| dc.description.abstract | This study proposes a rapid neural network-based camera distortion correction
(NCDC), based on a lightweight neural network, to accurately correct the distortion of
low-cost cameras. The NCDC is different from general camera distortion models
because it uses a neural network to simultaneously model numerous camera
distortions, including multi-lens and wide-angle distortion, as well as various
manufacturing flaws, in a low-cost camera. The proposed NCDC uses a neural
network with an error backpropagation training algorithm to map the complex
distortion surface. The optimal number of neurons was assigned as 4 to associate the
mapping model between the distortion image space and the correction image space. In
offline calibration processing, the NCDC calculates the distortion vector from a single
captured calibration image without requiring an estimation of the optical center. Two
different wide-angle lenses use the proposed neuron-base method to correct the
distortion. Results show that the maximal corrected error in a whole image is less than
2 pixels with 120° wide-angle lens, and that the mean square error (MSE) approaches
0.2050 between the corrected and ideal results. The NCDC is 429x more accurate
than the traditional polynomial method.
Simultaneously, this study proposes a VLSI architecture bases on the NCDC, and
built a verification system of FPGA that consists of microprocessor, bus, memory,
GbE and wide-angle camera to really capture a distortion image, and correct the
camera distortions. The neural network-based corrector can correct an over
1920x1080 resolution image from the low-price 120° wide-angle camera. To obtain a
more accurate correction, the neural network-based correction method was
implemented using a 24-bit fixed point, after which the difference in error between
the floating point in the software and the fixed point in the hardware was under 10
-6
.
The NCDC chip is sized 1.51 1.51 mm
2
, and contains 126K gates built using TSMC
90 nm CMOS technology. Working at 240 Mhz, this architecture can correct 30
frames of 4K2K-resolution video per second. | en_US |