Hyperspectral image classification using graph-based wavelet transform

ABSTRACT

Graph-based methods are developed to efficiently extract data information. In particular, these methods are adopted for high-dimensional data classification by exploiting information residing on weighted graphs. In this paper, we propose a new hyperspectral texture classifier based on graph-based wavelet transform. This recent graph transform allows extracting textural features from a constructed weighted graph using sparse representative pixels of hyperspectral image. Different measurements of spectral similarity between representative pixels are tested to decorrelate close pixels and improve the classification precision. To achieve the hyperspectral texture classification, Support Vector Machine is applied on spectral graph wavelet coefficients. Experimental results obtained by applying the proposed approach on Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) and Reflective Optics System Imaging Spectrometer (ROSIS) datasets provide good accuracy which could exceed 98.7%. Compared to other famous classification methods as conventional deep learning-based methods, the proposed method achieves better classification performance. Results have shown the effectiveness of the method in terms of robustness and accuracy.     

结合倒置特征金字塔和U-Net的高光谱图像分类

目的 地物分类是对地观测研究领域的重要任务。高光谱图像具有丰富的地物光谱信息,可用于提升遥感图像地物分类的准确度。如何对高光谱图像进行有效的特征提取与表示是高光谱图像分类应用的关键问题。为此,本文提出了一种结合倒置特征金字塔和U-Net的高光谱图像分类方法。方法 对高光谱数据进行主成分分析（principal component analysis,PCA）降维,获取作为网络输入的重构图像数据,然后使用U-Net逐层提取高光谱重构图像的空间特征。与此同时,利用倒置的特征金字塔网络抽取相应层级的语义特征;通过特征融合,得到既有丰富的空间信息又有较强烈的语义响应的特征表示。提出的网络利用注意力机制在跳跃连接过程中实现对背景区域的特征响应抑制,最终实现了较高的地物分类精度。结果 分析了PCA降维方法和输入数据尺寸对分类性能的影响,并在Indian Pines、Pavia University、Salinas和Urban数据集上进行了对比实验。本文方法在4个数据集上分别取得了98.91%、99.85%、99.99%和87.43%的总体分类精度,与支持向量机（support vector machine,SVM）等相关算法相比,分类精度高出1%~15%。结论 本文提出一种结合倒置特征金字塔和U-Net的高光谱图像分类方法,可以应用于有限训练样本下的高光谱图像分类任务,并在多个数据集上取得了较高的分类精度。实验结果表明倒置特征金字塔结构与U-Net结合的算法能够高效地实现高光谱图像的特征提取与表示,从而获得更精细的分类结果。     

Hyperspectral image classification via compact-dictionary-based sparse representation

Multimedia Tools and Applications - In this paper, a compact-dictionary-based sparse representation (CDSR) method is proposed for hyperspectral image (HSI) classification. The proposed dictionary...     

增强类可分性的高光谱图像分类

目的 在高光谱图像分类中,由于成像空间分辨率较低,混合像元大量存在。混合像元使得不同类别的光谱特征发生改变,失去原有的独特性,类内差异变大,类间差异变小。针对这一问题,本文提出基于分组滚动引导滤波的策略。同时针对高光谱图像中存在的“维数灾难”问题,提出了弹性网逻辑回归分类器的框架。方法 通过线性判别分析（linear discriminant analysis,LDA）算法生成具有判别性的引导图,对高光谱图像的每个波段执行滚动引导,从而让光谱曲线呈现类内凝聚、类间距离增大的趋势。通过构造逻辑回归目标函数的L1&L2范数正则项约束进行嵌入式波段选择,为每个类别选择出各自可分性强的波段,同时可以使高度相关性的波段保留下来作为分类依据。最后使用邻域优化策略对分类后结果进一步优化,提升分类精度。结果 分别在3个实验数据集上与其他分类算法进行对比,实验结果表明,本文算法的分类结果取得明显提升。本文算法的总体分类精度（overall accuracy,OA）在Indian Pines、Salinas和KSC （Kennedy Space Center）数据集上分别为96.61%、98.66%和99.04%,比其他算法平均分别高出4.8%、3%和1%,同时也在Indina Pines数据集中进行了对比实验以验证增强混合像元光谱可分性和波段可分性算法的有效性,对比实验结果表明本文算法改善了分类效果。结论 分别在光谱特性和波段选择两个环节增强类可分性,分类精度取得明显提升;同时,本文算法适合不同的数据集,并且在不同数量的训练样本下OA均表现较优,算法具有一定的鲁棒性。     

