A YOLOv7-based method for crown extraction from UAV imagery.

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Multispectral crown extraction faces challenges like complex background interference, poor feature utilization, and model redundancy.
• Existing methods struggle to effectively integrate spectral information and capture fine details for accurate segmentation.

Purpose of the Study:

• To propose an improved YOLOv7 architecture, Dual-YOLOv7, for enhanced multispectral crown detection and segmentation.
• To address limitations in feature extraction, detail capture, and bounding box regression accuracy.

Main Methods:

• A dual-branch feature extraction network integrating visible and infrared spectral data with an attention mechanism.
• Introduction of the D-SimSPPF module using depthwise separable convolution for optimized spatial pyramid pooling.
• Development of the CIoU-C loss function with a shape penalty for improved bounding box regression.

Main Results:

• The Dual-YOLOv7 model achieved detection and segmentation mAP50 scores of 91.6% and 90.1%, respectively.
• Significant improvements of 7.7% and 7.6% in detection and segmentation mAP50 over YOLOv7-seg.
• A 14.2% reduction in model parameter count after channel pruning, creating a lightweight solution.

Conclusions:

• The proposed Dual-YOLOv7 effectively overcomes challenges in multispectral crown extraction.
• The model demonstrates superior performance in accuracy and efficiency, suitable for UAV-based applications.
• The lightweight design makes it practical for real-time processing on resource-constrained platforms.