Semantic Edge Detection for Structurally Meaningful Boundary Extraction

Ngan Huynh

Co-Presenters: Tristan Wolenski

College: Hennings College of Science Mathematics and Technology

Major: BS.COMPUTER/SCI

Faculty Research Mentor: Chatterjee, Moitrayee  

Abstract:

Standard gradient-based edge detection methods, such as Canny and Sobel, are great at identifying intensity changes, but often fail at their fundamental purpose, accurately and reliably separating true object boundaries from noise. This proposes a major problem, particularly in high-stakes space applications, where accurate structural perception is absolutely essential. This project’s goal is to introduce an advanced AI-driven edge detection framework that overcomes the shortcomings of its predecessors by leveraging a combination of wireframe parsing and modern deep learning-based edge detection models like HED and DexiNed. These models produce significantly semantically cleaner edges, allowing precise edge detection for objects like solar panel boundaries and spacecraft silhouettes without the drawbacks of older edge detection methods. By training the model on labeled real-world imagery and augmenting it with synthetic CAD renders, the system will be able to align detected edges with corresponding geometric wireframes. This will enable significant NASA capabilities, such as autonomous spacecraft damage inspection, robust robot navigation via clean structural maps, precise pose estimation for orbital rendezvous, reliable docking alignment, and enhanced change detection in extraterrestrial environments. By improving edge detection so dramatically, it will enable more reliable and autonomous vision systems in space exploration.