Abstract
Object detection and image segmentation of regions of interest provide the foundation for numerous pipelines across disciplines. Robust and accurate computer vision methods are needed to properly solve image-based tasks. Multiple algorithms have been developed to solely detect edges in images. Constrained to the problem of creating a thin, one-pixel wide, edge from a predicted object boundary, we require an algorithm that removes pixels while preserving the topology. Thanks to skeletonize algorithms, an object boundary is transformed into an edge; contrasting uncertainty with exact positions. To extract edges from boundaries generated from different algorithms, we present a computational pipeline that relies on: a novel skeletonize algorithm, a non-exhaustive discrete parameter search to find the optimal parameter combination of a specific post-processing pipeline, and an extensive evaluation using three data sets from the medical and natural image domains (kidney boundaries, NYU-Depth V2, BSDS 500). While the skeletonize algorithm was compared to classical topological skeletons, the validity of our post-processing algorithm was evaluated by integrating the original post-processing methods from six different works. Using the state of the art metrics, precision and recall based Signed Distance Error (SDE) and the Intersection over Union bounding box (IOU-box), our results indicate that the SDE metric for these edges is improved up to 2.3 times. Our work provides guidance for parameter tuning and algorithm selection in the post-processing of predicted object boundaries.
Introduction
In modern computer vision, deep learning models often output "fuzzy" probabilistic boundaries rather than the crisp, one-pixel-wide lines required for high-precision applications like robotic surgery or augmented reality. This "last-mile" problem involves converting these blurry regions—where pixel intensity represents edge probability—into exact topological representations. Traditionally, post-processing steps like thinning or skeletonization are applied using fixed parameters, often ignoring the rich uncertainty information contained in the original output. This paper addresses the need for a mathematically optimized post-processing pipeline that preserves topology while maximizing edge accuracy.
Methodology
The proposed approach formalizes edge extraction as a three-parameter optimization problem involving thresholding ($\alpha$), skeletonization ($\beta$), and pruning ($\gamma$).
The Post-Processing Pipeline
The core of the methodology is a computational pipeline designed to extract edges from predicted boundaries. The process is defined by the function $\mathcal{E}(g, \alpha, \beta, \gamma)$, where $g$ is the input image.
- Thresholding ($\alpha$): Filters out low-intensity noise to isolate the primary boundary region.
- Skeletonization ($\beta$): The thinning step. The authors introduce Gray-Weighted Path Skeletonize (GWPS). Unlike standard methods that binarize images first, GWPS treats the boundary as a 3D topographic surface. It utilizes a Gray-Weighted Distance Transformation (GWDT) to identify the "ridge" of highest intensity, effectively locating the most probable position of the true edge.
- Pruning ($\gamma$): Removes "spurs" or small artifacts branching off the main edge to ensure a clean final result.
The researchers utilized combinatorial optimization (grid search) to determine the optimal parameter set for specific models and datasets, moving away from generic default settings.
Experimental Evaluation
The pipeline was validated across three distinct domains to ensure generalizability:
- Medical Imaging: 2,250 images of kidney boundaries from porcine laparoscopic surgeries.
- Robotics/Indoor Scenes: 1,449 images from the NYU-Depth V2 dataset.
- Natural Images: 500 images from the BSDS 500 dataset, a benchmark for edge detection.
The performance was measured using Signed Distance Error (SDE), which quantifies the distance between predicted and ground-truth edges, and Intersection over Union (IoU-box) for shape preservation.
Results and Discussion
The optimized pipeline demonstrated significant improvements over the original post-processing methods used in six prominent computer vision studies.
- Accuracy Gains: On the BSDS 500 dataset, the SDE was improved by up to 2.3 times.
- Algorithm Performance: The novel GWPS algorithm outperformed classical topological skeletons in 66.7% of the tested scenarios.
- Shape Fidelity: The IoU-box metric saw improvements of up to 1.6 times, indicating better preservation of the overall object shape.
The results highlight that "gray" information (probability values) is crucial for accurate ridge detection and that post-processing parameters must be task-specific.
Conclusion
This work demonstrates that treating post-processing as an optimized component of the computer vision pipeline significantly enhances edge quality. The introduction of the GWPS algorithm and the systematic tuning of $\alpha, \beta, \text{ and } \gamma$ provide a robust framework for future research. Future work will explore automating parameter selection via neural networks and integrating post-processing directly into the training phase of edge-detection models.
Summary updated