APF-YOLOV8: Enhancing Multiscale Detection and Intra-Class Variance Handling for UAV-Based Insulator Power Line...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Electrical Engineering

Background:

• Unmanned Aerial Vehicle (UAV)-based power line inspections face challenges in insulator detection, including varying object scales and visual similarities between insulator types.
• Traditional inspection methods are often less safe and efficient compared to automated UAV approaches.

Purpose of the Study:

• To develop an advanced deep learning model for accurate and efficient insulator detection in UAV-based power line inspections.
• To introduce a comprehensive dataset for training and evaluating insulator detection models under diverse real-world conditions.

Main Methods:

• Introduced APF-YOLO, a YOLOv8-based model incorporating an Adaptive Path Fusion (APF) neck and an Adaptive Feature Alignment Module (AFAM).
• AFAM utilizes local and global pathways with attention mechanisms to balance feature extraction for objects of different scales.
• Developed the Merged Public Insulator Dataset (MPID) to address the need for diverse training data, including occlusions and scale variations.

Main Results:

• APF-YOLO demonstrated superior performance over state-of-the-art models on the MPID, achieving a +2.71% increase in mAP@0.5:0.9 and a +1.24% increase in recall.
• The model maintained real-time performance suitable for practical applications, despite slightly increased computational demands.
• Evaluations confirmed the model's effectiveness in handling challenges like occlusions and scale variations inherent in real-world power line environments.

Conclusions:

• APF-YOLO, along with the MPID, provides a robust solution for UAV-based insulator detection, enhancing power line monitoring safety and efficiency.
• The developed model represents a significant advancement in automated infrastructure inspection technology.
• Future research will focus on optimizing APF-YOLO for edge devices to further broaden its applicability.