Low resolution remote sensing object detection with fine grained enhancement and swin transformer

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Object detection in remote sensing images is challenging due to small, dense targets and scale variations.
• Existing algorithms struggle with fine-grained details and multi-scale object representation in these scenarios.
• Low-resolution conditions further exacerbate detection difficulties.

Purpose of the Study:

• To develop a novel object detection framework addressing limitations in remote sensing image analysis.
• To enhance the detection of small, densely packed, and multi-scale objects.
• To improve overall detection accuracy and efficiency, particularly under low-resolution conditions.

Main Methods:

• Introduced the Fine-grained Enhanced Downsampling Network (FEDNet) for preserving target information during feature extraction.
• Developed the Swin Transformer-based Progressive Aggregation Network (STPANet) for advanced multi-scale feature fusion and contextual understanding.
• Incorporated the Shape-IoU loss function to optimize bounding box regression for improved small target accuracy.

Main Results:

• Achieved state-of-the-art performance on the DOTA dataset with a mean average precision (mAP@50) of 69.9%.
• Demonstrated superior results on the DIOR dataset, reaching an mAP@50 of 85.5%.
• The proposed method shows significant improvements in detecting small targets and handling low-resolution imagery.

Conclusions:

• The novel detection framework effectively overcomes key challenges in remote sensing object detection.
• FEDNet and STPANet integration provides robust feature representation and multi-scale fusion capabilities.
• The Shape-IoU loss function enhances bounding box regression accuracy, particularly for small objects.