Image saliency detection via multi-scale iterative CNN

Salient object detection has received increasingly more attention and achieved significant progress lately due to the powerful features learned by deep convolutional neural networks (CNNs). In this work, we propose a multi-scale iterative CNN for salient object detection, which has two complementary subnetworks at different spatial scales. For each subnetwork, we augment the CNN structures with an iterative learning process to learn the saliency map, where early stages of the CNN give a rough estimate of the saliency map and the remaining errors are gradually learned to refine the saliency map. By merging predictions of the two subnetworks, the training error can be reduced significantly and the estimated saliency map becomes more accurate. Unlike some previous CNN-based methods which often rely on superpixel segmentations, the proposed model is fully CNN and hence can estimate the saliency map much more efficiently. Extensive experiments on standard benchmarks demonstrate that our method outperforms some of the state-of-the-art methods in terms of both accuracy and speed and achieves as good performance as some recent state-of-the-art end-to-end methods under fair settings.