Machine learning techniques for music information retrieval

Tese de doutoramento, Informática (Engenharia Informática), Universidade de Lisboa, Faculdade de Ciências, 2015

Some Efficient Representations of 3D Information based on Octress

The thesis introduced the octree and addressed the complete nature of problems encountered, while building and imaging system based on octrees. An efficient Bottom-up recursive algorithm and its iterative counterpart for the raster to octree conversion of CAT scan slices, to improve the speed of generating the octree from the slices, the possibility of utilizing the inherent parallesism in the conversion programme is explored in this thesis. The octree node, which stores the volume information in cube often stores the average density information could lead to “patchy”distribution of density during the image reconstruction. In an attempt to alleviate this problem and explored the possibility of using VQ to represent the imformation contained within a cube. Considering the ease of accommodating the process of compressing the information during the generation of octrees from CAT scan slices, proposed use of wavelet transforms to generate the compressed information in a cube. The modified algorithm for generating octrees from the slices is shown to accommodate the eavelet compression easily. Rendering the stored information in the form of octree is a complex task, necessarily because of the requirement to display the volumetric information. The reys traced from each cube in the octree, sum up the density en-route, accounting for the opacities and transparencies produced due to variations in density.

A Prototype for a Multimodal Biometric Security system based on Face and Audio Signatures

Any automatically measurable, robust and distinctive physical characteristic or personal trait that can be used to identify an individual or verify the claimed identity of an individual, referred to as biometrics, has gained significant interest in the wake of heightened concerns about security and rapid advancements in networking, communication and mobility. Multimodal biometrics is expected to be ultra-secure and reliable, due to the presence of multiple and independent—verification clues. In this study, a multimodal biometric system utilising audio and facial signatures has been implemented and error analysis has been carried out. A total of one thousand face images and 250 sound tracks of 50 users are used for training the proposed system. To account for the attempts of the unregistered signatures data of 25 new users are tested. The short term spectral features were extracted from the sound data and Vector Quantization was done using K-means algorithm. Face images are identified based on Eigen face approach using Principal Component Analysis. The success rate of multimodal system using speech and face is higher when compared to individual unimodal recognition systems

Development of 2-D MRT based Image Compression Techniques

The thesis explores the area of still image compression. The image compression techniques can be broadly classified into lossless and lossy compression. The most common lossy compression techniques are based on Transform coding, Vector Quantization and Fractals. Transform coding is the simplest of the above and generally employs reversible transforms like, DCT, DWT, etc. Mapped Real Transform (MRT) is an evolving integer transform, based on real additions alone. The present research work aims at developing new image compression techniques based on MRT. Most of the transform coding techniques employ fixed block size image segmentation, usually 8×8. Hence, a fixed block size transform coding is implemented using MRT and the merits and demerits are analyzed for both 8×8 and 4×4 blocks. The N2 unique MRT coefficients, for each block, are computed using templates. Considering the merits and demerits of fixed block size transform coding techniques, a hybrid form of these techniques is implemented to improve the performance of compression. The performance of the hybrid coder is found to be better compared to the fixed block size coders. Thus, if the block size is made adaptive, the performance can be further improved. In adaptive block size coding, the block size may vary from the size of the image to 2×2. Hence, the computation of MRT using templates is impractical due to memory requirements. So, an adaptive transform coder based on Unique MRT (UMRT), a compact form of MRT, is implemented to get better performance in terms of PSNR and HVS The suitability of MRT in vector quantization of images is then experimented. The UMRT based Classified Vector Quantization (CVQ) is implemented subsequently. The edges in the images are identified and classified by employing a UMRT based criteria. Based on the above experiments, a new technique named “MRT based Adaptive Transform Coder with Classified Vector Quantization (MATC-CVQ)”is developed. Its performance is evaluated and compared against existing techniques. A comparison with standard JPEG & the well-known Shapiro’s Embedded Zero-tree Wavelet (EZW) is done and found that the proposed technique gives better performance for majority of images

Perceptually-based Comparison of Image Similarity Metrics

The image comparison operation ??sessing how well one image matches another ??rms a critical component of many image analysis systems and models of human visual processing. Two norms used commonly for this purpose are L1 and L2, which are specific instances of the Minkowski metric. However, there is often not a principled reason for selecting one norm over the other. One way to address this problem is by examining whether one metric better captures the perceptual notion of image similarity than the other. With this goal, we examined perceptual preferences for images retrieved on the basis of the L1 versus the L2 norm. These images were either small fragments without recognizable content, or larger patterns with recognizable content created via vector quantization. In both conditions the subjects showed a consistent preference for images matched using the L1 metric. These results suggest that, in the domain of natural images of the kind we have used, the L1 metric may better capture human notions of image similarity.

