Saliency Detection via CNN Coarse Learning and Compactness Based ELM Refinement

Ruirui Li, Shihao Sun, Lei Yang, Wei Hu

Abstract: Salient object detection has attracted a lot of research in computer vision. It plays a vital role in image retrieval, object recognition and other image processing tasks. Although varieties of methods have been proposed, most of them heavily depend on feature selection and fail in the case of complex scenes. We propose a processing framework for saliency detection which contains two main steps. It uses deep convolutional neural networks (CNNs) to find a coarse saliency region map that includes semantic clues. Then it refines the coarse saliency map by training an extreme learning machine (ELM) on a group of color and texture compactness features. To get final saliency objects, it synthesizes the coarse saliency region map and several multiscale saliency maps that are obtained by refining the coarse one together. The method achieves good experimental results and can be used to improve the existing salient object detection methods as well.