Image Encoding Using DNN Subjective Quality Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The quality of images decoded after a transformation and quantization process is often degraded due to errors, and increasing data compression to reduce this degradation can lead to a decrease in compression efficiency.
Innovation Solution
A method using a Deep Neural Network (DNN) to determine the subjective quality of images during compression, allowing for optimal compression of residual signals based on determined degrees of compression, and employing an autoencoder to encode and decode images efficiently.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If transformation and quantization process is applied to compress image data, then compression efficiency is improved, but image quality is degraded due to errors
Solution Approach 1:
The patent applies preliminary action by using a Deep Neural Network to predict subjective quality metrics before the actual encoding process. The DNN model predicts quality scores based on input features, allowing the system to pre-determine optimal encoding parameters that will maintain quality while achieving compression. This predictive approach enables quality preservation before the degradation can occur during quantization.
Solution Approach 2:
The patent implements parameter changes by dynamically adjusting encoding parameters based on DNN predictions. The system changes quantization parameters, bit allocation, and compression settings according to the predicted quality metrics. By modifying these parameters adaptively, the system maintains optimal balance between compression efficiency and image quality, preventing quality degradation while achieving effective compression.
2Manufacturing precision
If more data is encoded to reduce quality degradation, then image quality is improved, but compression efficiency is degraded
Solution Approach 1:
The patent applies local quality by using the DNN to predict quality metrics for different regions or frames independently. The system identifies areas where quality is more critical and allocates more bits to those regions, while applying higher compression to less critical areas. This localized approach ensures quality is maintained where needed without unnecessarily reducing compression efficiency in other areas.
Solution Approach 2:
The patent implements partial action by applying different compression levels to different parts of the image or video sequence. Instead of uniformly encoding all data at high quality, the system applies partial encoding at full quality only where the DNN predicts quality degradation would be noticeable, and uses higher compression elsewhere. This selective approach maintains quality where necessary while improving overall compression efficiency.
Data Source
Figure 1A
Figure 1B
Figure 2A
AI summary
Provided is a method of encoding an image, the method including: determining a subjective quality of the image when the image is compressed; determining at least one degree of compression that changes the subjective quality and is from among degrees of compression indicating how much the image is compressed; and encoding the image by compressing a residual signal of the image, based on compression information according to the determined degree of compression, wherein the subjective quality is determined for each frame by using a Deep Neural Network (DNN). Provided are an image decoding method and an image decoding apparatus for performing the image decoding method for decoding an image by using information encoded according to an image encoding method.