Visual Saliency And Sparse Learning Based Radar Image Target Detection

Characterized by its all-day and all-weather working property,synthetic aperture radar has become a major means of remote sensing for earth observation and found a wide range of applications in both military and civilian fields.In recent years,along with the development of synthetic aperture radar imaging technique,spaceborme and airborne platforms have ob-tained massive high-resolution(HR)earth observation imagery,which provides rich data for accurate scene perception.However,the current radar image interpretation capability can-not meet the processing demand of the collected massive data,so that valuable information cannot be effectively excavated to serve the specific application tasks.Traditional artificial interpretation method usually requires much time and effort,and how to enable computers to replace the human to accomplish the task of obtaining the key target information in the massive radar images is an important research topic in the field of radar remote sensing.Focusing on the task of real-time target detection from high-resolution synthetic aperture radar images,this dissertation proposes to design scene saliency information inspired intel-ligent radar target detection algorithms from the perspective of biological visual perception,in order to break through the existing bottleneck during the development of radar target detection technology and provide new ideas and methods for the real-time and accurate in-terpretation of massive data.Specifically,the research work will start with the modeling of biological visual attention mechanism and then use the designed scene saliency modeling methods as analysis tool to explore intelligent radar image target detection algorithms for complex scenes.The main innovations are summarized as follows:1.Aiming at the ill-posed nature of saliency modeling problem and its dependence on pri-or knowledge,a complex scene saliency map generation approach based on digraph and multi-scale Bayesian inference is proposed.According to the spatial closure and surround-edness property of salient objects,a superpixel based digraph path optimization model is designed for the measure of image region saliency.In addition,this model is extended to multiple scales to cope with the "small distance accumulation" problem and a multi-scale Bayesian inference model is designed to achieve pixel-level saliency fusion.The proposed method shows superior performance than classical saliency detection methods on MSRA-1000 dataset and meanwhile obtains satisfactory detection results on a mixed verification dataset with salient objects of different sizes,locations and numbers.2.To address the robustness problem of traditional artificial rules in complex scene salien-cy modeling,a saliency detection approach based on discriminative dictionary learning and joint Bayesian inference is proposed.A discriminative background dictionary learning mod-el based on independent subspace assumption is designed,which can actively learn discrim-inative saliency cues from the image.At the same time,a joint Bayesian inference model is deduced for multi-source saliency information fusion,which can achieve reliable saliency measure with heterogeneous visual information.The effectiveness of the proposed method is verified on benchmark saliency datasets and its application potential is demonstrated on the adaptive image resizing problem.3.In view of the limited generalization ability of single saliency cue in complex scene mod-eling,a multi-source saliency cue fusion model based on hybrid sparse learning is proposed.A new saliency cue called "minimum span distance" is proposed from the perspective of contour saliency and its realization discrete optimization model and iterative solution algo-rithm are also designed.On the basis of the obtained contour saliency information,a hybrid sparse learning model for multi-source cue fusion is established,which can combine sparse cognitive prior and image boundary prior to realize joint optimization of saliency map.Test-ing results on large scale saliency databases indicate the strong modeling capability of the proposed method under complex scenes.4.Aiming at the structural and weak scattering characteristics of targets appearing in high-resolution SAR imagery,a fast ship target detection method based on hierarchical sparse saliency modeling is proposed.A robust scene saliency modeling method based on random forest is designed to achieve accurate perception of regions of interest(ROI)in the scene.Then,on the basis of the extracted regions of interest,a dynamic contour model based on constant false alarm rate technique is designed,which can accurately locate the target area.Testing results on real high-resolution SAR data show the proposed method has better per-formance than traditional ship detection methods.5.To improve the performance and efficiency of target detection algorithms in high-resolution SAR imagery,a fast object-level CFAR ship detection method based on adaptive saliency search is proposed.A saliency search model under the guidance of Boolean map and Gestalt principle is established,which can rapidly extract high-quality target candidate regions from the scene,avoiding the huge computational cost of exhaustive search.Based on the extract-ed target candidate regions,an object-level CFAR detector is designed and derived to real-ize adaptive clutter modeling as well as accurate target localization.The proposed method achieves better localization accuracy than classical ship detection methods on real ocean and harbor SAR data.The above work investigates the problem of intelligent image understanding from the per-spective of brain inspired computing.Novel theoretical models in visual attention mecha-nism computation are proposed and new performance breakthroughs in radar image target detection are made,which together provide effective solutions for saliency modeling and target detection of complex scenes.The research work has both theoretical significance and practical value for the development of radar remote sensing technology.