Neural Network Adaptive Quantization for Image Encoding Efficiency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image encoding and decoding technologies, such as H.264 AVC and HEVC, struggle to efficiently utilize artificial intelligence (AI) for effective encoding and decoding, particularly in handling spatial and temporal redundancies within images.
Innovation Solution
Employing neural networks for adaptive quantization and inverse quantization in image encoding and decoding processes, including obtaining feature data, modifying probability data based on quantization data, and applying entropy encoding and decoding to improve efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional quantization methods are used in neural network-based encoding, then computational complexity is reduced, but encoding efficiency and image quality deteriorate
Solution Approach 1:
The patent applies preliminary action by performing adaptive quantization before entropy encoding. The neural network predicts optimal quantization parameters and probability data in advance, allowing the quantization process to be optimized beforehand rather than using fixed or simple adaptive methods. This preliminary optimization improves subsequent encoding efficiency while managing computational complexity through staged processing.
Solution Approach 2:
The patent implements parameter changes by dynamically adjusting quantization parameters based on neural network predictions. The system changes quantization step sizes and probability distributions adaptively for different image regions and content types, rather than using uniform quantization. This allows optimization of encoding efficiency for various image characteristics while controlling overall computational complexity through selective application.
2Manufacturing precision
If uniform quantization is applied, then device complexity is reduced, but image quality and compression efficiency worsen
Solution Approach 1:
The patent applies local quality by using region-based adaptive quantization where different quantization parameters are applied to different spatial regions of the image. The neural network analyzes local image characteristics and assigns appropriate quantization strength to each region, preserving quality in important areas while allowing more compression in less critical areas. This resolves the contradiction by achieving high overall quality without uniformly complex processing.
Solution Approach 2:
The patent implements dynamics by making quantization parameters adaptive rather than static. The system dynamically adjusts quantization parameters based on local image content, texture complexity, and importance maps generated by the neural network. This dynamic adaptation improves image quality preservation where needed while maintaining computational feasibility through selective complexity application.
3Productivity
If neural network-based adaptive quantization is used, then compression efficiency improves, but computational complexity and processing time increase
Solution Approach 1:
The patent applies segmentation by dividing the image processing into distinct stages: neural network feature extraction, probability data generation, adaptive quantization, and entropy encoding. Each stage processes specific aspects independently, allowing parallel processing and optimization of individual components. This segmentation improves overall compression efficiency while reducing total processing time through modular computation.
Solution Approach 2:
The patent uses an intermediary approach by introducing probability data as an intermediate representation between the neural network output and the final encoded bitstream. The probability data serves as a mediator that captures essential image characteristics without requiring full-precision neural network outputs to be processed, thereby improving compression efficiency while reducing the computational burden of subsequent encoding stages.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An image decoding method according to an embodiment may include obtaining, for first feature data obtained by neural network-based encoding of a current image, second feature data from a bitstream; obtaining quantization data and probability data by applying the second feature data to a neural network; modifying the probability data, based on sample values of the quantization data; obtaining quantized first feature data by applying entropy decoding based on the modified probability data for bits included in the bitstream; obtaining dequantized first feature data by dequantizing the quantized first feature data according to the sample values of the quantization data; and reconstructing the current image by neural network-based decoding of the dequantized first feature data.