Neural Network Image Compression via Learned Coefficient Patch Indices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional lossy image compression techniques struggle to balance visual quality and file size reduction, especially in motion video coding, where they often sacrifice important image details.
Innovation Solution
The use of neural networks for encoding and decoding images, where an encoder transforms images into coefficients, encodes them as integer patch indices into a set coding table, and a decoder reconstructs the images using jointly trained weights, with optional entropy coding to further compress the data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If traditional lossy image compression techniques (domain transform, quantization, entropy coding) are used, then file size is reduced, but visual quality and important image details are sacrificed
Solution Approach 1:
The patent replaces traditional mechanical compression algorithms (DCT, quantization matrices) with a neural network system that learns optimal compression through training. The encoder neural network and decoder neural network are trained jointly to minimize reconstruction error while achieving compression, substituting fixed mathematical transforms with adaptive learned transformations that preserve visual quality better than traditional methods
Solution Approach 2:
The patent changes the fundamental parameters of image compression by using learned transforms instead of fixed DCT bases, and by using learned quantization strategies instead of fixed quantization matrices. The neural networks learn optimal parameter representations that adapt to different image content, allowing better preservation of important visual information while achieving compression
2Manufacturing precision
If neural networks are trained to predict image information, then visual quality is improved, but computational complexity and training requirements increase
Solution Approach 1:
The patent performs preliminary training of the neural network models offline using large datasets before deployment. The encoder and decoder networks are pre-trained jointly to learn optimal compression representations. This preliminary action transfers the computational burden to the training phase, allowing the actual compression and decompression operations to be more efficient during runtime
Solution Approach 2:
The patent uses a codebook that stores pre-computed code vectors representing common image patterns. Instead of performing complex neural network operations for every compression task, the system copies relevant code vectors from the codebook that match the input image blocks, reducing computational complexity during actual compression while maintaining quality through the learned representations
Data Source
AI summary
Techniques for coding sets of images with neural networks include transforming a first image of a set of images into coefficients with an encoder neural network, encoding a group of the coefficients as an integer patch index into coding table of table entries each having vectors of coefficients, and storing a collection of patch indices as a first coded image. The encoder neural network may be configured with encoder weights determined by jointly with corresponding decoder weights of a decoder neural network on the set of images.