Skip to main content
main-content

Über dieses Buch

This book constitutes the refereed proceedings of the 4th Mexican Conference on Pattern Recognition, MCPR 2012, held in Huatulco, Mexico, in June 2012. The 31 revised full papers and 3 keynotes presented were carefully reviewed and selected from 64 submissions and are organized in topical sections on image processing; computer vision and image recognition; pattern recognition and neural networks; and document processing and speech recognition.

Inhaltsverzeichnis

Frontmatter

Keynote Addresses

From Linear Representations to Object Parts

The use of the skeleton for object decomposition in the framework of the structural approach to shape description is discussed. Object decomposition is obtained starting from a suitable partition of the skeleton. The elements of the skeleton partition are then used as seeds from which to recover the various regions into which the object is decomposed. A merging process is also accomplished so as to have a final decomposition in accordance with human perception and stable when the object is available in different poses or sizes.

Gabriella Sanniti di Baja, L. Serino, Carlo Arcelli

Perceptual Grouping Using Superpixels

Perceptual grouping plays a critical role in both human and computer vision. However, with the object categorization community’s preoccupation with object detection, interest in perceptual grouping has waned. The reason for this is clear: the object-independent, mid-level shape priors that form the basis of perceptual grouping are subsumed by the object-dependent, high-level shape priors defined by a target object. As the recognition community moves from object detection back to object recognition, a linear search through a large database of target models is intractable, and perceptual grouping will be essential for sublinear scaling. We review two approaches to perceptual grouping based on grouping superpixels. In the first, we use symmetry to group superpixels into symmetric parts, and then group the parts to form structured objects. In the second, we use contour closure to group superpixels, yielding a figure-ground segmentation.

Sven J. Dickinson, Alex Levinshtein, Cristian Sminchisescu

Automatic Design of Artificial Neural Networks and Associative Memories for Pattern Classification and Pattern Restoration

In this note we present our most recent advances in the automatic design of artificial neural networks (ANNs) and associative memories (AMs) for pattern classification and pattern recall. Particle Swarm Optimization (PSO), Differential Evolution (DE), and Artificial Bee Colony (ABC) algorithms are used for ANNs; Genetic Programming is adopted for AMs. The derived ANNs and AMs are tested with several examples of well-known databases. As we will show, results are very promising.

Humberto Sossa, Beatriz A. Garro, Juan Villegas, Carlos Avilés, Gustavo Olague

Image Processing

An Automatic Image Scaling Up Algorithm

A fully automatic scaling up algorithm is presented in the framework of interpolation methods. For any integer zooming factor

n

, the algorithm generates a magnified version of an input color image in one scan of the image. The computational complexity of the algorithm is O(N), where N is the size of the input image. The visual aspect of the magnified images is generally appealing also when considering large zooming factors. Peak Signal to Noise Ratio and Structural SIMilarity are used to evaluate the performance of the algorithm and to compare it with other scaling up algorithms.

Maria Frucci, Carlo Arcelli, Gabriella Sanniti di Baja

Sampling Techniques for Monte Carlo Matrix Multiplication with Applications to Image Processing

Randomized algorithms for processing massive data sets have shown to be a promising alternative to deterministic techniques. Sampling strategies are an essential aspect of randomized algorithms for matrix computations. In this work, we show that strategies that are effective or even optimal in the general case, can fail when applied to ill-conditioned matrices. Our experimental study suggests that there exists a relationship between sampling performance and conditioning of the matrices involved. We present an explanation for this behavior and propose a novel, efficient, and accurate sampling strategy for randomized multiplication of affinity matrices in image segmentation.

Humberto Madrid, Valia Guerra, Marielba Rojas

Extended Photometric Sampling for Surface Shape Recovery

Photometric sampling is a process where the surface normals of an object are estimated through the excitation of the object’s surface and a rotating light source around it. The method can be regarded as a special case of photometric stereo when extensive sampling is performed in order to calculate surface normals. The classic photometric sampling approach considers only variations around the azimuth angle of the moving light source. As a consequence, additional attention has to be be paid to the recovery of the light source directions and the removal of specular and shadowed regions. This paper investigates the effect of including variations around the zenith angle of the light source vector in a photometric sampling framework, developing a geometric approach to estimate the surface normal vectors. Experiments show that increasing the number of samples along the zenith variation benefits the estimation of the surface normals.

