Saliency Detection Via Logistic Label Propagation With Weight-adaptive Fusion

Saliency detection of images has been an important preprocessing step in computer vision field, whose principle aim is imitating the human eyes'process of tackling vision information, reducing redundant information, extracting the most salient region of an image and enabling it to catch intensive attention. A large amount of computation models have been proposed by plenty of saliency detection algorithms from different aspects, and have been applied in a variety of object recognition literatures.In this paper, we propose a saliency detection algorithm based on the combination of logistic label propagation and weight-adaptive fusion. First, we compute the objectness map and the backgroundness map of the input image, then multiply these two pre-maps to generate an initial saliency map. Second, we set a fixed threshold to obtain the foreground seeds, using the superpixels on four image boundaries as background seeds. In addition, we exploit the unsupervised clustering random forest algorithm to calculate the similarity matrix, with which we integrate into the process of logistic label propagation mechanism. In this way, we transmit the saliency values of seeds to the whole image. Third, we use a novel weight-adaptive fusion method to fuse the propagation result and the initial map. Finally, through the multi-scale training classifier, we optimize the fused result map, then generate the final saliency detection map.Extensive experiments on four publicly available datasets have been implemented, including MSRA5000, ECSSD, PASCAL-S and THUS 10000 dataset. Meanwhile, we have shown the comparison between our algorithm and other sixteen state-of-the-art methods. The experimental results have demonstrate that the proposed method performs well in terms of the Precision-Recall as well as the other three evaluation metrics when comparing with other methods.