Advanced Superpixel-Based Features and Machine Learning Based Saliency Detection

The saliency map simulates human perception and is useful for several image-processing applications. Many advanced saliency detection algorithms applied superpixel-based features instead of pixel-based features for saliency map generation. With superpixels, the high-level features can be extracted and a better saliency detection performance can be achieved. Recently, the convolutional neural network (CNN) has been thrived in computer vision. However, it was difficult to integrate it directly with superpixel-based method since the CNN required grid-like input while a superpixel generally has an irregular size and shape. In this work, several strategies are adopted to well apply machine learning techniques for superpixel-based saliency detection. First, instead of applying the CNN directly, the support vector machine (SVM) is applied as the classifier to well process the feature information. Instead, we apply the CNN ass the pre-processing procedure of superpixel merging. Moreover, several advanced superpixel-based features, including boundary connectivity and manifold rank features, are proposed to improve the accuracy of saliency detection. The simulation results on four datasets show that the proposed algorithm outperform both conventional and learning-based saliency detection methods.