Perceptual Compression Using Jacobian-Mapped Feature Vectors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for lossy data compression, particularly in image and video compression, struggle to accurately quantify the quality of compressed stimuli in a way that aligns with human perception, relying on subjective measures and imperfect metrics like RMS and PSNR, which fail to capture the complexities of human visual system interactions.
Innovation Solution
A novel system that uses hardware to project multivariate input data onto a perceptual manifold using Jacobian matrices, employing three methods to generate such matrices: approximating neural synaptic connectivity, replicating brain behavior, and creating task-specific approximations, enabling efficient electronic perceptual apparatuses that accelerate real-time computations without excessive power consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of substance
If traditional lossy compression formats (JPEG, MPEG) are used, then file size is reduced, but quality assessment accuracy deteriorates because they rely on imperfect metrics like RMS and PSNR that fail to capture human visual system interactions
Solution Approach 1:
The patent replaces traditional mechanical/mathematical compression metrics (RMS, PSNR) with a biologically-inspired perceptual model that simulates human visual system processing. This substitution uses neural network architectures trained on human quality assessments to predict perceived quality, replacing the inadequate mathematical metrics with a system that actually models human perception mechanisms.
Solution Approach 2:
The patent transforms the quality assessment from using standard mathematical parameters (RMS error, PSNR values) to using perceptual parameters derived from neural network processing. The system changes the measurement space from Euclidean mathematical distance to a perceptual manifold where distances reflect human judgment, fundamentally altering how quality is quantified.
2Measurement precision
If complex perceptual models are used to accurately assess stimulus quality, then measurement precision improves, but computational complexity and power consumption increase
Solution Approach 1:
The patent performs preliminary action by pre-training neural networks on large datasets of human quality assessments before deployment. The perceptual models are pre-computed and stored, allowing the system to quickly query pre-established perceptual manifolds rather than performing complex real-time neural network computations during actual quality assessment, significantly reducing operational power consumption.
Solution Approach 2:
The patent implements partial action by using simplified perceptual models for routine assessments and reserving full complex modeling only for critical quality decisions. The system adapts the level of computational detail based on the specific assessment context, using less computationally intensive methods when sufficient and only invoking full perceptual modeling when higher precision is genuinely required.
3Productivity
If real-time processing is required, then productivity increases, but measurement precision may deteriorate due to computational constraints
Solution Approach 1:
The patent segments the quality assessment process into distinct stages: preprocessing (resize, normalize), feature extraction (color space conversion, gradient calculation), and quality prediction (neural network inference). Each segment can be optimized independently for real-time performance, with critical path operations minimized and non-critical operations deferred or simplified, maintaining both speed and accuracy.
Data Source
AI summary
A method for compressing a digital representation of a stimulus includes encoding the digital representation as a feature vector within a feature space. The method also includes multiplying the feature vector with a Jacobian that maps the feature space to a non-Euclidean perceptual space according to a perceptual system that is capable of perceiving the stimulus. This multiplication generates a perceptual vector within the non-Euclidean perceptual space. The method also includes applying an update operator to the perceptual vector to move the perceptual vector in the perceptual space to an updated vector such that the updated vector has a lower entropy than the perceptual vector. The method also includes rounding the updated vector into a compressed vector that is smaller than the feature vector.