跨数据集评估的高光谱图像分类

目的 随着高光谱成像技术的飞速发展,高光谱数据的应用越来越广泛,各场景高光谱图像的应用对高精度详细标注的需求也越来越旺盛。现有高光谱分类模型的发展大多集中于有监督学习,大多数方法都在单个高光谱数据立方中进行训练和评估。由于不同高光谱数据采集场景不同且地物类别不一致,已训练好的模型并不能直接迁移至新的数据集得到可靠标注,这也限制了高光谱图像分类模型的进一步发展。本文提出跨数据集对高光谱分类模型进行训练和评估的模式。方法 受零样本学习的启发,本文引入高光谱类别标签的语义信息,拟通过将不同数据集的原始数据及标签信息分别映射至同一特征空间以建立已知类别和未知类别的关联,再通过将训练数据集的两部分特征映射至统一的嵌入空间学习高光谱图像视觉特征和类别标签语义特征的对应关系,即可将该对应关系应用于测试数据集进行标签推理。结果 实验在一对同传感器采集的数据集上完成,比较分析了语义—视觉特征映射和视觉—语义特征映射方向,对比了5种基于零样本学习的特征映射方法,在高光谱图像分类任务中实现了对分类模型在不同数据集上的训练和评估。结论 实验结果表明,本文提出的基于零样本学习的高光谱分类模型可以实现跨数据集对分类模型进行训练和评估,在高光谱图像分类任务中具有一定的发展潜力。     

应用级联多分类器的高光谱图像分类

目的 高光谱分类任务中,由于波段数量较多,图像中存在包含噪声以及各类地物样本分布不均匀等问题,导致分类精度与训练效率不能平衡,在小样本上分类精度低。因此,提出一种基于级联多分类器的高光谱图像分类方法。方法 首先采用主成分分析方法将高度相关的高维特征合成无关的低维特征,以加快Gabor滤波器提取纹理特征的速度;然后使用Gabor滤波器提取图像在各个尺寸、方向上的纹理信息,每一个滤波器会生成一张特征图,在特征图中以待分类样本为中心取一个d×d的邻域,计算该邻域内数据的均值和方差来作为待分类样本的空间信息,再将空间信息和光谱信息融合,以降低光线与噪声的影响;最后将谱—空联合特征输入级联多分类器中,得到预测样本关于类别的概率分布的平均值。结果 实验采用Indian Pines、Pavia University和Salinas 3个数据集,与经典算法如支持向量机和卷积神经网络进行比较,并利用总体分类精度、平均分类精度和Kappa系数作为评价标准进行分析。本文方法总体分类精度在3个数据集上分别达到97.24%、99.57%和99.46%,相对于基于径向基神经网络（RBF）核函数的支持向量机方法提高了13.2%、4.8%和5.68%,相对于加入谱—空联合特征的RBF-SVM （radial basis function-support vector machine）方法提高了2.18%、0.36%和0.83%,相对于卷积神经网络方法提高了3.27%、3.2%和0.3%;Kappa系数分别是0.968 6、0.994 3和0.995 6,亦有提高。结论 实验结果表明,本文方法应用于高光谱图像分类具有较优的分类效果,训练效率较高,无需依赖GPU,而且在小样本上也具有较高的分类精度。     

基于主动学习的高光谱图像分类方法

高光谱图像监督分类中,为了避免休斯效应需要大量的训练样本,但在实际应用中对样本进行标注成本非常高,因此,得到高质量的训练样本显得十分重要。提出一种基于主动学习的高光谱图像分类方法,通过对区域关注度的统计,有效地结合图像光谱和空间特性,基于主动学习方法获取信息量较大的训练样本,从而较大幅度提高了分类的精确度。实验结果表明,所提算法比传统的随机取样监督分类法和主动学习方法在分类精确度上有较大的优势。     

