Image Compression Using Learned Embeddings and GAN Reconstruction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional image compression techniques like JPEG suffer from pixilation in decompressed images due to lossy compression and lack of adaptability in bitrate, as they apply the same compression methods uniformly across all image areas, failing to prioritize important features.
Innovation Solution
A machine learning model is trained to generate adaptive bitrate compression by separating into encoder and decoder components, which can be installed on different devices, allowing for learned basis functions specific to image regions, prioritizing important areas with higher detail and less important areas with lower compression, using a bottleneck layer for dimensionality reduction and a GAN for improved image quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If conventional lossy compression techniques like JPEG are used to reduce file size, then the file size of the compressed image is reduced, but pixilation occurs in the decompressed image
Solution Approach 1:
The patent replaces conventional mechanical signal processing (discrete cosine transforms, frequency domain operations) with a machine learning model-based approach. The neural network learns optimal compression representations through training, substituting traditional mathematical transforms with learned feature embeddings that better preserve perceptual image quality while achieving compression.
Solution Approach 2:
The patent changes the fundamental parameters of image representation from fixed frequency-domain coefficients (JPEG) to adaptive learned embeddings. The system varies embedding dimensions and compression factors dynamically based on image content, allowing optimal balance between file size and quality for different regions and images rather than applying uniform compression.
2Ease of operation
If the same compression function is applied uniformly across all image portions, then the compression process is simple and consistent, but important features like faces and eyes are not preserved with higher priority
Solution Approach 1:
The patent implements local quality by allowing different portions of the image to have different compression levels. The machine learning model can allocate more embedding dimensions to important regions (faces, eyes) and fewer dimensions to less important regions, creating non-uniform compression that preserves critical features while maintaining overall efficiency.
Solution Approach 2:
The compression system transitions from static uniform compression to dynamic adaptive compression. The model dynamically adjusts compression factors based on local image content importance, learned during training and applied during inference, allowing the system to adapt to different regions and images rather than applying fixed compression rules.
Data Source
AI summary
A processing device receives a representation of an image, wherein the image has a first size and the representation has a second size that is smaller than the first size, the representation having been generated from the image by a first portion of a first trained machine learning model. The processing device processes the representation of the image using a second portion of the trained machine learning model to generate a reconstruction of the image and then outputs the reconstruction of the image.