Joint Multi-Image Saliency Analysis for Region of Interest Detection in Optical Multispectral Remote Sensing Images

Abstract

The automatic detection of regions of interest (ROI) is useful for remote sensing image analysis, such as land cover classification, object recognition, image compression, and various computer vision related applications. Recently, approaches based on visual saliency have been utilized for ROI detection. However, most existing methods focus on detecting ROIs from a single image, which generally cannot precisely extract ROIs against a complicated background or exclude images with no ROIs. In this paper, we propose a joint multi-image saliency (JMS) algorithm to simultaneously extract the common ROIs in a set of optical multispectral remote sensing images with the additional ability to identify images that do not contain the common ROIs. First, bisecting K-means clustering on the entire image set allows us to extract the global correspondence among multiple images in RGB and CIELab color spaces. Second, clusterwise saliency computation aggregating global color and shape contrast efficiently assigns common ROIs with high saliency, while effectively depressing interfering background that is salient only within its own image. Finally, binary ROI masks are generated by thresholding saliency maps. In addition, we construct an edge-preserving JMS model through edge-preserving mask optimization strategy, so as to facilitate the generation of a uniformly highlighted ROI mask with sharp borders. Experimental results demonstrate the advantages of our model in detection accuracy consistency and runtime efficiency.