Training Image Generation via Surface Luminance Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for generating training images for image recognition, particularly in industrial robot applications, face challenges in accurately representing object orientations and luminance distributions, requiring cumbersome adjustments and specialized equipment, which complicates the preparation of training images that reflect real-world environmental conditions.
Innovation Solution
An image processing apparatus that estimates surface luminance distribution using a luminance image and range image, generating training images based on model information and environmental conditions, allowing for the creation of images that approximate the object's surface luminance and orientation, thereby simplifying the generation of dictionaries for image recognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a CG image is generated by CAD to represent arbitrary orientations, then orientation flexibility is improved, but manufacturing precision deteriorates because edge extraction results vary greatly depending on material and ambient light
Solution Approach 1:
The patent uses a photographed image of the actual object as a template to copy its luminance distribution characteristics. Instead of generating luminance information from scratch via CAD, the system captures the real object's appearance under various orientations and uses this captured data to generate training images, ensuring the luminance characteristics accurately reflect the actual object's material properties and surface features
Solution Approach 2:
The patent introduces a photographed image as an intermediary between the CAD model and the final training image. The CAD model provides geometric information while the photographed image provides accurate luminance distribution. These two are combined through image processing to generate training images that have both correct geometry and realistic luminance characteristics
2Manufacturing precision
If rendering is used to generate training images close to photographed images, then manufacturing precision is improved, but device complexity worsens due to need for BRDF measurement equipment and ambient light measurement
Solution Approach 1:
Instead of using complex rendering with BRDF measurements, the patent directly copies the luminance distribution from photographed images of the actual object. This eliminates the need for specialized measurement equipment while maintaining accurate luminance characteristics that reflect the real object's material properties
Solution Approach 2:
The system uses the object itself to generate the training data by photographing the actual object under various orientations. The object's own appearance characteristics are captured and reused, eliminating the need for external measurement equipment and complex environmental controls
3Ease of manufacture
If environment mapping with a sphere is used to generate training images, then ease of manufacture is improved, but manufacturing precision deteriorates because it is difficult to prepare a sphere with the same reflection characteristic as the object
Solution Approach 1:
The patent copies the luminance distribution directly from photographed images of the actual object rather than using environment mapping with a sphere. This approach maintains reflection characteristic accuracy by using the object's own visual data while keeping the process simple through automated image capture and processing
Data Source
AI summary
In a case where generating a training image of an object to be used to generate a dictionary to be referred to in image recognition processing of detecting the object from an input image, model information of an object to be detected is set, and a luminance image of the object and a range image are input. The luminance distribution of the surface of the object is estimated based on the luminance image and the range image, and the training image of the object is generated based on the model information and the luminance distribution.