基于RGA的快速光学遥感图像舰船目标检测算法研究

发布时间：2018-01-18 07:04

本文关键词：基于RGA的快速光学遥感图像舰船目标检测算法研究　出处：《西南交通大学》2017年硕士论文　论文类型：学位论文

【摘要】：遥感技术具有侦查范围广,全天候,不受地理限制等优点,应用前景广阔。基于遥感图像的目标检测作为遥感图像应用中重要的一环,其对于资源调查、灾害检测以及军用侦查都具有重要的研究意义。由于遥感图像的复杂多样性,目标检测需要解决颜色纹理、旋转尺度变化、形似干扰物等一系列难点问题;同时随着遥感技术的发展,遥感信息数据的快速增长,依靠人工判别不能满足实时性的要求,这些都对遥感图像目标的检测提出了新的挑战。复杂背景下兼顾目标检测的精度和速度对于实时性应用具有非常重要的意义和价值。本文以光学遥感图像舰船目标为研究对象,围绕复杂背景下的目标检测算法的精度和效率进行研究。针对复杂背景下目标轮廓附近的噪声干扰、形似干扰物及目标部分遮挡等影响的问题,本文在RGA姿态一致性算法的基础上,设计了一种基于姿态回归的舰船检测方法。方法主要包括三个部分:(1)根据舰船模板轮廓点之间的位置关系和RGA分布,得到每个轮廓点及其邻域同姿态点,对被检目标轮廓点姿态估计时,将其与模板轮廓点及近邻同姿态点校验,抑制噪声点对目标中心的投票;(2)采用舰船局部连接结构加权的方法,提升具有舰船目标特征的整体投票比重,以增加V型设施、矩形等形似干扰物和目标之间的区分度;(3)在现有方法的基础上重新定义了轮廓命中率和最大非连续因子,对检测目标采用舰船模板轮廓命中率和最大连续丢失率进行修正,并对最后检测结果与回归的模板轮廓完整性进行综合判别,去除虚警。实验证明,本章的方法对目标轮廓附近的噪声具有良好的适应性,并且可以较好区分形似干扰物。在复杂背景下较目前最好的方法检测准确率提高了 8%左右。针对姿态回归舰船检测算法的时间复杂度过高问题,本文设计了一种基于显著性的快速舰船目标检测算法。首先,选用以超像素作为基本计算单位的对比度显著性检测方法,通过结合各超像素的颜色和空间距离差异得到对比度先验图,突出目标区域和背景区域的差异;其次,为了得到更准确的目标中心位置,使用超像素之间的差异值作为局部特征构建凸包确定目标的大致位置,对不同位置的超像素使用高斯模型赋予不同权重,得到中心先验图;同时为进一步抑制边界背景,在对比度先验图和中心先验图的基础上融合了边界背景先验图,通过三种先验信息融合的显著性检测方法快速精准的提取目标感兴趣区域,最后采用姿态回归的方法在感兴趣区域进行舰船目标检测。实验证明,复杂背景下本文算法有效快速去除背景区域的同时,检测准确率也得到一定提升。相比于RGA方法和姿态回归方法,该算法检测准确率分别提高了 12.9%和4.8%,检测时间降低了72%和78%。
[Abstract]:Remote sensing technology has the advantages of wide range of detection, all-weather, no geographical restrictions, and so on. As an important part of remote sensing image application, target detection based on remote sensing image is an important part of resource investigation. Disaster detection and military investigation are of great significance. Because of the complexity and diversity of remote sensing images, target detection needs to solve a series of difficult problems, such as color texture, rotation scale change, shape like interference object, etc. At the same time, with the development of remote sensing technology, the rapid growth of remote sensing information data, relying on manual discrimination can not meet the requirements of real-time. All of these put forward new challenges to target detection in remote sensing image. It is very important and valuable for real-time application to take into account the accuracy and speed of target detection in complex background. In this paper, the object of ship in optical remote sensing image is considered. Marked as the object of study. Focusing on the accuracy and efficiency of the target detection algorithm in complex background, aiming at the noise interference near the target contour, the shape of the jamming object and the partial occlusion of the target in the complex background, and so on. This paper is based on the RGA attitude consistency algorithm. A ship detection method based on attitude regression is designed. The method includes three parts: 1) according to the position relation and RGA distribution of ship template contour points. Each contour point and its neighborhood same attitude point are obtained. When the contour point attitude is estimated, it is checked with the template contour point and the nearest neighbor pose point to suppress the noise point voting on the target center. (2) the weighted method of local connection structure is used to increase the proportion of the whole voting with the characteristics of the ship's target, so as to increase the distinction between the V-shaped facilities, the rectangle and the object. 3) based on the existing methods, the contour hit ratio and maximum discontinuity factor are redefined, and the ship template contour hit ratio and the maximum continuous loss rate are corrected. Finally, the integrity of the final detection results and the regression template contour is comprehensively identified to remove false alarm. Experiments show that the method proposed in this chapter has a good adaptability to the noise near the target contour. Compared with the best method in complex background, the detection accuracy is improved by about 8%. Aiming at the problem of high time complexity of attitude regression ship detection algorithm. In this paper, we design a fast ship target detection algorithm based on saliency. Firstly, we choose the contrast significance detection method with super-pixel as the basic unit of calculation. By combining the color and spatial distance difference of each super-pixel, the contrast priori map is obtained to highlight the difference between the target region and the background area. Secondly, in order to get a more accurate target center position, the difference value between the super-pixels is used as the local feature to construct the convex hull to determine the approximate position of the target. Gao Si model is used to give different weights to superpixels in different positions, and a central prior map is obtained. At the same time, in order to further restrain the boundary background, the boundary background priori graph is fused on the basis of contrast prior graph and central prior graph. Through three priori information fusion salience detection methods quickly and accurately extract the region of interest of the target, and finally use the attitude regression method to detect the ship target in the region of interest. In the complex background, the algorithm can remove the background area effectively and quickly, and the detection accuracy is also improved, compared with the RGA method and attitude regression method. The detection accuracy of the algorithm is improved by 12.9% and 4.8, and the detection time is reduced by 72% and 78 respectively.
【学位授予单位】：西南交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP751

【参考文献】