Image super-resolution performance of multilayer feedforward neural networks

Abstract

Super-resolution is the process by which the bandwidth of a diffraction-limited spectrum is extended beyond the optical passband. Many algorithms exist which are capable of super-resolution; however most are iterative methods, which are ill-suited for real-time operation. One approach that has been virtually ignored in super-resolution research is the neural network approach. The Hopfield network has been a popular choice in image restoration applications, however it is also an iterative approach. We consider the feedforward architecture known as a Multilayer Perceptron (MLP), and present results on simulated binary and greyscale images blurred by a diffraction-limited OTF and sampled at the Nyquist rate. To avoid aliasing, the network performs as a nonlinear spatial interpolator while simultaneously extrapolating in the frequency domain. Additionally, a novel use of vector quantization for the generation of training data sets is presented. This is accomplished by training a nonlinear vector quantizer (NLIVQ), whose codebooks are subsequently used in the supervised training of the MLP network using Back-Propagation. The network shows good regularization in the presence of noise.