Improved Adaptive Spectrum Scale-Space in Frequency Domain for Saliency Detection

Abstract

Saliency detection is a process to find the significant region on the images, which is popular in the image processing area. It has been extended and used in many applications in computer vision systems. Many researches have attempted to propose effective model to detect the saliency area which is corresponding to human perception. Therefore, this research focuses on improving the saliency detection process by proposing the improved adaptive spectrum scale-space (IASSS). The main contributions of the proposed method include (i) scale-and-space Gaussian filter (AS²G filter) and (ii) the new method for saliency map selection based on local entropy. Firstly, the AS2G filter is used to suppress the non-saliency amplitude spectrum to extract the saliency map. Then, the best saliency map is selected from the results of the RGB and Lab color images by using the local entropy criteria. Then, the experimental results based on 235 images show that the overall performance of the proposed IASSS method outperforms the baseline methods including Itti, SR, SUN, SAL, HFT, and MSS methods.