Extraktion geschlossener Schiffs- und Objektkonturen mit 1-Pixel-Breite zur präzisen Segmentierung in Farb- und Infrarotbildern durch Deep Learning
Publication date
2025-11-19
Document type
Dissertation
Author
Kelm, André
Advisor
Referee
Frintrop, Simone
Granting institution
Helmut-Schmidt-Universität/Universität der Bundeswehr Hamburg
Exam date
2025-10-07
Organisational unit
Publisher
Universitätsbibliothek der HSU/UniBw H
Part of the university bibliography
✅
File(s)
Language
German
Keyword
Detailed ship segmentation in infrared and RGB images
One-pixel-wide contour, closed object shape extraction
Contour tracking with CNN
Walk the Lines algorithm
Object contour and edge detection
Thermal and RGB ship image dataset (unlabeled)
Abstract
This work focuses on the extraction of detailed and complex shapes of ships and other objects from RGB and thermal imagery. This task is important for accurate classification and identification and is not limited to maritime contexts. Ship segmentation from thermal imagery is particularly relevant to real-world maritime applications, but progress in this area is often not publicly available and generally lags behind RGB segmentation, underscoring the need for new benchmarks/datasets and methods. This study presents a new method for extracting 1-pixel wide, closed, highly detailed shapes using a novel object contour and edge detection technique. It achieves a state-of-the-art Optimal Dataset Scale (ODS) score of 0.752 on a refined PASCAL-val dataset and 0.824 on the BSDS500. Traditional post-processing with Non-Maximum Suppression (NMS) results in discontinuities in the contours, but the introduced Walk the Lines (WtL) algorithm overcomes this by extracting perfectly closed, detailed contours through a shallow CNN to ‘walk’ contour changes, making this approach both unique and innovative. Evaluations show that this method achieves higher IoU peaks with very high recall compared to existing methods, demonstrating the power and novelty of this approach. The WtL has the potential to partially replace NMS in computer vision, especially where runtime is not critical, thus avoiding negative artifacts while preserving details. In addition, WtL is a promising tool for creating highly detailed segmentations, especially when every detail counts and the object shapes are very challenging, which can also be the case in interdisciplinary imaging fields such as physics, medicine, and biology.
Version
Published version
Access right on openHSU
Open access