Neural Network Video Compression Decoding Speed
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current neural network-based video compression methods face challenges due to the inherent difficulty of efficiently removing inter-picture redundancy, leading to slow decoding processes due to the serial nature of entropy decoding.
Innovation Solution
The proposed solution decouples the arithmetic decoding process from the neural network-based modules, allowing it to be performed independently. This is achieved by using a second subnetwork to estimate probability parameters, which are then used for arithmetic decoding, thereby speeding up the decoding process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If arithmetic decoding is performed serially with neural network modules, then compression efficiency is maintained, but decoding speed becomes slow
Solution Approach 1:
The patent divides the decoding process into two independent segments: arithmetic decoding and neural network-based probability parameter estimation. By segmenting the serial processing into independent modules, arithmetic decoding can execute in parallel without waiting for neural network computations, thereby increasing decoding speed while maintaining compression efficiency through the coupled probability parameters.
2Productivity
If a single subnetwork estimates both mean and variance for arithmetic decoding, then device complexity is reduced, but decoding speed is limited by serial processing
Solution Approach 1:
The patent employs a second subnetwork to perform preliminary estimation of probability parameters (mean and variance) before arithmetic decoding. This preliminary action allows the arithmetic decoding to proceed independently and in parallel, reducing the processing time bottleneck and increasing decoding throughput by eliminating the serial dependency.
3Reliability
If probability parameters are estimated using neural networks, then compression efficiency improves, but decoding time increases due to computational complexity
Solution Approach 1:
The patent introduces probability parameters as intermediary elements that mediate between the neural network-based compression model and the arithmetic decoding process. These parameters serve as a bridge, allowing the neural network to provide compression efficiency through probability modeling while enabling fast arithmetic decoding to operate independently, thus reducing overall decoding time.
Data Source
AI summary
A mechanism for processing video data is in a neural network disclosed. The mechanism includes obtaining quantized residual latent samples. The quantized residual latent samples are processed to obtain processed quantized residual latent samples. A reconstructed latent sample can then be acquired based on the processed quantized residual latent sample.