Theses supervised by Prof. Dr. A. Enis Çetin
11 theses · İhsan Doğramacı Bilkent University
Orman yangını gözetleme amaçlı video işleme algoritmaları
We propose various image and video processing algorithms for wildfire surveillance. The proposed methods include; classifier fusion, online learning, real-time feature extraction, image registration and optimization. We develop an entropy functional based online classifier fusion framework. We use Bregman divergences as the distance measure of the projection operator onto the hyperplanes describing the output decisions of classifiers. We test the performance of the proposed system in a wildfire detection application with stationary cameras that scan predefined preset positions. In the second part of this thesis, we investigate different formulations and mixture applications for passive-aggressive online learning algorithms. We propose a classifier fusion method that can be used to increase the performance of multiple online learners or the same learners trained with different update parameters. We also introduce an aerial wildfire detection system to test the real-time performance of the analyzed algorithms. In the third part of the thesis we propose a real-time dynamic texture recognition method using random hyperplanes and deep neural networks. We divide dynamic texture videos into spatio-temporal blocks and extract features using local binary patterns (LBP). We reduce the computational cost of the exhaustive LBP method by using randomly sampled subset of pixels in the block. We use random hyperplanes and deep neural networks to reduce the dimensionality of the final feature vectors. We test the performance of the proposed method in a dynamic texture database. We also propose an application of the proposed method in real-time detection of flames in infrared videos. Using the same features we also propose a fast wildfire detection system using pan-tilt-zoom cameras and panoramic background subtraction. We use a hybrid method consisting of speeded-up robust features and mutual information to register consecutive images and form the panorama. The next step for multi-modal surveillance applications is the registration of images obtained with different devices. We propose a multi-modal image registration algorithm for infrared and visible range cameras. A new similarity measure is described using log-polar transform and mutual information to recover rotation and scale parameters. Another similarity measure is introduced using mutual information and redundant wavelet transform to estimate translation parameters. The new cost function for translation parameters is minimized using a novel lifted projections onto convex sets method. Keywords: Projections onto convex sets, classifier fusion, online learning, entropymaximization, wildfire detection, adaptive filtering, LMS, Bregman divergence,image Processing, infrared, mutual information, wavelet transform, image registration,log-polar transform, supporting hyperplanes..
Çokkipli işaret ve imge çözümleme tabanlı yangın tespit algoritmaları
Dynamic textures are common in natural scenes. Examples of dynamic tex-tures in video include fire, smoke, clouds, volatile organic compound (VOC)plumes in infra-red (IR) videos, trees in the wind, sea and ocean waves, etc.Researchers extensively studied 2-D textures and related problems in the fieldsof image processing and computer vision. On the other hand, there is very littleresearch on dynamic texture detection in video. In this dissertation, signal andimage processing methods developed for detection of a specific set of dynamictextures are presented.Signal and image processing methods are developed for the detection of flamesand smoke in open and large spaces with a range of up to 30m to the camera invisible-range (IR) video. Smoke is semi-transparent at the early stages of fire.Edges present in image frames with smoke start loosing their sharpness and thisleads to an energy decrease in the high-band frequency content of the image.Local extrema in the wavelet domain correspond to the edges in an image.The decrease in the energy content of these edges is an important indicatorof smoke in the viewing range of the camera. Image regions containing flames appear asfire-colored (bright) moving regions in (IR) video. In addition to motion andcolor (brightness) clues, the flame flicker process is also detected by using a Hid-den Markov Model (HMM) describing the temporal behavior. Image frames arealso analyzed spatially. Boundaries of flames are represented in wavelet domain.High frequency nature of the boundaries of fire regions is also used as a clue tomodel the flame flicker. Temporal and spatial clues extracted from the video arecombined to reach a final decision.Signal processing techniques for the detection of flames with pyroelectric (pas-sive) infrared (PIR) sensors are also developed. The flame flicker process of anuncontrolled fire and ordinary activity of human beings and other objects aremodeled using a set of Markov models, which are trained using the wavelet trans-form of the PIR sensor signal. Whenever there is