Improved YOLO-V3 with DenseNet for Multi-Scale Remote Sensing Target Detection.

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Remote sensing target detection is challenging due to varying object dimensions, dense distribution, and complex backgrounds.
• Existing models like YOLO-V3 may have limitations in accurately detecting remote sensing targets at different scales.

Purpose of the Study:

• To develop an improved You Only Look Once (YOLO)-V3-based model for enhanced remote sensing target detection.
• To address the limitations of standard YOLO-V3 in handling diverse target scales and complex environments.

Main Methods:

• A novel approach integrating DenseNet (Densely Connected Network) with YOLO-V3 to improve feature extraction capabilities.
• Expanding the detection scales of YOLO-V3 to four to accommodate targets of various sizes.
• Experimental validation using the Remote Sensing Object Detection (RSOD) and UCS-AOD datasets.

Main Results:

• The proposed model demonstrated superior accuracy compared to Faster-RCNN, SSD, YOLO-V3, and YOLO-V3 tiny on benchmark datasets.
• Achieved a significant increase in mean Average Precision (mAP) from 77.10% to 88.73% on the RSOD dataset.
• Showcased a notable 12.12% mAP improvement for detecting small targets, such as aircraft, without compromising detection speed.

Conclusions:

• The enhanced YOLO-V3 model effectively improves remote sensing target detection accuracy and scale adaptability.
• The integration of DenseNet and multi-scale detection offers a robust solution for complex remote sensing scenarios.
• The approach balances high accuracy with real-time performance, making it suitable for practical applications.