Salient Object Detection Algorithm Based on Dual-Layer Multi-Scale Neural Network

To further improve the accuracy of salient object detection, a novel Dual-Layer Multi-Scale Neural Network (DLMSNN) was proposed in this work. Different from existing deep model, the proposed deep model first learned deep features in a fine-to-coarse manner, and roughly located the salient object regions. Then, the model integrated multi-scale contextual information in a coarse-to-fine manner to precisely detect the entire salient object regions, and generated accurate saliency map of the input image. Finally, to further improve the performance, the dense conditional random field algorithm was used to refine the saliency map and produce the final result. The experimental results on several public benchmarks showed that the proposed algorithm outperformed traditional salient object detection methods and existing deep learning-based algorithms.