Detection of Specific Building in Remote Sensing Images Using a Novel YOLO-S-CIOU Model. Case: Gas Station...

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Traditional building classification from remote sensing is challenging due to complex environmental landscapes.
• Convolutional Neural Networks (CNNs) show promise in extracting spatial context for image processing tasks.

Purpose of the Study:

• To develop a novel CNN model, YOLO-S-CIOU, for improved specific building detection in remote sensing imagery.
• To enhance feature learning and bounding box regression for more accurate building identification.

Main Methods:

• Developed YOLO-S-CIOU, a novel CNN model based on YOLOv3.
• Replaced YOLOv3's Darknet53 module with SRXnet (superimposed SE-ResNeXt) for enhanced feature learning.
• Integrated Complete-IoU Loss (CIoU Loss) for improved bounding box regression.

Main Results:

• YOLO-S-CIOU achieved 97.62% average precision (AP) and 97.50% F1 score on a gas station dataset.
• The model reduced parameters by approximately 4% compared to YOLOv3 while improving AP by 2.23% and F1 score by 0.5%.
• Demonstrated significantly higher recall (50%) and precision (40%) than YOLOv3 in real-world gas station detection scenarios.

Conclusions:

• The proposed YOLO-S-CIOU model exhibits superior feature learning, robustness, and detection capabilities for specific buildings in remote sensing images.
• This advanced CNN architecture offers a more effective solution for smart city planning, management, and potentially military applications requiring precise building identification.