Theses supervised by Prof. Dr. Ahmet Enis Çetin
22 theses · İhsan Doğramacı Bilkent University
Çarpma işlemsiz sinir ağları
Artificial Neural Networks, commonly known as Neural Networks (NNs), have become popular in the last decade for their achievable accuracies due to their ability to generalize and respond to unexpected patterns. In general, NNs are computationally expensive. This thesis presents the implementation of a class of NN that do not require multiplication operations. We describe an implementation of a Multiplication Free Neural Network (MFNN), in which multiplication operations are replaced by additions and sign operations. This thesis focuses on the FPGA and ASIC implementation of the MFNN using VHDL. A detailed description of the proposed hardware design of both NNs and MFNNs is analyzed. We compare 3 different hardware designs of the neuron (serial, parallel and hybrid), based on latency/hardware resources trade-off. We show that one-hidden-layer MFNNs achieve the same accuracy as its counterpart NN using the same number of neurons. The hardware implementation shows that MFNNs are more energy efficient than the ordinary NNs, because multiplication is more computationally demanding compared to addition and sign operations. MFNNs save a significant amount of energy without degrading the accuracy. The fixed-point quantization is discussed along with the number of bits required for both NNs and MFNNs to achieve floating-point recognition performance.
Verimli enerjili Hadamard sinir ağları ile görüntü sınıflandırması
Deep learning has made significant improvements at many image processing tasks in recent years, such as image classification, object recognition and object detection. Convolutional neural networks (CNN), which is a popular deep learning architecture designed to process data in multiple array form, show great success to almost all detection & recognition problems and computer vision tasks. However, the number of parameters in a CNN is too high such that the computers require more energy and larger memory size. In order to solve this problem, we investigate the energy efficient network models based on CNN architecture. In addition to previously studied energy efficient models such as Binary Weight Network (BWN), we introduce novel energy efficient models. Hadamard-transformed Image Network (HIN) is a variation of BWN, but uses compressed Hadamard-transformed images as input. Binary Weight and Hadamard-transformed Image Network (BWHIN) is developed by combining BWN and HIN as a new energy efficient model. Performances of the neural networks with different parameters and different CNN architectures are compared and analyzed on MNIST and CIFAR-10 datasets. It is observed that energy efficiency is achieved with a slight sacrifice at classification accuracy. Among all energy efficient networks, our novel ensemble model outperforms other energy efficient models.
Fourier dönüşümünün fazı ve sınırlandırılmış enerji temelli imge ters evrişim yöntemleri
We developed deconvolution algorithms based on Fourier transform phase and bounded energy. Deconvolution is a major area of study in image processing applications. In general, restoration of original images from noisy filtered observation images is an ill-posed problem. We use Fourier transform phase as a constraint in developed image recovery methods. The Fourier phase information is robust to noise, which makes it suitable as a frequency domain constraint. One of our focus is microscopy images where the blur is caused by slight disturbances of the focus. Because of the symmetrical optical parameters, it may be assumed that the Point Spread Function (PSF) is symmetrical. This symmetry of PSF results in zero phase distortion in the Fourier transform coefficients of the original image. Since the convolution leads to multiplication in Fourier domain, we assume that the Fourier phase of some of the frequencies of observed image around the origin represents the Fourier phase of the original image in the same set of frequencies. Therefore the Fourier transform phases of the original image can be estimated from the phase of the observed image and this information can be used as a Fourier domain constraint. In order to complete the algorithm, we also use a Total Variation (TV) reduction based regularization in spatial domain. We embed the proposed Fourier phase relation and spatial domain regularization as additional constraints in well-known blind Ayers-Dainty deconvolution method. Another problem we focused on is the restoration of highly blurry Magnetic Particle Imaging (MPI) applications. In this study we developed a standalone iterative algorithm. The algorithm again relies on the symmetry property of the MPI PSF. The phase estimates of the true image are obtained from the observed image. In this case we employ an l1 projection based regularization algorithm. The l1 projection reduces the small coefficients to zero which is suitable for MPI application because the contrast between foreground and background is sufficiently large by nature. Finally, a more general restoration algorithm is developed for deconvolution of non-symmetrical filters. The algorithm uses the known Fourier phase properties of the PSF in order to estimate the Fourier transform phase of the original image. We also update the estimated Fourier transform magnitudes iteratively using the knowledge of observed image and the PSF. A TV reduction based regularization method completes the algorithm in spatial domain. Simulations and experimental results show that the proposed algorithm outperforms the Wiener filter. We also conclude that the addition of estimate of Fourier transform phase is useful in any deconvolution method.
