Showing 1 to 69 of 69 matching Articles
Results per page:
Export (CSV)
By
Sossa, Humberto; Garro, Beatriz A.; Villegas, Juan; Olague, Gustavo; Avilés, Carlos
Show all (5)
1 Citations
In this paper we describe how evolutionary computation can be used to automatically design artificial neural networks (ANNs) and associative memories (AMs). In the case of ANNs, Particle Swarm Optimization (PSO), Differential Evolution (DE), and Artificial Bee Colony (ABC) algorithms are used, while Genetic Programming is adopted for AMs. The derived ANNs and AMs are tested with several examples of wellknown databases.
more …
By
Sossa, Humberto; Garro, Beatriz A.; Villegas, Juan; Avilés, Carlos; Olague, Gustavo
Show all (5)
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 wellknown databases. As we will show, results are very promising.
more …
By
Vázquez, Roberto A.; Sossa, Humberto
8 Citations
In this paper we describe how associative memories can be applied to categorize images. If we present to an associative memory (AM) an image we would expect that the AM would respond with something that describes the content of the image; for example, if the image contains a tiger we would expect that the AM would respond with the word “tiger”. In order to achieve this goal, we first chose a set of images. Each image is next associated to the word that better describes the content of the image. With this information an AM is trained as in [10]. We then use the AM to categorize instances of images with the same content even if these images are distorted by some kind of noise. The accuracy of the proposal is tested using a set of images containing different species of flowers and animals.
more …
By
Gomez, Laura E.; Sossa, Humberto; Barron, Ricardo; Jimenez, Julio F.
Show all (4)
A new method for the retrieval of melodies from a database is described in this paper. For its functioning, the method makes use of Dynamic Neural Networks (DNN). During training a set ofDNN is first trained with information of the melodies to be retrieved. Instead of using traditional signal descriptors we use the matrix of synaptic weights that can be efficiently used for melody representation and retrieval. Most of the reported works have been focused on the symbolic representation of musical information. None of them have provided good results with original signals.
more …
By
Sossa, Humberto; Barrón, Ricardo; Cuevas, Francisco; Aguilar, Carlos; Cortés, Héctor
Show all (5)
We show how the binary αβ associative memories recently proposed by Yáñez in [1] can be extended to work now in the graylevel case. To get the desired extension we take the operators α and β, foundation of the αβ memories, and propose a more general family of operators among them the original operators α and β are a subset. For this we formulate a set of functional equations, solve this system and find a family of solutions. We show that the α and β originally proposed in [1] are just a particular case of this new family. We give the properties of the new operators. We then use these operators to build the extended memories. We provide the conditions under which the proposed extended memories are able to recall a pattern either from the pattern’s fundamental set or from altered versions of them. We provide real examples with images where the proposed memories show their efficiency.
more …
By
Vázquez, Roberto A.; Sossa, Humberto
2 Citations
Morphological associative memories (MAMs) are a special type of associative memory which exhibit optimal absolute storage capacity and onestep convergence. This associative model substitutes the additions and multiplications by additions/subtractions and maximums/minimums. This type of associative model has been applied to different pattern recognition problems including face localization and reconstruction of gray scale images. Despite of his power, it has not been applied to problems involving truecolor patterns. In this paper we describe how a Morphological Heteroassociative Memory (MHAM) can be applied in problems that involve truecolor patterns. In addition, a study of the behavior of this associative model in the reconstruction of truecolor images is performed using a benchmark of 14400 images altered by different type of noises.
more …
By
Sossa, Humberto; Barrón, Ricardo; Vázquez, Roberto A.
10 Citations
In this note we describe a new set of associative memories able to recall patterns in the presence of mixed noise. Conditions are given under which the proposed memories are able to recall patterns either from the fundamental set of patterns and from distorted versions of them. Numerical and real examples are also provided to show the efficiency of the proposal.
more …
By
Sossa, Humberto; Flusser, Jan
Geometric moments have been proven to be a very efficient tool for description and recognition of binary shapes. Numerous methods for effective calculation of image moments have been presented up to now. Recently, Sossa, Yañez and Díaz [Pattern Recognition, 34(2):271276, 2001] proposed a new algorithm based on a morphologic decomposition of the image into a set of closed disks. Their algorithm yields approximative results. In this paper we propose a refinement of their method that performs as fast as the original one but gives exact results.
