Extracting Regions of Interest from Annotated Images with Occluding Labels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods struggle to efficiently extract human knowledge from images where annotations are directly drawn, especially in cases where region of interest annotations are not continuous and perfectly closed, and where filtering techniques fail to separate regions of interest from textual labels.
Innovation Solution
A method involving the extraction of annotations from images using an excess color filter, followed by the separation of region of interest and text labels through erosion and subtraction, and the reconstruction of boundaries using bipartite matching and a greedy algorithm to create a filled-in polygon.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If filtering techniques are used to extract annotations, then extraction speed is improved, but the ability to separate regions of interest from textual labels deteriorates
Solution Approach 1:
The annotation extraction process is divided into distinct stages: initial filtering to isolate colored pixels, skeletonization to reduce boundary thickness, erosion to separate text from regions, and gap closing to complete boundaries. Each stage processes the data incrementally, maintaining both speed and accuracy.
Solution Approach 2:
The method performs preliminary skeletonization to reduce annotation boundaries to single-pixel lines before attempting separation. This preliminary structuring makes subsequent erosion and text separation operations more effective and reliable.
2Device complexity
If simple color filtering is used, then processing complexity is reduced, but the ability to handle non-continuous and imperfectly closed boundaries deteriorates
Solution Approach 1:
Boundary reconstruction is segmented into identifying dead-end vertices, finding complementary segments through bipartite matching, and joining segments to close gaps. This systematic approach handles non-continuous boundaries reliably while maintaining manageable complexity.
Solution Approach 2:
The skeletonized representation serves as an intermediary structure between the raw annotation and the final closed boundary. This intermediate form simplifies the identification of gaps and facilitates systematic reconstruction of complete boundaries.
3Difficulty of detecting and measuring
If image subtraction is used to lift annotations, then extraction capability is improved, but the ability to separate regions of interest from text labels and repair boundaries deteriorates
Solution Approach 1:
The method segments the annotation processing into distinct operational phases: extraction via subtraction, skeletonization for structure, erosion for separation, and gap closing for completion. This segmentation makes the complex overall task manageable and effective.
Solution Approach 2:
Skeletonization is performed as a preliminary action before separation and gap closing operations. This preliminary structuring of the annotation data enables subsequent operations to work more effectively with reduced complexity.
Data Source
AI summary
A system and method provide extractions of regions of interest from images hand annotated by reviewers by lifting annotations from images, filtering out text labels, reconstructing continuous closed boundaries, and marking the contained region.