Pasif bistatik radar sistemleri üzerinde hedef tespiti ve görüntülenmesi
Passive Bistatic Radar (PBR) systems have become more popular in recent years in many research communities and countries. Papers related to PBR systems have increasingly received significant attention in research. There are many tar- get detection methods for PBR system in the literature. This thesis assumes a system scenario based on stereo FM signals as transmitters of opportunity. Ambiguity function (AF) is a function that determines the locations of targets in range-Doppler map turns out to be noisy in practice. This can cause a problem with low SNR-valued targets because they cannot be visible. To solve this problem, compressive sensing (CS) and projection onto the epigraph set of the L1 ball (PES-L1) are used to denoise the range-Doppler map. Some CS methods are applied to the system scenario, which are Basis Pursuit (BP), Orthogonal Matching Pursuit (OMP), Compressed Sampling Matching Pursuit (CoSaMP), Iterative Hard Thresholding (IHT). In addition, AF is generally used to determine the similarities between two signals. Therefore, different correlation methods can be also used to compare the surveillance and time delayed frequency shifted replica of the reference signal. Maximal Information Coefficient (MIC), Pearson correlation coefficient, Spearman's rank correlation coefficient are used for the target detection. This thesis proposes a least squares (LS) based method which outperforms other correlation algorithms in terms of PSNR and SNR. Two LS coefficients are obtained from the real and imaginary parts of predicting the surveillance signal using the modulated reference signal. Norm of LS coefficients exhibit a peak at target locations. The proposed method detects close targets better than the ordinary AF method and decreases the number of sidelobes on multiple FM channels based the PBR system.
Seyreklik ve konveks programlama ile zaman-frekans işleme
In this thesis sparsity and convex programming-based methods for time-frequency (TF) processing are developed. The proposed methods aim to obtain high resolution and cross-term free TF representations using sparsity and lifted projections. A crucial aspect of Time-Frequency (TF) analysis is the identification of separate components in a multi-component signal. Wigner-Ville distribution is the classical tool for representing such signals but suffers from cross-terms. Other methods that are members of Cohen's class distributions also aim to remove the cross terms by masking the Ambiguity Function (AF) but they result in reduced resolution. Most practical signals with time-varying frequency content are in the form of weighted trajectories on the TF plane and many others are sparse in nature. Therefore the problem can be cast as TF distribution reconstruction using a subset of AF domain coefficients and sparsity assumption in TF domain. Sparsity can be achieved by constraining or minimizing the l1 norm. Projections Onto Convex Sets (POCS) based $l_1$ minimization approach is proposed to obtain a high resolution, cross-term free TF distribution. Several AF domain constraint sets are defined for TF reconstruction. Epigraph set of l1 norm, real part of AF and phase of AF are used during the iterative estimation process. A new kernel estimation method based on a single projection onto the epigraph set of l1 ball in TF domain is also proposed. The kernel based method obtains the TF representation in a faster way than the other optimization based methods. Component estimation from a multi-component time-varying signal is considered using TF distribution and parametric maximum likelihood (ML) estimation. The initial parameters are obtained via time-frequency techniques. A method, which iterates amplitude and phase parameters separately, is proposed. The method significantly reduces the computational complexity and convergence time.
