应用图像融合与多样性的舰船显著性检测

基于单源的图像显著性检测存在较大的虚警或漏检,文章提出了利用约简后的特征点和CPD算法对海面实拍船只图像进行多源图像匹配,获得图像间的变换投影方程并利用投影方程对单源图像的显著性检测结果进行叠加与虚警控制器分类,从而达到提高检测率与控制虚警的目的。显著性检测方面,文章分析了基于图等级多样性的显著性检测方法的不足和优点,引入了最大稳定区域检测方法对图像做前期处理,并对获得区域进行联合获得新区域,使得新区域能够最大限度地满足基于图的等级多样性显著性检测最优条件。对于检测获得的联合区域目标显著性不完整的情况,利用了区域的叠加性进行加权求和,最终获得了具有较好联通性的多舰船目标图像显著性检测结果。对于显著性检测结果中存在较大虚警的情况,文章进一步提出计算船只与浪花的多尺度分形维数,并结合Adaboost算法训练浪花虚警控制器。实验结果显示控制器能够消除一部分浪花带来的虚警,但是对于灰度与舰船极为相似的虚警无法消除。

Ship-target saliency detection via image fusion and graph-based manifold ranking

Objects saliency detected by single-source image always include a lot of false alarm and leak detection. This work proposes to simplify the feature points data first and then uses it to match the multi-source images, after which we can get a mapping function between the couple images. The function is used to match the saliency results between the images so as to improve the object detection rate and reduce the leak detection rate. For the saliency detection of the objects, this work propose to improve the saliency detection via graph-based manifold ranking method by MSER. We firstly detect the MSER regions and union the regions, which can make every region suit the request of that method. And after the saliency detection of every union regions we sum the saliency maps by weight W. The summing process unions the saliency objects well and can reduce the false alarm. But false alarm still exist. This work propose to compute a false alarm controller to lower the false alarm by associating the multi-fractal dim and Adaboost method, which works well for wave false alarm reduction but not so good for false alarms like the ships very much.