Hyperspectral image (HSI) is a three-dimensional data cube containing two spatial information dimensions and one spectral information dimension. The spectral vectors of different classes may have similar tendency and value that may bring about negative influences on classification. It is, therefore, important to introduce signal preprocessing techniques in the spatial domain to improve classification accuracy of HSIs. Assuming that local pixels in HSI have some correlations with each other, this paper proposes a spatial filtering model based on adaptive manifold (AM) for HSI. The AM for spatial filtering emphasizes the similar neighboring pixels and is robust to resist the noisy points with fast speed. The rich information in the filtered data is effective for improving the performance of the subsequent classification. The filtered data are classified by an extreme learning machine (ELM). The experimental results indicate that the framework built based on AM and ELM provides competitive performance. Specifically, by classifying the filtered data, the average accuracy of ELM can be improved as high as 30.54%, while performing tens to hundreds times faster than those state-of-the-art classifiers.
Although hyperspectral images (HSIs) captured by satellites provide much information in spectral regions, some bands are redundant or have large amounts of noise, which are not suitable for image analysis. To address this problem, we introduce a method for reconstructing the HSI with noise reduction and contrast enhancement using a matting model for the first time. The matting model refers to each spectral band of an HSI that can be decomposed into three components, i.e., alpha channel, spectral foreground, and spectral background. First, one spectral band of an HSI with more refined information than most other bands is selected, and is referred to as an alpha channel of the HSI to estimate the hyperspectral foreground and hyperspectral background. Finally, a combination operation is applied to reconstruct the HSI. In addition, the support vector machine (SVM) classifier and three sparsity-based classifiers, i.e., orthogonal matching pursuit (OMP), simultaneous OMP, and OMP based on first-order neighborhood system weighted classifiers, are utilized on the reconstructed HSI and the original HSI to verify the effectiveness of the proposed method. Specifically, using the reconstructed HSI, the average accuracy of the SVM classifier can be improved by as much as 19%.
In this paper, an automatic fractional coefficient setting method of fractional-order Darwinian particle swarm optimization (FODPSO) is proposed for hyperspectral image segmentation. The spectrum has been already taken into consideration by integrating various types of band selection algorithms, firstly. We provide a short overview of the hyperspectral image to select an appropriate set of bands by combining supervised, semi-supervised and unsupervised band selection algorithms. Some approaches are not limited in regards to their spectral dimension, but are limited with respect to their spatial dimension owing to low spatial resolution. The addition of spatial information will be focused on improving the performance of hyperspectral image segmentation for later fusion or classification. Many researchers have advocated that a large fractional coefficient should be in the exploration state while a small fractional coefficient should be in the exploitation, which does not mean the coefficient purely decrease with time. Due to such reasons, we propose an adaptive FODPSO by setting the fractional coefficient adaptively for the application of final hyperspectral image segmentation. In fact, the paper introduces an evolutionary factor to automatically control the fractional coefficient by using a sigmoid function. Therefore, fractional coefficient with large value will benefit the global search in the exploration state. Conversely, when the fractional coefficient has a small value, the exploitation state is detected. Hence, it can avoid optimization process get trapped into the local optima. Ultimately, the experimental segmentation results prove the validity and efficiency of our proposed automatic fractional coefficient setting method of FODPSO compared with traditional PSO, DPSO and FODPSO.
A novel approach to mammographic image segmentation, termed as PCNN-based level set algorithm, is presented in this paper. Just as its name implies, a method based on pulse coupled neural network (PCNN) in conjunction with the variational level set method for medical image segmentation. To date, little work has been done on detecting the initial zero level set contours based on PCNN algorithm for latterly level set evolution. When all the pixels of the input image are fired by PCNN, the small pixel value will be a much more refined segmentation. In mammographic image, the breast tumor presents big pixel value. Additionally, the mammographic image with predominantly dark region, so that we firstly obtain the negative of mammographic image with predominantly dark region except the breast tumor before all the pixels of an input image are fired by PCNN. Therefore, in here, PCNN algorithm is employed to achieve mammary-specific, initial mass contour detection. After that, the initial contours are all extracted. We define the extracted contours as the initial zero level set contours for automatic mass segmentation by variational level set in mammographic image analysis. What’s more, a new proposed algorithm improves external energy of variational level set method in terms of mammographic images in low contrast. In accordance with the gray scale of mass region in mammographic image is higher than the region surrounded, so the Laplace operator is used to modify external energy, which could make the bright spot becoming much brighter than the surrounded pixels in the image. A preliminary evaluation of the proposed method performs on a known public database namely MIAS, rather than synthetic images. The experimental results demonstrate that our proposed approach can potentially obtain better masses detection results in terms of sensitivity and specificity. Ultimately, this algorithm could lead to increase both sensitivity and specificity of the physicians’ interpretation of mammograms in clinical practice.
A novel approach for microcalcification clusters detection is proposed. At the first time, we make a short analysis of mammographic images with microcalcification lesions to confirm these lesions have much greater gray values than normal regions. After summarizing the specific feature of microcalcification clusters in mammographic screening, we make more focus on preprocessing step including eliminating the background, image enhancement and eliminating the pectoral muscle. In detail, Chan-Vese Model is used for eliminating background. Then, we do the application of combining morphology method and edge detection method. After the AND operation and Sobel filter, we use Hough Transform, it can be seen that the result have outperformed for eliminating the pectoral muscle which is approximately the gray of microcalcification. Additionally, the enhancement step is achieved by morphology. We make effort on mammographic image preprocessing to achieve lower computational complexity. As well known, it is difficult to robustly achieve mammograms analysis due to low contrast between normal and lesion tissues, there are also much noise in such images. After a serious preprocessing algorithm, a method based on blob detection is performed to microcalcification clusters according their specific features. The proposed algorithm has employed Laplace operator to improve Difference of Gaussians (DoG) function in terms of low contrast images. A preliminary evaluation of the proposed method performs on a known public database namely MIAS, rather than synthetic images. The comparison experiments and Cohen’s kappa coefficients all demonstrate that our proposed approach can potentially obtain better microcalcification clusters detection results in terms of accuracy, sensitivity and specificity.