Dışbükey maliyet fonksiyonları'nın epigraf kümesine dik izdüşümler kullanan imge restorasyonu ve yeniden inşa algoritmasi
This thesis focuses on image restoration and reconstruction problems. These inverse problems are solved using a convex optimization algorithm based on orthogonal Projections onto the Epigraph Set of a Convex Cost functions (PESC). In order to solve the convex minimization problem, the dimension of the problem is lifted by one and then using the epigraph concept the feasibility sets corresponding to the cost function are defined. Since the cost function is a convex function in RN, the corresponding epigraph set is also a convex set in RN+1. The convex optimization algorithm starts with an arbitrary initial estimate in RN+1 and at each step of the iterative algorithm, an orthogonal projection is performed onto one of the constraint sets associated with the cost function in a sequential manner. The PESC algorithm provides globally optimal solutions for different functions such as total variation, L1-norm, L2-norm, and entropic cost functions. Denoising, deconvolution and compressive sensing are among the applications of PESC algorithm. The Projection onto Epigraph Set of Total Variation function (PES-TV) is used in 2-D applications and for 1-D applications Projection onto Epigraph Set of L1-norm cost function (PES-L1) is utilized. In PES-L1 algorithm, fi rst the observation signal is decomposed using wavelet or pyramidal decomposition. Both wavelet denoising and denoising methods using the concept of sparsity are based on soft-thresholding. In sparsity-based denoising methods, it is assumed that the original signal is sparse in some transform domain such as Fourier, DCT, and/or wavelet domain and transform domain coefi cients of the noisy signal are soft-thresholded to reduce noise. Here, the relationship between the standard soft-thresholding based denoising methods and sparsity-based wavelet denoising methods is described. A deterministic soft-threshold estimation method using the epigraph set of `1-norm cost function is presented. It is demonstrated that the size of the `1-ball can be determined using linear algebra. The size of the L1-ball in turn determines the soft-threshold. The PESC, PES-TV and PES-L1 algorithms, are described in detail in this thesis. Extensive simulation results are presented. PESC based inverse restoration and reconstruction algorithm is compared to the state of the art methods in the literature.
Ters evrişim kullanarak pasif radarlarda menzil çözünürlüğü artırma
Passive radar (PR) systems attract interests in radar community due to its lower cost and power consumption over conventional radars. However, one of the main disadvantages of a PR system is its low range resolution. The reason for this is, the range resolution depends on the bandwidth of the transmitted waveform and in a PR scenario, it is impossible to change transmitted waveform properties of a commercial broadcast. In this thesis, a post processing scheme is proposed to improve the range resolution of an FM broadcast based PR system. In the post processing scheme, the output of the ambiguity function is re-expressed as convolution of the autocorrelation of the transmitted signal and a channel impulse response. Therefore, it is shown that it is possible to use deconvolution methods to compute the channel impulse response using the output of the ambiguity function and the autocorrelation of the transmitted signal. Thus, using deconvolution to solve the channel impulse response provides an increase in the range resolution of the PR system. The method successfully increases the target separation distance and range resolution of a PR system using single FM channel signal. The conventional ambiguity function is able to separate two targets when the targets have about 17 km between each other where as the deconvolution based post processing method can decrease this to about 10 km. The deconvolution based post processing methods also decreases the side lobes around the target when the system uses multi channel FM signals. For a scenario in which three FM channels are employed, the highest side lobe is -1.2 dB below the main target peak and after deconvolution, this highest side lobe decreases to about -10 dB below the main target peak.
H&E boyanmış karaciğer dokularında histopatolojik kanser kök hücre imgelerinin sınıflandırılması
Microscopic images are an essential part of cancer diagnosis process in modern medicine. However, diagnosing tis sues under microscope is a time-consuming task for pathologists. There is also a significant variation in pathologists' decisions on tissue labeling. In this study, we developed a computer-aided diagnosis (CAD) system that classifies and grades H&E stained liver tissue images for pathologists in order to speed up the cancer diagnosis process. This system is designed for H&E stained tissues, because it is cheaper than the conventional CD13 stain. The first step is labeling the tissue images for classification purposes. CD13 stained tissue images are used to construct ground truth labels, because in H&E stained tissues cancer stem cells (CSC) cannot be observed by naked eye. Feature extraction is the next step. Since CSCs cannot be observed by naked eye in H&E stained tissues, we need to extract distinguishing texture features. For this purpose, 20 features are chosen from nine different color spaces. These features are fed into a modified version of Principal Component Analysis (PCA) algorithm, which is proposed in this thesis. This algorithm takes covariance matrices of feature matrices of images instead of pixel values of images as input. Images are compared in the eigenspace and classifies them according to the angle between them. It is experimentally shown that this algorithm can achieve 76.0% image classification accuracy in H&E stained liver tissues for a three-class classification problem. Scale invariant feature transform (SIFT), local binary patterns (LBP) and directional feature extraction algorithms are also utilized to classify and grade H&E stained liver tissues. It is observed in the experiments that these features do not provide meaningful information to grade H&E stained liver tissue images. Since our aim is to speed up the cancer diagnosis process, computationally efficient versions of proposed modified PCA algorithm are also proposed. Multiplication-free cosine-like similarity measures are employed in the modified PCA algorithm and it is shown that some versions of the multiplication-free similarity measure based modified PCA algorithm produces better classification accuracies than the standard modified PCA algorithm. One of the proposed multiplication-free similarity measures achieves 76.0% classification accuracy in our dataset containing 454 images of three classes.
