Data Creation System Height Deformation for Object Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data augmentation techniques for machine learning, such as those described in Patent Literature 1, are insufficient for creating a wide variety of learning data, leading to a decline in object recognition performance, especially when objects need to be recognized locally.
Innovation Solution
A data creation system that generates second image data by deforming the height of first image data regions based on a reference point and boundary proximity, varying the height closer to the reference point and decreasing it closer to the boundary, to create more diverse learning data for improved object recognition models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If simple luminance and hue combination changes are applied to existing training data, then data augmentation is achieved, but the variety of learning data is insufficient for local object recognition
Solution Approach 1:
The patent introduces a new dimension of deformation about height to the existing training data. Instead of only changing luminance and hue in the 2D image plane, the system deforms the height dimension by calculating deformation amounts based on distance from reference points and boundaries, creating varied 3D-like perspectives from 2D images. This dimensional expansion significantly increases the variety of learning data for local object recognition.
Solution Approach 2:
The patent applies different deformation characteristics to different regions of the image. By calculating deformation amounts based on distance from reference points and proximity to boundaries, the system creates locally-adapted deformations where central regions undergo different transformations than boundary regions. This local quality approach ensures that deformation patterns are appropriate for local object recognition tasks.
2Reliability
If deformation about height is applied to create diverse learning data, then object recognition performance is improved, but data processing complexity increases
Solution Approach 1:
The patent segments the image into multiple regions with different deformation characteristics. By dividing the image based on reference points and boundaries, the system applies simplified deformation calculations to each segment rather than attempting complex global deformations. This segmentation reduces processing complexity while maintaining the ability to create diverse learning data for local recognition.
Solution Approach 2:
The patent applies partial deformation actions focused on specific regions rather than exhaustive processing of the entire image. By concentrating deformation operations on areas most relevant to local object recognition and using distance-based weighting, the system achieves improved recognition performance without the computational burden of processing every pixel uniformly.
Data Source
AI summary
A data creation system creates, based on first image data, second image data for use as learning data. A processor of the data creation system generates, based on the first image data including a first region as a pixel region representing the object and a second region adjacent to the first region, the second image data by causing deformation about height of the first region with respect to a reference plane. The processor generates the second image data such that the closer to a reference point within the first region a point of interest is, the greater a variation in the height of the first region with respect to the reference plane is and the closer to a boundary between the first region and the second region the point of interest is, the smaller the variation in the height of the first region with respect to the reference plane is.