光学遥感图像舰船目标检测技术的研究

发布时间：2018-12-18 20:05

【摘要】：舰船目标检测技术是遥感卫星图像处理与分析领域非常重要的课题,尤其对于高分辨率的光学遥感图像,其海量数据虽然提供了更加丰富的细节信息,但又严重制约了舰船目标的检测效率。因此,如何快速准确地获取舰船目标的位置信息已成为一个热点话题。针对光学遥感图像舰船目标的检测问题,本文重点研究了舰船目标候选区域提取和舰船目标的鉴别技术,使舰船目标检测的精度和效率得到了提高。本文主要工作包括以下几个方面:1、基于光学遥感图像的预处理,该环节主要包括图像的滤波、光照均衡化和去云雾干扰等步骤,其中重点研究了图像的光照均衡化处理和去云雾干扰等算法。该步骤旨在减弱噪声、光照不均匀和云雾等不利因素的干扰,更加突出目标信息。2、研究了舰船目标候选区域的提取方法,并提出了一种改进的PQFT舰船目标候选区域快速提取法。本文引入了PQFT方法,并在原有的PQFT算法中加入了小波变换,从不同尺度分析图像的显著特征,从中选择显著度最好的图像。实验表明,改进的PQFT显著性方法显著度要好于原尺度的效果,运算时间也得到了相应减少。3、为了判断所提取候选区域的目标是否为舰船目标,本文分别提取了舰船目标的形状特征、灰度特征、纹理特征和梯度方向直方图特征。并提出了一种改进的LBP特征提取算法,其抗干扰能力变强、计算复杂度降低,增加了LBP特征提取的可控性。通过融合舰船目标的多特征信息,能够更加准确地判别舰船目标和非舰船目标。4、将改进的极限学习机(Extreme Learning Machine,ELM)算法用于舰船目标的分类识别。ELM是一种神经网络算法,其特点为:网络是单隐藏层、隐藏层结点数人为设置、输入权值和隐藏层偏置随机产生,因此该方法具有计算时间短、泛化能力强和不易陷入局部最优等特点。而传统ELM算法的激活函数,如Sigmoid函数、sin函数和tanh函数等存在过饱和的缺点,本文提出了一种非线性修正的ELM算法,最后将本文改进的ELM算法用于舰船目标分类识别中。基于以上研究,运用MATLAB平台进行仿真验证,实验结果证明了以上改进算法的有效性,能够提高检测精度和检测效率。
[Abstract]:Ship target detection technology is a very important subject in the field of remote sensing satellite image processing and analysis, especially for high-resolution optical remote sensing image, its massive data provide more detailed information. However, the efficiency of ship target detection is seriously restricted. Therefore, how to obtain the position information of ship target quickly and accurately has become a hot topic. Aiming at the problem of ship target detection in optical remote sensing image, this paper focuses on the extraction of ship target candidate region and the identification technology of ship target, which improves the accuracy and efficiency of ship target detection. The main work of this paper includes the following aspects: 1. Based on the preprocessing of optical remote sensing image, this link mainly includes image filtering, illumination equalization and cloud and fog removal, etc. The algorithms of image illumination equalization and cloud-free interference are studied. The purpose of this step is to attenuate the interference of unfavorable factors, such as noise, uneven illumination, cloud and fog, and to highlight the target information. 2. An improved PQFT ship target candidate region extraction method is proposed. In this paper, the PQFT method is introduced, and wavelet transform is added to the original PQFT algorithm to analyze the salient features of the image from different scales and select the image with the best saliency. The experimental results show that the improved PQFT saliency method is more significant than the original scale, and the computational time is reduced accordingly. In order to judge whether the target of the candidate region extracted is a ship target, In this paper, the shape feature, grayscale feature, texture feature and gradient direction histogram feature of ship target are extracted. An improved LBP feature extraction algorithm is proposed, which has the advantages of strong anti-interference ability and low computational complexity, and increases the controllability of LBP feature extraction. By integrating the multi-feature information of ship target, we can distinguish ship target from non-ship target more accurately. 4. The improved extreme learning machine (Extreme Learning Machine, will be improved. ELM is a neural network algorithm, which is characterized by: the network is a single hidden layer, the number of hidden layer nodes are set, the input weight and the hidden layer bias are generated randomly. Therefore, this method has the advantages of short calculation time, strong generalization ability and difficulty to fall into local optimum. However, the activation function of traditional ELM algorithm, such as Sigmoid function, sin function and tanh function, has the disadvantage of supersaturation. In this paper, a nonlinear modified ELM algorithm is proposed, and the improved ELM algorithm is applied to the classification and recognition of ship targets. Based on the above research, the MATLAB platform is used for simulation verification. The experimental results show that the improved algorithm is effective and can improve the detection accuracy and efficiency.
【学位授予单位】：东华大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP751

【参考文献】