基于加权K近邻和卷积神经网络的高光谱图像分类

传统的谱空联合分类算法通常定义一个邻域空间作为空间信息,忽略空间中非邻域空间信息,且容易将异类像元也考虑在内。针对于高光谱图像分类问题,提出了一种加权K近邻算法能够自适应地提取空间信息,首先定义光谱和空间坐标组成的特征空间,利用该特征空间寻找目标像元的K个相似像元,并对这些像元根据特征空间进行加权;将加权后的像元按照一定方式组合成三维张量表示最终的谱空联合信息,使用三维卷积神经网络对其进行训练,得到最终分类结果。从实验结果来看,相对于改进前的算法,在总体分类精度上得到了一定的提升,与原始的三维卷积神经网络相比,在收敛速度上也得到大大提升,为高光谱图像的谱空联合分类提供了一种更加实用的方法。     

基于主动深度学习的高光谱影像分类

针对当前高光谱遥感影像分类人工标注样本费时费力,大量未标注样本未得到有效利用以及主要利用光谱信息而忽视空间信息等问题,提出了一种空-谱信息与主动深度学习相结合的高光谱影像分类方法。首先利用主成分分析对原始影像进行降维,在此基础上提取像素的一正方形小邻域作为该像素的空间信息并结合其原始光谱信息得到空谱特征。然后,通过稀疏自编码器得到原始数据的稀疏特征表达,并通过逐层无监督学习稀疏自编码器构建深度神经网络,输出原始数据的深度特征,将其连接到softmax分类器,利用少量标记样本以监督学习的方式完成模型的精调。最后,利用主动学习算法选择最不确定性样本对其进行标注,并加入至训练样本以提高分类器的分类效果。分别对PaviaU影像和PaviaC影像进行分类实验的结果表明,该方法在少量标记样本情况下,相对于传统方法能有效地提高分类精度。     

Fractional differentiation and non-Pareto multiobjective optimization for image thresholding

Various techniques have previously been proposed for single-stage thresholding of images to separate objects from the background. Although these global or local thresholding techniques have proven effective on particular types of images, none of them is able to produce consistently good results on a wide range of existing images. Here, a new image histogram thresholding method, called TDFD, based on digital fractional differentiation is presented for gray-level image thresholding. The proposed method exploits the properties of the digital fractional differentiation and is based on the assumption that the pixel appearance probabilities in the image are related. To select the best fractional differentiation order that corresponds to the best threshold, a new algorithm based on non-Pareto multiobjective optimization is presented. A new geometric regularity criterion is also proposed to select the best thresholded image. In order to illustrate the efficiency of our method, a comparison was performed with five competing methods: the Otsu method, the Kapur method, EM algorithm based method, valley emphasis method, and two-dimensional Tsallis entropy based method. With respect to the mode of visualization, object size and image contrast, the experimental results show that the segmentation method based on fractional differentiation is more robust than the other methods.     

Solving multiobjective problems using cat swarm optimization

This paper proposes a new multiobjective evolutionary algorithm (MOEA) by extending the existing cat swarm optimization (CSO). It finds the nondominated solutions along the search process using the concept of Pareto dominance and uses an external archive for storing them. The performance of our proposed approach is demonstrated using standard test functions. A quantitative assessment of the proposed approach and the sensitivity test of different parameters is carried out using several performance metrics. The simulation results reveal that the proposed approach can be a better candidate for solving multiobjective problems (MOPs).     

基于DS聚类的高光谱图像集成分类算法

针对高光谱遥感图像维数高、样本少导致分类精度低的问题,提出一种基于DS聚类的高光谱图像集成分类算法(DSCEA)。首先,根据高光谱数据特点,从整体波段中随机选择一定数量的波段,构成不同的训练样本;其次,分析图像的空谱信息,构造无向加权图,利用优势集(DS)聚类方法得到最大特征差异的波段子集;最后,根据不同样本,利用支持向量机训练具有差异的单个分类器,采用多数表决法集成最终分类器,实现对高光谱遥感图像的分类。在Indian Pines数据集上DSCEA算法的分类精度最高可达到84.61%,在Pavia University数据集上最高可达到91.89%,实验结果表明DSCEA算法可以有效的解决高光谱分类问题。     

基于空谱融合网络的高光谱图像分类方法

针对高光谱图像分类中提取的空-谱特征表达能力弱及维数较高的问题,提出一种基于空-谱融合网络（SSF-Net）的高光谱图像分类方法。首先,利用双通道卷积神经网络（Two-CNN）同时提取高光谱图像的光谱和空间特征;其次,使用多模态压缩双线性池化（MCB）将所提取的多模态特征向量的外积投射到低维空间,以此产生空-谱联合特征。该特征融合网络,既可以分析光谱特征和空间特征向量中元素之间的复杂关系,同时也避免对光谱和空间向量直接进行外积计算,造成维数过高、计算困难的问题。最终实验表明,与现有基于神经网络的分类方法相比,所提出的高光谱图像分类算法能够获得更高的像元分类精度,表明该网络所提取的空-谱联合向量对高光谱图像具有更强的特征表达能力。     

