Learned Inverse Transform Kernels for Video Quantization Compensation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video coding technologies face challenges in effectively compensating for quantization errors during the transformation and reconstruction of video data, leading to suboptimal image quality and compression efficiency.
Innovation Solution
The use of a modified inverse transform kernel that is trained to compensate for quantization effects, allowing for improved reconstruction of residual blocks by performing dequantization and inverse transform using a learned kernel to mitigate the impact of quantization on spectrum domain coefficients.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If a standard inverse transform kernel is used for reconstruction, then the decoding process is simple and fast, but quantization errors cause suboptimal image quality and compression efficiency
Solution Approach 1:
The patent modifies the inverse transform kernel parameters by training a neural network to predict optimized kernel values based on quantization parameter Qp and block size. This transforms the fixed standard kernel into an adaptive kernel that compensates for quantization effects, improving image quality while maintaining computational efficiency through learned parameter adjustments rather than complex recompute operations
Solution Approach 2:
The patent replaces the traditional mechanical/mathematical approach of using fixed inverse transform kernels with a data-driven neural network model. The neural network learns the relationship between quantized coefficients and optimal reconstruction parameters, substituting complex mathematical inversion operations with efficient neural network inference that adapts to quantization effects
2Productivity
If quantization is applied to reduce data size, then compression efficiency improves, but reconstruction accuracy deteriorates
Solution Approach 1:
The patent implements a feedback mechanism where the neural network takes quantized spectrum domain coefficients and quantization parameters as input, processes them through learned transformations, and outputs modified inverse transform kernels. This feedback loop continuously adapts the reconstruction process to account for quantization losses, improving accuracy without sacrificing compression efficiency
Solution Approach 2:
The patent performs preliminary training of the neural network model using training data that simulates various quantization scenarios. During training, the network learns to predict optimal reconstruction parameters before actual decoding occurs. This preliminary learning enables the system to compensate for quantization effects during real-time decoding without requiring additional computational resources
Data Source
AI summary
A bitstream includes coded information of a block, the coded information includes quantized spectrum domain coefficients corresponding to a residual block of the block, the residual block is transformed from a spatial domain to spectrum domain coefficients in a spectrum domain according to a transform kernel, the spectrum domain coefficients are quantized into the quantized spectrum domain coefficients according to a quantization parameter value. A modified inverse transform kernel is determined to be used for the block, the modified inverse transform kernel is different from an inverse of the transform kernel to compensate an influence of quantization on the spectrum domain coefficients. A dequantization is performed on the quantized spectrum domain coefficients to obtain dequantized spectrum domain coefficients. An inverse transform is performed on the dequantized spectrum domain coefficients according to the modified inverse transform kernel to obtain a reconstructed residual block for the reconstruction of the block.