Ultra High Fidelity Deep Image Decompression With l_∞-Constrained Compression.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Existing image compression methods struggle to maintain perceptual transparency at high compression ratios.
• Current deep learning-based restoration methods can sacrifice fine details or statistical outliers.

Purpose of the Study:

• To propose a novel asymmetric image compression system with enhanced rate-distortion performance.
• To leverage deep learning for superior compression artifact removal, focusing on preserving image fidelity.

Main Methods:

• An asymmetric system employing light L-infinity-constrained predictive encoding and a CNN-based soft decoder.
• A specialized restoration Convolutional Neural Network (CNN) designed for per-pixel error bound enforcement.

Main Results:

• Superior rate-distortion performance over BPG, WebP, FLIF, and other CNN codecs.
• Achieved high-quality reconstruction near or above the perceptual transparency threshold using both L2 and L-infinity error metrics.
• Effectively preserved small, distinctive image structures and statistical outliers, unlike mainstream CNN restoration techniques.

Conclusions:

• The proposed system offers significant coding gains through deep learning-based artifact removal.
• The per-pixel error bound enforcement in the CNN decoder is key to preserving image integrity.
• This approach advances the state-of-the-art in image compression, particularly for perceptually lossless results.