Approximately Invertible Neural Network for Learned Image Compression.

Area of Science:

• Machine Learning
• Computer Vision
• Image Processing

Background:

• Learned image compression utilizes coupled transforms for encoding and decoding images.
• Invertible neural networks (INNs) show promise for constructing these transforms.
• Quantization noise challenges the invertibility of INNs in compression.

Purpose of the Study:

• To propose a novel framework, Approximately Invertible Neural Network (A-INN), for learned image compression.
• To address the challenges of quantization noise and high-frequency information loss in INN-based compression.
• To provide a theoretical foundation for INN-based lossy image compression methods.

Main Methods:

• Developed an Approximately Invertible Neural Network (A-INN) framework.
• Incorporated a progressive denoising module (PDM) to mitigate quantization noise.
• Designed a Cascaded Feature Recovery Module (CFRM) for feature channel compression.
• Introduced a Frequency-enhanced Decomposition and Synthesis Module (FDSM) to preserve high-frequency details.

Main Results:

• The A-INN framework effectively reduces quantization noise during decoding.
• CFRM improved feature recovery from low-dimensional representations.
• FDSM enhanced the preservation of high-frequency image components.
• Experimental results demonstrated competitive or superior compression efficiency.

Conclusions:

• The proposed A-INN framework offers a robust approach to learned image compression.
• The integrated modules (PDM, CFRM, FDSM) significantly improve reconstruction quality.
• A-INN provides a strong theoretical and practical foundation for future INN-based compression research.