Browsing by Author "Mahbubur Rahman, Dr. S.M."

Now showing 1 - 4 of 4

Entropy-based image registration method using the curvelet transform
(Department of Electrical and Electronic Engineering (EEE), 2011-07) Mushfiqul Alam, Md.; Mahbubur Rahman, Dr. S.M.
Integration of images from different sources is increasingly used in different visual signal processing applications. But before the integration it is of utmost importance that the images are geometrically aligned and this aligning process is very often referred to as ‘image registration’. There are two approaches of registering the reference and distorted images, viz., feature-based and intensity-based. The intensity-based methods give better accuracy comparing to the feature-based methods by considering entire pixels of the images, instead of considering just few selected geometric features as in the latter method. Traditionally, image registration is carried out in different transform domains to incorporate the facilities that these transforms provide. Discrete wavelet transform (DWT) has widely been used for image registration. But it has poor directional selectivity. The complex wavelet, ridgelet and shearlet transforms proved slight improvement in directional selectivity over the DWT. Recently, the curvelet transform proved its superiority in directional selectivity among all these wavelet-like transforms, being able to properly detect the commonly occurred curve and edge singularities in images. Hence, in this thesis, an image registration algorithm is developed that uses the curvelet coefficients of images. Commonly-used probabilistic objective functions in the intensity-based registration algorithms include mutual information, joint entropy and cross-correlation of the transform coefficients of the distorted and reference images. None of these functions consider the conditional dependencies among the images which may exist as the images to be registered are usually captured from a same scene. In this thesis, a new conditional entropy-based objective function is developed using a suitable probabilistic modeling of the approximate level curvelet coefficients of images. The suitability of the probability distribution of the curvelet coefficients of images is validated with a standard statistical test of fit. For the purpose of alignment, a linear transformation, viz., affine transform is used. Extensive experimentations are carried out to test the performance of the proposed registration method as compared to other existing methods using commonly-used performance metrics.
Facial expression recognition using 2D gauss hermite moments of images
(Department of Electrical and Electronic Engineering (EEE), 2014-08) Saif Muhammad Imran; Mahbubur Rahman, Dr. S.M.
Automatic recognition of facial expressions can be an important component of natural human-machine interfaces; in behavioural sciences and in clinical practices. Expression recognition can be considered to consist of deformations of facial parts and their spatial relations or changes in the pigmentation of the face. The challenge of such recognition lies in classifying expressions both in the case of posed and spontaneous forms; where the former is an intentional expression and the latter is natural. Two major approaches of facial expression recognition include the holistic and landmark. This thesis deals with holistic based feature extraction since such approach considers the entire face images instead of selecting a few interested areas using computationally expensive algorithms for locating landmarks. Commonly used feature extraction techniques in holistic expression recognition methods include the principal component analysis (PCA), the linear discriminant analysis (LDA) and their variants, independent component analysis and even using the orthogonal moments. However, most of the existing approaches fail to consider either the local inherent spatial changes of the facial expression e.g., PCA or LDA. Although orthogonal moments carry local information of facial regions, the previously proposed methods select higher order moments heuristically without any justification. In this thesis, the Gauss-Hermite Moments (GHMs) are used for developing a holistic facial expression recognition algorithm since the GHMs are widely used in visual signal processing. Based on a novel concept of scattering ratio, moments are selected having higher discrimination power of expression in the GHM subspace. Further, due to the existence of significant correlations among certain expressions in the case of spontaneous form as compared to posed form, the GHM features are projected to a new expression subspace where the information are de-correlated using the PCA. Finally, these feature vectors are used to recognize the expressions using the well known support vector machine classifier. Experiments are carried out using two exhaustive databases, namely, the Cohn Kanade and Facial Recognition Grand Challenge, the former representing posed expressions while the latter spontaneous expressions. Experimental results on mutually exclusive subjects reveal that the proposed method can provide the recognition accuracies of at least 7 % and 4 % higher than the existing methods for posed and spontaneous expressions, respectively.
Gaussian-Hermite moment-based depth estimation from monocular image for stereo vision
(Department of Electrical and Electronic Engineering (EEE), 2015-07) Samiul Haque; Mahbubur Rahman, Dr. S.M.
Depth estimation has turned into an emerging and challenging eld of research in computer vision. The ubiquitous availability of stereo display technology has increased the demand of visual media with depth information. All major block- buster movies and games are now being released with stereo versions. Strict pa- rameters are required for generating stereo image in multiview set up. Thus the high-complexity of multiview imaging arrangement of capturing stereo image has motivated researchers to nd a robust method of estimating depth from conven- tional 2D imaging set up. In addition to that 3D display technology has evolved into an advanced stage, but huge amount of media are still in 2D, in such a case depth estimation from monocular image is the only solution for generating a stereo view. Hence, research e orts are ongoing to develop low-complexity depth estima- tion algorithm for a scene specially from its monoscopic images captured using a CCD camera. Existing depth calculation methods from monocular images include depth from motion, depth form geometry and depth estimation using a learned database. These methods are limited by object geometry, prior knowledge of the scene as well as highly prone to noise. So, there is still a search for robust and autonomous depth calculation algorithm which does not depend on speci c scene classes. Motivated by the noise robust and invariant properties of orthogonal moments to the geometry of objects, this thesis presents a new moment based depth estimation method, which is independent of any prior knowledge about the scene. In par- ticular, the Gaussian-Hermite moments (GHMs) which are very popular in visual signal processing are chosen to estimate the focus cue of a pixel from its neighbor- hood. It is known that there exists a signi cant correlation among the neighboring pixels in terms of depth information except for the sharp edges of an object. Hence a closed from expression of image matting is applied on the focus map of the im- age to generate the desired depth map. Extensive experiments are carried out in order to compare the proposed GHM-based depth estimation method with the existing methods using commonly-referred images in the literature. Performance comparisons of depth estimation in terms of visual quality, stereo generation and mean opinion score show that the proposed method performs signi cantly better than other methods.
Multiple time spatial images for video-based automatic tracking of vehicles
(Department of Electrical and Electronic Engineering (EEE), 2014-06) Niluthpol Chowdhury Mithun; Mahbubur Rahman, Dr. S.M.
Video-based vehicle tracking has become an active research area due to its numerous transportation related applications. Some common challenges in traditional video-based tracking methods include initialization of tracking, tracking an unknown number of targets, sensitivity to drift from true position due to the variations in lighting condition, scene conditions and camera position in long sequences, and absence of corrective mechanism. In this thesis, a novel approach for unsupervised vehicle tracking algorithms is developed by introducing multiple time-spatial images (MTSIs)-based detection in the Monte-Carlo Particle filter or Kalman filter based-tracking. Such a use of MTSIs in tracking algorithm provides the opportunity of reliable identification of a vehicular object automatically whenever it appears in a scene. Notably, the proposed tracking method employs the concept of multiple numbers of key vehicular frames (KVFs) for each of the vehicular-objects in the traffic. These KVFs allow an accurate estimate of the centroid position of a vehicle in the key frames, due to the fact that the relative sizes of the vehicles captured in the video are maintained in these KVFs. The spatial correspondence of a vehicle in KVFs is then integrated in Particle filter or Kalman filter-based tracking as a corrective measure to alleviate the common problem of drifting and thereby increasing the accuracy in tracking trajectory. Extensive experimentations are carried out in vehicular traffics of varying environments to evaluate the tracking performance of the proposed method as compared with the existing methods. Experimental results demonstrate that the proposed approach not only automates the initialization of tracking procedure, but also increases the accuracy of tracking trajectory evaluated by the closeness of centroids of a vehicular object both in the forward and backward tracking.