Deep Reconstruction of Least Significant Bits for Bit-Depth Expansion

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Bit-depth expansion (BDE) is crucial for displaying low bit-depth images on high bit-depth monitors.
• Existing BDE algorithms often produce perceivable artifacts due to traditional reconstruction methods.

Purpose of the Study:

• To develop an advanced BDE method using deep residual networks.
• To minimize artifacts and enhance visual quality in expanded bit-depth images.

Main Methods:

• A deep residual network architecture is proposed for BDE.
• Two distinct reconstruction channels are utilized for flat and non-flat image areas.
• A local adaptive adjustment preprocessing step is incorporated for flat areas.

Main Results:

• The proposed method demonstrates improved subjective visual quality, particularly in flat image regions.
• Experimental results show favorable visual quality and decent quantitative performance across various image sets.
• The integration of traditional debanding strategies with network-based reconstruction enhances flat area quality.

Conclusions:

• The deep residual network-based BDE method effectively addresses artifacts in low bit-depth images.
• The dual-channel approach and adaptive preprocessing contribute to superior image reconstruction.
• The proposed technique offers a promising solution for high-quality BDE in digital displays.