Color Space Transformation for Training Image Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Generating large quantities of training image data for machine learning models, particularly for target object detection, is challenging and requires improvement.
Innovation Solution
A method and system for generating multiple sets of training image data by acquiring object image data, dividing it into partial regions based on color ranges, modifying colors, and combining with background images to create diverse training images, facilitating efficient training of machine learning models for target object detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If large quantities of training image data are generated using conventional methods (3D computer graphics data and scene data), then the machine learning model can be trained for target object detection, but the data generation process is complex and difficult to scale
Solution Approach 1:
The patent uses copying by creating multiple color variations of a single object image through color space transformation. Instead of generating entirely new images using complex 3D rendering, the system copies the original object image and transforms its color representation in different color spaces (HSV, LAB, YCbCr), producing diverse training samples from a single source image.
Solution Approach 2:
The patent applies parameter changes by modifying color parameters within different color spaces. The system transforms the object image to multiple color spaces, then adjusts specific color parameters (hue, saturation, lightness in HSV; L*, a*, b* in LAB) to generate varied color versions. This allows efficient generation of diverse training data by changing numerical parameters rather than regenerating images.
2Adaptability or versatility
If training data includes only single-color object images, then the data generation process is simple, but the machine learning model cannot detect target objects with different colors effectively
Solution Approach 1:
The patent directly applies color changes by transforming object images into multiple color variations. The system converts the original RGB image to HSV, LAB, and YCbCr color spaces, then modifies color parameters (such as hue rotation in HSV, L* adjustments in LAB) to generate images with different colors while maintaining the same object structure, enabling the model to learn color-invariant detection.
Solution Approach 2:
The patent applies dimensionality change by transforming the image data from RGB color space to alternative color spaces (HSV, LAB, YCbCr). This dimensional transformation in color space allows the system to manipulate color properties independently and generate diverse color variations without altering the spatial structure of the object, thereby expanding training data diversity in a different dimensional domain.
3Measurement precision
If multiple color variations of training images are generated, then the machine learning model achieves better detection accuracy for different colored objects, but the data generation process requires more computational resources
Solution Approach 1:
The patent applies preliminary action by performing color space transformations and generating all color variations during the offline data preparation phase. By pre-generating diverse colored versions of training images before model training, the system eliminates the need for computationally intensive operations during the actual detection phase, improving both detection accuracy and real-time performance.
Data Source
AI summary
A method for generating a plurality of sets of training image data for training a machine learning model includes: (a) acquiring object image data representing an object image; (b) dividing the object image into T number of partial object images by dividing a region of the object image into T number of partial regions corresponding to respective ones of T number of partial color ranges; (c) generating a plurality of sets of color-modified object image data representing respective ones of a plurality of color-modified object images by performing an adjustment process on the object image data, the adjustment process including a color modification process to modify colors of at least one of the T number of partial object images; and (d) generating the plurality of sets of training image data using one or more sets of background image data and the plurality of sets of color-modified object image data.