【作者】 臧风妮；

【导师】 李庆忠；

【作者基本信息】 中国海洋大学 ， 港口、海岸及近海工程， 2014， 博士

【摘要】 我国具有辽阔的海岸线，海域十分广阔，因此，海事视频自动监测具有重要的现实意义和应用价值。在军用方面，可以维护海洋权益、加强海域监管、减少海事纠纷等；在民用方面，海面目标的自动检测技术可以高效实现港口、海湾等的海上交通管制、事故船只救援报警、海洋环境监测、海洋专属经济区保护，并对非法捕鱼、污染物倾倒、非法走私偷渡等起到强有力的监管作用。在海事智能视频监控方面，目前面临的难点就是当摄像机安装在晃动的海洋浮标或运动的海事船上时，如何实现在各种海况下快速地检测出运动的舰船目标，这是实现海域智能监测的关键技术之一。为此，本文对海洋动态视频监控中舰船目标快速检测算法以及海上雾天图像清晰化处理算法进行了深入研究，完成的主要研究工作如下：（1）提出了一种基于小波域视觉注意力模型的海面舰船目标快速检测算法。根据人类视觉观测特点，即大尺度下注重目标的轮廓，小尺度下注重目标的细节，首先利用利用提升小波变换在小波域建立了双尺度视觉选择注意模型，然后在粗分辨率低频子带上分别利用相位谱法和梯度法建立视觉显著图，并对两者进行有效融合形成综合视觉显著图，最后通过小波反变换得到原始高分辨率图像的视觉显著图，由此实现海面目标区域的检测与提取。实验结果表明：提出的算法能够快速、准确地实现海面舰船目标的自动检测，可用于基于海洋浮标的海事智能监测。（2）提出了一种基于海面背景纹理模型的舰船目标检测算法。根据海面纹理具有均匀性和一致性的特点，该算法首先在离散余弦变换（DCT）域提取图像子块的能量特征，实现了天空背景和海天线的快速检测。然后，对于海天线以下的海面区域，在DCT域提取图像子块的纹理特征向量，并利用自适应模糊C均值聚类方法建立了海面背景的混合纹理模型。最后，利用建立的海面纹理模型，实现了海面背景与舰船目标的分割。实验结果表明：提出的算法可以快速、鲁棒实现舰船目标的检测，尤其适合于高海况下基于运动监视平台的海事监测。（3）提出了一种基于轮廓波变换（Contourlet）的海上雾天图像清晰化处理算法。该算法首先将RGB图像转换为YUV图像，对亮度Y图像根据雾天成像物理模型，利用Contourlet变换将图像低频成分和高频成分有效分离；然后在低频子带上利用双边滤波技术实现了散射光成分的有效估计与去除，并消除了散射光去除过程引起的“光晕效应”；对Contourlet变换后的高频子带，进行自适应阈值去噪和纹理细节的增强处理；对Contourlet反变换重建后的亮度Y图像再进行直方图均衡化处理，以实现图像整体对比度的增强；最后将亮度Y图像的处理结果与颜色分量U、V合成得到清晰化的雾天彩色图像。实验结果表明：提出的算法既能显著降低处理时间；又能有效提高雾天图像的清晰度。更多还原

【Abstract】 China possesses a long coastline, so it is necessary to develop automatic monitoring andsurveillance system for its maritime space. In military area, it is of great significance forsafeguarding the maritime rights and interests, strengthening supervision and management ofimportant sea areas, and reducing the maritime disputes and so on. In civil area, it has manyapplications such as marine traffic management in port or bay, ocean environment monitoring,protection of exclusive economical zones, searching and rescuing of ships from wrecks, andtaking measures against illegal activities such as illegal fishing, smuggling, and immigration.Currently, the major impediment to intelligent video maritime monitoring is how to fast androbustlydetecttheseasurfacetargetsundervariousseastatesforbuoy-basedorship-basedvisualsurveillance, which isoneofthekeytechnologies for intelligent maritime visualmonitoring. Toovercometheabove-mentionedobstacle,thisthesisinvestigatesthefastshipdetectionalgorithmsandrestorationalgorithmofseafogdegradedimagesfordynamicmaritimevisualsurveillanceindepth,andhasaccomplishedthefollowingresearchwork.(1) A fast algorithm to detect sea surface ship targets based on the visual selective attentionmechanism in wavelet domain is proposed in the dissertation. According to the characteristics ofhumanvisualobservation,that is,thecontoursofobjectsishighlightedat largescaleandthedetailsof object is focused at small scale, a two-scale visual attention model is first established in thewaveletdomainbyusing liftingwavelettransform.Thentwo visualsaliencemapsaregeneratedbyemploying the phase spectrum method and gradient based method on the low-pass subband ofwavelet domain withcoarseresolutionrespectively, and asynthetic visualsalience map isobtainedby effective combination of the both obtained salient maps. Finally, the high resolution visualsaliencymapoforiginalimageisgeneratedbyinversewavelettransform,andtheseasurfaceobjectregionsareextracted fromthe final saliency map. Theexperimentalresults showthat theproposedalgorithm can detect the maritime targets quickly and accurately, so it can be used for maritimebuoy-basedintelligentmonitoring.(2) A novel algorithm for ship target detection based on texture model of sea surface ispresented inthedissertation. Accordingtothe factthatthetexture featureofseasurface isbasicallyuniform and consistent, a segmentation strategy from simple to complex is introduced in theproposed algorithm. Firstly, the simple sky background and horizon are obtained quickly using anenergyfeatureinDCTdomainofimageblocks.Secondly, inordertoseparateshiptargetsfromthecomplex sea background belowthe horizon, a newtexture Gaussian mixture modelofsea surfacebased on imageblocks inDCTdomain isdeveloped,then fromtheconstructed seasurface model,ship targets are segmented fromsea background. The experimentalresults showthat the proposedalgorithm can detect ship targets robustly and satisfy the real-time requirement of visual maritimesurveillance from non-stationary platforms, especially suitable for sea background with high sea state.(3) A visibility improving algorithm for sea fog images based on contourlet transform isdeveloped. Firstly, the RGB color space is transformed to the YUV color space. Then, for theluminance component Y, from physical formation model of fog images, the low frequencycomponent and the high frequency component is separated effectively from contourlet transform.Secondly,theamountofmediascattering light isestimatedandremovedbyusingadaptivebilateralfiltering on the low-pass subband of contourlet transform, thereby simultaneously eliminating thehalo effect;and forthe high-passsubbandsofcontourlet transform,thenoise is filtered byadaptivethresholding and thetexturedetails areenhanced adaptively. Thirdly, thereconstructed Yimage byinverse contourlet transform is further enhanced through histogram equalization to improve theglobalcontrast.Finally,thevisibilityenhancedimageisobtainedbycombinationtheresultedimageofYcomponentandthechrominancecomponentsofUandV. Theexperimentalresultsshowthatthe proposed algorithm can significantly reduce the processing time and effectively improve thevisibilityoffogdegradedimages.更多还原