FCN and Unit-Linking PCNN Based Image Saliency Detection

Abstract

Detecting salient regions of an image can significantly increase the efficiency of follow-up processing, and thus improve the performance of the whole system. In this paper, we proposed a novel model of image saliency detection, which combines pulse coupled neural network (PCNN) with a fully convolutional neural network (FCN). In our proposed model, an image is firstly fed into a unit-linking PCNN, and the segmentation result, providing topological properties of the objects, serves as an input channel of the FCN. Guided by the topological features, the deep neural network then provides a coarse saliency map. Finally, we use the segmentation result to refine the boundaries of the salient objects to generate a fine saliency map. Furthermore, in this model various techniques are introduced to the PCNN to refine the segmentation result, preserving structural integrity between different objects. Experimental results on several benchmarks show that our model outperforms other state-of-the-art approaches without retuning on different datasets.