Authors - Redwan Hasif Alvi, Rashedur M. Rahman Abstract - Image segmentation is an actively research field in the domain of deep learning and computer vision. Due to its use in a wide range of applications such as medical imaging problems, automated disease detection, self-driving vehicles and so on, image segmentation is widely used in dealing with building extraction tasks. Although deep learning models such as U-Net performs well with building extraction tasks that require semantic segmentation, presence of noisy data can still affect the performance of these models. Moreover, it is not unusual to have noisy labels present in datasets that are available online or collected automatically via the internet. As such, we commonly come across label noise present in datasets used for building extraction tasks that can decrease the performance of a learning task. In this paper, we implement multiple well-known loss functions used for image segmentation on an existing U-Net model for learning building extraction, and compare the results of each of the loss functions’ performance. Furthermore, we implement some of the widely used loss functions such as cross entropy (CE), Dice loss (DL) and Tversky loss (TL) on a noise-robust interme-diate false positive-false negative matrix model to evaluate the performances of these loss functions when dealing with noisy data.
Open Zoom