A Little Bit More: Bitplane-Wise Bit-Depth Recovery.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Imaging sensors capture images with 10-12 bits of dynamic range.
• Onboard processing quantizes images to 8 bits for standard encoding.
• Recovering lost bit depth is crucial for applications like high-bit-depth displays and photo editing.

Purpose of the Study:

• To develop an effective deep learning strategy for bit-depth reconstruction.
• To improve upon existing single-shot deep learning methods for image bit-depth enhancement.

Main Methods:

• Proposed a novel bitplane-wise learning framework for residual image reconstruction.
• Employed multiple levels of supervision during the training process.
• Utilized a simple network architecture for the bit-depth recovery task.

Main Results:

• Achieved state-of-the-art results in bit-depth reconstruction.
• Demonstrated significant performance improvements, ranging from 0.5dB to 2.3dB PSNR.
• Results showed superiority over prior methods across various quantization levels.

Conclusions:

• The bitplane-wise learning framework offers an advantageous approach for bit-depth reconstruction.
• The proposed method enables effective recovery of lost bit depth with enhanced image quality.
• This technique provides a robust solution for applications requiring high bit-depth imaging.