Weighted Loss Training for Accurate Image Block Partitioning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models for image compression, such as H.265/HEVC, often output block dividing patterns that significantly differ from the correct answers, leading to degraded image quality during encoding.
Innovation Solution
A method involving a determination unit to assess the similarity between output patterns and teacher data, and a learning unit to perform machine learning using a weighted loss function to improve the accuracy of block dividing patterns.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a learning model outputs block dividing patterns based on probability learning, then the model can generate diverse dividing patterns, but the output may significantly differ from correct answers leading to degraded image quality
Solution Approach 1:
The patent applies local quality by differentiating the treatment of different output nodes in the learning model. Nodes representing different block dividing patterns are assigned different weights based on their similarity to teacher data. This allows the system to maintain probability-based diversity for some patterns while enforcing higher accuracy for patterns closer to correct answers, thereby resolving the contradiction between versatility and precision.
2Adaptability or versatility
If the learning model outputs block dividing patterns largely different from correct answers, then the model shows high creativity in pattern generation, but encoding using these patterns significantly degrades image quality
Solution Approach 1:
The patent changes the parameter of loss function weighting dynamically during the learning process. By adjusting the weight parameter based on the similarity between output patterns and teacher data, the system can control the balance between encouraging diverse pattern generation and ensuring encoding quality. This parameter adjustment allows the model to explore creative patterns while maintaining reliability for patterns that matter most for image quality.
3Device complexity
If uniform weight is applied to all loss function values in machine learning, then the learning process is simple, but patterns significantly different from teacher data are not adequately corrected
Solution Approach 1:
The patent applies preliminary action by pre-calculating and storing teacher data that represents correct block dividing patterns before the actual learning process. During training, these pre-prepared teacher data serve as reference points for comparing model outputs and determining appropriate loss function weights. This preliminary preparation enables the system to efficiently guide the learning process toward accurate patterns without adding significant complexity during the actual training execution.
Data Source
AI summary
An apparatus that performs machine learning of a learning model that can increase a rate of outputting a block dividing pattern as a correct answer. A CPU of the apparatus determines a dividing pattern output from a learning model that outputs, using an image for learning as an input thereto, a dividing pattern obtained by dividing the image, and a weight associated with a similarity with teacher data when performing machine learning of the learning model. The CPU performs machine learning of the learning model using a loss function weighted by the determined weight.