Academic Highlights

Texture analysis is one of the key techniques of image understanding and processing with widespread applications from low level image segmentation to high level object recognition. Gaussian Markov random field (GMRF) is a particular model based texture feature extraction scheme which uses model parameters as texture features. In this study a novel robust texture descriptor based on GMRF is proposed specially for texture segmentation and classification. For these tasks, descriptive features are more favorable relative to the generative features. Therefore, in order to achieve more descriptive features, with the GMRFs, a localized parameter estimation technique is introduced here. The issues arising in the localized parameter estimation process, due to the associated small sample size, are addressed by applying Tikhonov regularization and an estimation window size selection criterion. The localized parameter estimation process proposed here can overcome the problem of parameter smoothing that occurs in traditional GMRF parameter estimation. Such a parameter smoothing disregards some important structural and statistical information for texture discrimination. The normalized distributions of local parameter estimates are proposed as the new texture features and are named as Local Parameter Histogram (LPH) descriptors. Two new rotation invariant texture descriptors based on LPH features are also introduced, namely Rotation Invariant LPH (RI-LPH) and Isotropic LPH (I-LPH) descriptors. The segmentation and classification results on large texture datasets demonstrate that these descriptors significantly improve the performance of traditional GMRF features and also demonstrate better performance in comparison with the state-of-the-art texture descriptors. Satisfactory natural image segmentation is also carried out based on the novel features. Furthermore, proposed features are employed in a real world medical application involving tissue recognition for emphysema, a critical lung disease causing lung tissue destruction. Features extracted from High Resolution Computed Tomography (HRCT) data are used in effective tissue recognition and pathology distribution diagnosis. Moreover, preliminary work on a Bayesian framework for integrating prior knowledge into the parameter estimation process is undertaken to introduce further improved texture features.

The method proposed in this project attempt to build a face recognition system with the use of a random codebook which is generated randomly without using a clustering algorithm. Instead, it performs a simple transformation against the extracted feature vector of a face for forming the histogram or the bag of key points. Therefore, the system will bypass the recursive computational load of a clustering algorithm. In the experiments freely available AT&T and Yale databases are used.

Initially, the feature vectors are created using the SIFT algorithm for a given face instances. The feature vectors inherit the problem of high dimensionality and also they are of variable lengths for each face. Hence feature vectors cannot be directly classified using standard methods of classification. Therefore, according to the proposed method these feature vectors are converted in to fixed dimensional vectors called transformed vectors which are going to be the abstract representation of all faces. For this conversion the random codebook approach is used. Transformed vector will be the bag of key points which can be classified using a suitable classifier. Even though the vector length variation can be addressed by above technique still the problem of high dimensional data exist when it comes to larger number of different face instances. Therefore, to reduce the dimensionality the Principle Component Analysis (PCA) is used. PCA is performed on transformed vectors which were created using the features of faces extracted by SIFT algorithm. In a way this face recognition system gets the combined advantage of SIFT and PCA. Identifying the effective classifying method which suited the classification of transformed vectors is also tested. The recognition efficiency of the proposed system will also depend on its classifier module. Therefore, Support Vector Machines (SVM) and k nearest neighbor classifier are used to analyze the performance.

This project suggests low cost multimode plastic optical fiber as a cost effective alternative for expensive glass fiber for short distance fiber communication links. An analysis of PMMA (Polymethylmetacrylate) and PFBVE (Perfluorobutenylvinylether) plastic fibers are carried out in order to evaluate their characteristics and to evaluate their capability to be used in some specific applications. Based upon the results, two copper to fiber media converters for RS232 and E1 level PDH standards, were developed. First the low cost RS232 optical link with optimal design of transmitter and receiver are implemented. Then the transmitter and receiver is upgraded to communicate E1 level data stream while keeping the simplicity and inexpensive qualities. These converters effectively convert electrical signals to required optical signals for transmission through plastic fiber medium. The results reveal that these optical fiber converters function very satisfactorily in the real world data transmission and can be implemented with a significantly lower budget compared to the currently used commercial media converters. These media converters can overwhelm the existing media converters due to their low cost, high reliability, simplicity, ability to plug and play, reduction of delays and reduction of size.

This project is an attempt to make an average power meter to measure power in a renewable energy system which is expected to be operated by wind energy. This solution is more cost effective and simple than many other available electronic power meters. The average power measurement is different from the instantaneous power reading which is comparatively easy to implement than the average power meter. The average power meter which was implemented is a prototype for measuring power in a scale down 230V, 1kW power system. The main target was to measure the average power produced by the wind power plant continuously. The meter is to be fixed inside the power station and directly connected to the power output lines from the wind power plant. The proposed power meter is a PIC microcontroller based digital power meter. The designing, implementing and testing of the component units of the power meter was carried out and the correct average power measurements were displayed successfully. This prototype power meter is intended to develop further to implement in the wind power station at Rasnayakepura, Nikewaratiya, Sri Lanka.

Accurate identification of plants is essential in large number of applications such as medical applications, natural productdevelopment, agricultural research purposes and daily household activities. A multitude of plants carry significant information which can be used to identify them. An automated leaf recognition system utilizes features of the leaves such as shape and texture in plant species recognition without requiring the human intervention. In this project an automated leaf recognition system to classify different plant species based on the shape and texture of the leaves is proposed. The proposed system maps acquired images of leaves to a texture and shape based feature space and use Linear Discriminant Analysis (LDA) to extract dominant features. Then k Nearest Neighbor (k-NN) is used for leaf classification. The results shows that the leaves can be classified to an accuracy of 95.6±1.7% on a leaf database consisting 32 leaf classes and 1907 leaf samples.

This paper proposes a deep neural network (DNN) approach for image texture classification. Optimal texture feature extraction to discriminate many texture classes is a challenging task. The choice of traditional texture features for texture classification is subjective and highly application dependent with lower generalization to other textures. Therefore, here we suggest a DNN approach which can overcome the problem of selecting a suitable and optimal texture feature extraction method. A DNN comprised of many hidden layers could extract low level to high level texture features automatically in an orderly fashion and able to use in multi class texture classification successfully. Classification results on publically available texture datasets show that deep learning techniques are applicable in the analysis of an image for identification through

texture features.