IEST), is proposed by incorporating these two approaches in a coarse-to-fine manner. Specifically, ST is employed to give a coarse estimation for the target light field, which is then refined by SepConv to improve the reconstruction quality of parallax views involving small disparity ranges. Experimental results show that IEST outperforms the other state-of-the-art light field reconstruction methods on nine challenging horizontalparallax evaluation SSLF datasets of different real-world scenes with moderate and large disparity ranges.