Vidya B. Manian

Improving Hyperspectral Image Classification Using Spatial Preprocessing

Spatial smoothing over the original hyperspectral data based on wavelet and anisotropic partial differential equations is incorporated using composite kernel in graph-based classifiers. The kernels combine spectral-spatial relationships using the smoothed and original hyperspectral images. Experiments with different real hyperspectral scenarios are presented. Comparison with recent graph-based methods shows that the proposed scheme gives better classification with lower computational cost.

SCALED AND ROTATED TEXTURE CLASSIFICATION USING A CLASS OF BASIS FUNCTIONS

Abstract Three classes of basis functions are considered for classifying scaled and rotated textured images. The first is the orthonormal, compactly supported Daubechies and the discrete Haar bases, the second is the biorthogonal basis and the third is the non orthogonal Gabor basis. Textures are scaled and rotated and the basis functions are used to expand them. Features are computed on a combination of inter-resolution coefficients. Experimental results show that the Daubechies orthonormal basis perform well in recognizing transformed textures, followed by the Haar basis. The concept of multiresolution representation and orthogonality are shown to be useful for invariant texture classificaiton.

Texture classification using logical operators

In this paper, a new algorithm for texture classification based on logical operators is presented. Operators constructed from logical building blocks are convolved with texture images. An optimal set of six operators are selected based on their texture discrimination ability. The responses are then converted to standard deviation matrices computed over a sliding window. Zonal sampling features are computed from these matrices. A feature selection process is applied and the new set of features are used for texture classification. Classification of several natural and synthetic texture images are presented demonstrating the excellent performance of the logical operator method. The computational superiority and classification accuracy of the algorithm is demonstrated by comparison with other popular methods. Experiments with different classifiers and feature normalization are also presented. The Euclidean distance classifier is found to perform best with this algorithm. The algorithm involves only convolutions and simple arithmetic in the various stages which allows faster implementations. The algorithm is applicable to different types of classification problems which is demonstrated by segmentation of remote sensing images, compressed and reconstructed images and industrial images.

Biodiversity Assessment Using Hierarchical Agglomerative Clustering and Spectral Unmixing over Hyperspectral Images

Hyperspectral images represent an important source of information to assess ecosystem biodiversity. In particular, plant species richness is a primary indicator of biodiversity. This paper uses spectral variance to predict vegetation richness, known as Spectral Variation Hypothesis. Hierarchical agglomerative clustering is our primary tool to retrieve clusters whose Shannon entropy should reflect species richness on a given zone. However, in a high spectral mixing scenario, an additional unmixing step, just before entropy computation, is required; cluster centroids are enough for the unmixing process. Entropies computed using the proposed method correlate well with the ones calculated directly from synthetic and field data.

Object segmentation in hyperspectral images using active contours and graph cuts

The interest in object segmentation on hyperspectral images is increasing and many approaches have been proposed to deal with this area. In this project, we developed an algorithm that combines both the active contours and the graph cut approaches for object segmentation in hyperspectral images. The active contours approach has the advantage of producing subregions with continuous boundaries. The graph cut approach has emerged as a technique for minimizing energy functions while avoiding the problems of local minima. Additionally, it guarantees continuity and produces smooth contours, free of self-crossing and uneven spacing problems. The algorithm uses the complete spectral signature of a pixel and also considers spatial neighbourhood for graph construction, thereby combining both spectral and spatial information present in the image. The algorithm is tested using real hyperspectral images taken from a variety of sensors, such as Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) and Hyperspectral Data Imagery Collection Experiment (HYDICE), and also taken by the SOC hyperspectral camera. This approach can segment different objects in an image. This algorithm can be applied in many fields and it should represent an important advance in the field of object segmentation.

Approaches to color- and texture-based image classification

A Gabor filtering method for texture-based classification of color images is presented. The algorithm is robust and can be used with different color representations. It involves a filter selection process based on texture smoothness. Unichannel and interchannel correlation features are computed. Two types of color representations have been considered: computing chromaticity values from xyY, HIS, and RGB spaces; and using the three channels of the perceptually uniform color spaces L*a*b* and HSV. The RGB space universally used in image processing can be used for color-texture-based classification by computing the rgb chromaticity values, which yield higher classification accuracies than the direct use of R, G, and B values. The wavelet transform methods have been extended to perform color image classifications with additional features. The two approaches, Gabor filtering and wavelet transform methods, are compared in terms of classification accuracy and efficiency. The pyramid wavelet transform (PWT) performs well with all color spaces. The tree-structured wavelet transform (TWT) is more suitable for smaller classification problems. The best color spaces for higher class problems with wavelet methods are L*a*b* and HSV spaces. The HSV space is found to be the best for application of both of these texture-based approaches. Computationally the Gabor method followed by PWT is fast and efficient.

Improving Hyperspectral Image Classification based on Graphs using Spatial Preprocessing

Spatial smoothing over the original hyperspectral data based on wavelet and partial differential equations (PDEs) are incorporated in the classifiers using composite kernel with kNN graphs. The kernels combine spectral-spatial relationships using the smoothed and original images. Experiments with real hyperspectral scenarios are presented. Comparison with recent graph based methods show that the proposed scheme improves existing methods.

Land cover and benthic habitat classification using texture features from hyperspectral and multispectral images

A texture-based method for classification of land cover and benthic habitats from hyperspectral and multispectral images is presented. The features considered in this work are a set of statistical and multiresolution texture features, including a 2-level wavelet transform that uses an orthonormal Daubechies filter. These features are computed over spatial extents from each band of the image. A stepwise sequential feature selection process is applied that results in the selection of optimal features from the original feature set. A supervised classification is performed with a distance metric. Results with AVIRIS hyperspectral and IKONOS multispectral images show that texture features perform well under different land cover scenarios and are effective in characterizing the texture information at different wavelengths. Results over coastal regions show that wavelet texture features computed over the reflectance spectrum can accurately detect the benthic classes.

Accelerating hyperspectral manifold learning using graphical processing units

Manifold learning has been widely studied in pattern recognition, image processing, and machine learning. A large number of nonlinear manifold learning methods have been proposed attempting to preserve a different geometrical property of the underlying manifold. In contrast, its application to hyperspectral images is computationally difficult due to the calculation of distances among spectral values in high-dimensional spaces. This paper compares feature extraction algorithms using isomap, Laplacian Eigenmaps, and local linear embedding in real hyperspectral images. They are implemented using massively parallel general purpose Graphical Processor Units (GPUs) to speed up computation. Their performance in classification of hyperspectral images and speed up of their computation is presented. Results using real and synthetic hyperspectral scenarios are presented. Additionally, a formulation including spatial information in these manifold learning algorithms is presented.

Hyperspectral image classification using spectral histograms and semi-supervised learning

In this paper, an algorithm that extracts regional texture information by computing spectral difference histograms over window extents in hyperspectral images is presented. The spectral angle distance is used as the spectral metric and different window sizes are explored for computing the histogram. The histograms are used in a semi-supervised learning framework that uses both labeled and unlabeled samples for training the support vector machine classifier, which is then tested with unlabeled samples. Results are presented with real and synthetic hyperspectral images. The method performs well with high spatial resolution images. The algorithm performs well under different noise levels.