Computer Vision Quality Verification Using Image Transformations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current labeled datasets for computer vision systems fail to address real-world image distortions, limiting their ability to evaluate the quality of computer vision systems effectively in diverse real-world scenarios.
Innovation Solution
Applying transformations that mimic real-world distortions to a reference dataset to generate a new, more diverse dataset of images, which allows for better evaluation of a computer vision system's quality and accuracy in real-world use cases.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If current labeled datasets are used for quality verification, then the verification process is simple and fast, but the evaluation accuracy and reliability are insufficient due to lack of real-world distortions
Solution Approach 1:
The patent applies image transformations (rotations, flips, zooms, distortions) to reference images before creating the labeled dataset. This preliminary action ensures that the dataset inherently contains diverse real-world variations, improving evaluation accuracy without requiring complex post-processing during verification
Solution Approach 2:
The patent systematically varies multiple image parameters simultaneously (orientation angles, scale factors, distortion types, lighting conditions) to generate training images. This multi-parameter approach creates comprehensive test coverage that accurately reflects real-world variability while maintaining structured dataset organization
2Reliability
If a diverse dataset with real-world distortions is created, then the computer vision system quality evaluation is more reliable, but the dataset creation process becomes more complex and time-consuming
Solution Approach 1:
The patent uses a smaller set of reference images and applies transformations to generate multiple test variants. This copying approach creates a large diverse dataset from a small reference set, ensuring reliable quality verification across various conditions without manually creating numerous unique images
Solution Approach 2:
The patent efficiently generates dataset diversity by systematically varying transformation parameters (rotation angles, scale factors, distortion types) applied to reference images. This parameter-based generation method creates realistic variations quickly through computational transformation rather than manual image creation
3Adaptability or versatility
If transformations are applied to generate diverse images, then the computer vision system performance is better evaluated, but the processing complexity and computational resources increase
Solution Approach 1:
The patent generates diverse test images by applying transformations to a limited set of reference images. This copying and transforming approach evaluates system adaptability across multiple conditions without requiring a proportionally large increase in processing complexity, as the same base images are reused with different transformations
Data Source
AI summary
Techniques for using image dataset transformations to verify the quality of a computer vision system are disclosed. In some example embodiments, a computer-implemented method comprises: accessing a database to obtain a reference image; generating a plurality of new images based on the reference image using a plurality of transformations, each one of the plurality of transformations being configured to change a corresponding visual property of the reference image; feeding the plurality of new images into an image classifier to generate a corresponding classification result for each one of the plurality of new images; determining that the image classifier does not satisfy one or more accuracy criteria based on the generated classification results for the plurality of new images; and based on the determining that the image classifier does not satisfy the one or more accuracy criteria, selectively performing a function.