Histopatolojik imgeler için imge işleme algoritmaları
Conventionally, a pathologist examines cancer cell morphologies under microscope. This process takes a lot of time and is subject to human mistakes. Computer aided diagnosis (CAD) systems and modules aim to help pathologists in their work to decrease the time consumption and the human mistakes. This thesis proposes a CAD module and algorithms which assist the pathologist in segmentation, detection and the classification problems in histopatholgic images. A multi-resolution super-pixel based segmentation algorithm is developed to measure the cell size, count the number of cells and track the motion of cells in Mesenchymal Stem Cell (MSC) images. The proposed algorithm is compared with Simple Linear Iterative Clustering (SLIC) algorithm. It is experimentally observed that in the segmentation stage, the cell detection rate is increased by 7% and the false alarm is decreased by 5%. In addition to this, two novel decision rules for merging similar neighboring super-pixels are proposed. One dimensional version of the Scale Invariant Feature Transform (SIFT) based merging algorithm is developed and applied to the histograms of the neighboring super-pixels to determine the similar regions. It is also shown that the merging process can be made with the use of wavelets. Moreover, it is shown that region covariance and codifference matrices can be used in detection of cancer stem cells (CSC) and a CAD module for the CSC detection in liver cancer tissue images are developed. The system locates CSCs in CD13 stained liver tissue images. The method has an online learning approach which improves the accuracy of detection. It is experimentally shown that, applying the proposed approach with the user guidance,increases the overall detection quality and accuracy up to 25% compared to using region descriptors alone. Also, the proposed module is compared with the similar plug-ins of ImageJ and Fiji. It is shown that, when the similar features are used, the implemented module achieves approximately 20% better classification results compared to the plug-ins of Imagej and Fiji. Furthermore, the proposed 1-D SIFT algorithm is expanded and used in classification of the cancer tissues images stained with Hematoxylin and Eosin (H&E) stain, which is a cost effective routine compared to the immunohistochemistry (IHC) procedure. The 1-D SIFT algorithm is able to classify healthy and cancerous tissue images with up to 91% accuracy in H&E stained images in our data set.
Doğrusal olmayan saçılma temelli oznitelik çıkarma kullanarak gemilerin akustik izlerinin sınıflandırılması
This thesis proposes a vessel recognition and classification system based on acoustic signatures. Conventionally, acoustic sounds are recognized by sonar operators who listen to audio signals received by ship sonars. The aim of this work is to replace this conventional human-based recognition system with an automatic feature-based classification system. Therefore, it can be regarded reasonable to adopt the speech recognition algorithms in classification of underwater acoustic signal recognition (UASR). The most widely used feature extraction methods of speech recognition are Linear Predictive Coding (LPC) and Mel Frequency Cepstral Coefficients (MFCC) and they are also used in UASR. In addition, the Scattering transform is used to obtain filter bank instead of mel-scale filter bank in MFCC algorithm. The scattering cascade decomposes an input signal into its wavelet modulus coefficients and various non-linearities are used between wavelet stages. The new proposed method is labeled as Scattering Transform Cepstral Coefficients (STCC). Sensitivity of human hearing system is not the same in all frequency bands and mel-scale filter bank in MFCC is more sensitive to small changes in low frequencies than high frequencies. Therefore, number of DWT decomposition levels is increased in low frequencies to determine accurate representation and experimental results shows that non-uniform filter banks provide better success rates. Non-linear Teager energy and hyperbolic tangent operators are used to increase the performance of classification in proposed features extraction methods. Non-linear operators and scattering transforms are used for the first time in UASR to identify the acoustic sounds of the platforms. Teager Energy Operator (TEO) estimates the true energy of the source of a resonance signal. TEO based MFCC, being more robust in noisy conditions than conventional MFCC, provides a better estimation of the platform energy. Although TEO has positive effect on MFCC, it decreases the performance of STCC. Different non-linear tanh operator is also applied to LPC, MFCC and STCC algorithms and experimental results show that tanh operator increases the performance of the classification in all feature extraction methods. This analysis and implementation was carried out with datasets of 24 different vessel signals recordings that belong to 10 separate classes of vessels. Artificial Neural Networks (ANN) and Support Vector Machines (SVM) are used as classifiers. Performance of the proposed methods is compared and experimental results demonstrate that STCC have the best performance and tanh based STCC achieves highest success rate with 98.50% accuracy in classification of vessel sounds.