more …
By
Valadez, Sergio; Sossa, Humberto; SantiagoMontero, Raúl; Guevara, Elizabeth
Show all (4)
1 Citations
In this work, an encoding of polysomnographic signals into a spike firing rate, based on the BSA algorithm, is used as a discriminant feature for sleep stage classification. This proposal obtains a better sleep staging compared with the mean power signals frequency. Furthermore, a comparison of classification results obtained by different algorithms  such as Support Vector Machines, Multilayer Perceptron, Radial Basis Function Network, Naïve Bayes, KNearest Neighbors and the decision tree algorithm C4.5  is reported, demonstrating that Multilayer Perceptron has the best performance.
more …
By
Jiménez, Julio; Sossa, Humberto; Cuevas, Francisco
A particle swarm optimization (PSO) based method to carry out fringe pattern demodulation is described. A particle swarm is codified with the parameters of the function that estimates the phase. A fitness function is established to evaluate the particles, which considers: (a) the closeness between the observed fringes and the recovered fringes, (b) the phase smoothness, (c) the prior knowledge of the object as its shape and size. The swarm of particles evolves until a fitness average threshold is obtained. We demonstrate that the method is able to successfully demodulate noisy fringe patterns and even a oneimage closedfringe pattern.
more …
By
Sossa, Humberto; Barrón, Ricardo; Cuevas, Francisco; Aguilar, Carlos; Cortés, Héctor
Show all (5)
1 Citations
In this paper we show how a binary memory can be used to recall graylevel patterns. Given a set of graylevel patterns to be first memorized: 1) Decompose each pattern into a set of binary patterns, and 2) Build a binary associative memory (one matrix for each binary layer) with each training pattern set (by layers). A given pattern or a distorted version of it is recalled in three steps: 1) Decomposition of the pattern by layers into its binary patterns, 2) Recovering of each one of its binary components, layer by layer also, and 3) Reconstruction of the pattern from the binary patterns already recalled in step 2. Conditions for perfect recall of a pattern either from the fundamental set or from a distorted version of one them are also given. Experiments are also provided.
more …
By
Vázquez, Roberto A.; Sossa, Humberto; Garro, Beatriz A.
1 Citations
Visual attention is a powerful mechanism that enables perception to focus on a small subset of the information picked up by our eyes. It is directly related to the accuracy of an object categorization task. In this paper we adopt those biological hypotheses and propose an evolutionary visual attention model applied to the face recognition problem. The model is composed by three levels: the attentive level that determines where to look by means of a retinal ganglion network simulated using a network of bistable neurons and controlled by an evolutionary process; the preprocessing level that analyses and process the information from the retinal ganglion network; and the associative level that uses a neural network to associate the visual stimuli with the face of a particular person. To test the accuracy of the model a benchmark of faces is used.
more …
By
Garro, Beatriz A.; Sossa, Humberto; Vázquez, Roberto A.
8 Citations
Due to their efficiency and adaptability, bioinspired algorithms have shown their usefulness in a wide range of different nonlinear optimization problems. In this paper, we compare two ways of training an artificial neural network (ANN): Particle Swarm Optimization (PSO) and Differential Evolution (DE) algorithms. The main contribution of this paper is to show which of these two algorithms provides the best accuracy during the learning phase of an ANN. First of all, we explain how the ANN training phase could be seen as an optimization problem. Then, we explain how PSO and DE could be applied to find the best synaptic weights of the ANN. Finally, we perform a comparison between PSO and DE approaches when used to train an ANN applied to different nonlinear problems.
more …
By
Sossa, Humberto; Guevara, Elizabeth
3 Citations
In this paper a modified dendrite morphological neural network (DMNN) is applied for recognition and classification of 3D objects. For feature extraction, the first two Hu’s moment invariants are calculated based on 2D binary images, as well as the mean and the standard deviation obtained on 2D grayscale images. These four features were fed into a DMNN for classification of 3D objects. For testing, COIL20 image database and a generated dataset were used. A comparative analysis of the proposed method with MLP and SVM is presented and the results reveal the advantages of the modified DMNN. An important characteristic of the proposed recognition method is that because of the simplicity of calculation of the extracted features and the DMNN, this method can be used in real applications.
more …
By
Vázquez, Roberto A.; Sossa, Humberto; Garro, Beatriz A.
