SEB-YOLO: An Improved YOLOv5 Model for Remote Sensing Small Target Detection

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Remote sensing target detection faces challenges with small objects and similar targets, leading to poor performance.
• Existing algorithms struggle with semantic information extraction and feature loss in complex remote sensing imagery.

Purpose of the Study:

• To propose an improved YOLOv5 algorithm (SEB-YOLO) for enhanced remote sensing image target detection.
• To address limitations in detecting small objects and distinguishing similar targets in remote sensing data.

Main Methods:

• Reconstructed the backbone network using Space-to-Depth (SPD) and convolution (Conv) layers (SPD-Conv) to preserve global features.
• Incorporated an attention mechanism in the backbone's final pooling layer for improved target identification and localization.
• Utilized a bidirectional feature pyramid network (Bi-FPN) with bilinear interpolation for enhanced feature fusion and cross-scale connections.
• Introduced a decoupled head to improve model convergence and resolve classification-regression task conflicts.

Main Results:

• SEB-YOLO achieved a mean Average Precision (mAP) of 93.5% on the NWPU VHR-10 dataset.
• SEB-YOLO achieved a mAP of 93.9% on the RSOD dataset.
• Demonstrated significant improvements of 4.0% and 5.3% over the original YOLOv5l algorithm on the respective datasets.

Conclusions:

• The proposed SEB-YOLO algorithm significantly improves target detection performance in complex remote sensing images.
• The integration of SPD-Conv, attention mechanisms, Bi-FPN, and a decoupled head enhances the model's ability to detect small and similar objects.
• SEB-YOLO offers a more robust solution for remote sensing image analysis and target identification.