水面目标检测与识别算法研究

发布时间：2018-04-29 18:16

本文选题：目标检测 + 水面目标　；参考：《华中科技大学》2016年硕士论文

【摘要】：最近几年,海洋经济总量持续快速增长,并成为拉动国民经济发展的重要引擎。海洋装备和港口设施对海洋经济的发展起着重要的促进作用。无人艇作为一种智能型海洋装备受到全世界的关注,并得到快速发展。本文重点关注水面目标的检测与识别算法研究,在实际应用中包括两个方面:基于无人艇的水面目标检测与识别算法研究和红外图像下泊港舰船检测算法研究。针对基于无人艇的水面目标检测与识别算法研究,由于水面气候条件复杂、背景复杂多变、目标类型众多,如果直接在原始图上进行检测,不仅难度较高,而且速度和准确率都很难得到保证。本文中提出了两种不同的解决思路,一种是通过将目标性和显著性相结合起来,首先,通过目标性分析得到目标候选区域,此时目标候选区域中会存在一定虚警信息,然后利用显著性分析得到显著性区域,最后将目标性与显著性相结合,剔除虚警,得到目标准确位置,该算法不带有特定目标类型信息,因此普适性较好;另外一种是将深度学习应用到水面目标的检测与识别中,并同时给出目标的具体类别信息和置信度。相比于现有的目标检测与识别算法,本文中提出的算法无论准确率上还是在速度方面上都有一定的提升,对无人艇的自动避障和自主航行具有重要的指导意义。针对红外图像下泊港舰船检测算法研究,由于红外成像的特点,想要通过泊港内舰船的边缘和纹理来直接做检测是非常困难的,并且实际场景中泊港内的背景十分复杂,有舰船、港口、海水、道路等多种目标类型,同时舰船的姿态、位置、大小、数量都是不确定的。考虑到以上几个难点,本文通过两步来实现泊港内舰船检测:港口区域检测和舰船检测,首先通过模板匹配将港口区域检测出来,然后再在港口区域内检测舰船。这样不仅可以缩小搜索的空间,提高算法的效率,还可以减少大量复杂的背景对算法的干扰,提高算法的准确率。最终通过实验证明,本文提出的算法不仅准确率比较高,同时虚警率还比较低。
[Abstract]:In recent years, the total amount of marine economy has continued to grow rapidly, and has become an important engine for the development of the national economy. Marine equipment and port facilities play an important role in promoting the development of marine economy. As an intelligent marine equipment, unmanned craft (UAV) has attracted worldwide attention and developed rapidly. This paper focuses on the research of the detection and recognition algorithm of the surface target, including two aspects in the practical application: the research of the detection and recognition algorithm of the surface target based on the unmanned craft and the research of the detection algorithm of the ship in the port under the infrared image. For the research of surface target detection and recognition algorithm based on unmanned craft, it is not only difficult to detect the surface target directly on the original map, but also because of the complex climatic conditions, complex background and many target types. And speed and accuracy are difficult to guarantee. In this paper, two different solutions are proposed. One is to combine the goal and salience. Firstly, the target candidate region is obtained by the target analysis, and there will be some false alarm information in the target candidate region. Then the significance region is obtained by significance analysis. Finally, the target and significance are combined to eliminate false alarm and get the accurate position of the target. The algorithm does not contain the information of specific target type, so it is more general. The other is to apply depth learning to the detection and recognition of surface targets, and to give the specific information and confidence of the target at the same time. Compared with the existing target detection and recognition algorithms, the algorithm proposed in this paper has a certain improvement in both accuracy and speed, which has an important guiding significance for automatic obstacle avoidance and autonomous navigation of unmanned craft. Aiming at the research of ship detection algorithm in port under infrared image, because of the characteristics of infrared imaging, it is very difficult to detect the ship directly through the edges and textures of the ship in the port, and the background in the actual scene is very complex. There are many target types, such as ship, port, sea water, road and so on. At the same time, the attitude, position, size and number of ships are uncertain. Considering the above difficulties, this paper realizes the ship detection in the port through two steps: the port area detection and the ship detection. Firstly, the port area is detected by template matching, and then the ship is detected in the port area. This can not only reduce the search space, improve the efficiency of the algorithm, but also reduce a large number of complex background interference to the algorithm, improve the accuracy of the algorithm. Finally, the experimental results show that the proposed algorithm not only has high accuracy, but also has a low false alarm rate.
【学位授予单位】：华中科技大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：U675.79;TP391.41

【参考文献】