Salient Object Detection based on Bayesian Surprise of Restricted Boltzmann Machine

Abstract

This article presents an algorithm for salient object detection by leveraging the Bayesian surprise of the Restricted Boltzmann Machine (RBM). Here an RBM is trained on patches sampled randomly from the input image. Due to this random sampling, the RBM is likely to get more exposed to background patches than that of the object. Thus, the trained RBM will minimize the free energy of its hidden states with respect to the background patches as opposed to the object. This, according to the free energy principle, implies minimizing Bayesian surprise which is a measure for saliency based on Kullback Leibler divergence between the input and reconstructed patch distribution. Hence, when the trained RBM is exposed to patches from the object region, it would have high divergence and in turn a high Bayesian surprise. Thus such pixels with high Bayesian surprise could be considered as salient pixels. For each pixel, a neighborhood (with the same size of training patch) is considered and is fed to the trained RBM to obtain the reconstructed patch. Thereafter, the Kullback Leibler divergence between the input and reconstructed neighborhood of each pixel is computed to measure the Bayesian surprise and is stored in the corresponding position in a matrix to form the saliency map. Experiments are carried out on three datasets namely MSRA-10K, ECSSD and DUTS. The results obtained depict promising performance by the proposed approach.