Image Processing Apparatus for Reducing Training Data Variation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The accuracy of annotation-attaching tasks for high-level recognition, such as identifying hazard regions, varies significantly among crowdsourcing workers, leading to inconsistent quality of training data when using existing methods, which negatively impacts machine learning performance.
Innovation Solution
An image processing method that acquires consecutive time-series images from an onboard camera, determines the positions of annotated regions, identifies images where these regions are not on the vehicle's path, and autonomously sets and annotates regions between them as hazard zones, reducing variation in training data quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If crowdsourcing is used to attach annotations to images, then the quantity of training data can be increased efficiently, but the quality variation among annotated data items increases due to differences in worker accuracy
Solution Approach 1:
The patent introduces an automated image processing apparatus as an intermediary that processes images between crowdsourcing workers and the final training dataset. The apparatus automatically attaches second annotations indicating hazard regions based on the first annotations provided by workers, thereby standardizing the quality of annotations while preserving the efficiency benefits of crowdsourcing for data collection.
Solution Approach 2:
The system enables self-service by having the automated processing apparatus work independently on images that meet specific criteria (multiple first regions within threshold distance on vehicle path). This automatic intervention occurs without requiring additional human review, maintaining productivity while improving quality consistency for problematic cases.
2Manufacturing precision
If automated processing is used to attach annotations, then quality consistency improves, but the complexity of the processing system increases
Solution Approach 1:
The automated processing apparatus applies quality control selectively rather than universally. It intervenes only for images meeting specific conditions (multiple first regions within threshold distance that are located on the vehicle's path), leaving other images to be processed normally. This localized approach improves quality consistency where needed while avoiding unnecessary complexity for cases that don't require intervention.
3Measurement precision
If all images are reviewed manually for high-level recognition, then annotation accuracy improves, but the time and cost required increases significantly
Solution Approach 1:
The system applies automated processing partially rather than universally. It performs automated second annotation attachment only for images that meet specific criteria indicating potential quality issues, while leaving other images to be processed through normal crowdsourcing workflows. This partial automation achieves sufficient accuracy for critical cases without incurring the time and cost of universal manual review.
Data Source
AI summary
An image processing method includes acquiring consecutive time-series images captured by an onboard camera of a vehicle, having a first annotation indicating two or more first regions, and at least including one or more images in which the two or more first regions are on a path of the vehicle and a distance therebetween is smaller than or equal to a threshold; determining, in reverse chronological order from an image of the last time point, positions of the two or more regions in each consecutive time-series image; identifying, from among the consecutive time-series images, the first image of a first time point in which none of the two or more first regions are located on the path, and setting, as a second region, a region between the two or more first regions in the identified first image; and attaching a second annotation to the first image corresponding to the first time point, the second annotation indicating the second region.