A hybrid neural classifier for dimensionality reduction and data visualization and its application to fault detection and classification of induction motors

In this paper, a hybrid neural classifier combining the auto-encoder neural network and the Lattice Vector Quantization (LVQ) model is described. The auto-encoder network is used for dimensionality reduction by projecting high dimensional data into the 2D space. The LVQ model is used for data visualization by forming and adapting the granularity of a data map. The mapped data are employed to predict the target classes of new data samples. To improve classification accuracy, a majority voting scheme is adopted by the hybrid classifier. To demonstrate the applicability of the hybrid classifier, a series of experiments using simulated and real fault data from induction motors is conducted. The results show that the hybrid classifier is able to outperform the Multi-Layer Perceptron neural network, and to produce very good classification accuracy rates for various fault conditions of induction motors.

Multi-scale local pattern co-occurrence matrix for textural image classification

Textural image classification technologies have been extensively explored and widely applied in many areas. It is advantageous to combine both the occurrence and spatial distribution of local patterns to describe a texture. However, most existing state-of-the-art approaches for textural image classification only employ the occurrence histogram of local patterns to describe textures, without considering their co-occurrence information. And they are usually very time-consuming because of the vector quantization involved. Moreover, those feature extraction paradigms are implemented at a single scale. In this paper we propose a novel multi-scale local pattern co-occurrence matrix (MS_LPCM) descriptor to characterize textural images through four major steps. Firstly, Gaussian filtering pyramid preprocessing is employed to obtain multi-scale images; secondly, a local binary pattern (LBP) operator is applied on each textural image to create a LBP image; thirdly, the gray-level co-occurrence matrix (GLCM) is utilized to extract local pattern co-occurrence matrix (LPCM) from LBP images as the features; finally, all LPCM features from the same textural image at different scales are concatenated as the final feature vectors for classification. The experimental results on three benchmark databases in this study have shown a higher classification accuracy and lower computing cost as compared with other state-of-the-art algorithms.

Classification of healthcare data using genetic fuzzy logic system and wavelets

Healthcare plays an important role in promoting the general health and well-being of people around the world. The difficulty in healthcare data classification arises from the uncertainty and the high-dimensional nature of the medical data collected. This paper proposes an integration of fuzzy standard additive model (SAM) with genetic algorithm (GA), called GSAM, to deal with uncertainty and computational challenges. GSAM learning process comprises three continual steps: rule initialization by unsupervised learning using the adaptive vector quantization clustering, evolutionary rule optimization by GA and parameter tuning by the gradient descent supervised learning. Wavelet transformation is employed to extract discriminative features for high-dimensional datasets. GSAM becomes highly capable when deployed with small number of wavelet features as its computational burden is remarkably reduced. The proposed method is evaluated using two frequently-used medical datasets: the Wisconsin breast cancer and Cleveland heart disease from the UCI Repository for machine learning. Experiments are organized with a five-fold cross validation and performance of classification techniques are measured by a number of important metrics: accuracy, F-measure, mutual information and area under the receiver operating characteristic curve. Results demonstrate the superiority of the GSAM compared to other machine learning methods including probabilistic neural network, support vector machine, fuzzy ARTMAP, and adaptive neuro-fuzzy inference system. The proposed approach is thus helpful as a decision support system for medical practitioners in the healthcare practice.

Structural classification of proteins through amino acid sequence using interval type-2 fuzzy logic system

This paper introduces a new multi-output interval type-2 fuzzy logic system (MOIT2FLS) that is automatically constructed from unsupervised data clustering method and trained using heuristic genetic algorithm for a protein secondary structure classification. Three structure classes are distinguished including helix, strand (sheet) and coil which correspond to three outputs of the MOIT2FLS. Quantitative properties of amino acids are used to characterize the twenty amino acids rather than the widely used computationally expensive binary encoding scheme. Amino acid sequences are parsed into learnable patterns using a local moving window strategy. Three clustering tasks are performed using the adaptive vector quantization method to derive an equal number of initial rules for each type of secondary structure. Genetic algorithm is applied to optimally adjust parameters of the MOIT2FLS with the purpose of maximizing the Q3 measure. Comprehensive experimental results demonstrate the strong superiority of the proposed approach over the traditional methods including Chou-Fasman method, Garnier-Osguthorpe-Robson method, and artificial neural network models.

Multi-output interval type-2 fuzzy logic system for protein secondary structure prediction

A new multi-output interval type-2 fuzzy logic system (MOIT2FLS) is introduced for protein secondary structure prediction in this paper. Three outputs of the MOIT2FLS correspond to three structure classes including helix, strand (sheet) and coil. Quantitative properties of amino acids are employed to characterize twenty amino acids rather than the widely used computationally expensive binary encoding scheme. Three clustering tasks are performed using the adaptive vector quantization method to construct an equal number of initial rules for each type of secondary structure. Genetic algorithm is applied to optimally adjust parameters of the MOIT2FLS. The genetic fitness function is designed based on the Q3 measure. Experimental results demonstrate the dominance of the proposed approach against the traditional methods that are Chou-Fasman method, Garnier-Osguthorpe-Robson method, and artificial neural network models.

