City Research Online

A Novel Focal Phi Loss for Power Line Segmentation with Auxiliary Classifier U-Net

Jaffari, R., Hashmani, M. and Reyes-Aldasoro, C. C. ORCID: 0000-0002-9466-2018 (2021). A Novel Focal Phi Loss for Power Line Segmentation with Auxiliary Classifier U-Net. Sensors, 21(8), 2803.. doi: 10.3390/s21082803

Abstract

The segmentation of power lines (PLs) from aerial images is a crucial task for the safe navigation of unmanned aerial vehicles (UAVs) operating at low altitudes. Despite the advances in deep learning-based approaches for PL segmentation, these models are still vulnerable to the class imbalance present in the data. The PLs occupy only a minimal portion (1-5%) of the aerial images as compared to the background region (95-99%). Generally, this class imbalance problem is addressed via the use of PL-specific detectors in conjunction with the popular class balanced cross entropy (BBCE) loss function. However, these PL-specific detectors do not work outside their application areas and a BBCE loss requires hyperparameter tuning for class-wise weights, which is not trivial. Moreover, the BBCE loss results in low dice scores and precision values and thus, fails to achieve an optimal trade-off between dice scores, model accuracy, and precision-recall values. In this work, we propose a generalized focal loss function based on the Matthews correlation coefficient (MCC) or the Phi coefficient to address the class imbalance problem in PL segmentation while utilizing a generic deep segmentation architecture. We evaluate our loss function by improving the vanilla U-Net model with an additional convolutional auxiliary classifier head (ACU-Net) for better learning and faster model convergence. The evaluation of two PL datasets, namely the Mendeley Power Line Dataset and the Power Line Dataset of Urban Scenes (PLDU), where PLs occupy around 1% and 2% of the aerial images area, respectively, reveal that our proposed loss function outperforms the popular BBCE loss by 16% in PL dice scores on both the datasets, 19% in precision and false detection rate (FDR) values for the Mendeley PL dataset and 15% in precision and FDR values for the PLDU with a minor degradation in the accuracy and recall values. Moreover, our proposed ACU-Net outperforms the baseline vanilla U-Net for the characteristic evaluation parameters in the range of 1-10% for both the PL datasets. Thus, our proposed loss function with ACU-Net achieves an optimal trade-off for the characteristic evaluation parameters without any bells and whistles. Our code is available at Github.

Publication Type: Article
Additional Information: © 2021 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/).
Publisher Keywords: power lines; semantic segmentation; Matthews correlation coefficient; loss function; data imbalance
Subjects: Q Science > QA Mathematics > QA75 Electronic computers. Computer science
T Technology > TK Electrical engineering. Electronics Nuclear engineering
Departments: School of Mathematics, Computer Science & Engineering > Computer Science > giCentre
Date available in CRO: 12 May 2021 12:16
Date deposited: 12 May 2021
Date of acceptance: 29 March 2021
Date of first online publication: 16 April 2021
URI: https://openaccess.city.ac.uk/id/eprint/26126
[img]
Preview
Text - Published Version
Available under License Creative Commons: Attribution International Public License 4.0.

Download (4MB) | Preview

Export

Downloads

Downloads per month over past year

View more statistics

Actions (login required)

Admin Login Admin Login