Wide-angle camera distortion correction using neural back mapping

This study proposes an efficient back mapping model that uses a lightweight neural network and virtual calibration plate to accurately correct the distortion of low quality wide angle camera. Unlike the radial model, the neural-based method uses non-linear functional mapping to model surface distortion, which consists of wide-angle distortion and various manufacturing errors in low-cost cameras. The proposed approach uses a lightweight multilayer feed-forward neural network (MFFNN) with error back-propagation training algorithm to map the complex distortion surface. The optimal number of neurons of hidden layer was assigned as 4 for associating the mapping model between the distortion image space (DIS) with the correction image space (CIS). This study uses a 105 degree wide-angle low-cost camera to test the proposed method. Results show that the maximal corrected error in a whole image is less than 2 pixels, and that the mean square error (MSE) approaches 0.2542 between the corrected and ideal results.