Study On Key Technology Of Typical Targets Recognition From Large-field Optical Remote Sensing Images

With the development of earth observation technology, remote sensing imagesare increasingly used both in military and civilian applications. Automatic targetrecognition from remote sensing images plays an important role in intelligencegathering for modern warfare, but the existing remote sensing informationprocessing technology has not yet reached the level of demand for practicalapplication. With the improvement of remote sensing image resolution, remotesensing imaging systems produce large amount of image data, and the size of asingle remote sensing image increases significantly. However, the correspondingdata processing technology cannot adapt to the the requirements of real-timeeffective applications. To improve the utilization rate of remote sensing image data,improve the efficiency and reliability of the automated processing of remote sensinginformation and enhance military reconnaissance and intelligence-gatheringcapabilities, this paper focuses on the problem of typical targets recognition fromhigh resolution remote sensing images. On the basis of study for imagesegmentation, regions of interest (ROI) detection and feature extraction, amulti-class targets detection and identification method was proposed to to overcomethe poor portability problem of existing remote sensing targets detection algorithmsand improve the efficiency and adaptability of the automatic target recognitionsystem.Object-oriented analysis is an important area for high-resolution remotesensing image target recognition, and its prerequisite is to segment the image intoregion objects with some semantic information. For the status of low efficiency andlow automation degree of existing segmentation algorithms while used for highresolution remote sensing image segmentation, a marker-controlled watershedsegmentation method based on a fast mean shift algorithm is proposed. Firstly,bilateral filtering is carried on the original image for denoising. Then, an improvedfast mean shift algorithm is performed to get initial segmentation. After that, thehomogeneous regions obtained by initial segmentation are extracted as markers toexecute the marker-controlled watershed transform. Finally, a region mergingprocedure based on the idea of object-oriented is adopted to achieve the finalsegmentation result. This method is an effective solution to the over-segmentationproblem of the tradional watershed segmentation algorithms, and multiscalesegmentation results can be obtained by adjusting the threshold of the makers area.Compared with other segmentation algorithms both in accuracy and efficiency, theproposed segmentation algorithm guarantees good accuracy as well as improves the implementation efficiency greatly.For the problem of massive data, redundant information and time-consumingtargets searching of high resolution remote sensing image with complex senes andlarge size, a visual saliency detection based ROI extraction strategy is proposed,which deals with targets of different kind and scale in different ways. To the linearstructure targets, a top-bottom saliency model based on the density distribution ofthe weighted length of the straight line segments is used in low resolution image toregard the regions having long straight lines of hight contrast as ROI, which is ableto quickly locate the candidate lagrge linear targets in large images. As to the blobtargets, a multiscale color histogram contrast based bottom-up saliency model isadopted on the corresponding regions in high resolution image to generate a fullresolution saliency map quickly, which is combined with the superpixelsegmentation results to output salient ROI easily attracting human visual attentionsby extracting the superpixels with higher saliency values, narrowing the targetssearching range. Visua saliency based methods of ROI extraction are capable offinding candidate targets efficiently, which improves the efficiency of targetsrecognition.According to the needs of remote sensing targets detection and identification,rapid extraction of lines, circles and geometrical characteristics of regions arestudied. First, an improved phase grouping method and perceptual grouping arecombined to constitute a fast straight line segments detection algotithm. After that,these line segments are gropued together by the principle of perceptual grouping toextract structures of parallel lines and perpendicular lines, which is the basis ofrecognition of airports, bridges, harbor and other linear structure targets. Then, animproved two-step hough transform circles detection algorithm is proposed, and itimproves the speed and accuracy of circular oil tanks identification while detectingoil depot targets. Finally, an efficient connected-components labelling algorithm isproposed, which lays the fundation for fast computation of shape characteristics ofregions. Mass experiments indicate the efficiency of our algorithms.Based on the study of image segmentation, ROI detection and featureextraction, a shape features based approach is proposed to recognize multi-classtargets from high resolution remote sensing images. This method firstly use thedensity distribution of weighted line segment length based saliency model to extractROIs of linear targets in low resolution image and map them to the high resolutionimage. An object-oriented method is then adopted on the corresponding region toclassify it as water and land, providing context information for later recognitionprocess. And then, linear structural features are extracted to construct a featurevector for each ROI, and a decision tree support vector machine is used to identifyairports, ports and large bridges targets. Finally, in the areas corresponding to airports and harbors on high-resolution image, the multi-scale color histogramcontrast based saliency model is used to detect ROI for blob targets. And then, theoil depot targets are recognized according to the spacial distribution of the circularoil tanks, and the airplanes are recognized by a SIFT feature and moment invariantsbased SVM classifier. Experimental results indicate the effectiveness of theproposed multi-class targets recognization system, which has higher recognition rateand efficiency.