Salient Object Detection Based on Amplitude Spectrum Optimization

Abstract

Saliency detection is prerequisite for many computer vision tasks. The existing frequency domain models can not always detect a complete object. We propose a novel salient object detection model based on an optimized amplitude spectrum. This model computes saliency map in two steps. Firstly, we optimize amplitude spectrum by smoothing the peaks in log amplitude spectrum. The raw saliency maps are computed by combining the optimized amplitude spectrum and the original phase spectrum according to different thresholds. Secondly, we compute the entropy of raw saliency maps and select the raw saliency map with the smallest value of entropy as the final saliency map. Our model detects more complete object region. By testing on the databases ASD, MSRA10K, DUT-OMRON and SED2, experiments demonstrate that the proposed model outperforms the state-of-the-art models.