فهرست مطالب hassan ghassemian

SAR images are used in many applications such as building detection. Extracting the building is very challenging due to the radar nature of the SAR images. However, due to the advantages of radar images such as day and night imaging, building extraction from SAR images is a hot topic. In this context, one of the main challenges is the effect of building orientation on the profile created in the SAR image. Also, the two geometric distortions of shadow and layover affect the SAR image. In most building extraction methods, shadow and double bounce are used as two main parameters in building detection. In this paper, different morphological profiles for detecting shadow index and double-bounce index (DMPSIDI) method have been developed using its combination with the method based on statistical features for building extraction. The DMPSIDI method is a morphological-based method that extracts buildings from SAR images independent of changing their profile. The proposed method is also robust to different data using weighting in the main parameters of shadow and double bounce.

Pansharpening is the fusion of panchromatic (PAN) and multispectral (MS) images to obtain a high spectral and spatial resolution image. Various metrics are introduced to assess the performance of different algorithms of pansharpening. This paper proposes a new metric for spectral quality evaluation of fused images. In the proposed method, spectrum vector of each pixel of fused image is compared to corresponding spectrum of reference image. Area of difference between two spectra is measured, and by applying this process to all pixel vectors of the fused image and taking an average over obtained values, spectral distortion of whole image is obtained. To investigate the efficiency of the proposed index, deliberate spectral distortion is applied to fused image and the proposed metric's ability to detect distortion is examined. Experimental results on real remote sensing images demonstrate the superior performance of the proposed metric compared to other existing metrics.

Dimensionality reduction methods transform or select a low dimensional feature space to efficiently represent the original high dimensional feature space of data. Feature reduction techniques are an important step in many pattern recognition problems in different fields especially in analyzing of high dimensional data. Hyperspectral images are acquired by remote sensors and human face images are one of the high dimensional data types. Because of limitation in the number of training samples, feature reduction is the important preprocessing step for classification of these types of data. Face recognition is one of the main interesting studies in human computer interaction applications. Face is among the most significant biometric characteristics which are used for identification of individuals. Before face recognition, feature reduction is an important processing step. In this paper, we apply the new feature extraction methods, which have been firstly proposed for feature reduction of hyperspectral imagery remote sensing, on the face databases for the first time. In this research, we compare the performance of seven new feature extraction methods with four state-of-the-art feature extraction methods. The proposed methods are Nonparametric Supervised Feature Extraction (NSFE), Clustering Based Feature Extraction (CBFE), Feature Extraction Using Attraction Points (FEUAP), Cluster Space Linear Discriminant Analysis (CSLDA), Feature Space Discriminant Analysis (FSDA), Feature Extraction using Weighted Training samples (FEWT), and Discriminant Analysis- Principal Component 1 (DA-PC1). The experimental results on two face databases, Yale and ORL, show the better performance of some new feature extraction methods, from the recognition accuracy point of view compared to methods such as linear discriminant analysis (LDA), non-parametric weighted feature extraction (NWFE), median-mean line discriminant analysis (MMLDA), and supervised locality preserving projection (LPP).

A fusion method for spectral-spatial classification of hyperspectral images is proposed in this paper. In the proposed framework, at first, the dimension of hyperspectral image is reduced by several state-of-the-art spectral feature extraction methods, i.e., Binary Coding Based Feature Extraction (BCFE), Clustering Based Feature Extraction (CBFE), Feature Extraction Based on Ridge Regression (FERR), Feature Extraction Using Attraction Points (FEUAP), Feature Extraction using Weighted Training samples (FEWT), and Feature Space Discriminant Analysis (FSDA). Then, the spatial features are calculated from the spectral features extracted from each spectral feature extraction method individually using the proposed smoothing filters and morphological operators. Finally, majority voting decision rule is used to obtain the final classification map. The proposed framework, in addition to removing the useless spatial information such as noise and distortions, adds useful spatial information such as shape and size of objects presented in scene image. The use of complement information obtained from six spectral feature extraction methods with different ideas for class discrimination, significantly improves the classification results. The proposed framework provides in average 6.64%, 7.07%, 8.23%, 7.52% and 20.52% improvement in classification results of three real hyperspectral images compared to generalized composite kernel (GCK), multiple feature learning (MFL), weighted joint collaborative representation (WJCR), original hyperspectral bands stacked on extended morphological profile (HS+EMP) and original hyperspectral bands (HS), respectively in terms of overall accuracy.

