A novel instance segmentation framework for accurate counting of slender and dense objects
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Abstract Accurate counting of slender, densely stacked objects, such as welding wire stubs and welding electrode stubs, is crucial for the effective management of automated welding plants. However, the extreme aspect ratios and indistinct inter-object boundaries make the task highly challenging. To address these challenges, we propose a novel instance segmentation framework for accurately counting slender and dense objects, such as welding-wire stubs or welding electrode stubs. Specifically, its detection head adopts an instance-segmentation head with oriented bounding boxes, which adds an angular offset to the standard horizontal box, ensuring that each anchor tightly envelops the slender object. Subsequently, we propose an edge feature fusion feature pyramid network as the neck to enhance boundary awareness. It fuses edge information of objects and boosts the network’s capacity for extracting boundary features. During loss computation, we propose an edge-guided cross-entropy loss. It combines the distance error between the edges of the predicted mask and the ground-truth mask with the binary cross-entropy loss, enabling the network to explicitly optimize mask boundaries and significantly improve both accuracy and convergence speed. Finally, comprehensive accuracy benchmarking and error evaluation conducted on both our self-collected dataset and established public datasets corroborate the proposed algorithm’s validity and robustness, achieving an even more superior performance. We simultaneously developed an intelligent system for counting welding electrode stubs, demonstrating its practical applicability in welding workshops. These results demonstrate the effectiveness and robustness of the proposed novel instance-segmentation network.
Fields of Study & Subjects
Advanced Neural Network Applications
Industrial Vision Systems and Defect Detection
Image and Object Detection Techniques
Publication details
- Journal / source
- Measurement Science and Technology
- Volume
- 37
- Issue
- 15
- Pages
- 155003–155003
- Publication date
- 30 March 2026
- Language
- EN