an activity within the viewingrange of the PIR sensor, the sensor signal is analyzed in the wavelet domain andthe wavelet signals are fed to a set of Markov models. A fire or no fire decision ismade according to the Markov model producing the highest probability.Smoke at far distances (> 100m to the camera) exhibits different temporal andspatial characteristics than nearby smoke and fire. This demands specific methodsexplicitly developed for smoke detection at far distances rather than using nearbysmoke detection methods. An algorithm for vision-based detection of smoke dueto wild fires is developed. The main detection algorithm is composed of foursub-algorithms detecting (i) slow moving objects, (ii) smoke-colored regions, (iii)rising regions, and (iv) shadows. Each sub-algorithm yields its own decision as azero-mean real number, representing the confidence level of that particular sub-algorithm. Confidence values are linearly combined for the final decision.Another contribution of this thesis is the proposal of a framework for activefusion of sub-algorithm decisions. Most computer vision based detection algo-rithms consist of several sub-algorithms whose individual decisions are integratedto reach a final decision. The proposed adaptive fusion method is based on theleast-mean-square (LMS) algorithm. The weights corresponding to individualsub-algorithms are updated on-line using the adaptive method in the training(learning) stage. The error function of the adaptive training process is definedas the difference between the weighted sum of decision values and the decisionof an oracle who may be the user of the detector. The proposed decision fusionmethod is used in wildfire detection.
Kızılberisi algılayıcılarla uçucu organik bileşenler kaçağı tespiti
Advances in technology and industry leads to a rise in the living standards ofpeople. However, this has also introduced a variety of serious problems, suchas the undesired release of combustible and toxic gases which have become anessential part of domestic and industrial life. Therefore, detection andmonitoring of VOC gases have become a major problem in recent years. In thisthesis, we propose novel methods for detection and monitoring VOC gas leaksby using a Pyro-electric (or Passive) Infrared (PIR) sensor and a thermopilesensor. A continuous time analog signal is obtained for both of the sensors andsent to a PC for signal processing. While using the PIR sensor, we have HiddenMarkov Models (HMM) for each type of event to be classified. Then, by usinga probabilistic approach we determine which class any test signal belongs to. Inthe case of a thermopile sensor, in addition to Hidden Markov Modelingmethod, we also use a method based on the period of the sensor signal. Thefrequency of the output signal of the thermopile sensor increases with thepresence of VOC gas leak. By using this fact, we control whether the period ofa test signal is below a predefined threshold or not. If it is, our system triggersan alarm. Moreover, we present different methods to find the periods of a givensignal.
Radarda otomatik hedef sınıflandırma için yöntemler
Automatictarget recognition (ATR) using radar is an active research area. Inthis thesis, we develop new automatic radar target classificationmethods. We focus on two specific problems: (i) Synthetic ApertureRadar (SAR) target classification and (ii)Pulse-doppler radar (PDR)target classification. SAR and PDR target classification areextensively used for ground and battlefield surveillance tasks.In the first part of the thesis, a novel descriptive featureparameter extraction method from Synthetic Aperture Radar (SAR)images is proposed. Feature extraction and classification methodswhich were developed to handle optical images are usuallyinappropriate for SAR images because of the multiplicative nature ofthe severe speckle noise and imaging defects. In addition, SARimages of the same object taken at different aspect angles showgreat differences, which makes it hard to obtain satisfactoryresults. Consequently, feature parameter extraction method based ontwo-dimensional cepstrum is proposed and its object recognitionresults are compared with principal component analysis (PCA) andindependent component analysis (ICA) methods. The extracted featureparameters are classified using Support Vector Machines (SVMs).Experimental results are presented.In the second part of the thesis, the automatic classificationexperiments over ground surveillance Pulse-doppler radar echo signalare investigated in order to overcome the limitations of humanoperators and improve the classification accuracy. Covariance methodapproach is introduced for PDR echo signal classification. To thebest our knowledge, the use of covariance method-basedclassification is not investigated in radar automatic targetclassification problems. Furthermore, different approaches whichinvolves SVMs are developed. As feature parameters, cepstrum andMFCCs are used. Performances of these two approaches are comparedwith the Gaussian Mixture Models (GMM) based classification scheme.Experimental results and conclusions are presented.
