卷积神经网络在遥感目标识别中的应用研究

发布时间：2018-04-28 14:27

本文选题：目标识别 + 高空间分辨率遥感影像　；参考：《中国科学院大学(中国科学院遥感与数字地球研究所)》2017年硕士论文

【摘要】：遥感目标识别旨在定位并识别出遥感影像中的感兴趣目标,它是遥感图像处理的核心问题之一。随着搭载平台与传感器技术的飞速发展,遥感影像空间分辨率不断提高,与自然图像在视觉上的差异也越来越小。越来越多的计算机视觉方法可被应用于高空间分辨率遥感影像目标识别。2012年,Krizhevsky等首次将卷积神经网络应用于大规模自然图像分类,分类精度显著优于传统的图像分类方法,从而掀起了卷积神经网络的研究热潮。2014年,Girshick等提出了R-CNN(Regions with Convolutional Neural Network Features)模型,奠定了使用卷积神经网络进行目标识别的基础。目前,基于卷积神经网络的自然图像目标识别算法,在精度上已经接近了人眼的识别精度,速度上也达到了准实时处理。遥感图像空间分辨率的提高以及卷积神经网络在自然图像分类、目标识别等领域取得的成功,极大地推动了卷积神经网络在遥感图像处理领域的应用研究。基于卷积神经网络的高分辨率遥感影像分类、目标识别、图像分割等研究已成为当前研究热点之一。本文主要研究卷积神经网络在高空间分辨率遥感影像目标识别中的应用,完成了以下研究工作:(1)明确了卷积神经网络特征在图像识别中的有效性。本文在相同的遥感影像场景分类框架下,对比了卷积神经网络和词袋模型两种方法提取的特征性能。使用卷积神经网络特征在UcMerced数据集上分类时,平均准确率均值达到了87%,在WhU数据集上优于97%,显著高于词袋模型,证明了卷积神经网络特征在图像识别领域比传统人工特征更具有优势。(2)提出了一种由粗到精的高空间分辨遥感影像单类目标识别方法。该方法首先采用基于HOG的级联AdaBoost算法快速提取少量的感兴趣区域,然后使用基于卷积神经网络特征的支持向量机对感兴趣区域进行分类。本文提出的方法在飞机识别实验中准确率和识别率分别为98.4%和87.1%。与R-CNN模型相比,本文提出的方法不但准确率和识别率有所提高,而且速度是R-CNN模型的10倍,在实际应用中具有重要意义。(3)研究了如何使用Faster-RCNN模型进行遥感多类目标识别。通过固定卷积神经网络低层卷积层和微调高层卷积层的方式,缩小了Faster-RCNN模型需要训练的参数规模。在NWPU VHR-10数据集上,使用Faster R-CNN模型进行遥感多类目标识别的平均准确率均值达到了72.6%,说明了Faster R-CNN是一种有效的遥感多类目标识别方法。本文研究结果表明,卷积神经网络在遥感影像目标识别中具有巨大的应用潜力,是当前和未来的研究热点。
[Abstract]:Remote sensing target recognition, which is one of the core problems in remote sensing image processing, aims to locate and identify interested objects in remote sensing images. With the rapid development of platform and sensor technology, the spatial resolution of remote sensing image is increasing, and the visual difference between remote sensing image and natural image is becoming smaller and smaller. More and more computer vision methods can be applied to target recognition of high spatial resolution remote sensing images. In 2012, convolution neural network was applied to large-scale natural image classification for the first time, and the classification accuracy was significantly better than that of traditional image classification methods. The R-CNN(Regions with Convolutional Neural Network Features (R-CNN(Regions with Convolutional Neural Network Features) model was proposed by Girshick in 2014, which laid the foundation for target recognition using convolutional neural networks. At present, the natural image target recognition algorithm based on convolution neural network has approached the recognition accuracy of human eyes and achieved quasi-real-time processing in speed. The improvement of spatial resolution of remote sensing image and the success of convolution neural network in natural image classification and target recognition greatly promote the application of convolution neural network in remote sensing image processing. The research of high resolution remote sensing image classification, target recognition and image segmentation based on convolution neural network has become one of the research hotspots. In this paper, the application of convolution neural network in high spatial resolution remote sensing image target recognition is studied. The following research work is done: 1) the effectiveness of convolution neural network feature in image recognition is clarified. In this paper, under the same remote sensing image scene classification framework, the feature performance of convolution neural network and word bag model extraction is compared. When the convolutional neural network features are used to classify UcMerced data sets, the average accuracy is 87% and 97% in WhU data sets, which is significantly higher than the word bag model. It is proved that convolutional neural network features have more advantages than traditional artificial features in image recognition. Firstly, the concatenated AdaBoost algorithm based on HOG is used to extract a small number of regions of interest, and then SVM based on convolution neural network features is used to classify the regions of interest. The accuracy and recognition rate of the proposed method are 98.4% and 87.1%, respectively. Compared with the R-CNN model, the method proposed in this paper not only improves the accuracy and recognition rate, but also is 10 times faster than the R-CNN model. It is of great significance in practical application. By means of fixed convolution layer and fine-tuned high-level convolution layer, the parameter scale of Faster-RCNN model need to be trained is reduced. On the NWPU VHR-10 data set, the average accuracy of multi-class remote sensing target recognition based on Faster R-CNN model is 72.6, which shows that Faster R-CNN is an effective method for multi-class remote sensing target recognition. The research results show that convolutional neural network has great application potential in remote sensing image target recognition, and it is a hot research topic in the future.
【学位授予单位】：中国科学院大学(中国科学院遥感与数字地球研究所)
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：P237;TP751

【相似文献】