Reproduce “BitNet: Learning-Based Bit-Depth Expansion” on OpenResearch

About this paper

Junyoung Byun, Kyujin Shim, Changick Kim

BitNet, a convolutional neural network developed by KAIST, performs learning-based bit-depth expansion, achieving state-of-the-art objective quality (e.g., PSNR 33.426 for 3-bit to 16-bit expansion) and near real-time processing speeds (over 25 fps for 768x512 images) by simultaneously removing false contours and restoring visual details and accurate colors.

Key insights

• An end-to-end learning-based CNN approach can effectively address both false contour removal and simultaneous detail/color restoration, surpassing traditional heuristic and channel-wise methods in performance.
• Joint processing of RGB channels within the CNN leads to more natural and accurate color restoration, particularly for heavily quantized images, compared to processing channels independently.
• Explicitly providing bit-depth information to the network helps it adapt its processing and improves overall performance, especially in maintaining global consistency.
• Architectural elements like dilated convolutions for expanded receptive fields and the multi-scale feature integration module are crucial for robust bit-depth expansion and effective artifact handling.

Summary by alphaXiv.