On Biologically Inspired Stochastic Reinforcement Deep Learning: A Case Study on Visual Surveillance

Abstract

Here, we present a biologically inspired visual network (BIVnet) for image processing tasks. The proposed model possesses similarities with its neural counterpart and is trained by a stochastic algorithm which employs a partially observable Markov decision process to execute a reinforcement learning strategy. The network was tested on a collection of available datasets in surveillance-related tasks and showed superior performance compared with the state-of-the-art architectures. An average improvement of 15.2% in accuracy on a collection of publicly available image datasets is shown in our experimental results. © 2013 IEEE.