2 Citations
Recently, it was shown how some metaphors, adopted from the infant vision system, were useful for face recognition. In this paper we adopt those biological hypotheses and apply them to the 3D object recognition problem. As the infant vision responds to low frequencies of the signal, a lowfilter is used to remove high frequency components from the image. Then we detect subtle features in the image by means of a random feature selection detector. At last, a dynamic associative memory (DAM) is fed with this information for training and recognition. To test the accuracy of the proposal we use the Columbia Object Image Library (COIL 100).
more …
By
Vázquez, Roberto A.; Sossa, Humberto
1 Citations
An associative memory is a particular type of neural network for recalling output patterns from input patterns that might be altered by noise. During the last 50 years, several associative models have emerged and they only have been applied to solve problems where input patterns are images. Most of these models have several constraints that limit their applicability in complex problems. Recently in [13] it was introduced a new associative model based on some aspects of the human brain. This model is robust under different type of noises and image transformations, and useful in complex problems such as face and 3d object recognition. In this paper we adopt this model and apply it to problems that not involve images patterns, we applied to speech recognition problems. In this paper it is described a novel application where an associative memory works as a voice translator device performing a speech recognition process. In order to achieve this, the associative memory is trained using a corpus of 40 English words with their corresponding translation to Spanish. Each association used during training phase is composed by a voice signal in English and a voice signal in Spanish. Once trained our EnglishSpanish translator, when a voice signal in English is used to stimulate the associative memory we expect that the memory recalls the corresponding voice signal in Spanish. In order to test the accuracy of the proposal, a benchmark of 14500 altered versions of the original voice signals were used.
more …
By
Vázquez, Roberto A.; Sossa, Humberto; Garro, Beatriz A.
4 Citations
Hebbian heteroassociative learning is inherently asymmetric. Storing a forward association from pattern A to pattern B enables the recalling of pattern B given pattern A. This, in general, does not allow the recalling of pattern A given pattern B. The forward association between A and B will tend to be stronger than the backward association between B and A. In this paper it is described how the dynamical associative model proposed in [10] can be extended to create a bidirectional associative memory where forward association between A and B is equal to backward association between B and A. This implies that storing a forward association, from pattern A to pattern B, would enable the recalling of pattern B given pattern A and the recalling of pattern A given pattern B. We give some formal results that support the functioning of the proposal, and provide some examples were the proposal finds application.
more …
By
Osuna, Valentín; Cuevas, Erik; Sossa, Humberto
1 Citations
Acute lymphoblastic leukemia is a blood cancer that can be cured if it is detected at early stages; however, the analysis of smear blood by a human expert is tired and subject to errors. In such a sense, diagnostic of the disease is costly and time consuming. Considering that situation, several automatic segmentation methods have been proposed, some of them containing combinations of classic image analysis tools, as thresholding, morphology, color segmentation and active contours, only to mention some. In this paper is proposed the use of Hellinger distance as an alternative to Euclidean distance in order to estimate a Gaussian functions mixture that better fits a graylevel histogram of blood cell images. Two evolutionary methods (Differential Evolution and Artificial Bee Colony) are used to perform segmentation based on histogram information and an estimator of minimum distance. The mentioned techniques are compared with classic Otsu’s method by using a qualitative measure of the resulting segmentation and groundtruth images. Experimental results show that the three methods performed almost in a similar fashion, but the evolutionary ones evaluate almost 75 % less the objective function compared with Otsu’s. Also, was found that the use of a minimum distance estimator constructed with Hellinger distance and evolutionary techniques is robust and does not need a penalization factor as the needed when an Euclidean distance is used.
more …
By
Arce, Fernando; Zamora, Erik; Hernández, Gerardo; Antelis, Javier M.; Sossa, Humberto
Show all (5)
A braincomputer interface provides individuals with a way to control a computer. However, most of these interfaces remain mostly utilized in research laboratories due to the absence of certainty and accuracy in the proposed systems. In this work, we acquired our own dataset from seven ablebodied subjects and used Deep MultiLayer Perceptrons to classify motor imagery encephalography signals into binary (Rest vs Imagined and Left vs Right) and ternary classes (Rest vs Left vs Right). These Deep MultiLayer Perceptrons were fed with power spectral features computed with the Welch’s averaged modified periodogram method. The proposed architectures outperformed the accuracy achieved by the stateoftheart for classifying motor imagery bioelectrical brain signals obtaining 88.03%, 85.92% and 79.82%, respectively, and an enhancement of 11.68% on average over the commonly used Support Vector Machines.
more …
By
Vázquez, Roberto A.; Sossa, Humberto; Garro, Beatriz A.