Contribuições aos Processos de Clustering com Base em Métricas não-Euclidianas

In this work we present a new clustering method that groups up points of a data set in classes. The method is based in a algorithm to link auxiliary clusters that are obtained using traditional vector quantization techniques. It is described some approaches during the development of the work that are based in measures of distances or dissimilarities (divergence) between the auxiliary clusters. This new method uses only two a priori information, the number of auxiliary clusters Na and a threshold distance dt that will be used to decide about the linkage or not of the auxiliary clusters. The number os classes could be automatically found by the method, that do it based in the chosen threshold distance dt, or it is given as additional information to help in the choice of the correct threshold. Some analysis are made and the results are compared with traditional clustering methods. In this work different dissimilarities metrics are analyzed and a new one is proposed based on the concept of negentropy. Besides grouping points of a set in classes, it is proposed a method to statistical modeling the classes aiming to obtain a expression to the probability of a point to belong to one of the classes. Experiments with several values of Na e dt are made in tests sets and the results are analyzed aiming to study the robustness of the method and to consider heuristics to the choice of the correct threshold. During this work it is explored the aspects of information theory applied to the calculation of the divergences. It will be explored specifically the different measures of information and divergence using the Rényi entropy. The results using the different metrics are compared and commented. The work also has appendix where are exposed real applications using the proposed method

Técnicas neuronais de agrupamentos aplicadas à detecção de anomalias em mamogramas digitais

This work proposes the development of an intelligent system for analysis of digital mammograms, capable to detect and to classify masses and microcalcifications. The digital mammograms will be pre-processed through techniques of digital processing of images with the purpose of adapting the image to the detection system and automatic classification of the existent calcifications in the suckles. The model adopted for the detection and classification of the mammograms uses the neural network of Kohonen by the algorithm Self Organization Map - SOM. The algorithm of Vector quantization, Kmeans it is also used with the same purpose of the SOM. An analysis of the performance of the two algorithms in the automatic classification of digital mammograms is developed. The developed system will aid the radiologist in the diagnosis and accompaniment of the development of abnormalities

Algoritmo SOM com estrutura hierárquica e dinâmica aplicado a compressão de imagens

ln this work the implementation of the SOM (Self Organizing Maps) algorithm or Kohonen neural network is presented in the form of hierarchical structures, applied to the compression of images. The main objective of this approach is to develop an Hierarchical SOM algorithm with static structure and another one with dynamic structure to generate codebooks (books of codes) in the process of the image Vector Quantization (VQ), reducing the time of processing and obtaining a good rate of compression of images with a minimum degradation of the quality in relation to the original image. Both self-organizing neural networks developed here, were denominated HSOM, for static case, and DHSOM, for the dynamic case. ln the first form, the hierarchical structure is previously defined and in the later this structure grows in an automatic way in agreement with heuristic rules that explore the data of the training group without use of external parameters. For the network, the heuristic mIes determine the dynamics of growth, the pruning of ramifications criteria, the flexibility and the size of children maps. The LBO (Linde-Buzo-Oray) algorithm or K-means, one ofthe more used algorithms to develop codebook for Vector Quantization, was used together with the algorithm of Kohonen in its basic form, that is, not hierarchical, as a reference to compare the performance of the algorithms here proposed. A performance analysis between the two hierarchical structures is also accomplished in this work. The efficiency of the proposed processing is verified by the reduction in the complexity computational compared to the traditional algorithms, as well as, through the quantitative analysis of the images reconstructed in function of the parameters: (PSNR) peak signal-to-noise ratio and (MSE) medium squared error

Contribuições a técnicas de agrupamento e visualização de dados multivariados utilizando mapas auto-organizáveis

Self-organizing maps (SOM) are artificial neural networks widely used in the data mining field, mainly because they constitute a dimensionality reduction technique given the fixed grid of neurons associated with the network. In order to properly the partition and visualize the SOM network, the various methods available in the literature must be applied in a post-processing stage, that consists of inferring, through its neurons, relevant characteristics of the data set. In general, such processing applied to the network neurons, instead of the entire database, reduces the computational costs due to vector quantization. This work proposes a post-processing of the SOM neurons in the input and output spaces, combining visualization techniques with algorithms based on gravitational forces and the search for the shortest path with the greatest reward. Such methods take into account the connection strength between neighbouring neurons and characteristics of pattern density and distances among neurons, both associated with the position that the neurons occupy in the data space after training the network. Thus, the goal consists of defining more clearly the arrangement of the clusters present in the data. Experiments were carried out so as to evaluate the proposed methods using various artificially generated data sets, as well as real world data sets. The results obtained were compared with those from a number of well-known methods existent in the literature

Lattice constellations and codes from quadratic number fields

We propose new classes of linear codes over integer rings of quadratic extensions of Q, the field of rational numbers. The codes are considered with respect to a Mannheim metric, which is a Manhattan metric modulo a two-dimensional (2-D) grid. In particular, codes over Gaussian integers and Eisenstein-Jacobi integers are extensively studied. Decoding algorithms are proposed for these codes when up to two coordinates of a transmitted code vector are affected by errors of arbitrary Mannheim weight. Moreover, we show that the proposed codes are maximum-distance separable (MDS), with respect to the Hamming distance. The practical interest in such Mannheim-metric codes is their use in coded modulation schemes based on quadrature amplitude modulation (QAM)-type constellations, for which neither the Hamming nor the Lee metric is appropriate.