Obje ve doku tespiti ve sınıflandırması
Object and texture recognition are two important subjects in computer vision. An efficient and fast algorithm to compute a short and efficient feature vector for classification of images is crucial for smart video surveillance systems. In this thesis, feature extraction methods for object and texture classification are investigated, compared and developed.A method for object classification based on shape characteristics is developed. Object silhouettes are extracted from videos by using the background subtraction method. Contour of the objects are obtained from these silhouettes and this 2-D contour signals are transformed into 1-D signals by using a type of radial transformation. Discrete cosine transformation is used to acquire the frequency characteristics of these signals and a support vector machine (SVM) is employed for classification of objects according to this frequency information. This method is implemented and integrated into a real time system together with object tracking.For texture recognition problem, we defined a new computationally efficient operator forming a semigroup on real numbers. The new operator does not require any multiplications. The codifference matrix based on the new operator is defined and an image descriptor using the codifferencematrix is developed. Texture recognition and license plate identification examples based on the new descriptor are presented. We compared our method with regular covariance matrix method. Our method has lower computational complexity and it is experimentally shown that it performs as well as the regular covariance method.
Hareket vektörleri ile içerik tabanlı kopya video sezimi
In this thesis, we propose a motion vector based Video Content Based Copy Detection (VCBCD) method. Detecting the videos violating the copyright of the owner comes into question by growing broadcasting of digital video on different media. Unlike watermarking methods in VCBCD methods, the video itself is considered as a signature of the video and representative feature parameters are extracted from a given video and compared with the feature parameters of a test video. Motion vectors of image frames are one of the signatures of a given video. We first investigate how well the motion vectors describe the video.We use Mean value of Magnitudes of Motion Vectors (MMMV) and Mean value of Phases of Motion Vectors (MPMV) of macro blocks, which are the main building blocks of MPEG-type video coding methods. We show that MMMV and MPMV plots may not represent videos uniquely with little motion content because the average of motion vectors in a given frame approaches zero.To overcome this problem we calculate the MMMV and MPMV graphs in a lower frame rate than the actual frame rate of the video. In this way, the motion vectors may become larger and as a result robust signature plots are obtained. Another approach is to use the Histogram of Motion Vectors (HOMV) that includes both MMMV and MPMV information.We test and compare MMMV, MPMV and HOMV methods using test videos including copies and the original movies.
SAR imgelerinde hedef tespit yöntemleri
Automatic recognition and classification of man-made objects in SAR (Synthetic Aperture Radar) images have been an active research area because SAR sensors can produce images of scenes in all weather conditions at any time of the day which is not possible with infrared and optical sensors [1, 2]. In this thesis, different feature parameter extraction methods from SAR images are proposed. The new approach is based on region covariance (RC) method which involves the computation of a covariance matrix of a ROI (region of interest). Entries of the covariance matrix are used in target detection. In addition, the use of computationally more efficient region codifference matrix for target detection in SAR images is also introduced. Simulation results of target detection in MSTAR (Moving and Stationary Target Recognition) database are presented. The RC and region codifference methods deliver high detection accuracies and low false alarm rates. The performance of these methods is investigated with various distance metrics and Support Vector Machine (SVM) classifiers. It is also observed that the region codifference method produces better results than the commonly used Principle Component Analysis (PCA) method which is used together with SVM.The second part of the thesis offers some techniques to decrease the computational cost of the proposed methods. In this approach, ROIs are filtered by directional filters (DFs) at first as a pre-processing stage. Images are categorized according to the filter outputs. The proposed RC and codifference methods are applied within the categories determined by these filters. Simulation results of target detection in MSTAR database are presented through decisions made with l1 norm distance metric and SVM. The number of comparisons made with the training images using l1 norm distance measure decreases as these images are distributed into categories. Therefore, the computational cost of the previous algorithm is significantly reduced. SAR image classification results based on l1 norm distance metric are better than the results obtained using SVM and they show that the two-stage approach does not reduce the performance rate of the previously proposed method much, especially when codifference features are used.
