Blind First-Order Perspective Distortion Correction Using Parallel Convolutional Neural Networks.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Perspective distortion is a common image artifact.
• Existing methods like generative adversarial networks (GANs) and encoder-decoder networks have limitations in correcting perspective distortion and preserving image scale.

Purpose of the Study:

• To develop a novel network architecture for effective removal of perspective distortion in images.
• To improve upon existing methods by recovering the original scale and proportion of images.

Main Methods:

• A network architecture utilizing three parallel convolutional neural networks (CNNs) was designed.
• Each CNN predicts specific elements of the 3x3 transformation matrix (M).
• The corrected image is generated by applying the inverse transformation (M^-1) to the distorted input image. The training dataset was generated using KITTI images.

Main Results:

• The proposed parallel CNN method demonstrates significant promise in correcting perspective distortions.
• Experimental results show that the method outperforms current state-of-the-art techniques.
• The approach successfully recovers the intended scale and proportion of the image, a capability lacking in other methods.

Conclusions:

• The parallel CNN architecture offers an effective solution for perspective distortion removal.
• This method represents an advancement in image correction by preserving image scale and proportion.
• The approach shows potential for various computer vision applications requiring accurate image geometry.