Improved YOLOv10: A Real-Time Object Detection Approach in Complex Environments.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Object detection of small and occluded targets is crucial for intelligent systems but remains challenging.
• Existing models often struggle with complex scenarios, limiting their real-world applicability.

Purpose of the Study:

• To improve the YOLOv10 algorithm for enhanced detection of small and occluded objects.
• To introduce a generalizable optimization framework for YOLO-series models.

Main Methods:

• Implemented Mosaic-9 data augmentation to increase small target density.
• Replaced PANet with Bidirectional Feature Pyramid Network (BiFPN) for optimized feature fusion.
• Integrated Squeeze-and-Excitation (SE) channel attention into the CSPDarknet backbone.

Main Results:

• Achieved 69.5% mAP@0.5, a 7.7% increase over YOLOv10n, with 12.1 ms inference speed.
• Mosaic-9 improved small target perception, BiFPN boosted mAP@0.5 by 5.7%, and SE enhanced occlusion robustness by 4.8%.
• Experiments conducted on a self-constructed dataset of 6508 images.

Conclusions:

• The proposed multi-module optimization framework significantly advances lightweight object detection.
• The enhanced YOLOv10 algorithm demonstrates superior performance in complex scenarios with small and occluded targets.
• This work contributes to the development of more robust and efficient intelligent vision systems.