فهرست مطالب حسین پورقاسم

Fire is one of the dangers that can endanger human health in a short time and if it is not controlled in time, it will cause a lot of damage. Therefore, timely and accurate identification of the location of the fire can prevent the consequences of its expansion. In this research, a new method for fire detection is proposed based on the extraction of its temporal-spatial features in video frames. In the proposed algorithm, a multiscale convolutional neural network along with a YOLO (you only look once) network is used to extract spatial features and identify fire candidate regions. Then, fractal analysis based on the temporal blanket method is then used to remove non-moving textures similar to fire and to examine the temporal features of the candidate region. Finally, the fire region is separated from the other parts of the image by fusion the results of the two steps. The evaluation results of the proposed method on three data sets show that the accuracy of fire detection is about 96.1%, while the precision and recall values are 92% and 96.9%, respectively. Experimental results show that the proposed method performs better than existing algorithms and thus confirms the ability of this method for efficient use in the real world.

Diabetes is one of the most common diseases in the world, adversely affects different body organs. One of the most common causes of eye problems is diabetes. Analyzing retinal damage is one of the best ways to diagnose diabetes so one of the best ways to diagnose diabetes is to look at the damage to the retina. Hence, first, a highly applicable and effective method, which is a combination of the Wiener filter and the discrete wavelet transform (DWT), is used for the removal of noise from images. Afterward, the k-means clustering algorithm is used to remove the bad image sections including very light and very dark areas of the image. Next, the image color and shape features are extracted. We transfer the images to the lab space, which fits the eye more, to extract the image color features. To extract the image shape features, first the images are converted into grey images and then the shape features are extracted. After extracting the features, the number of features is reduced using the Principal Component Analysis (PCA) algorithm. Besides, the best and most effective features are also selected. Finally, the support vector machine classifier with different kernel is used to classify the features and images into two categories, namely the healthy participants and patients. The accuracy resulting from this algorithm using the test images is over 90%.

Human diseases such as diabetes, high blood pressure and the cerebral source disorders have effects on the retina vessels of human’s eyes. By classifying the retina vessels as two sets of arteries and veins, it can be evaluated the progress and symptoms of mentioned diseases. In this paper, a retinal blood vessel classification algorithm based on structural, directional and frequency features along with feature optimization using Tagouchi genetic algorithm is proposed. For this purpose, to classify the vessels in fundus images, at first the vessels are segmented. In this algorithm, to extract simultaneously information related to direction, diameter and dynamical behavior of the blood vessel, a novel feature based on wavelet transform using entropy contents of DWT and Directional Wavelet Entropy (DWE), Fourier transform using Fourier descriptors have been presented. Also 2-D Frequency Similarity Sectors (2DFSS) is introduced to represent and describe the variations of thickness and direction of the blood vessel. After extracting the feature vector using hybrid model of Genetic algorithm and Tagouchi strategy, the optimal features are selected. Then by employing the multi-layer neural network classifier, the vessels are recognized into arteries and veins classes. With these represented attributes, the classification is performed based on the structure and direction of vessels. Ultimately, the accuracy rate of 82.09% and precision rate of 81.58% are simultaneously obtained in problem of the retinal vessel recognition on a database consisting of 40 images.

In this paper, the design and simulation of a transmission impedance amplifier using 0.18 µ CMOS technology for use in the amplifier section of optical telecommunication receivers are presented. The idea presented in this paper is based on the use of a parallel-parallel feedback structure with a cascade interest rate. This structure is quasi-differential and is designed to increase the allowable range of input power and prevent amplifier saturation, a circuit with a variable gain. The simulation results show a maximum gain of 64 dB, a bandwidth of 1.8 GHz, power consumption of 20mw, and a spectral intensity of the noise current referred to as the PA / 9 input. Also, by using variable gain control, the gain of the amplifier transmission impedance can be reduced to 46 dBΩ. The results indicate that the proposed design is very suitable for a 2.5 Gb / s optical telecommunication system.

In this study, a new algorithm for dermoscopy image classification into two types of malignant and benign is presented. At first, one preprocessing step to remove noise and also enhance image quality is performed. After that using Otsu thresholding, the lesion is separated from the healthy area. Then color and shape features are extracted from the segmented image. The colored features based on statistical moments of quantized grayscale and quantized color histogram are defined. These features demonastrare distribution of color components. Moreover, the shape features are extracted information of the segmented regions with two scenarios. In the first scenario, the features are represented the expantion of region and in the second scenario, the features are represented the edge variations of the extracted regions. Finally, the classification procedure is performed using K-Neasert Neighbor (KNN), Decision Tree, Support Vector Machine (SVM) and Adaboost. The proposed algorithm is evaluated on a standard database consisting of 200 images. The results show that classification using Adaboost classifier obtains the precision rate, accuracy rate and sensitivity rate of 96%, 96.7% and 95%, respectively.