Işık değişimlerine dayanıklı video işleme yöntemleri
Moving shadows constitute problems in various applications such as image segmentation, smoke detection and object tracking. Main cause of these problems is the misclassification of the shadow pixels as target pixels. Therefore, the use of an accurate and reliable shadow detection method is essential to realize intelligent video processing applications. In the first part of the thesis, a cepstrum based method for moving shadow detection is presented. The proposed method is tested on outdoor and indoor video sequences using well-known benchmark test sets. To show the improvements over previous approaches, quantitative metrics are introduced and comparisons based on these metrics are made.Most video processing applications require object tracking as it is the base operation for real-time implementations such as surveillance, monitoring and video compression. Therefore, accurate tracking of an object under varying scene and illumination conditions is crucial for robustness. It is well known that illumination variations on the observed scene and target are an obstacle against robust object tracking causing the tracker lose the target. In the second part of the thesis, a two dimensional (2D) cepstrum based approach is proposed to overcome this problem. Cepstral domain features extracted from the target region are introduced into the covariance tracking algorithm and it is experimentally observed that 2D-cepstrum analysis of the target region provides robustness to varying illumination conditions. Another contribution is the development of the co-difference matrix based object tracking instead of the recently introduced co-variance matrix based method.One of the problems with most target tracking methods is that they do not have a well-established control mechanism for target loss which usually occur when illumination conditions suddenly change. In the final part of the thesis, a confidence interval based statistical method is developed for target loss detection. Upper and lower bound functions on the cumulative density function (cdf) of the target feature vector are estimated for a given confidence level. Whenever the estimated cdf of the detected region exceeds the bounds it means that the target is no longer tracked by the tracking algorithm. The method is applicable to most tracking algorithms using features of the target image region.
İmge öznitelik çıkarımı için kepstral yöntemler
Image feature extraction is one of the most vital tasks in computer vision and pattern recognition applications due to its importance in the preparation of data extracted from images.In this thesis, 2D cepstrum based methods (2D mel- and Mellin-cepstrum) are proposed for image feature extraction. The proposed feature extraction schemes are used in face recognition and target detection applications. The cepstral features are invariant to amplitude and translation changes. In addition, the features extracted using 2D Mellin-cepstrum method are rotation invariant. Due to these merits, the proposed techniques can be used in various feature extraction problems.The feature matrices extracted using the cepstral methods are classified by Common Matrix Approach (CMA) and multi-class Support Vector Machine (SVM). Experimental results show that the success rates obtained using cepstral feature extraction algorithms are higher than the rates obtained using standard baselines (PCA, Fourier-Mellin Transform, Fourier LDA approach). Moreover, it is observed that the features extracted by cepstral methods are computationally more efficient than the standard baselines.In target detection task, the proposed feature extraction methods are used in the detection and discrimination stages of a typical Automatic Target Recognition (ATR) system. The feature matrices obtained from the cepstral techniques are applied to the SVM classifier. The simulation results show that 2D cepstral feature extraction techniques can be used in the target detection in SAR images.