Hyperspectral (HS) imaging is a significant tool in remote sensing applications. HS sensors measure the reflected light from the surface of objects in hundreds or thousands of spectral bands, called HS images. Increasing the number of these bands produces huge data, which have to be transmitted to a terrestrial station for further processing. In some applications, HS images have to be sent instantly to the station requiring a high bandwidth between the sensors and the station. Most of the time, the bandwidth between the satellite and the station is narrowed limiting the amount of data that can be transmitted, and brings the idea of Compressive Sensing (CS) into the minds. In addition to the large amount of data, in these images, mixed pixels are another issue to be considered. Despite of their high spectral resolution, their spatial resolution is low causing a mixture of spectra in each pixel, but not a pure spectrum. As a result, the analysis of mixed pixels or Spectral Unmixing (SU) technique has been introduced to decompose mixed pixels into a set of endmembers and abundance fraction maps. The endmembers are extracted from spectral signatures related to different materials, and the abundance fractions are the proportions of the endmembers in each pixel. In recent years, due to the large amount of data and consequently the difficulties of real-time signal processing, and also having the ability of image compression, methods of Compressive Sensing and Unmixing (CSU) have been introduced. Two assumptions have been considered in these the finite number of elements in each pixel and the low variation of abundance fractions.
HYCA algorithm is one of the methods trying to compress these kinds of data with their inherent features. One of the sensible characteristics of this algorithm is to utilize spatial information for better reconstruction of the data. In fact, HYCA algorithm splits the data cube into non-overlapping square windows and assumes that spectral vectors are similar inside each window. In this study, a real-time method is proposed, which uses the spectral information (non-neighborhood pixels) in addition to the spatial information. The proposed structure can be divided into two parts: transmitting information into the satellites and information recovery into the stations.
In the satellites, firstly, to utilize the spectral information, a new real-time clustering method is proposed, wherein the similarity between the entire pixels is not restricted to any specific form such as square window. Figure 3 shows a segmented real HS image. It can be seen that the considering square form limits the capability of the HYCA algorithm and the similarity can be found in the both neighborhood and non-neighborhood pixels. Secondly, to utilize similarity in each cluster, different measurement matrices are used. By doing this, various samples can be achieved for each cluster and further information are extracted. On the other hand, usage of different measurement matrices may affect the system stability. As a matter of fact, generating the different measurement matrices is not simple and increases complexity into the transmitters. Therefore, it conflicts with the aim of CS theory, reducing complexity into the transmitters. As a result, in the proposed method, the number of the clusters is determined by the number of the producible measurement matrices. Figure 4 shows the schematic of the proposed structure in the satellites.
In the stations, we follow HYCA procedure in equation 8 and 9, but the different similar pixels are applied to the both equations. By doing this, we reach to the improved HYCA algorithm. Finally, the proposed structure is shown in the Table 1.
To evaluate the proposed method, both real and simulated data have been used in this article. In addition, normalized mean-square error is considered as an error criteria. For the simulated data, in constant measurement sizes, the effects of the additive noise, and for real data, the effects of measurement sizes have been investigated. Besides, the proposed method has been compared with HYCA and C-HYCA and some of the traditional CS based methods. The experimental results show the superiority of the proposed method in terms of signal to noise ratios and the measurement sizes, up to in the simulated data and in the real data, which makes it suitable in the real-world applications.

In this paper, a new feature reduction technique for hyperspectral data classification problem is proposed based on extracting new features with smaller dimension with respect to the original data. For each pixel of a hyperspectral image, a specific rational function approximation is developed to fit its own spectral response curve (SRC) and the coefficients of the numerator and denominator polynomials of this function are considered as new extracted features. The method focuses on geometrical nature of SRCs and relies on the fact that the sequence discipline - ordinance of reflectance coefficients in spectral response curve - contains some information which has not been addressed by many other existing methods that are based on statistical analysis of data. Maximum likelihood classification results demonstrate that our method provides better classification accuracies compared to many competing feature extraction algorithms. In addition, the proposed algorithm has the possibility to be applied to all pixels of image individually and simultaneously.

The purpose of unmixing in hyperspectral images is extraction of the endmembers spectral signatures and estimation of their related abundance fractions. Most algorithms used for endmember extraction (EE) process, are established on spectral information without any attention to spatial context and correlation of image pixels. Recently, several algorithms have been developed which utilize spatial and spectral information with the aim of improving EE and unmixing accuracy. In this paper, a novel spatial spectral preprocessor is proposed which exploits class map obtained by unsupervised clustering technique and 8th neighborhood window in order to identify pixels located in border regions between two or more clusters and discards not spatially homogenous regions. Afterwards, it calculates spectral purity weight of not border pixels in order to look for spatially homogenous and spectrally pure ones using otsu threshold. Endmembers can be extracted rapidly and accurately by means of coupling our proposal with EEs. Our distinct scheme can reduce RMSE of reconstructed image and EE processing time as well as improve a new criterion known as Efficiency regarding the state-of-the-art preprocessors on real hyperspectral images.

