DoG-Loss for Semantic Segmentation

Ali Solak, Cihan Topal

Abstract: Semantic segmentation is an important machine vision problem with many applications. It aims to classify images based on pixels and label each pixel. One of the main challenges of this problem is to ensure that the contours of the objects are accurate and the areas they cover are detected in a holistic manner. In addition, the successful learning of low-frequency classes in the datasets by the model and the preservation of object integrity also significantly affect the success. In this study, a difference of Gaussian (DoG) based loss function is proposed to improve segmentation accuracy and class estimation. In this way, the segmentation model focuses on the contours of the objects to better preserve their shape integrity. Experiments show that the proposed DoG loss function achieves up to %3.9 better results than the commonly used segmentation loss functions.

External IDs:dblp:conf/siu/SolakT23a