Object Recognition Model Training via Simulated Image Augmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image recognition models face challenges in achieving high recognition accuracy when only a small volume of images is available, particularly in specific categories or fields like auto-piloting and smart retail, where obtaining large volumes of images is difficult.
Innovation Solution
An image recognition method and system that generates simulated sample images by adjusting physical parameters such as object rotation, displacement, shading, covering ratio, lens distortion, scene, and light source position, and automatically marks object range and type information, allowing for machine learning-based training of object recognition models even with limited original images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a large volume of original images is used for training, then recognition accuracy is improved, but it becomes difficult to obtain sufficient images in specific categories or fields
Solution Approach 1:
The patent creates simulated sample images by copying and transforming existing original sample images through parameter adjustments. This allows generating large volumes of training data from limited original images, resolving the contradiction between needing many images for accuracy and the inability to obtain them in specific fields.
Solution Approach 2:
The patent adjusts physical parameters (rotation, displacement, shading, covering ratio, lens distortion, scene, light source position) of original images to generate diverse simulated samples. This transforms a small set of original images into a large training dataset, enabling high recognition accuracy without requiring abundant original images.
2Quantity of substance
If simulated sample images are generated by adjusting physical parameters, then the volume of training data is increased, but the complexity of the image processing system increases
Solution Approach 1:
The patent replaces complex manual image annotation and data collection processes with automated computational methods. The marking unit automatically marks object range and type information in simulated images without manual intervention, reducing system complexity despite generating large volumes of training data.
Solution Approach 2:
The system performs self-service by automatically generating simulated sample images and their corresponding annotations from original images. The marking unit autonomously identifies and marks objects in simulated images, eliminating the need for manual annotation of each generated image and reducing overall system complexity.
3Measurement precision
If manual marking is performed on each simulated sample image, then accurate object information is obtained, but the time and labor required increase significantly
Solution Approach 1:
The marking unit implements self-service by automatically marking object range and type information in simulated sample images based on the original sample images and applied transformations. This eliminates manual marking of each simulated image while maintaining accurate object information, resolving the time-accuracy tradeoff.
Solution Approach 2:
The system performs preliminary action by marking objects in original sample images before generating simulated images. The marking information is then automatically transferred to simulated images based on parameter transformations, avoiding the need to re-mark each simulated image and significantly reducing time consumption.
Data Source
AI summary
An image recognition method, a training system for an object recognition model and a training method for an object recognition model are provided. The image recognition method includes the following steps. At least one original sample image of an object in a field and an object range information and an object type information in the original sample image are obtained. At least one physical parameter is adjusted to generate plural simulated sample images of the object. The object range information and the object type information of the object in each of the simulated sample images are automatically marked. A machine learning procedure is performed to train an object recognition model. An image recognition procedure is performed on an input image.