Segmentation is necessary to determine the boundaries of the vessel. Intravascular ultrasound imaging (IVUS) is used for the diagnosis of coronary artery diseases. In this study, a new method is proposed for segmentation of IVUS images. First preprocessing is done to convert images from Cartesian coordinates to polar coordinates, remove the catheter in images and speckle noise with Nonlinear Anisotropic Diffusion Filtering. Then, texture features of an image are extracted using Gabor filter, and the image segmentation and determining the vessels boundary will be discussed using active contour without edge for vector value model. Calcium plaques have been determined using phase clustering and the exact boundary of calcium plaques is extracted using active contour model. This method has been tested on thirty images, and the results of the image segmentation have been validated by an expert. The area diffusion between the internal border and the expert’s opinion is 0.4310.236, and the area diffusion between the external border and the expert’s opinion is 0.6530.723. Area diffusion of calcium plaque extracted by the proposed algorithm compared with virtual histology images has been achieved equal to 5.90 percent.

In common background subtraction method usually the shadow of objects is extracted as the moving objects that cause some errors in the performance of Intelligent Transportation Systems (ITS). In this paper, an effective algorithm based on combination of HSV color space and Principal Component Analysis is proposed. In this algorithm, the candidate shadow region is detected by using HSV color space. In this step, some part of vehicles may be detected as the moving shadow. So, to compensate and improve the performance of moving shadow detection, principal components analysis algorithm is applied to recognize the automobile object by modeling the automobile based on orthogonal eigen vectors of database. Our proposed algorithm is evaluated on real and operational videos of ITS. The obtained results demonstrate the efficiently and effectiveness of our proposed algorithm in the ITS applications.

One of the important problems in EEG signal analysis is epilepsy detection. In this paper, an epilepsy detection and classification algorithm is proposed based on fusion of the extracted Hartley transform features. In this algorithm, in addition to usual frequency and time features such as spectrum entropy and power spectral density, a novel feature is defined based on fusion of the extracted Hartley transform features. To define this feature, the extracted features from Hartley transform are fused based on a kernel matrixes scenario. To optimize and reduce the dimensions of extracted feature vector from EEG signals, a hybrid model is applied based on Genetic algorithm using Memetic learning strategy and multilayer back propagation neural network. The final classification is carried out on the optimal feature vectors by a Perceptron neural network with a hidden layer and average accuracy rate of 95.325% is obtained.

Seizure detection using brain signal (EEG) analysis is the important clinical methods in drug therapy and the decisions before brain surgery. In this paper, after signal conditioning using suitable filtering, the Gamma frequency band has been extracted and the other brain rhythms, ambient noises and the other bio-signal are canceled. Then, the wavelet transform of brain signal and the map of wavelet transform in multi levels are computed. By dividing the color map to different epochs, the histogram of each sub-image is obtained and the statistics of it based on statistical momentums and Negentropy values are calculated. Statistical feature vector using Principle Component Analysis (PCA) is reduced to one dimension. By EMD algorithm and sifting procedure for analyzing the data by Intrinsic Mode Function (IMF) and computing the residues of brain signal using spectrum of Hilbert transform and Hilbert – Huang spectrum forming, one spatial feature based on the Euclidian distance for signal classification is obtained. By K-Nearest Neighbor (KNN) classifier and by considering the optimal neighbor parameter, EEG signals are classified in two classes, seizure and non-seizure signal, with the rate of accuracy 76.54% and with variance of error 0.3685 in the different tests.

Vehicle tracking is an important issue in Intelligence Transportation Systems (ITS) to estimate the location of vehicle in the next frame. In this paper، a feature-based vehicle tracking algorithm using Kanade-Lucas-Tomasi (KLT) feature tracker is developed. In this algorithm، a merge and split-based hierarchical two-stage grouping algorithm is proposed to represent vehicles from the tracked features. In the proposed grouping algorithm، with defining measures such as distance، spread and also blob analysis، initial grouping results formed by K-means clustering algorithm are refined. Moreover، to modify the performance of KLT tracker and also optimized utilization from grouping results obtained by proposed algorithm، an effective group matching algorithm based on a merging and splitting scheme is employed to match the tracked groups from a frame to the next frame. The proposed tracking algorithm is evaluated on different test videos with various illumination conditions such as day، night and shadow. The obtained results show that our proposed tracking algorithm covers the most challenges of tracking in the ITS applications.