Omurga seslerinin omurga sağlığı değerlendirmesi amacıyla analizi
This thesis proposes a spinal health assessment system based on acoustic biosignals. The aim of this study is to offer an alternative to the conventional spinal health assessment techniques such as MR, CT or x-ray scans. As conventional methods are time-consuming, expensive and harmful (radiation risk caused by medical scanning techniques), a cheap, fast and harmless method is proposed. It is observed that individuals with spinal health problems have unusual sounds. Using automatic speech recognition (ASR) algorithms, a diagnosis algorithm was developed for classifying joint sounds collected from the vertebrae of human subjects. First, feature parameters are extracted from spinal sounds. One of the most popular feature parameters used in speech recognition are Mel Frequency Cepstrum Coeffcients (MFCC). MFCC parameters are classified using Artificial Neural Networks (ANN). In addition, the scattering transform cepstral coefficients (STCC) algorithm is implemented as an alternative to the mel filterbank in MFCC. The correlation between the medical history of the subjects and the "click" sound in the collected sound data is the basis of the classification algorithm. In the light of collected data, it is observed that "click" sounds are detected in the individuals who have suffered low back pain (slipped disk) but not in healthy individuals. The identification of the "click" sound is carried out by using MFCC/STCC and ANN. The system has 92.2% success rate of detecting "click" sounds when MFCC based algorithm is used. The success rate is 83.5% when STCC feature extraction scheme is used.
Ortak fark öznitelikleri kullanarak görsel nesne takibi
Visual object tracking has been one of the widely studied computer vision tasks which has a broad range of applications in various areas from surveillance to medical studies. There are different approaches proposed for the problem in the literature. While some of them use generative methods where an appearance model is built and used for localizing the object on the image, others use discriminative approaches that models the object and background as two different classes and turns the tracking task into a binary classification problem. In this study, we propose a novel object tracking algorithm based on co-difference matrix and compare its performance with the recent state-of-the-art tracking algorithms on two specific applications. Experiments on a large class of datasets show that the proposed co-difference based object tracking algorithm has successful results in terms of track maintenance, success rate and localization accuracy. The proposed algorithm uses co-difference matrix as the image descriptor. Extraction of co-difference features is similar to the well known covariance method. However the vector product operator is redefined in a multiplication-free manner. The new operator yields a computationally efficient implementation for real time object tracking applications. For our experiments, we prepared a comparison framework that contains over 70000 annotated images for visual object tracking task. We conducted experiments for two different application areas seperately. The first one is infrared surveillance sytems. For this application, we used a thermal image dataset that contains various objects such as humans, cars and military vehicles. The second application area is cell tracking on time-lapse microscopy images. Image sequences for the second application contain cells of different shapes and sizes. For both applications, datasets include a considerable amount of rotation and background clutter. Performance of the tracking algorithms are evaluated quantitatively based on three different metrics. These metrics measure the track maintenance score, success rate and localization accuracy of an algorithm. Experiments indicate that the proposed co-difference based tracking algorithm is among the best performing methods by having the highest localization accuracy and success rate for the surveillance dataset, and the highest track maintenance score for the cell motility dataset.
İmge ve video işleme uygulamaları için faz tabanlı yaklaşımlar
In this thesis, phase information is utilized to address several issues in image processing applications; namely image quality assessment, image contrast enhancement, and visual object tracking. The classical two-dimensional (2D) mel-cepstrum features, which ignore the phase information by design, are enhanced with image phase to form the 2D complex mel-cepstrum features. While integrating the phase information with the existing cepstral features, the unwrapping of phase information is carried out. The 2D complex mel-cepstrum features are fed into a regression scheme to map the feature matrices to subjective scores for the assessment of image quality. A Fourier domain approach for contrast enhancement of microscopy images is developed. The enhancement framework determines the frequency components in which the phase transitions are significant. The significant spectrum components are amplified by a factor depending on the level of transitions. In this way, phase variations are translated into amplitude changes which directly contribute to the enhancement process. Selective variation, which is an extension to the classical total variation framework, is introduced to determine the appropriate parameter set for the enhancement framework. The selective variation scheme evaluates the variations of the image in the high-frequency regions. A visual object tracking scheme based on image phase information is proposed. The main aim of the proposed scheme is to reduce the computational complexity of cross-correlation based matching frameworks. Starting from the derivation of normalized cross-correlation function, the tracking solution is simplified to a phase minimization problem under certain assumptions. The utilization of look-up tables for phase shifts enables a further decrease in computational cost.
