YOLO-ISTD: An infrared small target detection method based on YOLOv5-S

Area of Science:

• Computer Vision
• Artificial Intelligence
• Infrared Imaging

Background:

• Infrared target detection is crucial for industrial applications like environmental monitoring and autonomous driving.
• Detecting small, weak targets in infrared imagery presents significant challenges due to limited information and context.

Purpose of the Study:

• To develop an improved method for infrared small target detection (ISTD).
• To enhance the accuracy and reliability of detecting weak targets in infrared images.

Main Methods:

• Proposed YOLO-ISTD, an enhanced YOLOv5-S framework incorporating a novel Shuffle Attention CSP (SACSP) feature extraction module.
• Introduced a Non-Local SPPF (NL-SPPF) module for improved long-range feature capture and correlation.
• Developed a modified K-means clustering algorithm (K-means_DIOU) for better anchor generation.
• Modified detection heads for finer-grained feature analysis.

Main Results:

• YOLO-ISTD achieved significant improvements on the NUST-SIRST dataset, with mAP@0.5 increasing by 8.568% and mAP@0.95 by 8.618%.
• The method effectively reduced missed detections and false alarms compared to existing models.
• Demonstrated substantial gains in precision, recall, and model convergence speed.

Conclusions:

• The proposed YOLO-ISTD method offers a robust solution for infrared small target detection.
• The integrated SACSP and NL-SPPF modules, along with K-means_DIOU, significantly enhance detection performance.
• This work contributes to advancing the capabilities of infrared surveillance and monitoring systems.