DCN-YOLO: A Small-Object Detection Paradigm for Remote Sensing Imagery Leveraging Dilated Convolutional Networks

Area of Science:

• Computer Vision
• Remote Sensing Technology
• Artificial Intelligence

Background:

• Optical remote sensing images are crucial for military reconnaissance, environmental monitoring, and urban planning.
• Conventional convolutional methods struggle to extract features from small objects due to low pixel counts, fuzzy features, and complex backgrounds.

Purpose of the Study:

• To enhance the detection accuracy and robustness of small objects in remote sensing images.
• To address the limitations of conventional convolutions in feature extraction for small objects.

Main Methods:

• Proposed a novel object detection method, DCN-YOLO, incorporating multi-scale dilated convolutions.
• Introduced a Dilated Convolutional Residual (DCR) module for high-level feature extraction.
• Developed a context aggregation (CONTEXT) module for global semantic understanding using remote interaction.

Main Results:

• Achieved an AP50 of 56.6 on the AI-TOD dataset.
• Demonstrated significant improvement in detecting small objects in remote sensing images.
• Enhanced model robustness for small object detection tasks.

Conclusions:

• The DCN-YOLO method effectively improves small object detection in remote sensing.
• Multi-scale dilated convolutions and contextual aggregation are key to enhancing feature extraction and semantic understanding.
• This work offers a new technical approach for small object detection in remote sensing applications.