Felipe Hernández-Rodríguez, Mario Castelán

A Segmentation Method for Tree Crown Detection and Modelling from LiDAR Measurements

A watershed segmentation algorithm is proposed for automatic extraction of tree crowns from LiDAR data to support 3-d modelling of forest stands. A relatively sparse LiDAR point cloud was converted to a surface elevation in raster format and a canopy height model (CHM) extracted. Then, the segmentation method was applied on the CHM and results combined with the original point cloud to generate models of individual tree crowns. The method was tested in 200 circular plots (400

m

2

) located over 50 sites of a conservation area in Mexico City. The segmentation method exhibited a moderate to perfect detection rate on 66% of plots tested. One major factor for a poor detection was identified as the relatively low sampling rate of LiDAR data with respect to crown sizes.

José Luis Silván-Cárdenas

Texture Analysis for Skin Probability Maps Refinement

In this paper a new method for skin regions detection and segmentation is proposed. To improve the conventional color-based skin models, skin probability maps are subject to texture analysis using discriminative statistical features. Although the texture was utilized for skin detection in some of the existing methods, the main contribution of the work reported here is that the probability maps rather than the original color images are processed. The method has been validated in a series of experiments using two data sets. The obtained results are reported in the paper, and they confirm that the method is competitive.

Michal Kawulok

Composite Correlation Filters for Detection of Geometrically Distorted Objects Using Noisy Training Images

Correlation filters for object detection use information about the appearance and shape of the object of interest. Therefore, detection performance degrades when the appearance of the object in the scene differs from the appearance used in the filter design process. This problem has been approached by utilizing composite filters designed from a training set containing known views of the object of interest. However, common composite filter design is usually carried out under the assumption that the ideal appearance and shape of the target are known. In this work we propose an algorithm for composite filter design using noisy training images. The algorithm is a modification of the class synthetic discriminant function technique that uses arbitrary filter impulse responses. Furthermore, filters can be adapted to achieve a prescribed discrimination capability for a class of backgrounds if a representative sample is known. Computer simulation results obtained with the proposed algorithm are presented and compared with those of common composite correlation filters.

Pablo M. Aguilar-González, Vitaly Kober

Adaptive Spatial Concealment of Damaged Coded Images

The transmission over error-prone networks of still images or videos coded by block-based techniques like JPEG and MPEG respectively, may lead to block loss degrading, particularly the visual quality of images. Working under this environment, such as wireless communication where retransmission may be not feasible, application of error concealment techniques is consequently required to reduce degradation caused by the missing information. This paper surveys algorithms for spatial error concealment and proposes an adaptive and effective method based on edge analysis that performs well in current situations where significant loss of information is present and the data of the past reference images are not also available. The proposed method and the reviewed algorithms were implemented, tested and compared. Experimental results show that the proposed approach outperforms existing methods by up to 8.6 dB on average.

Alejandro A. Ramírez-Acosta, Mireya S. García-Vázquez, Mariko Nakano

Computer Vision and Image Recognition

Human Sign Recognition for Robot Manipulation

This paper addresses the problem of recognizing signs generated by a person to guide a robot. The proposed method is based on video color analysis of a moving person making signs. The analysis consists of segmentation of the middle body, arm and forearm location and recognition of the arm and forearm positions. The proposed method was experimentally tested on videos with different target colors and illumination conditions. Quantitative evaluations indicate 97.76% of correct detection of the signs in 1807 frames.

Leonardo Saldivar-Piñon, Mario I. Chacon-Murguia, Rafael Sandoval-Rodriguez, Javier Vega-Pineda

Fuzzy Sets for Human Fall Pattern Recognition

Vision-based fall detection is a challenging problem in pattern recognition. This paper introduces an approach to detect a fall as well as its type in infrared video sequences. The regions of interest of the segmented humans are examined image by image though calculating geometrical and kinematic features. The human fall pattern recognition system identifies true and false falls. The fall indicators used as well as their fuzzy model are explained in detail. The fuzzy model has been tested for a wide number of static and dynamic falls.

Marina V. Sokolova, Antonio Fernández-Caballero

Vision System for 3D Reconstruction with Telecentric Lens

This paper addresses 3D object reconstruction from images acquired by camera-telecentric lense array. Firstly, we present a geometric model of an array camera-telecentric lens. Then we developed and implemented the calibration process using a planar checkerboard pattern. At the same time, we developed a three-dimensional reconstruction system based on contour extraction on objects with dimensions less than 50mm of diameter. Finally an analysis of the uncertainty model parameters and performance reconstruction of 3D objects are presented.

