Learning Model Generation via Optical Distortion Conversion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image recognition methods face high processing loads due to the need for continuous conversion of captured images, and limited accuracy due to differences in optical characteristics between learning and recognition devices, particularly with wide-angle lenses like fisheye lenses.
Innovation Solution
An information processing device and method that converts learning images to match the distortion characteristics of the recognition device's optical system, generating a learning model based on these converted images to improve accuracy and reduce processing load.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conversion processing is performed on captured images every time image recognition is performed, then image recognition accuracy is improved, but processing load increases
Solution Approach 1:
The patent applies preliminary action by pre-converting training images to the target optical characteristics before model training. This allows the conversion processing to be performed once during the learning phase rather than repeatedly during inference, significantly reducing the processing load at recognition time while maintaining accuracy.
Solution Approach 2:
The patent segments the image processing workflow into distinct phases: training image conversion, model training, and inference. By separating the conversion operation from the recognition operation, the system performs heavy processing only when necessary (during training) and enables fast recognition during deployment.
2Productivity
If images for learning are captured using an image capturing device for image recognition, then processing load is reduced, but the number of collectable images is limited and learning model accuracy is insufficient
Solution Approach 1:
The patent changes the optical characteristic parameter of training images by converting them to match the target device's optical properties. This allows using images from any source (including devices with different optical characteristics) while adapting them to the specific device being trained, thereby expanding the available training data pool without increasing processing load during inference.
Solution Approach 2:
The patent creates converted copies of training images that simulate the optical characteristics of the target device. These copied and transformed images serve as synthetic training data that mimics what the target device would capture, enabling model training without requiring physical images from the specific device.
3Device complexity
If a learning model is generated without considering distortion characteristics of the optical system, then device complexity is reduced, but learning model accuracy deteriorates
Solution Approach 1:
The patent applies preliminary action by pre-adjusting training images to match the target optical characteristics before model training. This preliminary conversion ensures the model learns the correct distortion patterns without requiring complex real-time adjustment mechanisms, maintaining device simplicity while improving accuracy.
Data Source
AI summary
Provided is an information processing device configured to generate a learning model for performing image recognition on a first image acquired by a first imaging device including an optical system having a first optical characteristic, including: a conversion unit configured to convert a second image for learning to generate a third image having a distortion characteristic based on the first optical characteristic; and a generation unit configured to generate the learning model based on the third image. The second image is an image acquired by a second imaging device including an optical system having a second optical characteristic different from the first optical characteristic.