Swin-HSTPS: Research on Target Detection Algorithms for Multi-Source High-Resolution Remote Sensing Images

Area of Science:

• Remote Sensing
• Computer Vision
• Deep Learning

Background:

• Accurate detection of traffic port stations (airports, ports) in high-resolution remote sensing images is challenging due to complex multi-target scenarios.
• Current deep learning methods, while advanced for single-target detection, struggle with recognizing intricate multi-target complexes in remote sensing data.
• High-precision positioning requires comprehensive analysis of feature information from multiple small targets within these stations.

Purpose of the Study:

• To develop a novel detection model, Swin-HSTPS, for accurately identifying traffic port stations in high-resolution remote sensing images.
• To enhance the recognition accuracy of multi-target complexes within these images.
• To achieve high-precision positioning of traffic port stations by analyzing combined feature information of multiple small targets.

Main Methods:

• Constructed a novel Swin-HSTPS model for traffic port station detection.
• Integrated the MixUp hybrid enhancement algorithm for improved image feature information during preprocessing.
• Incorporated the PReLU activation function into the Swin Transformer's forward network, creating a ResNet-like residual network to strengthen pixel block information interaction.

Main Results:

• The Swin-HSTPS model achieved an optimal average precision of 85.3%, an improvement of approximately 8% over the standard Swin Transformer detection model.
• Demonstrated superior target prediction accuracy compared to the Swin Transformer model.
• Experimental results confirmed the model's ability to accurately locate traffic port stations like airports and ports.

Conclusions:

• The Swin-HSTPS model effectively addresses the challenge of multi-target complex recognition in high-resolution remote sensing images.
• The model offers superior target prediction capabilities for traffic port stations compared to mainstream models like R-CNN and YOLOv5.
• Swin-HSTPS provides a robust solution for high-precision positioning of traffic port stations in remote sensing applications.