Filter-Flow-Based Bit-Depth Expansion

Abstract

Bit-depth expansion is a technique for recovering a high-bit-depth (HBD) image from a low-bit-depth (LBD) image. Although recent high-performance displays can display each color in 10-bit or 12-bit depth to express more detailed color gradations, most existing image and video sources still have an 8-bit depth. Simple restoration from an LBD image into an HBD image leads to artifacts such as missing high-frequency information and false contours, so high-quality bit-depth expansion approaches have been explored. In recent years, deep convolutional neural network (CNN) techniques have provided excellent performance in many image processing tasks such as super-resolution, classification, deblurring, and denoising. In this paper, we propose a novel bit-depth expansion approach using predictive filter flow (PFF). PFF predicts a spatially variable filter using the CNN and reconstructs the target image by filtering it with the input image. The proposed method trains the PFF network that predicts the filter to transform an LBD image into an HBD image. Experimental results confirm that the proposed method provides better results than state-of-the-art bit-depth expansion algorithms.