Machine Learning Data Generation for Uniform Color Distribution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning techniques face reduced learning accuracy due to deviations in color attributes' distribution, particularly hue, in teacher image groups, leading to potential false patterns during deep learning processes like demosaic networks, where robustness is compromised.
Innovation Solution
An information processing apparatus generates a teacher image group with uniform distribution characteristics by analyzing and compensating sparse hue distributions using Computer Graphics (CG) images, ensuring balanced hue, saturation, and luminance attributes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If teacher images are generated by changing image-inherent components (camera parameters, light source parameters), then a sufficient number of teacher images can be secured, but deviation occurs in the distribution of color attributes (hue, saturation, luminance) in the obtained teacher image group
Solution Approach 1:
The patent performs preliminary analysis of the color attribute distribution in the generated teacher image group before using them for training. By analyzing the hue, saturation, and luminance distributions in advance, the system identifies deviations and can take corrective actions such as generating additional images with specific color characteristics or adjusting the image generation parameters to achieve more uniform distribution before the training process begins.
Solution Approach 2:
The patent adjusts the parameters used in teacher image generation to control and balance the distribution of color attributes. By modifying camera parameters, light source parameters, and other image-inherent components systematically, the system aims to produce a teacher image group with more uniform distribution of hue, saturation, and luminance, thereby resolving the contradiction between quantity and distribution uniformity.
2Productivity
If deep learning is performed using teacher image group with deviation in color attributes distribution, then learning can be conducted with sufficient data, but false patterns occur and robustness of trained model is reduced
Solution Approach 1:
The patent introduces a feedback mechanism where the distribution characteristics of color attributes in the teacher image group are analyzed and evaluated. Based on this analysis, the system provides feedback to adjust the image generation process or select appropriate subsets of teacher images, ensuring that the training data has balanced color attribute distribution. This feedback loop helps maintain both learning efficiency and model robustness.
Data Source
AI summary
An object is to obtain a trained model whose robustness is high in a case of learning a network in an image signal processing system. For teacher images consisting an acquired teacher image group, a distribution characteristic of at least one of three attributes of color is analyzed and based on analysis results, a teacher image group uniform in the distribution characteristic is generated. Learning of a network is performed by generating a data set including a set of a teacher image and a pupil image from a new teacher image group thus obtained.