基于深度学习的遥感图像检索方法研究

发布时间：2019-01-04 22:59

【摘要】：遥感图像数据的海量性、多样性和复杂性等特点对遥感图像检索的速度和精度提出了更高的要求,基于内容的遥感图像检索方法是近年来研究的重点,其中特征提取方法是影响遥感图像检索效果的关键。传统的特征提取方法多是提取遥感图像的底层视觉特征来进行检索,存在着底层特征难以表述图像语义信息的缺点。本文提出基于深度学习的遥感图像检索方法,通过神经网络的训练,建立图像底层特征和高层语义之间的映射关系,采用两种不同的深度学习方法进行遥感图像检索实验:(1)提出一种基于深度学习的半监督遥感图像检索方法,该方法首先对遥感图像进行预处理;然后基于稀疏自动编码的方法在大量未标注的遥感图像上进行特征学习得到特征字典,并基于卷积神经网络的思想,使用训练出来的特征字典对遥感图像进行卷积和池化得到每幅图像的特征图;接下来使用特征图训练Softmax分类器;最后对待检索图像分类,在同一类别中计算特征间的距离,进而实现遥感图像的检索。实验结果表明,该方法能够有效提高遥感图像检索的速度和准确度。(2)提出一种基于卷积神经网络的遥感图像检索方法,该方法基于深度学习keras框架实现,包括CNN特征提取网络层和Softmax分类层,通过深度较大的卷积神经网络提取遥感图像的高层语义特征,引入dropout层提高模型的泛化能力,进一步提高了遥感图像的检索精度。本文的方法(1)在遥感图像检索实验中取得的检索准确度为90.6%,检索时间为7.1844s,本文的方法(2)在遥感图像检索试验中取得的检索准确度为98.8%,检索时间为9.138s,本文方法也存在以下不足:当待检索图像分类错误时,检索准确度较低。
[Abstract]:The characteristics of mass, diversity and complexity of remote sensing image data put forward higher requirements for the speed and accuracy of remote sensing image retrieval. Content-based remote sensing image retrieval method is the focus of research in recent years. The feature extraction method is the key to the effect of remote sensing image retrieval. Most of the traditional feature extraction methods are to extract the underlying visual features of remote sensing images for retrieval, which has the disadvantage that the underlying features are difficult to express the semantic information of the image. In this paper, a method of remote sensing image retrieval based on deep learning is proposed. Through the training of neural network, the mapping relationship between image bottom features and high-level semantics is established. Two different depth learning methods are used for remote sensing image retrieval. (1) A semi-supervised remote sensing image retrieval method based on depth learning is proposed. Then the feature dictionary is obtained by feature learning on a large number of unlabeled remote sensing images based on sparse automatic coding method and based on the idea of convolution neural network. Using the trained feature dictionary to convolution and pool the remote sensing image, the feature map of each image is obtained. Then the feature map is used to train the Softmax classifier, and finally to classify the retrieval image, the distance between the features is calculated in the same category, and then the remote sensing image retrieval is realized. The experimental results show that this method can effectively improve the speed and accuracy of remote sensing image retrieval. (2) A remote sensing image retrieval method based on convolution neural network is proposed. It includes CNN feature extraction network layer and Softmax classification layer. The high level semantic feature of remote sensing image is extracted by convolution neural network. The generalization ability of remote sensing image is improved by introducing dropout layer, and the retrieval accuracy of remote sensing image is further improved. Methods in this paper (1) the retrieval accuracy in remote sensing image retrieval experiment is 90.6 and the retrieval time is 7.1844 s, and the retrieval accuracy in remote sensing image retrieval experiment is 98.8. The retrieval time is 9.138s, and this method also has the following shortcomings: when the image classification error is to be retrieved, the retrieval accuracy is low.
【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】：TP751

【参考文献】