A small target detection algorithm based on improved YOLOv5 in aerial image.

Area of Science:

• Computer Vision
• Remote Sensing Technology
• Artificial Intelligence

Background:

• Uncrewed aerial vehicle (UAV) aerial photography faces challenges in small target detection due to target size and complex environmental conditions.
• Accurate and rapid identification of target categories in remote sensing imagery is crucial but difficult.

Purpose of the Study:

• To propose an improved remote sensing target detection algorithm based on the YOLOv5 architecture.
• To enhance detection speed and accuracy for small targets in complex aerial environments.

Main Methods:

• Introduced Distribution Focal Loss function to accelerate network convergence and improve focus on annotated data within the YOLOv5s model.
• Modified the Cross Stage Partial (CSP) network structure, including adjusting convolution kernel size and adding a stack-separated convolution module.
• Designed a new attention mechanism for effective feature fusion across different hierarchical feature maps.

Main Results:

• The proposed algorithm achieved a 3.5% increase in detection accuracy on the Remote Sensing Object Detection (RSOD) dataset compared to the original YOLOv5 algorithm.
• Demonstrated significant performance improvements in precision for remote sensing target detection.

Conclusions:

• The improved YOLOv5 algorithm effectively enhances the precision of remote sensing small target detection.
• The algorithm shows potential for practical applications in industrial and military sectors requiring accurate aerial surveillance.