Dalgacık dönüşümü kullanarak videoda insan yüzü ve gözlerin yerlerinin tespiti
Human face detection and eye localization problems have received signiï¬cantattention during the past several years because of wide range of commercial andlaw enforcement applications. In this thesis, wavelet domain based human facedetection and eye localization algorithms are developed. After determining allpossible face candidate regions using color information in a given still image orvideo frame, each region is ï¬ltered by a high-pass ï¬lter of a wavelet transform.In this way, edge-highlighted caricature-like representations of candidate regionsare obtained. Horizontal, vertical and ï¬lter-like edge projections of the candi-date regions are used as feature signals for classiï¬cation with dynamic program-ming (DP) and support vector machines (SVMs). It turns out that the proposedfeature extraction method provides good detection rates with SVM based clas-siï¬ers. Furthermore, the positions of eyes can be localized successfully usinghorizontal projections and proï¬les of horizontal- and vertical-crop edge image re-gions. After an approximate horizontal level detection, each eye is ï¬rst localizedhorizontally using horizontal projections of associated edge regions. Horizontaledge proï¬les are then calculated on the estimated horizontal levels. After deter-mining eye candidate points by pairing up the local maximum point locationsin the horizontal proï¬les with the associated horizontal levels, the veriï¬cation isalso carried out by an SVM based classiï¬er. The localization results show thatthe proposed algorithm is not aï¬ected by both illumination and scale changes.iiiivKeywords: Human face detection, eye detection, eye localization, wavelet trans-form, edge projections, classiï¬cation, support vector machines, dynamic program-ming.
3-boyutlu modellerin imge sıkıştırma yöntemleriyle sıkıştırılması
A Connectivity-Guided Adaptive Wavelet Transform (CGAWT) based mesh compres-sion algorithm is proposed. On the contrary to previous work, the proposed methoduses 2D image processing tools for compressing the mesh models. The 3D models areï¬rst transformed to 2D images on a regular grid structure by performing orthogonalprojections onto the image plane. This operation is computationally simpler than pa-rameterization. The neighborhood concept in projection images is diï¬erent from 2Dimages because two connected vertex can be projected to isolated pixels. Connectiv-ity data of the 3D model deï¬nes the interpixel correlations in the projection image.Thus the wavelet transforms used in image processing do not give good results onthis representation. CGAWT is deï¬ned to take advantage of interpixel correlations inthe image-like representation. Using the proposed transform the pixels in the detailsubbands are predicted from their connected neighbors in the low-pass subbands ofthe wavelet transform. The resulting wavelet data is encoded using either ?Set Parti-tioning In Hierarchical Trees? (SPIHT) or JPEG2000. SPIHT approach is progressivebecause diï¬erent resolutions of the mesh can be reconstructed from diï¬erent partitionsof SPIHT bitstream. On the other hand, JPEG2000 approach is a single rate coder.The quantization of the wavelet coeï¬cients determines the quality of the reconstructediiimodel in JPEG2000 approach. Simulations using diï¬erent basis functions show thatlazy wavelet basis gives better results. The results are improved using the CGAWTwith lazy wavelet ï¬lterbanks. SPIHT based algorithm is observed to be superior toJPEG2000 based mesh coder and MPEG-3DGC in rate-distortion.Keywords: 3D Model Compression, Image-like mesh representation, Connectivity-Guided Adaptive Wavelet Transformiv
Videoda dinamik doku analizi ve alev, duman, uçucu organik bileşik buharı bulmaya uygulanması
Dynamic textures are moving image sequences that exhibit stationary characteristics in time such as fire, smoke, volatile organic compound (VOC) plumes, waves, etc. Most surveillance applications already have motion detection and recognition capability, but dynamic texture detection algorithms are not integral part of these applications. In this thesis, image processing based algorithms for detection of specific dynamic textures are developed. Our methods can be developed in practical surveillance applications to detect VOC leaks, fire and smoke. The method developed for VOC emission detection in infrared videos uses a change detection algorithm to find the rising VOC plume. The rising characteristic of the plume is detected using a hidden Markov model (HMM). The dark regions that are formed on the leaking equipment are found using a background subtraction algorithm. Another method is developed based on an active learning algorithm that is used to detect wild fires at night and close range flames. The active learning algorithm is based on the Least-Mean-Square (LMS) method. Decisions from the sub-algorithms, each of which characterize a certain property of the texture to be detected, are combined using the LMS algorithm to reach a final decision. Another image processing method is developed to detect fire and smoke from moving camera video sequences. The global motion of the camera is compensated by finding an affine transformation between the frames using optical flow and RANSAC. Three frame change detection methods with motion compensation are used for fire detection with a moving camera. A background subtraction algorithm with global motion estimation is developed for smoke detection.