Brain hemorrhage due to head trauma is one of the most common causes of death. Early diagnosis of location and type of brain hemorrhage is crucial. For saving the patients completely, it is necessary to detect the correct location and type of the hemorrhage in an early stage. In this study, we introduced an automatic brain hemorrhage detection and classification algorithm to improve and accelerate the process of physicians’ decision-making. To achieve the purpose, at first a segmentation algorithm was usedto detect and separate the hemorrhage regions from other parts of the brain. Then, a number of appropriate features from each detected hemorrhage region were extracted and then by using genetic algorithm, the most convenient features were selected. The utilized computed tomography (CT-scan) images in this research were collected from Kashani hospital CT-scan center (Isfahan, Iran) and were of 70 men and women between the ages of 15-60 years. Using the proposed segmentation and classification algorithm, the segmentation accuracy for different types of hemorrhages [epidural (EDH), intracerebral (ICH) and subdural (SDH)] were obtained as 96.87%, 96.10%, 92.15%, respectively. Also, intraventricular hemorrhage (IVH) was detected and separated from other types of hemorrhage with the accuracy rate of 91.82% and classified with accuracy rate of 94.13%. In this research, an independent and automatic brain hemorrhage detection and classification algorithm was assessed. Our proposed algorithm is an attempt to improve and accelerate the process of physicians’ decision-making to save the patients’ lives. By using the proposed algorithm, we were able to detect and classify four kinds of dangerous hemorrhages in CT-scan images and accelerat the process of the victim’s treatment.

The purpose of this study is to design a content-based medical image retrieval system and provide a new method to reduce semantic gap between visual features and semantic concepts. Generally performance of the retrieval systems based on only visual contents decrease because these features often fail to describe the high level semantic concepts in user’s mind. In this paper this problem is solved using a new approach based on projection of relevant and irrelevant images in to a new space with low dimensionality and less overlapping in relevance feedback level. For this purpose، first we change the feature space using Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA) techniques and then classify the feedback images applying Support Vector Machine (SVM) classifier. The proposed framework has been evaluated on a database consisting of 10،000 medical X-ray images of 57 semantic classes. The obtained results show that the proposed approach significantly improves the accuracy of retrieval system.

Dermoscopy is one of the major imaging techniques used in diagnoses of Melanoma and other skin diseases. Because of difficulties and subjectivity of human interpretation، automatic and computerized analysis of dermoscopic images has opened an important research area. Automatic lesion detection is one of the main steps in analysis of these images. Finding an optimal threshold for segmenting the lesion is a severe task in image processing. Different methods for thresholding already exist. In this research a novel thresholding approach according to well-known thresholding methods and reinforcement algorithm for segmenting dermoscopic images is presented. The reinforced agent learns optimal weights for different thresholding methods and finally segments the dermoscopic image with optimal threshold. A reward function is designed for achieving the similarity ratio between the binary output image and original gray level image and calculating reward/punish signal which should be exerted to reinforced agent. We use three thresholding methods، Otsu، Kittler and Kapur، for combining in the reinforced agent and the detected lesions are compared with the ground-truth which is determined dermatologists and the border error is calculated. The results are also compared with other well-known automatic methods which indicate that the proposed method yields to more accuracy and less border error in detection of lesion in dermocopy images.

Recently image fusion has prominent and applicable roles in medical image processing. Digital subtraction angiography (DSA) image is applied to display map of blood vessels. In this essay، a new fusion algorithm for DSA serial images based on discrete wavelet transform coefficients is proposed. Fusion of high frequency coefficients is based on proposed fusion map and four evaluation criteria which introduce level of coefficient''s energy. Our algorithm will be compared for different wavelet transforms and activity criteria for high frequency coefficients. The comparisons are based on the objective evaluation criteria which show measure of noise existence، sharpness and correlation between the fusion result and reference image. Finally، Meyer discrete wavelet transform is resulted as the best wavelet transform، and sum of modified Laplacian، local energy are introduced as activity level measurment for high and low frequency coefficients in process of brain vessel map extraction.

Watermarking is used to protect copyright proof. Robustness index is the most important parameter that evaluates watermarking algorithm against different attacks such as noise and compression. In this paper، a novel semi- blind image watermarking algorithm based on joint wavelet transform (WT) and singular value decomposition (SVD) transform is proposed. In this algorithm، a new strategy is used to joint WT and SVD effectively. The most important advantage of this algorithm is robustness against a number of common attacks. Experimental results show that the proposed algorithm improves evaluation parameters more than other methods presented previously in the literature.