Mikroskopik görüntülerin bilgisayar destekli yorumlanması için imge işleme yöntemleri
Image processing algorithms for automated analysis of microscopic images havebecome increasingly popular in the last decade with the remarkable growth incomputational power. The advent of high-throughput scanning devices allowsfor computer-assisted evaluation of microscopic images, resulting in a quick andunbiased image interpretation that will facilitate the clinical decision-making process.In this thesis, new methods are proposed to provide solution to two imageanalysis problems in biology and histopathology.The first problem is the classification of human carcinoma cell line images.Cancer cell lines are widely used for research purposes in laboratories all overthe world. In molecular biology studies, researchers deal with a large numberof specimens whose identity have to be checked at various points in time. Anovel computerized method is presented for cancer cell line image classification.Microscopic images containing irregular carcinoma cell patterns are representedby subwindows which correspond to foreground pixels. For each subwindow,a covariance descriptor utilizing the dual-tree complex wavelet transform (DTCWT)coefficients as pixel features is computed. A Support Vector Machine(SVM) classifier with radial basis function (RBF) kernel is employed for finalclassification. For 14 different classes, we achieve an overall accuracy of 98%,which outperforms the classical covariance based methods.Histopathological image analysis problem is related to the grading of follicularlymphoma (FL) disease. FL is one of the commonly encountered cancer types inthe lymph system. FL grading is based on histological examination of hematoxilinand eosin (H&E) stained tissue sections by pathologists who make clinical decisionsby manually counting the malignant centroblast (CB) cells. This gradingmethod is subject to substantial inter- and intra-reader variability and samplingbias. A computer-assisted method is presented for detection of CB cells in H&EstainedFL tissue samples. The proposed algorithm takes advantage of the scalespacerepresentation of FL images to detect blob-like cell regions which reside inthe scale-space extrema of the difference-of-Gaussian images. Multi-stage falsepositive elimination strategy is employed with some statistical region propertiesand textural features such as gray-level co-occurrence matrix (GLCM), gray-levelrun-length matrix (GLRLM) and Scale Invariant Feature Transform (SIFT). Thealgorithm is evaluated on 30 images and 90% CB detection accuracy is obtained,
Mikroskopik imge işleme, analiz, sınıflandırma ve sıkıştırma için yeni yöntemler
Microscopic images are frequently used in medicine and molecular biology. Many interesting image processing problems arise after the initial data acquisition step, since image modalities are manifold. In this thesis, we developed several algorithms in order to handle the critical pipeline of microscopic image storage/compression and analysis/classification more efficiently.The first step in our processing pipeline is image compression. Microscopic images are large in size (e.g. 100K-by-100K pixels), therefore finding efficient ways of compressing such data is necessary for efficient transmission, storage and evaluation.We propose an image compression scheme that uses the color content of a given image, by applying a block-adaptive color transform. Microscopic images of tissues have a very specific color palette due to the staining process they undergo before data acquisition. The proposed color transform takes advantage of this fact and can be incorporated into widely-used compression algorithms such as JPEG and JPEG 2000 without creating any overhead at the receiver due to its DPCM-like structure. We obtained peak signal-to-noise ratio gains up to 0.5 dB when comparing our method with standard JPEG.The next step in our processing pipeline is image analysis. Microscopic image processing techniques can assist in making grading and diagnosis of images reproducible and by providing useful quantitative measures for computer-aided diagnosis. To this end, we developed several novel techniques for efficient feature extraction and classification of microscopic images.We use region co-difference matrices as inputs for the classifier, which have the main advantage of yielding multiplication-free computationally efficient algorithms. The merit of the co-difference framework for performing some important tasks in signal processing is discussed.We also introduce several methods that estimate underlying probability density functions from data. We use sparsity criteria in the Fourier domain to arrive at efficient estimates. The proposed methods can be used for classification in Bayesian frameworks.We evaluated the performance of our algorithms for two image classification problems: Discriminating between different grades of follicular lymphoma, a medical condition of the lymph system, as well as differentiating several cancer cell lines from each another. Classification accuracies over two large data sets (270 images for follicular lymphoma and 280 images for cancer cell lines) were above 98%.