Real-world UAV image deblurring by PSF calibration

Area of Science:

• Computer Vision
• Robotics
• Image Processing

Background:

• Unmanned aerial vehicles (UAVs) are vital for environmental monitoring, agriculture, and infrastructure inspection.
• Image blur in UAV data degrades visual quality and hinders high-level vision tasks like object detection and 3D reconstruction.
• Camera-related factors are a significant source of blur in UAV imagery.

Purpose of the Study:

• To address image blur in UAV-acquired imagery caused by camera-specific factors.
• To develop a novel deblurring method that does not require optical design specifications.
• To improve the sharpness and utility of UAV imagery for various applications.

Main Methods:

• Camera calibration was used to estimate the point spread function (PSF) across different fields of view.
• A synthetic blurred dataset was generated using the estimated PSFs.
• A deep neural network was trained on the synthetic dataset for image deblurring.
• Existing deblurring techniques were evaluated and compared against the proposed method.

Main Results:

• The proposed PSF-driven deep learning approach effectively restored sharpness in real-world UAV images.
• Experimental results demonstrated superior performance compared to several existing deblurring techniques.
• The method successfully deblurred images acquired under various camera conditions and fields of view.

Conclusions:

• The developed PSF-driven deblurring method offers a robust solution for enhancing UAV imagery.
• Accurate PSF estimation through camera calibration is crucial for effective image deblurring.
• This approach significantly improves the quality of UAV data for downstream applications.