Image Dataset Cleanup Using IoU Filtering for Semantic Redundancy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for semantic redundancy filtering in image data sets for machine learning models are inefficient and do not effectively eliminate unnecessary repetitions and pixel-level overlaps, particularly in applications like automatic unpacking machines and autonomous vehicles.
Innovation Solution
A method utilizing an intersection-over-union (IoU) filter to compare images and remove redundant images by determining similarity based on a predetermined threshold, focusing on image foreground dynamics in static or quasi-static backgrounds.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If semantic redundancy filtering is applied to reduce data volume, then data processing efficiency is improved, but information loss may occur
Solution Approach 1:
The patent applies different filtering strategies to different regions of the image data. By using intersection-over-union filtering for semantic redundancy reduction while maintaining local image quality and content, the method selectively removes only truly redundant images while preserving informative ones, thus balancing data volume reduction with information retention.
Solution Approach 2:
The patent performs preliminary filtering actions by pre-processing the image data to identify and remove obviously redundant images before the main training process. This preliminary action reduces the overall data volume and processing time while the subsequent training processes continue to refine the model using the filtered data, minimizing information loss.
2Speed
If image data is processed in rapid succession from camera streams, then real-time processing capability is improved, but data redundancy increases
Solution Approach 1:
The patent merges multiple redundant images into representative samples by using intersection-over-union filtering. Instead of processing each individual image separately, the method combines semantically similar images from rapid camera streams into consolidated data representations, reducing the quantity of data while maintaining the real-time processing capability through efficient filtering algorithms.
3Quantity of substance
If existing redundancy filtering methods are used, then data volume is reduced, but pixel-level overlaps remain
Solution Approach 1:
The patent replaces traditional mechanical or simple algorithmic filtering with intersection-over-union filtering that operates on semantic and spatial relationships. This substitution allows for more precise control over pixel-level overlaps by using mathematical intersection and union operations on image masks and segmentation data, achieving both data volume reduction and high precision overlap removal.
Data Source
AI summary
A method for cleaning up an image data set used for training, validating, and/or testing a machine learning model includes providing the image data set that includes a plurality of images. The method also includes comparing a predetermined comparison image of the plurality of images with at least a portion of remaining images of the plurality of images by applying an intersection-over-union filter. Based on the comparison, the method includes determining at least one redundant image with respect to the predetermined comparison image in at least the portion of remaining images of the plurality of images, and cleaning up the image data set by removing the at least one redundant image from the plurality of images.