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

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Accurate power line (PL) segmentation from aerial images is vital for safe unmanned aerial vehicle (UAV) navigation.
• Deep learning models struggle with significant class imbalance, where PLs constitute only 1-5% of aerial imagery.
• Existing solutions like class balanced cross entropy (BBCE) loss require complex tuning and yield suboptimal performance.

Purpose of the Study:

• To develop a generalized focal loss function based on the Matthews correlation coefficient (MCC) to address class imbalance in PL segmentation.
• To improve a generic deep segmentation architecture (U-Net) with an auxiliary classifier head (ACU-Net) for enhanced learning and convergence.
• To achieve an optimal trade-off between segmentation accuracy, precision, and recall without extensive hyperparameter tuning.

Main Methods:

• Proposed a novel generalized focal loss function utilizing the MCC for class imbalance in PL segmentation.
• Implemented an enhanced U-Net model (ACU-Net) with an auxiliary convolutional classifier head.
• Evaluated the proposed loss function and ACU-Net on two public PL datasets (Mendeley and PLDU) with minimal PL representation.

Main Results:

• The proposed focal loss function significantly outperformed BBCE loss, achieving 16% higher PL dice scores and improved precision/false detection rate (FDR) by 15-19% on both datasets.
• ACU-Net demonstrated superior performance over the baseline U-Net, with improvements in key evaluation parameters ranging from 1-10%.
• The approach achieved a balanced trade-off across evaluation metrics, including dice scores, accuracy, and precision-recall values.

Conclusions:

• The generalized focal loss function effectively tackles class imbalance in aerial PL segmentation using generic deep learning architectures.
• The ACU-Net model provides enhanced learning and faster convergence, leading to improved segmentation performance.
• This work offers a robust and efficient solution for critical PL segmentation tasks in UAV applications.