All Title Author
Keywords Abstract

-  2018 

基于CNN的无人机遥感影像质量评价

DOI: 10.13360/j.issn.2096-1359.2018.05.019

Keywords: 卷积神经网络, 遥感图像, 质量评价, 无人机影像, 深度学习
convolution neural network
, remote sensing images, quality evaluation, UAV images, deep learning

Full-Text   Cite this paper   Add to My Lib

Abstract:

运用无人机的遥感影像来调查林地状态是一种有效的途径,为了进一步提升遥感图像质量的评价精度,笔者提出了一种基于卷积神经网络(convolutional neural network, CNN)的无人机遥感图像质量评价方法,主要包括图像采集与预处理、数据扩增、模型训练和测试4个阶段。首先对无人机采集到的遥感图像进行主观质量打分,分别获取同一区域不同阶段图像的质量分数; 然后运用图像旋转和剪裁等方法对遥感图像进行数据扩增,将扩增后的图片和原始图片融合作为实验数据集; 其次在Caffe深度学习框架中构建基于CNN深层特征的回归模型,并训练; 最后,根据已建立好的深度回归模型和学习到的参数,预测无人机遥感图像的质量分数。结果表明,提出的方法可以取得较准确的评分效果,在保证客观打分的同时,能基本保持和人眼视觉的感受一致。
Remote sensing images processing for Unmanned Aerial Vehicle(UAV)is an effective approach in woodland survey, and convolution neural network(CNN)is one of the most representative techniques of deep learning methods. The quality evaluation accuracy of remote sensing images could be improved using these technologies. In this paper, a novel method for quality evaluation of the UAV remote sensing images is presented, which was based on the CNN. The processing phase was divided into following four stages, i.e., image acquisition and preprocessing, data augmentation, CNN model training and performance testing. In particular, firstly quality scores for the collected remote sensing images were created, which were from different stages of the same region. The quality score of these images was divided into following 5 grades, namely, very good, good, general, poor and bad, and their corresponding scores were 5, 4, 3, 2 and 1, respectively. The scoring results established by the 10 experts were further processed, and the Mean Opinion Score(MOS)was taken as the final quality scores of these images by removing the obvious discrete values. Then, remote sensing images were augmented using rotation and clipping, and each remote sensing image was rotated clockwise before the forward shear. Augmented images and the original images were fused together as the experimental data in this work. After that, the fused images were randomly divided into the training set, the validation set, and the testing set with the proportion of 10:1:1. Then the regression model based on the CNN hierarchy features and subjective quality scores was constructed in the Caffe deep learning framework. The training set and the validation set were used to train the established CNN model and calculate the training performance. Finally, according to the trained regression CNN model and parameters, the testing set was used to calculate the testing accuracy by the real quality scores and regression values. Experimental results showed that the proposed method can achieve a high-quality scoring accuracy, meanwhile can maintain the objective scoring and almost achieve the same visual perception as the human. The accuracy and universality of the

Full-Text

comments powered by Disqus

Contact Us

service@oalib.com

QQ:3279437679

微信:OALib Journal