Unrolling Graph Total Variation for Light Field Image Denoising
Rino Yoshida, Kazuya Kodama, Huy Vu, Gene Cheung, Takayuki Hamamoto · 2022 IEEE International Conference on Image Processing (ICIP) · 2022
A light field (LF) image is composed of multiple sub-aperture images (SAIs) from slightly offset viewpoints. To denoise a noise-corrupted LF image, leveraging recent development in deep algorithm unfolding, we pursue a hybrid graph-model-based / data-driven approach. Specifically, we first connect each pixel in a target patch of an SAI to neighboring pixels within the patch, and to pixels in co-located "similar" patches in adjacent SAIs. Given graph connectivity, we formulate a maximum a posteriori (MAP) problem using graph total variation (GTV) as signal prior. We then unroll the iterations of a corresponding optimization algorithm into a sequence of neural layers. In each unrolled layer, we learn relevant features per pixel from data using a convolutional neural net (CNN) in a supervised manner, so that edge weights can be computed as functions of feature distances. Each neural layer can be interpreted as a graph low-pass filter for a 4D LF image patch. Experiments show that our proposal outperformed two model-based and two deep-learning-based implementations in numerical and visual comparisons.