Neural Image Compression via Non-linear Quantization and Substitution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional hybrid video codec frameworks optimize machine-learning models separately, leading to increased rate-distortion loss and computing costs, resulting in lower performance for image and video compression.
Innovation Solution
An end-to-end neural image compression framework using non-linear quantization with substitution, where a neural network generates a substitute image, which is then compressed and quantized using a non-linear quantizer for improved compression performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional hybrid video codec frameworks optimize machine-learning models separately, then each model can be trained independently, but rate-distortion loss increases and computing cost increases
Solution Approach 1:
The patent merges multiple separately trained machine-learning models into a unified end-to-end neural network framework. The encoder and decoder models are combined and trained jointly using a single loss function that optimizes rate-distortion performance across the entire compression system, eliminating the need for separate model training and reducing overall computing costs.
Solution Approach 2:
The unified neural network framework performs multiple functions simultaneously: it conducts end-to-end optimization, generates substitute images for non-linear quantization, and performs compression all within a single system. This multi-functional approach replaces the need for multiple specialized models, reducing computing overhead while maintaining compression performance.
2Reliability
If traditional hybrid video codec frameworks optimize machine-learning models separately, then model training is simplified, but rate-distortion loss increases
Solution Approach 1:
The patent combines multiple models into a unified end-to-end framework that optimizes rate-distortion performance jointly. By merging the encoder and decoder into a single trainable system with a unified loss function, the framework achieves better rate-distortion performance while managing complexity through integrated design rather than separate model optimization.
Solution Approach 2:
The patent introduces non-linear quantization with substitution as a parameter change in the compression pipeline. By transforming the quantization process and introducing substitute images, the system achieves improved rate-distortion performance without proportionally increasing framework complexity, as the changes are integrated into the end-to-end training process.
3Reliability
If linear quantization is used in neural image compression, then the quantization process is simple, but compression performance is limited
Solution Approach 1:
The patent changes the quantization parameter from linear to non-linear, specifically using logarithmic quantization. This parameter change improves compression performance by better matching the perceptual characteristics of images and the probability distribution of pixel values, while the complexity is managed through integration with the end-to-end training framework.
Solution Approach 2:
The patent introduces substitute images as an intermediary between the original image and the quantization process. These substitute images are generated by the neural network and serve as intermediaries that are better suited for non-linear quantization, improving compression performance while keeping the quantization process manageable through the mediating role of the substitute images.
Data Source
AI summary
Method, apparatus, and non-transitory storage medium for end-to-end neural image compression using non-linear quantization with substitution, including receiving one or more input images, generating a substitute image associated with the input image based on the input image using a neural network based substitute feature generator, compressing the substitute image, quantizing the compressed substitute image to obtain a quantized representation of the input image with higher compression performance by using a non-linear quantizer, and entropy encoding the quantized substitute image using a neural network based encoder to generate a compressed representation of the quantized representation.