航空遥感光电图像预处理与目标特征提取技术研究

RESEARCH ON AERIAL REMOTE SENSING PHOTOELECTRIC IMAGE PREPROCESSING AND TARGETFEATURE EXTRACTION TECHNOLOGY

李晓峰

11749193

硕士

计算机技术

韩松

2019-05-30

2019-06-28

哈尔滨工业大学

深圳

遥感技术在近几十年来得到了高速的发展，遥感光电图像资源正成为宝贵的财富。如何科学、高效地处理遥感图像，使之多维度、多信息的优势得到充分发挥，是当前研究的重点。一方面，对于遥感图像预处理的需求依然旺盛。飞行器在高空采集的图像常常受到云雾的干扰，图像质量下降，目标模糊不清晰，不利于测绘、侦察等任务的开展。另一方面，传统的先验特征提取法的目标识别准确率不高，在实际应用场景中起到的作用十分有限。 深度学习神经网络的出现给问题的解决提供了新思路。本文基于图像去雾算法和卷积神经网络方法，围绕遥感图像预处理和特征提取技术中存在的问题，开展相关研究，并提出创新： （1）目前基于模型的去雾算法虽然能够从大气散射原理上出发去除云雾，但对于天空等大面积的高亮度区域表现不佳，容易出现亮度过大等失真问题，且图像整体对比度难以调控；基于非模型的方法计算量较小，处理速度更快，但方法适用性不足，无法满足多场景下的效果需求。针对这两种算法的特点，本文提出了基于灰度损失函数的图像去雾算法，改进并设计灰度损失函数模型，将图像去雾前后的灰度进行量化对比，对图像去雾的程度和效果实现定量评价。该算法将模型法与非模型法的特点结合起来，充分发挥图像复原和图像增强方法各自的优势。通过实验数据对比，能够证明本算法对图像去雾具有实际意义。 （2）在传统图像特征提取领域，通常基于先验的知识提取图像目标特征，例如纹理、边缘、颜色特征。基于先验的方法占用的计算资源较少，已经能够在部分飞行器等部分机载平台上使用。随着应用需求的不断提高和应用场景的扩大，传统方法已显得不从心。近年来，深度学习神经网络正成为图像特征提取的主流。本文在卷积神经网络现有的研究基础上，提出了轻量化的遥感图像特征提取网络，将图像的深层特征与浅层特征相融合，降低了特征在网络传输中的风险损失。同时，应用传统图像特征提取方法，对图像进行多尺度、多方向Gabor变换，得到Gabor纹理特征向量。将先验特征与卷积神经网络特征有机融合，减轻深层网络带来的计算负担，并有效提升提取到特征的可靠性。通过大量数据集的测试验证，本算法对图像特征提取具有优化提升效果。

Remote sensing technology has developed at a high speed in recent years, and the resources of remote sensing photoelectric images are becoming a valuable asset. Nowadays researchers are focusing on making good deal with remote sensing images scientifically and efficiently, so that the advantage of multi-dimensional information can be fully taken. On the one hand, there still has great demand for remote sensing image preprocessing. The images collected by aircrafts at high altitude are often interfered by clouds and fog, which degrade the image quality and make the target unclear, affect the tasks like geographic mapping and reconnaissance. On the other hand, traditional image feature extraction method which based on prior knowledge is limited in most scenarios and can’t meet the needs of tasks. Neural network in deep learning provides a new solution for solving problems above. This paper focuses on the remote sensing image’s preprocessing and feature extraction, proposes innovative method based on the image dehazing and convolutional neural network: (1) At present, the model-based dehazing algorithm can remove the cloud based on the principle of atmospheric scattering, but it does not perform well in bright areas such as the sky, which easily leading to image distortion, and the contrast of the image is difficult to control. The non-model method’s calculation amount is small, with the processing speed is faster than model-based. But the applicability of non-model method is insufficient in multiple different scenarios. In order to solve the problem, this paper proposes a improved image dehazing algorithm that make use of the image’s pixel loss function. The pixel loss between hazed image and dehazed image is used to be an evaluation standard for dehazing quality. The algorithm combines the model-based method with the non-model method to take the advantages of image restoration and image enhancement. The experimental data of several algorithm can prove that the algorithm proposed has significance for image dehazing. (2) In the field of traditional image feature extraction, image target features such as texture, edge, and color features are usually extracted based on priori knowledge. These priori methods consume less computing resources and has been able to be used on some airborne platforms such as drone. As the demand of image processing grows and market expands, the traditional methods can’t be satisfactory any more. Deep learning and neural network are becoming the mainstream in image feature extraction research. Based on the existing research of convolutional neural network, this paper proposes a lightweight remote sensing image feature extraction network, which combines the deep abstract features with the shallow figurative features, reduces the loss of the feature in network transmission process. At the same time, the algorithm extracts the multi-scale and multi-directional Gabor feature vector using the traditional priori-based method. Compared with the deep convolutional neural networks, the network which is integrated with Gabor feature can alleviate the computational burden. The algorithm is proved to perform better than single convolution neural network through the experiments which include large number of data sets.

图像去雾 特征提取 卷积神经网络 Gabor变换

image dehazing feature extraction convolutional neural network Gabor transform

中文

联合培养

南方科技大学

李晓峰. 航空遥感光电图像预处理与目标特征提取技术研究[D]. 深圳. 哈尔滨工业大学,2019.