Neural Network Feature Tensor Compression for Video Redundancy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for compressing feature tensors from neural networks do not effectively address spatial-temporal redundancy and inter-view redundancy in video images, leading to inefficiencies in video signal coding.
Innovation Solution
A method and device for compressing feature tensors using a neural network-based approach, involving quantization and entropy encoding, with adaptive quantization size derivation based on encoding information, distribution information, and iterative error correction to achieve target bit rates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of energy
If feature tensor compression is performed using conventional methods, then coding efficiency is maintained at baseline levels, but spatial-temporal redundancy and inter-view redundancy in video images are not effectively removed
Solution Approach 1:
The video data is segmented into multiple feature tensors representing different spatial, temporal, and inter-view dimensions. Each tensor is processed independently through quantization and entropy encoding, allowing targeted compression of specific redundancy types while preserving essential information.
Solution Approach 2:
The patent applies adaptive quantization by dynamically adjusting the quantization parameter based on the statistical characteristics of each feature tensor. This parameter adaptation enables optimal balance between compression ratio and information preservation for different video content regions and types.
2Productivity
If quantization is applied to feature tensors, then compression ratio is improved, but coding precision is reduced
Solution Approach 1:
The quantization process is made dynamic through adaptive quantization parameter selection. The system adjusts quantization strength based on local feature importance, signal statistics, and rate-distortion considerations, allowing finer precision to be maintained in critical regions while applying stronger compression in less important areas.
Solution Approach 2:
Different quantization strategies are applied to different regions of the feature tensors based on their local characteristics. Important features receive gentler quantization to preserve precision, while redundant regions undergo more aggressive compression, achieving overall high compression ratio without uniform precision loss.
3Productivity
If adaptive quantization size derivation is implemented, then encoding efficiency is enhanced, but device complexity increases
Solution Approach 1:
The system performs self-adaptation by automatically deriving quantization parameters from the statistical properties of the input feature tensors themselves. This self-service mechanism eliminates the need for external complex control systems, as the data inherently guides the quantization process through its own distribution characteristics.
Solution Approach 2:
The adaptive quantization process incorporates feedback loops where encoding results are evaluated and used to adjust subsequent quantization decisions. The system monitors encoding efficiency metrics and dynamically refines quantization parameters to optimize performance, creating a self-improving encoding pipeline.
Data Source
AI summary
A method and a device for processing an image on the basis of a neural network, according to an embodiment of the present invention, can acquire a feature tensor from an input image by using a first neural network including a plurality of neural network layers, acquire a quantized feature tensor by quantizing the acquired feature tensor on the basis of the quantization size, and generate a bitstream by performing entropy encoding on the quantized feature tensor.