José Guadalupe Rico Espino, José-Joel Gonzalez-Barbosa, Roberto Augusto Gómez Loenzo, Diana Margarita Córdova Esparza, Ricardo Gonzalez-Barbosa

A Tool for Hand-Sign Recognition

We present a software tool created for human-computer interaction based on hand gestures. The underlying algorithm utilizes computer vision techniques. The tool is able to recognize in real-time six different hand signals, captured using a web cam. Experiments conducted to evaluate the system performance are reported.

David J. Rios Soria, Satu Elisa Schaeffer

Improving the Multiple Alignments Strategy for Fingerprint Verification

Developing accurate fingerprint verification algorithms is an active research area. A large amount of fingerprint verification algorithms are based on minutiae descriptors. An important component of these algorithms is the alignment strategy. The single alignment strategy, with

O

(

n

2

) time complexity, uses the local matching minutiae pair that maximizes the similarity value to align the minutiae. Nevertheless, even if the selected minutiae pair is a true matching pair, it is not necessarily the best pair to carry out fingerprint alignment. The multiple alignments strategy alleviates these limitations by performing multiple minutiae alignments, increasing the time complexity to

O

(

n

4

). In this paper, we improve the multiple alignment strategy, reducing its complexity while still achieving a high accuracy. The new strategy is based on the rationale that most minutiae descriptors from one fingerprint correspond with their most similar descriptors from the other fingerprint. To test the new strategy behavior, we adapt three well known algorithms to a traditional multiple alignment strategy and to our strategy. Several experiments in the FVC2004 database show that our strategy outperforms both the single and the multiple alignments strategies.

Miguel Angel Medina-Pérez, Milton García-Borroto, Andres Eduardo Gutierrez-Rodríguez, Leopoldo Altamirano-Robles

Breaking reCAPTCHAs with Unpredictable Collapse: Heuristic Character Segmentation and Recognition

In this paper we present a novel approach for automatic segmentation and recognition of reCAPTCHA in Web sites. It is based on CAPTCHA image preprocessing with character alignment, morphological segmentation with three-color bar character encoding and heuristic recognition. The original proposal consists in exploiting three-color bar code for characters in CAPTCHA for their robust segmentation with presence of random collapse overlapping letters and distortions by particular patterns of waving rotation. Additionally, a novel implementation of SVM-based learning classifier for recognition of combinations of characters in training corpus has been proposed that permits to increment more than twice the recognition success rate without time extension of system response. The main goal of this research is to reduce vulnerability of CAPTCHA from spam and frauds as well as to provide a novel approach for recognizing either handwritten or degraded and damaged texts in ancient manuscripts. Our designed framework implementing the proposed approach has been tested in real-time applications with sites used CAPTCHAS achieving segmentation success rate about of 82% and recognition success rate about of 94%.

Claudia Cruz-Perez, Oleg Starostenko, Fernando Uceda-Ponga, Vicente Alarcon-Aquino, Leobardo Reyes-Cabrera

Using Short-Range Interactions and Simulated Genetic Strategy to Improve the Protein Contact Map Prediction

Protein contact map prediction is one of the most important intermediate steps of the protein folding prediction problem. In this research we want to know how a decision tree predictor based on short-range interactions can learn the correlation among the covalent structures of a protein residues. The proposed solution predicts protein contact maps by the combination of a forest of 400 decision trees with an input codification for short-range interactions and a genetic-based edition method. The method’s performance was satisfactory, improving the accuracy instead using all information of the protein sequence. For a globulin data set the method can predict contacts with a maximal accuracy of 43%. The presented predictive model illustrates that short-range interactions play a predominant role in determining protein structure.

Cosme E. Santiesteban Toca, Milton García-Borroto, Jesus S. Aguilar Ruiz

Pattern Recognition and Neural Networks

Associative Model for Solving the Wall-Following Problem

A navigation system for a robot is presented in this work. The Wall-Following problem has become a classic problem of Robotics due to robots have to be able to move through a particular stage. This problem is proposed as a classifying task and it is solved using an associative approach. In particular, we used Morphological Associative Memories as classifier. Three testing methods were applied to validate the performance of our proposal: Leave-One-Out, Hold-Out and K-fold Cross-Validation and the average obtained was of 91.57%, overcoming the neural approach.

