深海海底飞机残骸检测算法研究

发布时间：2018-06-05 21:28

本文选题：深海海底图像 + 飞机残骸识别　；参考：《大连海事大学》2017年硕士论文

【摘要】：近年来,空难事故频发,与陆地上空难不同,飞机在海上发生空难时机身会摔成碎片并沉入海底,在深海海底对飞机残骸碎片进行探测对飞机黑匣子的打捞具有重要意义。搜寻水下目标如同大海捞针,特别是在深海环境中,变得尤为困难。通常需要通过AUV搭载声呐和水下相机进行目标搜索,并利用水下相机图像进行取证,最终由人眼进行目标确认。由于AUV在水下工作时间较长,水下相机拍摄的图像数量十分巨大,而包含飞机残骸的图像数量却很少,如何滤除大量无效图像,提高检测的效率,是本文研究的核心问题。本文研究了深海海底背景和飞机残骸的特性,提出了飞机残骸检测方法,该方法先检测图像中的疑似目标区域,而后对疑似区域进行判决。在疑似目标区域检测中,首先,针对飞机残骸具有明显的形状和线条的特点,利用Hough变换直线检测算法来检测图像中的直线,并将检测结果标记在图像中以增强有效边缘,而后再利用基于图论的图像显著性算法(Graph Based Visual Saliency,GBVS)获取该图像的显著度图,将显著度最高的几组区域标记为目标疑似区域。在确认疑似目标区域是否为飞机残骸时,采用支持向量机(Support Vector Machine,SVM)分类器,针对深海海底背景图像特性和飞机残骸图像特性,提出了平均亮度、对比度、边缘密度和纹理方差四个指标作为支持向量机分类器的特征向量,并利用深海背景图像和飞机残骸图像制作了训练图像库,训练了支持向量机分类器,利用该分类器可实现对目标疑似区域的判决。为验证本文算法,开展了深水水池成像实验和近海海底成像实验,尽量模拟深海的景物特点和工作环境,获取了深海海底模拟图像数据,并对采集图像进行了目标检测,计算了漏警率,和有效图像在总图像中所占比率。实验检测结果表明,本文提出的算法具有低漏警率的特点,在保留有效图像的同时,可以大量滤除无效图像,从而大幅降低需人工判读的图像数量。
[Abstract]:In recent years, there are frequent air accidents, which are different from those on land. The fuselage will fall into pieces and sink to the bottom of the sea when there is an air accident on the sea. It is of great significance to detect the debris of aircraft wreckage in the deep sea and salvage the black box of the aircraft. Searching for underwater targets is like looking for a needle in a haystack, especially in deep-sea environments. Usually, it is necessary to carry out target search by AUV sonar and underwater camera, and use underwater camera image to obtain evidence, and finally confirm the target by human eyes. Because of the long working time of AUV under water, the number of images taken by underwater camera is very large, but the number of images containing airplane wreckage is very small. How to filter a large number of invalid images and improve the efficiency of detection is the core problem of this paper. In this paper, the background of deep sea and the characteristics of aircraft wreckage are studied, and a detection method of aircraft wreckage is proposed. The method first detects the suspected target area in the image, and then judges the suspected area. In the detection of the suspected target area, firstly, aiming at the obvious shape and line characteristics of the wreckage, the Hough transform line detection algorithm is used to detect the straight line in the image, and the detection result is labeled in the image to enhance the effective edge. Then, the graph Based Visual salience graph is used to obtain the salience graph of the image, and some regions with the highest saliency are marked as suspected regions. When confirming whether the suspected target area is the wreckage of the aircraft, the support vector machine support Vector Machine (SVM) classifier is used to propose the average brightness and contrast for the background image characteristics of the deep sea floor and the image characteristics of the wreckage. The edge density and texture variance are used as the feature vectors of SVM classifier, and the training image database is made by using deep-sea background image and airplane wreckage image, and the SVM classifier is trained. By using this classifier, the target suspected area can be judged. In order to verify this algorithm, the deep water pool imaging experiment and the offshore seabed imaging experiment are carried out to simulate the scene characteristics and working environment of the deep sea as far as possible, and obtain the deep sea bottom simulation image data, and carry on the target detection to the collected image. The false alarm rate and the ratio of the effective image to the total image are calculated. The experimental results show that the proposed algorithm has the characteristics of low false alarm rate and can filter out invalid images at the same time of preserving effective images, thus greatly reducing the number of images that need manual interpretation.
【学位授予单位】：大连海事大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：V328;TP391.41

【相似文献】