Genaralizing Generative Models: Application to Image Super-Resolution
Yanyan Mu, Roussos G. Dimitrakopoulos, Frank P. Ferrie · 2016
Generative models such as neural networks or markov random fields are difficult to extend beyond a limited spatial extent due to the large configuration spaces involved. Their usage in applications such as super resolution is thus problematic, despite their capacity for learning rich descriptions from local receptive fields. For example, due to model complexity, the number of parameters grows exponentially with respect to the size of the receptive field. In this paper we show how to deal with this limitation using a Convolutional Deep Boltzmann Machine (ConvDBM) for modelling distributions on large receptive fields with a controllable number of parameters. In particular, we show that i) by weight sharing and joint training over the second hidden layer, the prior distribution on a large receptive field can be represented properly using a small number of parameters, ii) scaling up to high resolution images can be achieved by applying the resulting ConvDBM sequentially with tiled weights. Experimental results are presented that show successful application of this approach to the problem of super-resolution.