Rodolfo Navarro, Elena Acevedo, Antonio Acevedo, Fabiola Martínez

The List of Clusters Revisited

One of the most efficient index for similarity search, to fix ideas think in speeding up

k

-nn searches in a very large database, is the so called

list of clusters

. This data structure is a counterintuitive construction which can be seen as extremely unbalanced, as opposed to balanced data structures for exact searching. In practical terms there is no better alternative for exact indexing, when every search return all the incumbent results; as opposed to approximate similarity search. The major drawback of the list of clusters is its quadratic time construction.

In this paper we revisit the list of clusters aiming at speeding up the construction time without sacrificing its efficiency. We obtain similar search times while gaining a significant amount of time in the construction phase.

Eric Sadit Tellez, Edgar Chávez

A Heuristically Perturbation of Dataset to Achieve a Diverse Ensemble of Classifiers

Ensemble methods like Bagging and Boosting which combine the decisions of multiple hypotheses are among the strongest existing machine learning methods. The diversity of the members of an ensemble is known to be an important factor in determining its generalization error. We present a new method for generating ensembles, named CDEBMTE (Creation of Diverse Ensemble Based on Manipulation of Training Examples), that directly constructs diverse hypotheses using manipulation of training examples in three ways: (1) sub-sampling training examples, (2) decreasing/increasing errorprone training examples and (3) decreasing/increasing neighbor samples of error-prone training examples.

The technique is a simple, general meta-learner that can use any strong learner as a base classifier to build diverse committees. Experimental results using two well-known classifiers (1) decision-tree induction and (2) multilayer perceptron as two base learners demonstrate that this approach consistently achieves higher predictive accuracy than both the base classifier, Adaboost and Bagging. CDEBMTE also outperforms Adaboost more prominent when training data size is becomes larger.

We propose to show that CDEBMTE can be effectively used to achieve higher accuracy and to obtain better class membership probability estimates.

Experimental results using two well-known classifiers as two base learners demonstrate that this approach consistently achieves higher predictive accuracy than both the base classifier, Adaboost and Bagging. CDEBMTE also outperforms Adaboost more prominent when training data size is becomes larger.

Hamid Parvin, Sajad Parvin, Zahra Rezaei, Moslem Mohamadi

Compact and Efficient Permutations for Proximity Searching

Proximity searching consists in retrieving the most similar objects to a given query. This kind of searching is a basic tool in many fields of artificial intelligence, because it can be used as a search engine to solve problems like

$kN\!N$

searching. A common technique to solve proximity queries is to use an index. In this paper, we show a variant of the permutation based index, which, in his original version, has a great predicting power about which are the objects worth to compare with the query (avoiding the exhaustive comparison). We have noted that when two permutants are close, they can produce small differences in the order in which objects are revised, which could be responsible of finding the true answer or missing it. In this paper we pretend to mitigate this effect. As a matter of fact, our technique allows us both to reduce the index size and to improve the query cost up to 30%.

Karina Figueroa Mora, Rodrigo Paredes

NURBS Parameterization: A New Method of Parameterization Using the Correlation Relationship between Nodes

NURBS (Non-uniform rational B-splines) has become the industry standard tools for the representation, design and data exchange of geometric information to be processed and used by computers because of their useful geometrical properties. The problem of the parameterization of data points in NURBS curve/surface has been considered by several of researchers. We propose in this paper a new parameterization method for NURBS approximation. The current methods of parameterization such as centripetal method uses only the previous knot vector to calculate the recent knot. In this paper, we give a new parameterization method based on the correlation of the nodes. This approach inherits the advantages of the relation and position of the knots.

Sawssen Jalel, Mohamed Naouai, Atef Hamouda, Malek Jebabli

Genetic Algorithm for Multidimensional Scaling over Mixed and Incomplete Data

Multidimensional scaling maps a set of

n

-dimensional objects into a lower-dimension space, usually the Euclidean plane, preserving the distances among objects in the original space. Most algorithms for multidimensional scaling have been designed to work on numerical data, but in soft sciences, it is common that objects are described using quantitative and qualitative attributes, even with some missing values. For this reason, in this paper we propose a genetic algorithm especially designed for multidimensional scaling over mixed and incomplete data. Some experiments using datasets from the UCI repository, and a comparison against a common algorithm for multidimensional scaling, shows the behavior of our proposal.

