YOLO-FR: A YOLOv5 Infrared Small Target Detection Algorithm Based on Feature Reassembly Sampling Method.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Infrared Imaging

Background:

• Dim-small target detection in infrared imagery is crucial but challenging due to feature loss during network sampling.
• Existing models often struggle with preserving spatial and semantic information, leading to reduced detection accuracy.

Purpose of the Study:

• To propose YOLO-FR, a novel YOLOv5-based model designed to mitigate feature loss in infrared dim-small target detection.
• To enhance feature representation and fusion strategies for improved detection performance.

Main Methods:

• Developed a feature reassembly sampling approach to scale feature maps without altering feature information content.
• Introduced an STD Block to minimize feature loss during down-sampling by encoding spatial information into channels.
• Utilized the CARAFE operator to upscale feature maps while preserving feature mapping characteristics.
• Improved the neck network to fuse backbone-extracted features with semantic information for a small receptive field detection head.

Main Results:

• The YOLO-FR model achieved a mean Average Precision (mAP50) of 97.4%, a 7.4% improvement over the original network.
• Demonstrated superior performance compared to existing methods like J-MSF and YOLO-SASE.
• Effectively reduced feature loss during the sampling process, enhancing detection capabilities.

Conclusions:

• The proposed YOLO-FR model significantly enhances infrared dim-small target detection accuracy by addressing feature loss issues.
• Feature reassembly sampling, STD Block, CARAFE operator, and improved neck network fusion are effective strategies for this task.
• YOLO-FR represents a substantial advancement in the field of infrared object detection.