Deep Rotation Correction Without Angle Prior.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Image tilt is a common issue due to imperfect capture conditions.
• Manual correction is time-consuming and requires user expertise.
• Existing methods often require knowing the rotation angle.

Purpose of the Study:

• To develop an automatic Rotation Correction method for images with unknown tilt angles.
• To achieve high content fidelity during automatic image straightening.
• To create a robust and easily integrable solution for image editing software.

Main Methods:

• A neural network predicts optical flows to warp tilted images.
• A robust elastic warp strategy is employed, combining mesh deformation and residual optical flows.
• Pixel-wise optical flow estimation is enhanced for stability, especially with large tilt angles.

Main Results:

• The proposed method successfully corrects image tilt with high fidelity.
• The algorithm outperforms state-of-the-art methods that require angle priors.
• A comprehensive dataset for rotation correction evaluation was created.

Conclusions:

• The developed Rotation Correction technique offers an effective, automated solution for straightening tilted images.
• The method's robustness and performance, even without angle information, make it valuable for practical applications.
• The released code and dataset facilitate further research and development in image correction.