The Saliency Estimation Based On The Objectness

Human eyes can quickly focus on interest targets in a complex scene with the help of the visual system.Simulation of human visual system to obtain a salient region in the image,which is saliency detection,has become one of the hot field in computer vision.Lots of studies have shown that the saliency is governed by two different mechanisms: a bottom-up sensory-driven mechanism and a top-down semantic-driven mechanism.At present,the bottom-up saliency is nearly mature,the research hot spot is moving to the top-down semantic-driven method.One of the challenge in this direction is how to obtain high-level semantic information and how to use it in saliency estimation.In this paper,we propose a method of saliency detection based on the objectness in an image.Objectness is an abstract high-level semantic information,which reflects the distribution of all objects in the image.The objectness can guide the saliency distribution from semantics,thereby improving the detection results.In this paper,the framework of saliency estimation is based on the autoencoder-based saliency estimation,objectness is then introduced to improve the estimation of saliency.Saliency estimation is obtained by training a deep stacked autoencoder network to represent the input image.The trained network can inference the central patch from its surrounding patch for each pixel.The reconstruction error of the central patch can be used to estimate the saliency.The objectness is the possibility of a pixel or region being inside an object,which is determined by the three factors,color contrast,edge density,and superpixels straddling.In implementation,a number of windows are sampled from the multiple representation of the input image,where the three factors are evaluated on.In this paper,two ways of combining autoencoder-based saliency and objectness are evaluated.On the one hand,the objectness is used as prior distributions for extracting image patches to train the deep autoencoder to get a better representation of the imput image.Experiments show that the improvement to saliency is very limited using this method.On the other hand,linear combination of autoencoder-based saliency and the objectness is also evaluated.Experiments show that linear combination can significantly improve the final saliency.This paper shows that in the top-down semantic-driven saliency estimation model,the high-level semantic features of the image has abundant a priori information,which can significantly improve the saliency detection.