基于KMNF和BP神经网络的高光谱遥感影像分类

为了对高维非线性的高光谱影像进行降维及信息提取,提出了高光谱影像核最小噪声分离变换(kernel minimum noise fraction,KMNF)特征提取后利用BP神经网络分类的方法.以高光谱影像KMNF特征提取后的前几个特征分量作为BP神经网络的输入,进行BP神经网络分类,并与单独的高光谱影像BP神经网络分类进行比较.美国内华达州CUPRITE矿区AVIRIS数据的实验结果表明,基于KMNF和BP神经网络的高光谱影像分类较单独BP神经网络分类总体精度及时间性能均得到提高.     

Color image segmentation based on multiobjective artificial bee colony optimization

This paper presents a new color image segmentation method based on a multiobjective optimization algorithm, named improved bee colony algorithm for multi-objective optimization (IBMO). Segmentation is posed as a clustering problem through grouping image features in this approach, which combines IBMO with seeded region growing (SRG). Since feature extraction has a crucial role for image segmentation, the presented method is firstly focused on this manner. The main features of an image: color, texture and gradient magnitudes are measured by using the local homogeneity, Gabor filter and color spaces. Then SRG utilizes the extracted feature vector to classify the pixels spatially. It starts running from centroid points called as seeds. IBMO determines the coordinates of the seed points and similarity difference of each region by optimizing a set of cluster validity indices simultaneously in order to improve the quality of segmentation. Finally, segmentation is completed by merging small and similar regions. The proposed method was applied on several natural images obtained from Berkeley segmentation database. The robustness of the proposed ideas was showed by comparison of hand-labeled and experimentally obtained segmentation results. Besides, it has been seen that the obtained segmentation results have better values than the ones obtained from fuzzy c-means which is one of the most popular methods used in image segmentation, non-dominated sorting genetic algorithm II which is a state-of-the-art algorithm, and non-dominated sorted PSO which is an adapted algorithm of PSO for multi-objective optimization.     

Evolutionary multiobjective optimization using an outranking-based dominance generalization

One aspect that is often disregarded in the current research on evolutionary multiobjective optimization is the fact that the solution of a multiobjective optimization problem involves not only the search itself, but also a decision making process. Most current approaches concentrate on adapting an evolutionary algorithm to generate the Pareto frontier. In this work, we present a new idea to incorporate preferences into a multi-objective evolutionary algorithm (MOEA). We introduce a binary fuzzy preference relation that expresses the degree of truth of the predicate “x is at least as good as y”. On this basis, a strict preference relation with a reasonably high degree of credibility can be established on any population. An alternative x is not strictly outranked if and only if there does not exist an alternative y which is strictly preferred to x. It is easy to prove that the best solution is not strictly outranked. For validating our proposed approach, we used the non-dominated sorting genetic algorithm II (NSGA-II), but replacing Pareto dominance by the above non-outranked concept. So, we search for the non-strictly outranked frontier that is a subset of the Pareto frontier. In several instances of a nine-objective knapsack problem our proposal clearly outperforms the standard NSGA-II, achieving non-outranked solutions which are in an obviously privileged zone of the Pareto frontier.     

Computer aided design of regulators using multiobjective optimization

The problem of designing linear regulators subject to quadratic cost functions is addressed using multiobjective optimization. Instead of representing conflicting performance measures in a scalar objective function, these measures are treated individually in an attempt to simultaneously optimize the objectives. Although there is no unique solution the approach offers the designer the opportunity of investigating trade-offs in a systematic way. A software package, based on the goal attainment method, is presented and applied to design examples.     

Fuzzy multiobjective optimization of truss-structures using genetic algorithm

In this paper, a method for solving fuzzy multiobjective optimization of space truss with a genetic algorithm is proposed. This method enables a flexible method for optimal system design by applying fuzzy objectives and fuzzy constraints. The displacement, tensile stress, fuzzy sets, membership functions and minimum size constraints are considered in formulation of the design problem. An algorithm was developed by using MATLAB programming. The algorithm is illustrated on 56-bar space truss system design problem and the results are discussed.