Salient regions detection in satellite images using the combination of MSER local features detector and saliency models

Nowadays, due to quality development of satellite images, automatic target detection on these images has been attracted many researchers' attention. Remote-sensing images follow various geospatial targets; these targets are generally man-made and have a distinctive structure from their surrounding areas. Different methods have been developed for automatic target detection.  In most of these methods, target searching was used for features extraction or matching in order to detect the geospatial targets. Hence, in this paper, in order to improve computational time in target detection process, the areas of images with  high probability of geospatial target existence, were selected. This will significantly improve the automation level of the process and computational time in following processes. For this purpose, a combination of saliency models and the of-local features detector's algorithm were used. The proposed method consists of three main steps, including local feature extraction by applying MSER algorithm, saliency maps generation by applying AWS and WMAP models and regions determination with high probability of geospatial target existence. In this paper, a threshold was defined by calculating the saliency values in the whole image. Salient regions were detected by applying the threshold for each extracted area. This method was implemented on six satellite images by different sensors and the mentioned six satellite images derived from Google Earth software that consist several geospatial targets with various backgrounds. Two criteria were used in order to quantitatively assess the results. Moreover, this method was compared with the mean shift segmentation algorithm in the color space images that was applied to detect saliency regions of an image. The results of the proposed method showed that the average of the detected areas was 5.1% of the total area of the images in which 98.28% of the targets were located in these regions. In addition, the average amount of computational time was 22.1 seconds. The results showed the superiority of the proposed method in terms of accuracy and computational speed.