Deep Neural Network Image Encoding Using DNN Update Information
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image processing technologies face challenges in efficiently encoding and decoding high-resolution/high-definition images, particularly in reducing bitrate while maintaining image quality, especially when using deep neural networks (DNNs) for image up-scaling and down-scaling processes.
Innovation Solution
The proposed solution involves a method and apparatus for image encoding and decoding using deep neural networks (DNNs), where DNN update information is managed to determine updated setting parameters for AI up-scaling, ensuring efficient bitrate reduction and quality maintenance by jointly training first and second DNNs for down-scaling and up-scaling processes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If traditional codecs (e.g., HEVC) are used for encoding high-resolution images, then device complexity is reduced and ease of operation is maintained, but bitrate reduction efficiency and image quality are insufficient
Solution Approach 1:
The patent applies preliminary action by pre-training deep neural networks for both down-scaling and up-scaling operations before actual image processing. The DNN models are trained offline to learn optimal transformation patterns, so that during runtime, the pre-trained models can efficiently process images with minimal real-time computation overhead. This resolves the contradiction by preparing the complex DNN structures in advance, making them ready for deployment without adding significant runtime complexity.
Solution Approach 2:
The patent utilizes parameter changes by dynamically adjusting DNN model parameters (such as scaling factors, resolution targets, and model architecture configurations) based on specific processing requirements. Different DNN models with varying parameters are selected or trained for different up-scaling ratios and image types, allowing the system to optimize between image quality and computational complexity by choosing appropriate parameter configurations for each scenario.
2Loss of information
If deep neural networks are used for AI up-scaling, then image quality is improved and bitrate is reduced, but processing time and computational resources increase
Solution Approach 1:
The patent applies preliminary action by pre-training deep neural networks for both down-scaling and up-scaling operations before actual image processing. The DNN models are trained offline to learn optimal transformation patterns, so that during runtime, the pre-trained models can efficiently process images with minimal real-time computation overhead. This resolves the contradiction by preparing the complex DNN structures in advance, making them ready for deployment without adding significant runtime complexity.
Solution Approach 2:
The patent applies dynamics by implementing adaptive DNN model selection and configuration based on processing requirements. The system can dynamically choose between different pre-trained DNN models with varying complexity levels, adjust processing resolution stages, and optimize computational resource allocation based on the specific up-scaling task at hand, thereby balancing processing time and image quality.
3Adaptability or versatility
If DNN setting information is updated frequently, then image quality and adaptability are improved, but data transmission overhead and processing complexity increase
Solution Approach 1:
The patent applies the taking out principle by extracting and separately managing DNN model parameters and configuration data from the main image processing workflow. The DNN setting information (including weights, biases, and architectural parameters) is extracted as distinct configurable elements that can be updated independently. This allows the system to update only necessary DNN parameters without retransmitting entire model structures, reducing data overhead while maintaining adaptability.
Solution Approach 2:
The patent applies preliminary action by pre-defining multiple DNN configuration profiles and models during system setup. These pre-configured DNN settings include various model architectures, hyperparameters, and processing configurations that can be selected and activated based on specific processing needs. This preliminary preparation reduces the need for frequent real-time updates of DNN setting information, as the system can switch between pre-configured profiles rather than generating new configurations dynamically.
Data Source
AI summary
Methods and apparatuses for image encoding and image decoding are provided. The image decoding method includes: obtaining deep neural network (DNN) update permission information indicating whether one or more pieces of DNN setting information are updated; based on the DNN update permission information indicating that the one or more pieces of the DNN setting information are updated, obtaining DNN update information necessary for determining one or more pieces of the DNN setting information that are updated; determining the one or more pieces of the updated DNN setting information according to the DNN update information; and obtaining a third image by performing artificial intelligence (AI) up-scaling on a second image according to the one or more pieces of the updated DNN setting information.