Fiber Image Segmentation Using Instance-Semantic Region Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image segmentation technologies struggle to accurately detect individual fibers, especially when they are entangled or overlapping, and are prone to erroneously identifying line-shaped dirt as fibers.
Innovation Solution
An image processing system utilizing a combination of instance and semantic segmentation models, specifically Mask R-CNN and DeepLab, to generate and correct individual-object segmentation results by merging them with category segmentation results through logical conjunction and region correction techniques such as dilation and smoothing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a single instance segmentation model is used to detect individual fibers, then the model can identify separate fiber objects, but it fails to accurately detect fibers in entangled or overlapping regions and may erroneously detect line-shaped dirt as fibers
Solution Approach 1:
The patent merges instance segmentation (Mask R-CNN) and semantic segmentation (DeepLab) models to create a hybrid segmentation system. The instance segmentation model identifies individual fiber objects, while the semantic segmentation model accurately delineates fiber regions including entangled and overlapping areas. By combining both segmentation results, the system achieves both object-level and region-level accuracy, resolving the contradiction between detecting individual fibers and maintaining reliability in complex regions.
Solution Approach 2:
The patent introduces a correction mechanism as an intermediary between the two segmentation models. The semantic segmentation result serves as a reference to correct the instance segmentation result, particularly in entangled and overlapping regions where instance segmentation alone fails. This intermediary correction process improves detection reliability without compromising the ability to identify individual fiber objects.
2Area of stationary object
If the segmentation model processes complex images with entangled fibers, then it can capture all fiber regions, but it cannot accurately identify individual fibers in those regions
Solution Approach 1:
The patent applies segmentation at two different levels: instance segmentation divides the image into individual fiber objects, while semantic segmentation divides the image into fiber and non-fiber regions. By performing segmentation at multiple levels, the system can both cover all fiber regions (including entangled areas) and accurately identify individual fibers where applicable, resolving the contradiction between comprehensive coverage and precise identification.
3Measurement precision
If the segmentation model is sensitive enough to detect all potential fibers, then it can identify individual fibers, but it also detects line-shaped dirt and other non-fiber objects as fibers
Solution Approach 1:
The patent implements a feedback mechanism where the semantic segmentation result is used to verify and correct the instance segmentation result. The semantic segmentation model provides ground truth region information that feedbacks to the instance segmentation model, allowing it to distinguish true fibers from false positives like line-shaped dirt. This feedback loop maintains high detection precision while reducing false positive detections through cross-validation.
Data Source
Figure 1~2
Figure 3
Figure 4
AI summary
An image processing device includes an image acquisition part configured to acquire a target image in which fibers are captured, a first segmentation part configured to generate an individual-object segmentation result detecting each of the fibers included in the target image using a trained individual-object segmentation mode, a second segmentation part configured to generate a category segmentation result recognizing regions where the fibers are captured in the target image using a trained category segmentation model, a region correction part configured to correct the individual-object segmentation result with the category segmentation result, and a result output part configured to output a correction result of the individual-object segmentation result.