Visual Train Positioning Test Images for Motion Blur Validation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing visual train positioning methods based on deep learning face challenges in generating comprehensive and reproducible test images that accurately reflect real-world scenarios, particularly in motion-blurred conditions, leading to safety concerns due to unexplainable model outputs.
Innovation Solution
A method and system utilizing a GAN-based image generation network to create diverse blurred images by training on real environment images, classifying blur types, and using structural similarity to filter test cases, ensuring images are of the same scenario and blur type as the original.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If real environment images are used directly as test cases, then the test images are easy to obtain, but the test completeness is insufficient due to uneven distribution, poor diversity and high similarity
Solution Approach 1:
The patent uses a GAN-based image generation network to create synthetic test images that copy the characteristics of real environment images. The generator network learns from real blurred images and generates new test cases with diverse blur types, scenarios, and conditions, solving the problem of insufficient diversity while maintaining realism.
Solution Approach 2:
The patent applies parameter changes by introducing various blur parameters (motion blur, defocus blur, depth of field) and environmental parameters into the image generation process. By systematically varying these parameters, the system generates comprehensive test cases that cover different scenarios, improving test completeness while maintaining ease of generation.
2Measurement precision
If deep learning models are used for visual positioning, then positioning accuracy is improved, but safety reliability deteriorates due to unexplainable model outputs and potential wrong results with high confidence
Solution Approach 1:
The patent applies preliminary action by conducting extensive testing and validation of the deep learning model using diverse synthetic test cases before deployment. The GAN-generated test images simulate various failure scenarios and edge cases, allowing the model to be thoroughly evaluated and validated in advance, ensuring safety reliability while maintaining positioning accuracy.
Solution Approach 2:
The patent implements feedback mechanisms through the adversarial training process where the discriminator provides feedback to the generator about the realism of generated images. This feedback loop ensures that generated test cases accurately reflect real-world conditions, enabling comprehensive validation of the positioning model's safety and reliability.
3Reliability
If motion-blurred images are used for testing, then real-world scenarios are better reflected, but the images appear less frequently and are difficult to reproduce in real environment
Solution Approach 1:
The patent uses copying by creating synthetic motion-blurred images through the GAN framework. The generator learns to copy the characteristics of real motion-blurred images and reproduces them systematically, making it easy to generate diverse test cases with consistent blur patterns that reflect real-world scenarios without requiring actual motion-blurred captures.
Solution Approach 2:
The patent replaces the mechanical system of capturing real motion-blurred images (which requires specific physical conditions) with a computational system. The GAN-based image generation network substitutes the physical capture process, allowing motion-blurred test images to be generated on demand through algorithmic processing rather than requiring specific real-world conditions.
Data Source
AI summary
A method and system for generating test cases for visual train positioning are provided and relate to the technical field of train positioning. The method includes: first obtaining real environment images around a train, and classifying the real environment images based on blur types; training a generative adversarial network (GAN)-based image generation network with each non-blurred training image, each blurred training image of any blur type in a same scenario and a preset blur type as inputs, and a reconstructed blurred image and a corresponding blur type as outputs, to obtain an image generation model; then inputting each real non-blurred image and target reference data into the image generation model to generate a target reconstructed blurred image, and then deleting target reconstructed blurred images with structural similarities lower than a set threshold.