Salient Object Detection Based on Deep Multi-level Cascade Network

Fully convolutional neural networks (FCNs) have played an important role in current saliency detection task, since the multi-layer structure describe the depth features of an image in different scales. To reasonably and effectively aggregate and utilize the hierarchical features, we propose a novel multi-level convolution feature cascade model with an end-to-end way in this paper. Our model consists of two modules: one is the multi-level depth feature extraction module via our improved FCNs; the other module aim at combining the characteristics of multiple resolutions with coarse semantics and fine details via \(\mathbf {C}\)ascade, \(\mathbf {U}\)psampling and \(\mathbf {D}\)econvolution operations, named as CUD module. Finally, the output of CUD module is used to predict the saliency map through further learning. The proposed model can efficiently and flexibly aggregate multi-layer convolution features and provides accurate saliency maps. The extensive experiments show that our method achieves satisfactory results compared with some current representative methods.