Lens Adaptation via Transfer Learning and Feedback Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Generating an estimation model for image content estimation across different lenses is challenging due to variations in feature appearance caused by lens characteristics, such as field angle, which requires a large amount of diverse learning data.
Innovation Solution
A processing system that generates sample images from a first image captured using a specific lens, uses these images to input into an estimation model trained on data from a second lens with different characteristics, estimates the relative positional relationship of the sample images, determines the correctness of the estimation, and corrects the model parameters when necessary to improve estimation accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If an estimation model is generated using learning data from a specific lens type, then estimation accuracy for that lens type is improved, but the model cannot accurately estimate images captured by other lens types with different characteristics
Solution Approach 1:
The patent changes the parameter of lens characteristics by capturing training images with multiple lens types having different field angles and optical properties. This allows the estimation model to learn and adapt to various lens characteristics, enabling accurate estimation across different lens types while maintaining high estimation accuracy for each specific lens.
Solution Approach 2:
The patent creates a universal estimation model that can handle multiple lens types through transfer learning. The model trained on images from one lens type can be transferred and adapted to estimate images from other lens types, making the model multi-functional and adaptable to different lens characteristics without requiring completely separate models for each lens type.
2Measurement precision
If learning data is collected for each lens type to achieve accurate estimation, then estimation accuracy is improved, but the complexity and effort of data preparation increases significantly
Solution Approach 1:
The patent performs preliminary action by collecting and organizing training images from multiple lens types in advance. This pre-prepared diverse training data enables the estimation model to be trained comprehensively, reducing the need for additional data collection and preparation work when deploying the model for different lens types later.
Solution Approach 2:
The patent implements feedback mechanisms where the estimation model's performance is evaluated across different lens types, and the training process is iteratively improved. This feedback loop allows the model to learn from estimation errors and continuously enhance its accuracy for various lens characteristics, reducing the need for extensive manual data preparation for each lens type.
Data Source
AI summary
The present invention provides a processing system (10) including: a sample image generation unit (11) that generates a plurality of sample images being each associated with a partial region of a first image generated using a first lens; an estimation unit (12) that generates an image content estimation result indicating a content for each of the sample images using an estimation model generated by machine learning using a second image generated using a second lens differing from the first lens; a task execution unit (14) that estimates a relative positional relationship of a plurality of the sample images in the first image; a determination unit (15) that determines whether an estimation result of the relative positional relationship is correct; and a correction unit (16) that corrects a value of a parameter of the estimation model when the estimation result of the relative positional relationship is determined to be incorrect.