A system is presented which combines deep neural networks with discrete inference techniques for the successful recognition of an image. The system presented builds upon the classical sliding window method but applied in parallel over an entire input image. The result is discretized by treating each classified window as a node in a markov random field and applying a minimization of its associated energy levels. Two important benefits are observed with this system: a gain in performance by virtue of the system’s parallel nature, and an improvement in the localization precision due to the inherent connectivity between classified windows.