Differentiable Boundary Regularization
2024 IEEE International Geoscience and Remote Sensing Symposium, IGARSS 2024, Athens, Yunanistan, 7 - 12 Temmuz 2024, ss.4662-4665, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/igarss53475.2024.10641244
- Basıldığı Şehir: Athens
- Basıldığı Ülke: Yunanistan
- Sayfa Sayıları: ss.4662-4665
- Anahtar Kelimeler: boundary, contour, curve, linearization, polygonalization, regularization, vectorization
- Ankara Yıldırım Beyazıt Üniversitesi Adresli: Evet
Özet
This study presents a new boundary regularization algorithm for linearizing closed boundaries defined as adjacent vertices. By minimizing a novel cost function consisting of a L2-norm data fidelity term and non-differentiable L0-norm boundary regularization term, given noisy boundary is linearized. Data fidelity term enforces similarity to the given boundary while L0-norm enforces linearized boundary with sharp corners. Proposed cost function is made differentiable via quadratic approximation, leading to a linear system characterized by a sparse symmetric positive definite matrix. Obtained linear system is efficiently solved using an iterative solver. Finally, applicability of the method is demonstrated with the tests on real-world and synthetic data sets. In addition to its standalone application, owing to its differentiability, the proposed method can also serve as a layer within deep neural networks, enabling the creation of loss functions tailored specifically for vector outputs which is useful in image and point cloud analysis.