DNN Decoder Update Parameters for Quality Adaptation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image and video coding schemes using Deep Neural Networks (DNNs) face challenges in balancing objective and perceptual quality, with objective metrics providing higher fidelity but poorer visual results, while subjective metrics offer more pleasing results but are harder to define accurately, and existing methods typically train networks for either objective or perceptual metrics, lacking adaptability for different bitrate levels and specific content types.
Innovation Solution
A method and apparatus for updating a Deep Neural Network-based decoder by obtaining and encoding update parameters from a first training configuration using an auto-encoder, allowing the decoder to be modified for both objective and perceptual quality, with the ability to adapt layers and training configurations for specific bitrates and content types, enabling efficient compression and decompression.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a DNN-based decoder is trained for objective quality metrics, then fidelity is improved, but visual quality deteriorates
Solution Approach 1:
The patent implements dynamic adaptability by enabling the DNN-based decoder to switch between different training configurations (objective metrics vs. perceptual metrics) based on bitrate levels and content types. This allows the system to dynamically select the most appropriate quality optimization strategy rather than being fixed to a single training regime.
Solution Approach 2:
The patent changes the training configuration parameters of the DNN-based decoder to optimize for different quality metrics. By adjusting parameters such as loss function weights and training data selection, the system can transition between optimizing for objective fidelity metrics versus perceptual visual quality metrics.
2Object-affected harmful factors
If a DNN-based decoder is trained for perceptual quality metrics, then visual quality is improved, but definition accuracy deteriorates
Solution Approach 1:
The system dynamically adjusts the training configuration based on operational requirements. When perceptual visual quality is prioritized, the decoder uses training configurations optimized for subjective metrics, while maintaining the capability to switch to objective metric optimization when definition accuracy becomes more important.
Solution Approach 2:
The patent modifies training parameters such as loss function composition and evaluation metrics to balance perceptual quality and definition accuracy. By changing these parameters, the system can emphasize either visual pleasingness or metric fidelity depending on the application context.
3Productivity
If the decoder structure is fixed for a specific training configuration, then training efficiency is improved, but adaptability deteriorates
Solution Approach 1:
The patent creates a universal DNN-based decoder architecture that can function under multiple training configurations. Rather than creating separate decoders for different quality metrics, the system designs a single decoder structure capable of being trained and adapted to various bitrate levels and content types, achieving multi-functionality.
Solution Approach 2:
The decoder structure incorporates dynamic elements that allow it to adapt its behavior and parameters based on different operational conditions. This includes the ability to modify layer configurations, activation functions, and training parameters dynamically to suit different bitrate levels and content characteristics.
4Loss of energy
If update parameters are not transmitted, then bandwidth consumption is reduced, but compression efficiency deteriorates
Solution Approach 1:
The patent extracts and transmits only the essential update parameters needed to adapt the decoder to different bitrate levels and content types. Rather than transmitting complete decoder models or all possible parameters, the system identifies and transmits only the critical update parameters, reducing bandwidth consumption while maintaining compression efficiency.
Solution Approach 2:
The system applies local quality optimization by transmitting update parameters selectively based on specific needs. Different parts of the decoder receive different update parameters tailored to the specific bitrate level and content type, rather than applying uniform updates across the entire system.
Data Source
AI summary
A method and an apparatus for decoding at least one part of at least one image is disclosed. The method comprises decoding at least one update parameter and modifying a deep neural network-based decoder based on said decoded update parameter. The method further comprises decoding at least one part of at least one image using at least said modified decoder.