ProxIQA: A Proxy Approach to Perceptual Optimization of Learned Image Compression.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Traditional neural network loss functions (lp norms) lack perceptual consistency for visual information.
• Existing methods for image analysis and compression do not fully align with human visual perception.

Purpose of the Study:

• To develop a novel "proximal" optimization approach for image analysis networks.
• To create a loss layer that mimics quantitative perceptual models for improved image quality assessment.
• To enhance deep image compression networks using perceptual optimization.

Main Methods:

• Constructed a proxy network (ProxIQA) to emulate perceptual models.
• Integrated ProxIQA as a loss layer within neural networks.
• Applied the optimization framework to train an end-to-end image compression network.

Main Results:

• Demonstrated the effectiveness of the ProxIQA framework in optimizing image analysis networks.
• Achieved up to 31% bitrate reduction in deep image compression compared to Mean Squared Error (MSE) optimization.
• Maintained specified perceptual quality levels (e.g., VMAF).

Conclusions:

• The ProxIQA approach offers a more perceptually aligned method for optimizing image analysis and compression.
• This framework enables significant bitrate savings in image compression without compromising visual quality.
• ProxIQA represents a promising advancement in perceptually driven deep learning for image processing.