Fusing Self-Attention and CoordConv to Improve the YOLOv5s Algorithm for Infrared Weak Target Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) excel in target detection but struggle with small targets and complex backgrounds.
• Infrared weak target detection in challenging environments remains a significant hurdle for existing algorithms.
• Real-time detection requirements add complexity to improving accuracy for these scenarios.

Purpose of the Study:

• To enhance the accuracy of infrared weak target detection in complex scenes.
• To address the limitations of standard CNNs in identifying small targets with low contrast.
• To develop a real-time detection solution that maintains high performance.

Main Methods:

• Modified the YOLOv5s algorithm by integrating the Bottleneck Transformer structure and CoordConv.
• Introduced a novel loss function utilizing a 2D Gaussian distribution and Normalized Gaussian Wasserstein Distance (NWD).
• Evaluated the improved model against mainstream detection algorithms on relevant datasets.

Main Results:

• The enhanced YOLOv5s model achieved a mean Average Precision (mAP) of 96.7%.
• Demonstrated a 2.2 percentage point improvement in mAP compared to the original YOLOv5s.
• Showcased superior performance in detecting weak targets within complex backgrounds.

Conclusions:

• The proposed improvements significantly enhance infrared weak target detection accuracy.
• The integration of Bottleneck Transformer, CoordConv, and the NWD loss function effectively addresses detection challenges.
• The refined algorithm offers a promising solution for real-time, high-accuracy weak target detection.