Piroelektrik kızılberisi algılayıcı tabanlı olay tespiti
Pyroelectric Infra-red (PIR) sensors have been extensively used in indoor andoutdoor applications as they are low cost, easy to use and widely available. PIRsensors respond to IR radiating objects moving in its viewing range. The currentsensors give an output of logical one when they detect a hot object's motion anda logical zero when there is no moving hot object. In this method, only movingobjects can be detected and the rate of false alarm is high.New types of PIR sensors are more sophisticated and more capable. Theyhave a lower false alarm ratio compared to classical ones. Although they candistinguish pets and humans, again they can only be used for detection of hotobject motions due to the limitations caused by the usage of the simple comparatorstructure inside. This structure is unalterable, not flexible for development, and not suitable for implementing algorithms.A new approach is developed to use PIR sensors by modifying the sensorcircuitry. Instead of directly using the output of a classical PIR sensor, an analogsignal is extracted from the PIR output and it is sampled. As a result,intelligent signal processing algorithms can be developed using the discrete-timesensor signal. In this way, it is possible to develop human, pet and flame detection methods. It is also possible to find the direction of moving objects and estimate their distances from the sensor. Furthermore, the path of a moving target can be estimated using a PIR sensor array.We focus on object and event classification using sampled PIR sensor signals.Pet, human and flame detection methods are comparatively investigated.Different human motion events are modeled and classifed using Hidden MarkovModels (HMM) and Conditional Gaussian Mixture Models (CGMMs). The sampleddata is wavelet transformed for feature extraction and then fed into HMMsfor analysis. The final decision is reached according to the Markov Model producingthe highest probability. Experimental results demonstrate the reliabilityof the proposed HMM based decision and event classification algorithm.
Resimlerde ve videolarda ateş ve alev tespiti yöntemleri
In this thesis, automatic fire detection methods are studied incolor domain, spatial domain and temporal domain. We firstinvestigated fire and flame colors of pixels. Chromatic Model,Fisher's linear discriminant, Gaussian mixture color model andartificial neural networks are implemented and tested for flamecolor modeling. For images a system that extracts patches andclassifies them using textural features is proposed. Performance ofthis system is given according to different thresholds and differentfeatures. A real-time detection system that uses information incolor, spatial and temporal domains is proposed for videos. Thissystem, which is develop by modifying previously implementedsystems, divides video into spatiotemporal blocks and uses featuresextracted from these blocks to detect fire.
Gündüz ve kızılötesı kamera videolarında alev tespit algoritmalarının CUDA tabanlı gerçekleştirilmesi
Automatic fire detection in videos is an important task but it is a challenging problem. Video based high performance fire detection algorithms are important for the detection of forest fires. The usage area of fire detection algorithms can further be extended to the places like state and heritage buildings, in which surveillance cameras are installed. In uncontrolled fires, early detection is crucial to extinguish the fire immediately. However, most of the current fire detection algorithms either suffer from high false alarm rates or low detection rates due to the optimization constraints for real-time performance. This problem is also aggravated by the high computational complexity in large areas, where multi-camera surveillance is required. In this study, our aim is to speed up the existing color video fire detection algorithms by implementing in CUDA, which uses the parallel computational power of Graphics Processing Units (GPU). Our method does not only speed up the existing algorithms but it can also reduce the optimization constraints for real-time performance to increase detection probability without affecting false alarm rates. In addition, we have studied several methods that detect flames in infrared video and proposed an improvement for the algorithm to decrease the false alarm rate and increase the detection rate of the fire.
