Fisheye Image Detection of Trees Using Improved YOLOX for Tree Height Estimation

Area of Science:

• Forestry science
• Computer vision
• Remote sensing

Background:

• Tree height is a critical forestry metric, but direct measurement is challenging due to factors like wind disturbance.
• Existing measurement techniques often lack the desired accuracy and efficiency.

Purpose of the Study:

• To develop an accurate and efficient tree height measurement method using fisheye images.
• To improve the performance of the lightweight target detection network YOLOX-tiny for forestry applications.

Main Methods:

• Proposed an enhanced YOLOX-tiny network incorporating the CBAM attention mechanism, transfer learning, and data augmentation.
• Utilized fisheye images for tree height extreme point extraction.
• Applied deep learning techniques to improve measurement accuracy and stability.

Main Results:

• The improved YOLOX-tiny network demonstrated enhanced detection performance, with significant improvements in recall rate, F1 score, and AP.
• Achieved a highest relative error of 4.06% and an average relative error of 1.62% in tree height measurements.
• The proposed method outperformed previous studies across all measurement stages.

Conclusions:

• Deep learning, specifically the enhanced YOLOX-tiny model, offers a viable solution for accurate and efficient tree height measurement from fisheye images.
• The integration of attention mechanisms and advanced training strategies significantly boosts the model's performance in forestry research.