Image Prediction Learning Using High-Frequency Edge Separation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image compression methods, particularly those using rule-based and machine learning models, suffer from low prediction performance, especially when dealing with edges in images, leading to inefficient compression and increased noise in steep edge regions.
Innovation Solution
A learning device performs component separation on frequency components of reference pixels using feature vectors to learn a model that outputs prediction values based on high-frequency vectors, enhancing prediction accuracy through a neural network model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If a rule-based simple machine learning model is used for prediction, then the model complexity is low and ease of operation is high, but prediction performance deteriorates especially for edge regions
Solution Approach 1:
The patent segments the frequency components of reference pixels into multiple bands (low-frequency and high-frequency components) and processes them separately. The high-frequency components are specifically extracted to capture edge information, while the low-frequency components provide the base prediction. This segmentation allows the model to handle different spatial frequencies appropriately, improving prediction accuracy at edges without excessive complexity.
Solution Approach 2:
The patent transforms the prediction problem from the spatial domain to the frequency domain by applying Fourier transform. This dimensional transformation enables the model to capture periodic patterns and edge structures more effectively. The high-frequency vectors in the frequency domain provide additional information about edge characteristics, enhancing prediction performance while maintaining manageable model complexity.
2Measurement precision
If existing machine learning models are used for prediction, then prediction performance is better than rule-based models, but compression noise increases in steep edge regions
Solution Approach 1:
The patent applies different processing strategies to different frequency components locally. High-frequency components are specifically targeted for extraction and separate processing to preserve edge sharpness, while low-frequency components are processed separately for smooth regions. This local differentiation prevents the introduction of compression noise in edge regions while maintaining overall prediction accuracy.
Solution Approach 2:
The patent introduces high-frequency vectors as an intermediary representation that captures edge information without directly using the original pixel values. These high-frequency vectors serve as a mediator between the reference pixels and the prediction, allowing the model to infer edge characteristics indirectly and reduce compression artifacts in steep edge regions.
3Measurement precision
If frequency component separation is performed on reference pixels, then prediction accuracy for edges is improved, but calculation complexity increases
Solution Approach 1:
The patent performs frequency component separation and extraction of high-frequency vectors as preliminary actions during the learning phase. By pre-processing the reference pixels into frequency components and storing the high-frequency vectors, the actual prediction process during encoding/decoding becomes simpler and faster. This preliminary action shifts the computational burden to the learning phase rather than the inference phase.
Data Source
AI summary
A learning device according to the present disclosure includes: a first filter processing unit that performs component separation on frequency components included in reference pixels based on feature vectors of the reference pixels in the vicinity of a pixel to be predicted in image data; and a learning unit that learns a model that outputs a prediction value of the pixel to be predicted by using, as learning data, a set of a high-frequency vector, which is a feature vector of a high-frequency component among frequency components obtained by the component separation, and high-frequency information, which relates to a high-frequency component among frequency components included in the pixel to be predicted.