Efficient target detection method based on wavelet transform and progressive feature pyramid network: a case study of...

Area of Science:

• Computer Vision
• Deep Learning
• Electrical Engineering

Background:

• High-voltage transmission lines are vital for economies and public welfare.
• Foreign objects pose significant risks to power system stability and safety.
• Existing object detection models struggle with small, low-contrast foreign objects in complex environments.

Purpose of the Study:

• To develop an enhanced YOLOv11 object detection model for high-voltage transmission line inspection.
• To improve the detection accuracy and robustness of foreign objects, especially small and low-contrast ones.
• To provide a practical, real-time solution for power-grid inspection scenarios.

Main Methods:

• Integrated Wavelet-Transform Convolution (WTConv) block to enhance feature decomposition and detail preservation.
• Developed a Progressive Feature Pyramid Network (PFPN) for multi-scale feature fusion and refinement.
• Introduced an Inner-EIoU loss function to focus on precise localization of small targets.

Main Results:

• The enhanced detector improved mAP₀.₅ from 0.841 to 0.872 and mAP₀.₅:₀.₉₅ from 0.620 to 0.640 on the TLFO dataset.
• Achieved higher Precision (0.962 vs 0.918), reduced parameters (4.83M vs 5.97M), and increased inference speed (28.5 FPS vs 24.1 FPS).
• Demonstrated generalizability with a +1.6 point mAP₀.₅:₀.₉₅ improvement on the MS COCO dataset.

Conclusions:

• The proposed WTConv, PFPN, and Inner-EIoU combination effectively addresses the limitations of traditional models for foreign-object detection.
• The enhanced YOLOv11 offers a practical and efficient solution for real-time power-grid inspection.
• The model's performance on MS COCO indicates its potential applicability beyond the power-grid domain.