Correcting image distortion with deep learning and classical calibration parameters

Image distortion is a fundamental imperfection found in all imaging systems, often manifesting as the bending of straight lines in photographs. Traditional correction methods, such as camera calibration using known grid patterns and distortion coefficients, are effective but limited in adaptability and generalization. This paper explores a hybrid approach that leverages both classical calibration and machine learning to dynamically predict and correct lens distortion. By training a convolutional neural network on a large dataset of synthetically distorted grid images, the model learns to accurately estimate the inverse distortion coefficients in the Brown–Conrady model (k 1 , p 1 , p 2 ) required to restore images to their undistorted form using OpenCV's standard tools. The proposed method achieves a low mean absolute error, validating its potential for effective distortion correction. This novel pipeline bridges classical and learning-based approaches, offering a promising foundation for future development.