AI Encoding Decoding Deep Neural Networks Bitrate Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image encoding and decoding technologies struggle to efficiently handle high-resolution and high-quality images, leading to increased bitrate and reduced efficiency in encoding and decoding processes.
Innovation Solution
The use of artificial intelligence (AI) encoding and decoding methods, specifically employing deep neural networks (DNNs) for AI down-scaling and up-scaling of images, to achieve a low bitrate by transforming and reconstructing images based on AI data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If conventional encoding and decoding methods are used for high-resolution images, then image quality can be maintained, but bitrate increases and processing efficiency decreases
Solution Approach 1:
The patent replaces conventional mechanical encoding/decoding systems with AI-based deep neural networks. The encoder uses a first DNN to perform intelligent downscaling and the decoder uses a second DNN to perform intelligent upscaling, substituting traditional compression algorithms with learning-based models that achieve better quality-bitrate tradeoffs
Solution Approach 2:
The patent changes the fundamental parameters of the encoding/decoding process by using trained neural network models with learned weights and biases. The DNNs are trained to optimize the transformation between high-resolution and low-resolution images, changing the mathematical parameters from fixed algorithmic operations to adaptive learned transformations
2Manufacturing precision
If conventional encoding and decoding methods are used for high-resolution images, then image quality can be maintained, but processing efficiency decreases
Solution Approach 1:
The patent replaces conventional mechanical encoding/decoding systems with AI-based deep neural networks. The encoder uses a first DNN to perform intelligent downscaling and the decoder uses a second DNN to perform intelligent upscaling, substituting traditional compression algorithms with learning-based models that achieve better quality-bitrate tradeoffs
Solution Approach 2:
The patent applies preliminary action by training the deep neural networks in advance on large datasets of high-resolution images. The encoder DNN is pre-trained to learn optimal downscaling transformations and the decoder DNN is pre-trained to learn corresponding upscaling transformations, so that during actual encoding/decoding operations, the pre-trained models can process images efficiently without requiring real-time complex computations
3Quantity of substance
If AI down-scaling and up-scaling is performed using jointly trained DNNs, then bitrate is reduced and processing efficiency is improved, but system complexity increases
Solution Approach 1:
The patent merges the training process of the encoder and decoder DNNs by performing joint training on matched pairs of high-resolution and downscaled images. The loss function combines both encoding and decoding quality metrics, merging the optimization objectives of both networks into a unified training framework that learns coordinated transformations
Solution Approach 2:
The patent creates universal DNN models that can handle multiple functions: the encoder DNN performs both compression and quality optimization, while the decoder DNN performs both reconstruction and quality enhancement. These multi-functional models replace separate specialized components, reducing overall system complexity despite the advanced AI capabilities
Data Source
Figure 1a
Figure 1b
Figure 1c
AI summary
An artificial intelligence (AI) decoding apparatus includes a memory storing one or more instructions, and a processor configured to execute the stored one or more instructions, to obtain image data corresponding to a first image that is encoded, obtain a second image corresponding to the first image by decoding the obtained image data, determine whether to perform AI up-scaling of the obtained second image, based on the AI up-scaling of the obtained second image being determined to be performed, obtain a third image by performing the AI up-scaling of the obtained second image through an up-scaling deep neural network (DNN), and output the obtained third image, and based on the AI up-scaling of the obtained second image being determined to be not performed, output the obtained second image.