P. Tecuanhuehue-Vera, Jesús Ariel Carrasco-Ochoa, José Fco. Martínez-Trinidad

Experimental Validation of an Evolutionary Method to Identify a Mobile Robot’s Position

A method to determine the position of a mobile robot using machine learning strategies was introduced in [1]. The method raises the possibility to decrease the size of database that holds the images that describe an area where a robot will localize itself. The present work does a statistical validation of the approach by calculating the Hamming and Euclidean distances between all the images using on the one hand all their pixels and on the other hand the reduced set of pixels obtained by the GA as described in [1]. To perform the analysis, a new series of images were taken from a specific position at several angles in both horizontal (pan) and vertical (tilt). These images were compared using two different measures: a) the Hamming distance and b) the Euclidean distance to determine how similar are one from another.

Angel Kuri-Morales, Ignacio Lopez-Peña

Up and Down Trend Associations in Analysis of Time Series Shape Association Patterns

The method of recognition of shape association patterns with direct and inverse relationships is proposed. This method is based on a new time series shape association measure based on Up and Down trend associations. The application of this technique to analysis of associations between well production data in petroleum reservoirs is discussed.

Ildar Batyrshin

Unsupervised Linkage Learner Based on Local Optimums

Genetic Algorithms (GAs) are categorized as search heuristics and have been broadly applied to optimization problems. These algorithms have been used for solving problems in many applications, but it has been shown that simple GA is not able to effectively solve complex real world problems. For proper solving of such problems, knowing the relationships between decision variables which is referred to as linkage learning is necessary. In this paper a linkage learning approach is proposed that utilizes the special features of the decomposable problems to solve them. The proposed approach is called Linkage Learner based on Local Optimums and Clustering (LLLC). The LLLC algorithm is capable of identifying the groups of variables which are related to each other (known as linkage groups), no matter if these groups are overlapped or different in size. The proposed algorithm, unlike other linkage learning techniques, is not done along with optimization algorithm; but it is done in a whole separated phase from optimization search. After finding linkage group information by LLLC, an optimization search can use this information to solve the problem. LLLC is tested on some benchmarked decomposable functions. The results show that the algorithm is an efficient alternative to other linkage learning techniques.

Hamid Parvin, Sajad Parvin

A Modified Back-Propagation Algorithm to Deal with Severe Two-Class Imbalance Problems on Neural Networks

In this paper we propose a modified back-propagation to deal with severe two-class imbalance problems. The method consists in automatically to find the over-sampling rate to train a neural network (NN), i.e., identify the appropriate number of minority samples to train the NN during the learning stage, so to reduce training time. The experimental results show that the performance proposed method is a very competitive when it is compared with conventional SMOTE, and its training time is lesser.

R. Alejo, P. Toribio, R. M. Valdovinos, J. H. Pacheco-Sanchez

Computing #2SAT and #2UNSAT by Binary Patterns

We present some results about the parametric complexity of #2SAT and #2UNSAT, which consist on counting the number of models and falsifying assignments, respectively, for two Conjunctive Forms (2-CF’s) . Firstly, we show some cases where given a formula

F

, #2SAT(

F

) can be bounded above by considering a binary pattern analysis over its set of clauses. Secondly, since #2SAT(

F

) = 2

n

-#2UNSAT(

F

) we show that, by considering the constrained graph

G

F

of

F

, if

G

F

represents an acyclic graph then, #UNSAT(

F

) can be computed in polynomial time. To the best of our knowledge, this is the first time where #2UNSAT is computed through its constrained graph, since the inclusion-exclusion formula has been commonly used for computing #UNSAT(

F

).

Guillermo De Ita Luna, J. Raymundo Marcial-Romero

Document Processing and Speech Recognition

A New Document Author Representation for Authorship Attribution

This paper proposes a novel representation for Authorship Attribution (AA), based on Concise Semantic Analysis (CSA), which has been successfully used in Text Categorization (TC). Our approach for AA, called Document Author Representation (DAR), builds document vectors in a space of authors, calculating the relationship between textual features and authors. In order to evaluate our approach, we compare the proposed representation with conventional approaches and previous works using the c50 corpus. We found that DAR can be very useful in AA tasks, because it provides good performance on imbalanced data, getting comparable or better accuracy results.

