基于光学遥感图像的飞机目标检测算法研究

发布时间：2018-07-06 14:41

本文选题：遥感图像 + 目标检测　；参考：《西南交通大学》2014年硕士论文

【摘要】：随着遥感技术的发展,遥感图像由于覆盖范围广、信息量大、观测周期短等特点受到了广泛的关注和研究,遥感图像目标检测作为遥感图像解析的一个重要研究方向,其对于资源调查、灾害监测、资源勘探以及军事目标的识别判读等都具有重要的意义。由于遥感图像数据量巨大,依靠传统人工判读的方式从中提取特定目标的信息难以适应遥感技术的发展趋势,如何从这些数据中快速、准确的提取所需要的信息成为当今遥感图像解析的重点与难点。本文以光学遥感图像的目标检测算法为研究对象,主要围绕如何提高目标检测算法效率与目标中心的预测精度进行研究。本文在综述遥感图像目标检测算法的基础上对其归纳和总结,阐述算法的基本原理和检测过程,并对经典的目标检测算法进行仿真实验和分析。文中首先阐述了特征提取对目标检测的重要性,然后重点介绍了目标检测中常用的几种特征以及基于这些特征的飞机目标检测算法,并对算法进行仿真实验,分析算法中存在的不足之处。其次针对显著图应用在遥感图像中缺少客观评价标准的问题,本文将显著图的量化评价标准引入到遥感图像中。文中仿真对比分析了常用的三种显著图算法,并结合遥感图像以及遥感图像中飞机目标的特性,从主观和客观评价方面对显著图的突显效果进行评价。在综合考虑突显精度和显著区域突显完整程度的情况下,Itti算法对于遥感图像中飞机小目标的突显效果优于GBVS算法和SR算法。通过对显著图与遥感图像模板匹配相结合的算法进行仿真实验,验证了基于显著图算法的目标检测可以舍弃百分之七十以上的背景区域,缩小目标搜索空间是提高目标检测算法效率的一种有效途径。最后,基于随机森林目标检测算法在霍夫投票阶段是平均分配投票权重的,针对平均权重对目标位置预测不够准确的问题,本文将原算法基于随机森林叶子节点信息的平均权重改为基于样本目标特征字典的指数分布投票权重。在训练阶段将样本的目标像素特征向量存储起来作为样本字典,检测阶段对于随机森林预测为样本的像素,计算该像素点特征与字典中特征之间的欧式距离,根据该距离计算对应的指数函数分配权重,并对目标中心进行投票。通过仿真实验分析了改进投票权重前后霍夫图像的特点,并验证这种改进能够提高目标检测算法的性能。
[Abstract]:With the development of remote sensing technology, remote sensing image has received extensive attention and research because of its wide coverage, large amount of information and short observation period. Target detection of remote sensing image is an important research direction in remote sensing image analysis. It is of great significance for resource investigation, disaster monitoring, resource exploration and military target recognition. Because of the huge amount of remote sensing image data, it is difficult to adapt to the development trend of remote sensing technology by relying on the traditional manual interpretation method to extract the information of specific target. Accurate extraction of information is becoming the focus and difficulty of remote sensing image analysis. In this paper, the object detection algorithm of optical remote sensing image is studied, which focuses on how to improve the efficiency of the target detection algorithm and the prediction accuracy of the target center. Based on the summarization of remote sensing image target detection algorithm, this paper summarizes its basic principle and detection process, and carries on the simulation experiment and analysis to the classical target detection algorithm. In this paper, the importance of feature extraction to target detection is first expounded, and then several commonly used features in target detection and aircraft target detection algorithms based on these features are introduced, and the simulation experiments are carried out. The shortcomings of the algorithm are analyzed. Secondly, aiming at the lack of objective evaluation criteria for salient maps in remote sensing images, this paper introduces the quantitative evaluation criteria of salient maps into remote sensing images. In this paper, three salient image algorithms are compared and analyzed, and the salient effect of salient map is evaluated from subjective and objective aspects by combining the characteristics of remote sensing images and aircraft targets in remote sensing images. Considering the accuracy of salience and the degree of salience integrity, Itti algorithm is superior to GBVS algorithm and SR algorithm in highlighting small targets in remote sensing images. Through the simulation experiment on the algorithm of combining salient map and template matching of remote sensing image, it is proved that the target detection based on salient map algorithm can discard more than 70% background area. Reducing target search space is an effective way to improve the efficiency of target detection algorithm. Finally, based on the stochastic forest target detection algorithm, the voting weight is distributed equally in the Hough voting stage, so the average weight is not accurate enough to predict the target position. In this paper, the average weight of the original algorithm based on random forest leaf node information is changed to an exponential distribution voting weight based on sample target feature dictionary. In the training stage, the target pixel feature vector of the sample is stored as a sample dictionary, and the Euclidean distance between the pixel feature and the dictionary feature is calculated in the detection stage. According to the distance, the corresponding exponential function is calculated to distribute the weight, and the target center is voted. The characteristics of Hough image before and after the improved voting weight are analyzed by simulation experiments, and it is verified that the improved algorithm can improve the performance of the target detection algorithm.
【学位授予单位】：西南交通大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP751

【参考文献】