Deep Learning Image Coding Method Optimizing Compression Efficiency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current video signal compression methods fail to optimize coding efficiency effectively, particularly in addressing temporal and spatial redundancies and inter-view redundancies.
Innovation Solution
A deep learning-based image coding method and device utilizing a feature map prediction neural network for rate control, which includes multiple synthetic neural networks learned to minimize bits per pixel, peak signal-to-noise ratio, or structural similarity index measure at various compression rates, and involves dequantization and synthesis of feature maps using convolution, correlation, and other neural network layers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional video compression methods are used, then temporal and spatial redundancy removal is performed, but coding efficiency is not optimized effectively
Solution Approach 1:
The patent replaces conventional mechanical compression algorithms with deep learning-based neural networks that automatically learn optimal compression characteristics from training data, enabling the system to adapt to various compression rates and optimize coding efficiency without manual parameter adjustment
Solution Approach 2:
The patent implements multiple synthetic neural networks, each trained for specific compression rates, allowing the system to dynamically switch between different compression characteristics based on the required balance between coding efficiency and reconstruction quality
2Productivity
If multiple synthetic neural networks for different compression rates are implemented, then coding efficiency is improved, but device complexity increases
Solution Approach 1:
The patent creates a dynamic system where the appropriate synthetic neural network is selected based on the required compression rate, allowing the system to adapt its complexity level to match the operational requirements rather than always running at maximum complexity
Solution Approach 2:
The patent divides the complex compression task into multiple specialized synthetic neural networks, each handling a specific compression rate range, which allows for more efficient processing by avoiding unnecessary computational complexity when lower compression rates are not required
Data Source
AI summary
A deep learning-based signal processing method according to the present invention may: obtain a quantized feature map from a bitstream; reconstruct the feature map by performing inverse quantization on the quantized feature map; and synthesize the reconstructed feature map on the basis of a neural network.