A Saliency Detection Model Based on Local and Global Kernel Density Estimation

Visual saliency is an important and indispensable part of visual attention. We present a novel saliency detection model using Bayes' theorem. The proposed model measures the pixel saliency by combining local kernel density estimation of features in center-surround region and global density estimation of features in the entire image. Based on the model, a saliency detection method is presented that extracts the intensity, color and local steering kernel features and employs feature level fusion method to obtain the integrated feature as the corresponding pixel feature. Experimental results show that our model outperforms the current state-of-the-art models on human visual fixation data.