Automatic Chromatic Aberration Correction Using Convolutional Neural Networks with a U-Net Architecture

Abstract

Chromatic aberration is a common optical defect in simple lens systems that degrades image quality by causing unwanted color fringes and reduced sharpness. In this paper, we propose a method based on a convolutional neural network with a U-Net architecture to automatically correct chromatic aberration. We generated a synthetic dataset of 80,000 images from Google Open Images by simulating aberration using perchannel Gaussian blur and lateral chromatic shift. Trained for 100 epochs with a combined loss function (MSE + 0.1.L1), the network achieved a PSNR of 30.4 dB and an SSIM of 0.9418 on the test set. The results demonstrate the feasibility of this approach for restoring degraded images.