Fusing Local Texture Description Of Saliency MAP And Enhanced Global Statistics For Ship Scene Detection And Object Extraction In High-resolution Remote Sensing Image

With the development of remote sensing technology,remote sensing image data shows a trend with magnanimity,multi-resolution,multi-source and multi-band.Facing the arrival of big data of remote sensing,how to improve the ability of remote sensing image analysis to accommodate the level of image acquisition with high-speed growth,meanwhile,to obtain valuable information in big data is a difficult problem to be solved.Among them,the ship target detection in remote sensing image has always been an extremely important strategic significance in the civilian and military fields and become a hot issue in the field of image processing.With the increasing complexity and intensification of marine issues,the strategic objective “To Build Ocean Power” is proposed at first time in the18 th CPC National Congress.As a result,the importance of ship target detection in remote sensing has risen to a new height.The coarse to fine strategy is usually adopted for ship target detection of high-resolution remote sensing images with complicated background of sea and land.For the high-resolution remote sensing images,people are concerned about the targets which are account for only small parts of the whole image.For example,compared to the ocean as well as land and other complex background areas,the ship target is relatively so small.In order to solve the problem of how to quickly detect the region of interest,inspired by the visual attention mechanism and the Gist of scene,a new feature representation based on fusing local texture description of saliency map and enhanced global statistics for ship scene detection in very high-resolution remote sensing images with inland,coastal,and oceanic regions is introduced in this paper.Moreover,the combination of ROI saliency map and WTA network is applied to the ship ROI for extraction of object contour.Thus,the proposed method forms a new framework from scene detection to objection extraction.First,two low computational complexity methods are adopted: ITTI attention model and LBP.The ITTI attention model is used to extract saliency map,from which local texture histograms are extracted by LBP with uniform pattern.Meanwhile,Gabor filters with multi-scale and multi-orientation are convolved with the input image to extract Gist,means and variances which are used to form the enhanced global statistics.The strategy of feature fusion is the full integration of the saliency features,local texture features and global features.Second,sliding window-based detection is applied to obtain local image patches and extract the fusion of local and global features.SVM is then used for training and classification.Our detection manner could remove coastal and oceanic regions effectively.Moreover,the interest region of ship scene can be detected accurately.Finally,WTA mechanism is applied to saliency maps of ship scene to achieve fast object extraction.Experiments on 20 large very high-resolution remote sensing images collected by Google Earth shows that the fusion feature has significantly advantages than LBP,meanwhile,the fusion feature has about 3%,5%,and 2% advantages than Saliency map-based LBP,Gist and Gabor textures,respectively.Furthermore,compared to these single features,desirable results with higher true positive rate and lower false positive rate can be obtained in ship scene detection.Furthermore,the desirable performance also can be obtained for ship extraction.