According to many medical and biometric applications of retinal images، the automatic and accurate extraction of the retinal blood vessels is very important. In this paper، an effective method is introduced to extract the blood vessels from the background of colored images of retina. In this algorithm، by applying the equalizer function on the retinal images، the brightness of the images is considerably uniformed. Because of high ability of Curvelet transform in introducing image borders in various scale and directions، borders and، consequently the contrast of retinal images can be enhanced. Therefore، the enhanced retinal image can be prepared for the extraction of blood vessels by improving Curvelet coefficients of the retinal images، adaptively and locally. Since the blood vessels in retinal images are distributed in various directions، we use the adaptive weighted morphology operators to extract the blood vessels from retinal images. Morphology operators based on reconstruction are used to refine the appeared frills with the size of smaller than arterioles in images properly. Finally، by analyzing the connected component in the images and applying adaptive filter on the components locally، all residual frills are refined from the images. The proposed algorithm in this paper has been evaluated by the images in the DRIVE database. The results how that the blood vessels are extracted from background of the retinal images of DRIVE database with the high accuracy of 96. 15%، which in turn shows the high ability of the proposed algorithm in extracting the retinal blood vessels.

Most of the existing image resolution enhancement algorithms assume that the image is clean and noise free، but this assumption is not practically valid. One strategy for interpolation of noisy images is to denoise the image first and then interpolate the denoised image. However، this strategy does not lead to satisfying results because denoising may smooth image details and also other artifacts such as blurring and blocking introduced due to image denoising will be amplified in the following interpolation stage. Thus، in this paper we propose a joint salt and pepper noise removal and resolution enhancement algorithm using dual-tree complex wavelet transform and feedforward neural networks. In this algorithm، the wavelet subbands corresponding to noise free high resolution image are estimated from noisy low resolution image by multi-layer perceptron (MLP). Therefore the noise free high resolution image is obtained by complex wavelet reconstruction of the estimated subbands. Takeing advantages of complex wavelet transform such as nearly shift invariance and directional selectivity the subband estimation by neural networks is done with high accuracy. As it is verified in the experimental results، the proposed algorithm has better performance both subjectively and objectively and is able to maintain the image fine structures well.

Logo detection and recognition module is a vital requirement in official automation systems for document image archiving and retrieval applications. In this paper، we present a novel framework for logo detection and recognition based on sequential segmentation and classification strategy of document image. In this framework، using a two-stage segmentation algorithm (consisting of wavelet-based and threshold-based segmentation algorithms) and hierarchical classification by two multilayer Perceptron (MLP) classifiers and a k-nearest neighbor (KNN) classifier، a document image divides to text، pure picture and logo candidate regions. Ultimsately، in final decision، class of logo candidate region is determined based on pre-defined classes. In the hierarchical classification and logo recognition stages، the best feature space is selected by forward selection algorithm from a perfect set of texture and shape features. The proposed structure is evaluated on a variety and vast database consisting of the document and non-document images with Persian and international logos. The obtained results show efficiency of the proposed framework in the real and operational conditions.

Color and shape are basic characteristics which are used to recognize traffic signs. In this paper, a new speed limit sign detection method in various conditions is proposed. In this method, color image is segmented based on a thresholding technique in HSI color space. Then, corner features are detected using convolution masks and found the location of sign. The first advantage of this method is high accuracy to detect the location of sign. So, the object can be detected with 30% noise level, 30 meters for distances of signs, and for rotated signs. The second advantage of the proposed method is high speed in sign detection. Utimately, the type of sign can be recognized with eliminate redundant information and match between extracted image and database image. If the illumination conditions be ideal, the recognition rate is obtained to 89%.

This paper presents a new classification framework for Iranian license plate character recognition. In this framework, a set of robust features are calculated from license plate characters based on directional projections,kirsch edge detector and local means. The characters are then classified using mixture of experts which use the multilayer Perceptrons (MLPs) as expert and gating networks. The proposed recognition algorithm is evaluated on a database of Iranian license plate characters consisting of 14256 binary images, and the recognition rate of 99.42% is achieved. The proposed algorithm yields better performance of the Iranian license plate character recognition in comparison with conventional methods which use a single MLP neural network.

Relevance feedback approaches is used to improve the performance of content-based image retrieval systems. In this paper، a novel relevance feedback approach based on similarity measure modification in an X-ray image retrieval system based on fuzzy representation using fuzzy attributed relational graph (FARG) is presented. In this approach، optimum weight of each feature in feature vector is calculated using similarity rate between query image and relevant and irrelevant images in user feedback. The calculated weight is used to tune fuzzy graph matching algorithm as a modifier parameter in similarity measure. The standard deviation of the retrieved image features is applied to calculate the optimum weight. The proposed image retrieval system uses a FARG for representation of images، a fuzzy matching graph algorithm as similarity measure and a semantic classifier based on merging scheme for determination of the search space in image database. To evaluate relevance feedback approach in the proposed system، a standard X-ray image database consisting of 10000 images in 57 classes is used. The improvement of the evaluation parameters shows proficiency and efficiency of the proposed system.