Multi-scale Graph Fusion for Co-saliency Detection

The key challenge of co-saliency detection is to extract discriminative features distinguish the common salient foregrounds from backgrounds in a group relevant images. In this paper, we propose new framework which includes two strategies improve ability features. Specifically, on one hand, segment each image semantic superpixel clusters as well generate different scales/sizes images for input by VGG-16 model. Different scales capture patterns As result, multi-scale can various among all many kinds perspectives. Second, method Graph Convolutional Network (GCN) fine-tune features, aiming at capturing information and private or complementary feature scale. Moreover, proposed GCN jointly conducts fine-tune, graph learning, learning unified framework. We evaluated our three benchmark data sets, compared state-of-the-art methods. Experimental results showed that outperformed comparison methods terms evaluation metrics.