Multi-Scale DenseNets-Based Aircraft Detection from Remote Sensing Images.

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Deep learning methods for aircraft detection face challenges due to diverse imaging modes causing variations in aircraft size and view.
• Standard deep convolutional neural networks (DCNNs) can lose crucial location information and submerge small target features.

Purpose of the Study:

• To propose a compact multi-scale dense convolutional neural network (MS-DenseNet) for enhanced aircraft detection in remote sensing.
• To improve the detection of small aircraft and address feature loss in deep learning models.

Main Methods:

• Utilized DenseNet for feature extraction to enhance propagation and reuse of high-resolution, bottom-level features.
• Integrated Feature Pyramid Network (FPN) with DenseNet to create MS-DenseNet for multi-scale feature learning, focusing on small objects.
• Developed compact architectures (MS-DenseNet-41, -65, -77) by compressing convolutional layers.

Main Results:

• The compact MS-DenseNet-65 demonstrated significant improvements in detecting small aircraft.
• Achieved state-of-the-art performance with 94% recall and 92.7% F1-score.
• MS-DenseNet-65 required less computational time compared to other methods.

Conclusions:

• The proposed MS-DenseNet architectures, particularly MS-DenseNet-65, offer superior performance for aircraft detection in remote sensing images.
• The method shows good transferability across different datasets (UCAS-AOD, RSOD), indicating robustness.
• MS-DenseNet effectively addresses the challenges of multi-resolution imaging and small object detection.