One of the most preprocessing steps before the classification of hyperspectral images is supervised feature extraction. Because obtaining the training samples is hard and time consuming¡ the number of available training samples is limited. We propose a supervised feature extraction method in this paper that is efficient in small sample size situation. The proposed method¡ which is called weighted feature line embedding (WFLE)¡ uses the feature line concepts for production of virtual training samples and then¡ uses them for estimation of within-class and between-class scatter matrices. The new idea of WFLE is based on more correction on the non-appropriate and abnormal samples through weighting process in estimation of scatter matrices. The WFLE is compared with some popular and state-of-the-art feature extraction methods such as LDA¡ GDA¡ NWFE¡ NPE¡ LPP and NFLE. The experimental results show the good performance of WFLE in comparison with other methods in small sample size situation.

In the present research we have used gray level co-occurrence matrices (GLCM) and Gabor filters to extract texture features in order to classify satellite images. The main drawback of GLCM algorithm is its time-consuming nature. In this work, we proposed a fast GLCM algorithm to overcome the mentioned weakness of the traditional GLCM. The fast GLCM is capable of extracting approximately the same features as the traditional GLCM does, but in much less time (about 200 times faster). The other weakness of the traditional GLCM is its lower accuracy in the regions near the class borders. Since features extracted using Gabor filters are more accurate in boundary regions, we combined Gabor features with GLCM features. In this way we could compensate the latter mentioned weakness of GLCM. Experimental results show good capabilities of the proposed fast GLCM and the feature fusion method in classification of very high resolution remote sensing images.

Regarding to the high between-band correlation and large volumes of hyperspectral data, feature reduction (either feature selection or extraction) is an important part of classification process for this data type. A variety of feature reduction methods have been developed using spectral and spatial domains. In this paper, a feature extracting technique is proposed based on rational function curve fitting. For each pixel of a hyperspectral image, a specific rational function approximation is developed to fit the spectral response curve of that pixel. Coefficients of the numerator and denominator polynomials of these functions are considered as new extracted features. This new technique is based on the fact that the sequence discipline - ordinance of reflectance coefficients in spectral response curve - contains some information which has not been considered by other statistical analysis based methods, such as Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA) and their nonlinear versions. Also, we show that naturally different curves can be approximated by rational functions with equal form, but different amounts of coefficients. Maximum likelihood classification results demonstrate that the Rational Function Curve Fitting Feature Extraction (RFCF-FE) method provides better classification accuracies compared to competing feature extraction algorithms. The method, also, has the ability of lossy data compression. The original data can be reconstructed using the fitted curves. In addition, the proposed algorithm has the possibility to be applied to all pixels of image individually and simultaneously, unlike to PCA and other methods which need to know whole data for computing the transform matrix.

When the number of training samples is limited, feature reduction plays an important role in classification of hyperspectral images. In this paper, we propose a supervised feature extraction method based on discriminant analysis (DA) which uses the first principal component (PC1) to weight the scatter matrices. The proposed method, called DA-PC1, copes with the small sample size problem and has not the limitation of linear discriminant analysis (LDA) in the number of extracted features. In DA-PC1, the dominant structure of distribution is preserved by PC1 and the class separability is increased by DA. The experimental results show the good performance of DA-PC1 compared to some state-of-the-art feature extraction methods.

Congenital heart disease is now the most common severe congenital abnormality found in live births and the cause of more than half the deaths from congenital anomalies in childhood. Heart murmurs are often the first signs of pathological changes of the heart valves, and they are usually found during auscultation in the primary health care. Auscultation is widely applied in clinical activity; nonetheless sound interpretation is dependent on clinician training and experience. Distinguishing a pathological murmur from a physiological murmur is difficult and prone to error. To address this problem we have devised a simplified approach to pediatric cardiac scanning. This will not detect all forms of congenital heart disease but will help in the diagnosis of many defects. Cardiac auscultatory examinations of 93 children were recorded, digitized, and stored along with corresponding echocardiographic diagnoses, and automated spectral analysis using discrete wavelet transforms was performed. Patients without heart disease and either no murmur or an innocent murmur (n = 40) were compared to patients with a variety of cardiac diagnoses and a pathologic systolic murmur present (n = 53). A specificity of 100% and a sensitivity of 90.57% were achieved using signal processing techniques and a k-nn as classifier.

