SED-YOLO based multi-scale attention for small object detection in remote sensing.

Area of Science:

• Computer Vision
• Remote Sensing Image Processing
• Artificial Intelligence

Background:

• Object detection in remote sensing is vital but challenged by small objects, noise, and clutter.
• Existing methods struggle with the accurate identification of diminutive targets in complex aerial imagery.

Purpose of the Study:

• To develop an improved object detection network, SED-YOLO, specifically for enhancing small object detection in remote sensing images.
• To boost the accuracy and efficiency of small object identification in challenging remote sensing scenarios.

Main Methods:

• The proposed SED-YOLO network is based on YOLOv5s, incorporating Switchable Atrous Convolution (SAC) for enhanced feature extraction.
• An Efficient Multi-Scale Attention (EMA) mechanism and an adaptive Concat method are integrated for efficient multi-scale learning and feature fusion.
• The detection head is expanded to four scales, including a small object layer and the Dynamic Head (DyHead) module for adaptive attention.

Main Results:

• SED-YOLO achieved a mean Average Precision (mAP) of 71.6% on the DOTA dataset.
• This represents a 2.4% improvement over the original YOLOv5s model.
• The network demonstrated superior performance in detecting small objects within cluttered remote sensing imagery.

Conclusions:

• The SED-YOLO network effectively addresses the challenges of small object detection in remote sensing.
• The integration of SAC, EMA, adaptive Concat, and DyHead significantly enhances detection accuracy and adaptability.
• SED-YOLO offers a promising solution for precise small object identification in complex remote sensing applications.