A Metric Learning-Based Improved Oriented R-CNN for Wildfire Detection in Power Transmission Corridors

Area of Science:

• Computer Vision
• Artificial Intelligence
• Electrical Engineering

Background:

• Wildfire detection in power transmission corridors is critical for grid stability and safety.
• Challenges include distinguishing smoke from background clutter, diverse target appearances, and detecting small smoke/flame objects.
• Existing methods struggle with accuracy in complex environments.

Purpose of the Study:

• To develop an advanced object detection model for accurate wildfire detection in power transmission corridors.
• To improve the recognition of small-scale smoke and flame targets.
• To enhance the overall stability and safety of power transmission infrastructure.

Main Methods:

• An improved Oriented R-CNN model incorporating metric learning was proposed.
• A multi-center metric loss (MCM-Loss) module was introduced to improve feature differentiation.
• ResNeXt and FPN-CARAFE modules were integrated to enhance feature extraction and multi-scale representation.

Main Results:

• The MCM-Loss module improved smoke target average precision (AP) by 2.7%.
• Replacing ResNet with ResNeXt increased mean average precision (mAP) by 0.6%.
• The FPN-CARAFE module boosted fire target AP by 8.1%, achieving a final mAP of 90.4% (6.4% improvement).

Conclusions:

• The proposed model demonstrates superior performance for wildfire detection in power transmission corridors.
• The integration of metric learning and advanced network modules effectively addresses detection challenges.
• This research provides valuable support for wildfire monitoring and power grid safety.