Adrián Pastor López-Monroy, Manuel Montes-y-Gómez, Luis Villaseñor-Pineda, Jesús Ariel Carrasco-Ochoa, José Fco. Martínez-Trinidad

A Machine-Translation Method for Normalization of SMS

Normalization of SMS is a very important task that must be addressed by the computational community because of the tremendous growth of services based on mobile devices, which make use of this kind of messages. There exist many limitations on the automatic treatment of SMS texts derived from the particular writing style used. Even if there are suficient problems dealing with this kind of texts, we are also interested in some tasks requiring to understand the meaning of documents in different languages, therefore, increasing the complexity of such tasks. Our approach proposes to normalize SMS texts employing machine translation techniques. For this purpose, we use a statistical bilingual dictionary calculated on the basis of the IBM-4 model for determining the best translation for a given SMS term. We have compared the presented approach with a traditional probabilistic method of information retrieval, observing that the normalization model proposed here highly improves the performance of the probabilistic one.

Darnes Vilariño, David Pinto, Beatriz Beltrán, Saul León, Esteban Castillo, Mireya Tovar

Reduced Universal Background Model for Speech Recognition and Identification System

Minimal Enclosing Ball (MEB) has a limitation for dealing with a large dataset in which computational load drastically increases as training data size becomes large. To handle this problem in huge dataset used for speaker recognition and identification system, we propose two algorithms using Fuzzy C-Mean clustering method. Our method uses divide-and-conquer strategy; trains each decomposed sub-problems to get support vectors and retrains with the support vectors to find a global data description of a whole target class. Our study is experimented on Universal Background Model (UBM) architectures in speech recognition and identification system to eliminate all noise features and reducing time training. For this, the training data, learned by Support Vector Machines (SVMs), is partitioned among several data sources. Computation of such SVMs can be efficiently achieved by finding a core-set for the image of the data.

Lachachi Nour-Eddine, Adla Abdelkader

GA Approaches to HMM Optimization for Automatic Speech Recognition

Hidden Markov Models (HMMs) have been widely used for Automatic Speech Recognition (ASR). Iterative algorithms such as Forward - Backward or Baum-Welch are commonly used to locally optimize HMM parameters (i.e., observation and transition probabilities). However, finding more suitable transition probabilities for the HMMs, which may be phoneme-dependent, may be achievable with other techniques. In this paper we study the application of two Genetic Algorithms (GA) to accomplish this task, obtaining statistically significant improvements on un-adapted and adapted Speaker Independent HMMs when tested with different users.

Yara Pérez Maldonado, Santiago Omar Caballero Morales, Roberto Omar Cruz Ortega

Phonetic Unification of Multiple Accents for Spanish and Arabic Languages

Languages like Spanish and Arabic are spoken over a large geographic area. The people that speak these languages develop differences in accent, annotation and phonetic delivery. This leads to difficulty in standardization of languages for education and communication (both text and oral). The problem is addressed by phonetic dictionaries to some extent. They provide the correct pronunciation for a word. But, they contribute little to standardize or unify the language for a learner. Our system is to provide unification of different accents and dialects. It creates a standard for learning and communication.

Saad Tanveer, Aslam Muhammad, A. M. Martinez-Enriquez, G. Escalada-Imaz

Environmental Sound Recognition by Measuring Significant Changes in the Spectral Entropy

Automatic identification of activities can be used to provide information to caregivers of persons with dementia for identifying assistance needs. Environmental audio provides significant and representative information of the context, making microphones a choice to identify activities automatically. However, in real situations, the audio captured by microphones comes from overlapping sound sources, making its identification a challenge for audio analysis and retrieval. In this paper we propose a succinct representation of the signal by measuring the multiband spectral entropy of the signal frame by frame, followed by a cosine transform and binary codification, we call this the Cosine Multi-Band Spectral Entropy Signature (CMBSES). To test our proposal, we created a database of a mix-up of triples from a collection of nine environmental sounds in four different signal-to-noise ratios (SNR). We codified both the original sounds and the triples and then searched all the original sounds in the mix-up collection. To establish a ground truth we also tested the same database with 48 people of assorted ages. Our feature extraction outperforms the state-of-the-art Mel Frequency Cepstral Coefficients (MFCC) and it also surpass humans in the experiment.

Jessica Beltrán-Márquez, Edgar Chávez, Jesús Favela

Backmatter

Weitere Informationen