Deep-Based Film Grain Removal and Synthesis.

Area of Science:

• Computer Vision
• Digital Image Processing
• Video Compression

Background:

• Film grain is a natural artifact in analog film and is often added to digital content for aesthetic reasons.
• The random nature of film grain poses challenges for video compression, making it difficult to preserve and expensive to store.
• Current video coding methods struggle with efficient film grain management, impacting both quality and file size.

Purpose of the Study:

• To develop deep learning-based methods for effective film grain removal and realistic synthesis.
• To enhance video coding efficiency by intelligently handling film grain.
• To provide controllable film grain restoration and generation capabilities.

Main Methods:

• An encoder-decoder architecture was utilized for the film grain removal model.
• A conditional generative adversarial network (cGAN) was employed for the film grain synthesis model.
• Both models were trained on extensive datasets of clean and grainy image pairs.

Main Results:

• The film grain removal model demonstrated effectiveness in filtering grain at various intensities in both non-blind and blind configurations.
• The film grain synthesis model successfully reproduced realistic film grain with adjustable intensity levels.
• Evaluations confirmed the proposed models' capabilities in both quantitative and qualitative aspects.

Conclusions:

• Deep learning offers a powerful approach to managing film grain in video coding.
• The proposed models provide efficient solutions for both removing unwanted grain and synthesizing realistic grain.
• These techniques have the potential to improve video quality and compression efficiency for film-like content.