Detection of plane in remote sensing images using super-resolution

Area of Science:

• Computer Vision
• Remote Sensing Technology
• Artificial Intelligence

Background:

• Object detection in remote sensing images faces challenges like small objects, noise, and cloud cover, leading to low accuracy and high error rates.
• Existing models struggle with hyper-parameter sensitivity and modal collapse, hindering reliable detection.
• The Feature Pyramid Network (FPN) in YOLOv3 has limitations in effectively integrating features from different layers.

Purpose of the Study:

• To propose a novel object detection model, SR-YOLO, for enhanced performance in remote sensing image analysis.
• To address the limitations of SRGAN, such as hyper-parameter sensitivity and modal collapse.
• To improve the feature fusion capabilities within the YOLOv3 architecture.

Main Methods:

• Developed SR-YOLO by combining Super-Resolution Generative Adversarial Network (SRGAN) with YOLOv3.
• Replaced the FPN network in YOLOv3 with the Path Aggregation Network (PANet) to shorten the distance between feature layers.
• Utilized an enhanced path in PANet to enrich layer characteristics, improving model robustness and detection capabilities.

Main Results:

• SR-YOLO demonstrated excellent performance on the UCAS-High Resolution Aerial Object Detection Dataset.
• Achieved a significant increase in average precision (AP) from 92.35% to 96.13% compared to YOLOv3.
• Reduced the log-average miss rate (MR-2) from 22% to 14% and increased the Recall rate from 91.36% to 95.12%.

Conclusions:

• SR-YOLO offers a robust and highly effective solution for object detection in challenging remote sensing imagery.
• The integration of SRGAN and PANet within the YOLOv3 framework significantly boosts detection accuracy and reliability.
• The proposed model shows substantial improvements over standard YOLOv3, particularly for datasets with small objects and occlusions.