Physical Reqularization Loss: Integrating Physical Knowledge to lmage Segmentation

Yan Ding, Shuang Li, Huafeng Li, Guanqiu Qi, Baisen Cong, Yunpeng Gong, Zhiqin Zhu

Abstract:

This paper presents a novel image segmentation optimization approach that incorporates physical constraints through
anisotropic diffusion equations during the segmentation process. It points out that most current models rely on data-driven
approaches to perform image segmentation tasks, requiring large amounts of high-quality training data. When faced with
limited datasets or significant domain shifts, these models often struggle to learn comprehensive features, affecting segmentation accuracy. Moreover, these models often require dataset-specific architectural modifications, resulting in inconsistent
cross-dataset performance. To resolve this issue, this paper proposes a novel method incorporating a new regulator derived
from anisotropic diffusion equations which is designated as the physical regulation loss (PRL) function . Through constraints
of PRL based on observable results, PRL enhances image segmentation models’ extrapolation and generalization capabilities,
improving their overall performance. The experiment results prove the theoretical feasibility of PRL, and this paper also
presents an efficient numerical algorithm for calculation. The proposed method demonstrates improvements across diverse
segmentation tasks, including urban scene analysis Cityscapes, natural image processing datasets Pascal VOC 201, and medical imaging applications such as dermatology datasets ISIC 2017 & 2018, and brain tumor segmentation datasets BraTS
2020.