Small target detection with remote sensing images based on an improved YOLOv5 algorithm.

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Small target detection in remote sensing is difficult due to target size and background complexity.
• Existing algorithms struggle with the subtle features of small objects in diverse environments.

Purpose of the Study:

• To develop a robust algorithm for accurate small target detection in remote sensing imagery.
• To enhance the feature extraction and fusion capabilities for improved performance.

Main Methods:

• A modified YOLOv5s architecture incorporating a novel lightweight and stable feature extraction module (LCB).
• Data augmentation using mosaic operations to diversify training datasets.
• Multi-scale feature fusion using Res2 unit, Dres2, and Spatial Pyramid Pooling Small (SPPS).
• Efficient Intersection over Union (EIoU) loss function for improved convergence and regression accuracy.

Main Results:

• The proposed method demonstrates effectiveness in small target detection tasks.
• Ablation and comparative experiments validate the algorithm's performance against established methods.
• The enhanced feature extraction and fusion strategies contribute to improved detection accuracy.

Conclusions:

• The developed algorithm offers a significant advancement in small target detection for remote sensing.
• The integration of specialized modules and loss functions enhances model robustness and accuracy.
• This work provides a valuable tool for analyzing remote sensing data with small targets.