In this paper a novel unsupervised feature extraction method for detection of melanoma in skin images is presented. First of all، normal skin surrounding the lesion is removed in a segmentation process. In the next step، some shape and texture features are extracted from the output image of the first step: GLCM، GLRLM، the proposed directional-frequency features، and some parameters of Ripplet transform are used as texture features; Also، NRL features and Zernike moments are used as shape features. Totally، 63 texture features and 31 shape features are extracted. Finally، the number of extracted features is reduced using PCA method and a proposed method based on Fisher criteria. Extracted features are classified using the Perceptron Neural Networks، Support Vector Machine، 4-NN، and Naïve Bayes. The results show that SVM has the best performance. The proposed algorithm is applied on a database that consists of 160 labeled images. The overall results confirm the superiority of the proposed method in both accuracy and reliability over previous works.

The fusion of valuable spectral and spatial features can significantly improve the performance of high resolution hyperspectral images classification. In this paper, we propose a spectral and spatial feature extraction method based on discriminant analysis. To increase the class discrimination, we maximize the between-class scatters and minimize the within-class scatters. To include the spatial information in the feature extraction process, we estimate the spatial scatters in a spatial neighborhood window with multi-scale fashion. We compare our proposed method, which is called spectral-spatial discriminant analysis (SSDA), with some spatial-spectral feature extraction methods included original spectral bands plus Gabor filters, gray level co-occurance matrix (GLCM), and morphology profiles and also with some popular spectral feature extraction methods such as nonparametric weighted feature extraction (NWFE) and locality preserving projection (LPP). Moreover, we compare SSDA with some recently proposed spectral-spatial classification approaches. The experimental results on two real hyperspectral images show the good performance of SSDA compared to the competitor methods.

This paper aims to propose an effective paroxysmal atrial fibrillation (PAF) predictor which is based on the analysis of the heart rate variability (HRV) signal. Predicting the onset of PAF, based on non-invasive techniques, is clinically important and can be invaluable in order to avoid useless therapeutic intervention and to minimize the risks for the patients. This method consists of four steps: preprocessing, feature extraction, feature reduction, and classification. In the first step, the QRS complexes are detected from the electrocardiogram (ECG) signal and then the HRV signal is extracted. In the next step, the recurrence plot (RP) of HRV signal is obtained and six features are extracted to characterize the basic patterns of the RP. These features consist of length of longest diagonal segments, average length of the diagonal lines, entropy, trapping time, length of longest vertical line, and recurrence trend. In the third step, these features are reduced to three features by the linear discriminant analysis (LDA) technique. Using LDA not only reduces the number of the input features but also increases the classification accuracy by selecting most discriminating features. Finally, a support vector machine-based classifier is used to classify the HRV signals. The performance of the proposed method in prediction of PAF episodes was evaluated using the Atrial Fibrillation Prediction Database (AFPDB) which consists of both 30-minutes ECG recordings end just prior to the onset of PAF and segments at least 45 min distant from any PAF events. The obtained sensitivity, specificity, and positive predictivity were 96.55%, 100%, and 100%, respectively.

In this paper, we propose a novel method for fully automatic detection and tracking of human heads and faces in video sequences. The proposed algorithm consists of two modules: a face detection module and a face tracking module. The Detection module, detects the face region and approximates it with an ellipse at the first frame using a modified version of AdaBoost cascaded classifier. The detection module is capable of considering the 2-D head pose rotation. The tracking module utiliyes a combination of deformable mesh energy minimization and feature matching approaches. In order to track a face, features are extracted in the face region to tessellate the human face with triangular unstructured meshes. For tracking a mesh, it is necessary to define mesh energies including internal and external energies. We have used new energy definitions for both the internal and the external energies which can accurately track rigid and non-rigid motions of a face and facial features at subsequent frames. We tested the proposed method with different video samples like cluttered backgrounds, partial illumination variations, put on glasses, and 2-D and/or 3-D rotating and translating heads. The experimental results showed that the algorithm is rotation insensitive and has high accuracy, stability and also has convergence for face detection and tracking

Contrast resolution and detail resolution are two important parameters in ultrasound imaging. This paper presents a new method to enhance these parameters, simultaneously. A parallel auxiliary beamformer has been employed whose weightings are such that an estimation of the leaked signal through the main beamformer is obtained. Then the output of main beamformer is modified according to the estimated leaked signal. The efficiency of our adaptive method is demonstrated by applying it over an experimental data set and provided an enhancement of about 22 percent in lateral resolution and 15-20 dB in contrast resolution. This method also has the advantages of simplicity and possibility of real time implementation.