Deep supervised visual saliency model addressing low-level features

Deep neural networks detect visual saliency with semantic information. These high-level features locate salient regions efficiently but pay less attention to structure preservation. In our paper, we emphasize crucial low-level features for deep neural networks in order to preserve local structure and integrity of objects. The proposed framework consists of an image enhancement network and a saliency prediction network. In the first part of our model, we segment the image with a superpixel based unit-linking pulse coupled neural network (PCNN) and generate a weight map representing contrast and spatial properties. With the help of these low-level features, a fully convolutional network (FCN) is employed to compute saliency map in the second part. The weight map enhances the input channels of the FCN, meanwhile refines the output prediction with polished details and contours of salient objects. We demonstrate the superior performance of our model against other state-of-the-art approaches through experimental results on five benchmark datasets. © 2019, Springer-Verlag GmbH Germany, part of Springer Nature.