Spatio-Temporal Film Grain Removal in Video Using Attention-Guided Neural Networks

Authors:

Abstract
Film grain is a feature of analog film content and is present in digital content as well to enhance perceived quality. However, compressing content with film grain is challenging due to the random and complex nature of grain patterns. Thus, there is a need to remove film grain effectively from content while maintaining temporal consistency. Existing methods for film grain removal, especially in images, often struggle to maintain consistency when applied to videos. Furthermore, most video denoising techniques focus on removing Additive White Gaussian Noise (AWGN), which has fundamentally different characteristics from film grain. In this paper, we present a novel deep learning-based approach for temporally consistent film grain removal in video sequences. We introduce a noise strength estimation network that adapts to the spatio-temporally varying levels of film grain, determining the film grain parameters to guide the removal process. By integrating self-attention mechanisms in both the noise strength estimation and film grain removal networks, our method enhances the accuracy of film grain removal while preserving fine details and temporal consistency. We compare our approach against state-of-the-art video denoising methods and demonstrate its superior performance both visually and quantitatively, with improvements of approximately 4 dB in PSNR while maintaining temporal stability across varying grain levels.