Aralık dışbükey programlama ve seyreklik kullanan imge ve sinyal işleme algoritmaları
In this thesis, signal and image processing algorithms based on sparsity and interval convex programming are developed for inverse problems. Inverse signal processing problems are solved by minimizing the L1 norm or the Total Variation (TV) based cost functions in the literature. A modified entropy functional approximating the absolute value function is defined. This functional is also used to approximate the L1 norm, which is the most widely cost function in sparse signal processing problems. The modified entropy functional is continuous, differentiable and convex. As a result, it is possible to develop iterative, globally convergent algorithms for compressive sensing, denoising and restoration problems using the modified entropy functional. Iterative interval convex programming algorithms are constructed using Bregman's D-Projection operator. In sparse signal processing, it is assumed that the signal can be represented using a sparse set of coefficients in some transform domain. Therefore, by minimizing the total variation of the signal, it is expected to realize sparse representations of signals. Another cost function that is introduced for inverse problems is the Filtered Variation (FV) function, which is the generalized version of the Total Variation (VR) function. The TV function uses the differences between the pixels of an image or samples of a signal. This is essentially simple Haar filtering. In FV, high-pass filter outputs are used instead of differences. This leads to flexibility in algorithm design adapting to the local variations of the signal. Extensive simulation studies using the new cost functions are carried out. Better experimental restoration and reconstruction results are obtained compared to the algorithms in the literature.
Çoklu algılayıcı tabanlı çevre destekli yaşam sistemi
An important goal of Ambient Assisted Living (AAL) research is to contribute to the quality of life of the elderly and handicapped people and help them to maintain an independent lifestyle with the use of sensors, signal processing and the available telecommunications infrastructure. From this perspective, detection of unusual human activities such as falling person detection has practical applications. In this thesis, a low-cost AAL system using vibration and passive infrared (PIR) sensors is proposed for falling person detection, human footstep detection, human motion detection, unusual inactivity detection, and indoor flooding detection applications. For the vibration sensor signal processing, various frequency analysis methods which consist of the discrete Fourier transform (DFT), mel-frequency cepstral coefficients (MFCC), discrete wavelet transform (DWT) with different filter-banks, dual-tree complex wavelet transform (DT-CWT), and single-tree complex wavelet transform (ST-CWT) are compared to each other to obtain the best possible classification result in our dataset. Adaptive-threshold based Markov model (MM) classifier is preferred for the human footstep detection. Vibration sensor based falling person detection system employs Euclidean distance and support vector machine (SVM) classifiers and these classifiers are compared to each other. PIR sensors are also used for falling person detection and this system employs two PIR sensors. To achieve the most reliable system, a multi-sensor based falling person detection system which employs one vibration and two PIR sensors is developed. PIR sensor based system has also the capability of detecting uncontrolled flames and this system is integrated to the overall system. The proposed AAL system works in real-time on a standard personal computer or chipKIT Uno32 microprocessors without computers. A network is setup for the communication of the Uno32 boards which are connected to different sensors. The main processor gives final decisions and emergency alarms are transmitted to outside of the smart home using the auto-dial alarm system via telephone lines. The resulting AAL system is a low-cost and privacy-friendly system thanks to the types of sensors used. Keywords: Ambient assisted living, vibration sensor, passive infrared sensor, complex wavelet transform, support vector machines, falling person detection, Markov models, human footstep detection, unusual inactivity detection, indoor flooding detection.
İmge işleme için kullanılan çok ölçekli yönsel öznitelik çıkarma yöntemlerinin karşılaştırılması
Almost all images that are presented in classi cation problems regardless of area of application, have directional information embedded into its texture. Although there are many algorithms developed to extract this information, there is no `golden' method that works the best every image. In order to evaluate performance of these developed algorithms, we consider 7 different multi-scale directional feature extraction algorithms along with our own multi-scale directional ltering framework. We perform tests on several problems from diverse areas of application such as font/style recognition on English, Arabic, Farsi, Chinese, and Ottoman texts, grading of follicular lymphoma images, and stratum corneum thickness calculation. We present performance metrics such as k-fold cross validation accuracies and times to extract feature from one sample, and compare with the respective state of art on each problem. Our multi-resolution computationally efficient directional approach provides results on a par with the state of the art directional feature extraction methods.