Neuron-Based VLSI Architecture for Real-Time Camera Distortion Correction

This paper propose an efficient VLSI architecture of camera distortion correction based on neural camera distortion model (NCDM). Different from conventional imaging method uses over two kind models to correct the camera and lenses distortions, the NCDM uses a single model to correct at once the geometry distor-tion and the unsymmetrical manufacture errors. The NCDM with four neurons perform the wide-angle dis-tortion correction, results show that the maxima whole l corrected error in a image is less than 1.1705 pixels, and that the MSE approaches 0.1743 between the corrected and ideal results. The distortion correction by NCDM is 429x more accurate than the conventional approach. The chip size of NCDM is 1.51x1.51 mm2 that contains 126K gates by using TSMC 90 nm CMOS technology process. As working at 240Mhz, this ar-chitecture can correct 30 frames and Full-HD resolution video per second. Results show that the maximal corrected error in a whole image is less than 1.4 pixels, and that the mean square error approaches 0.0376 be-tween the corrected and ideal results.