7 Citations
In this paper we propose a viewbased method for 3D object recognition based on some biological aspects of infant vision. The biological hypotheses of this method are based on the role of the response to low frequencies at early stages, and some conjectures concerning how an infant detects subtle features (stimulating points) from an object. In order to recognize an object from different images of it (different orientations from 0° to 100°) we make use of a dynamic associative memory (DAM). As the infant vision responds to low frequencies of the signal, a lowfilter is first used to remove high frequency components from the image. Then we detect subtle features in the image by means of a random feature selection detector. At last, the DAM is fed with this information for training and recognition. To test the accuracy of the proposal we use the Columbia Object Image Library (COIL 100) database.
more …
By
Campos, Yerania; Rodner, Erik; Denzler, Joachim; Sossa, Humberto; Pajares, Gonzalo
Show all (5)
1 Citations
We provide an alternative methodology for vegetation segmentation in cornfield images. The process includes two main steps, which makes the main contribution of this approach: (a) a lowlevel segmentation and (b) a class label assignment using Bag of Words (BoW) representation in conjunction with a supervised learning framework. The experimental results show our proposal is adequate to extract green plants in images of maize fields. The accuracy for classification is 95.3 % which is comparable to values in current literature.
more …
By
Vega, Roberto; Guevara, Elizabeth; Falcon, Luis Eduardo; SanchezAnte, Gildardo; Sossa, Humberto
Show all (5)
8 Citations
Blood vessel segmentation is the first step in the process of automated diagnosis of cardiovascular diseases using retinal images. Unlike previous work described in literature, which uses rulebased methods or classical supervised learning algorithms, we applied Lattice Neural Networks with Dendritic Processing (LNNDP) to solve this problem. LNNDP differ from traditional neural networks in the computation performed by the individual neuron, showing more resemblance with biological neural networks, and offering high performance on the training phase (99.8% precision in our case). Our methodology requires four steps: 1)Preprocessing, 2)Feature computation, 3)Classification, 4)Postprocessing. We used the Hotelling T^{2} control chart to reduce the dimensionality of the feature vector from 7 to 5 dimensions, and measured the effectiveness of the methodology with the F_{1}Score metric, obtaining a maximum of 0.81; compared to 0.79 of a traditional neural network.
more …
By
Cruz, Benjamín; Sossa, Humberto; Barrón, Ricardo
1 Citations
Associative memories (AM’s) have been extensively used during the last 40 years for pattern classification and pattern restoration. In this paper Conformal Geometric Algebra (CGA) is used to develop a new associative memory (AM). The proposed AM makes use of CGA and quadratic programming to store associations among patterns and their respective classes. An unknown pattern is classified by applying an inner product between the pattern and the build AM. Numerical and real examples are presented to show the potential of the proposal.
more …
By
Barrón, Ricardo; Sossa, Humberto; Cruz, Benjamín
In this work we present an algorithm for training an associative memory based on the socalled multilayered morphological perceptron with maximal support neighborhoods. We compare the proposal with the original one by performing some experiments with real images. We show the superiority of the new one. We also give formal conditions for correct classification. We show that the proposal can be applied to the case of graylevel images and not only binary images.
more …
By
OcampoVega, Ricardo; SanchezAnte, Gildardo; FalconMorales, Luis E.; Sossa, Humberto
Show all (4)
RadialBasis Function Neural Networks (RBFN) are a well known formulation to solve classification problems. In this approach, a feedforward neural network is built, with one input layer, one hidden layer and one output layer. The processing is performed in the hidden and output layers. To adjust the network for any given problem, certain parameters have to be set. The parameters are: the centers of the radial functions associated to the hidden layer and the weights of the connections to the output layer. Most of the methods either require a lot of experimentation or may demand a lot of computational time. In this paper we present a novel method based on a partition algorithm to automatically compute the amount and location of the centers of the radialbasis functions. Our results, obtained by running it in seven public databases, are comparable and even better than some other approaches.
more …
By
Sossa, Humberto; Barrón, Ricardo; Vázquez, Roberto A.
14 Citations
Most results (lemmas and theorems) providing conditions under which associative memories are able to perfectly recall patterns of a fundamental set are very restrictive in most practical applications. In this note we describe a simple but effective procedure to transform a fundamental set of patterns (FSP) to a canonical form that fulfils the propositions. This way pattern recall is strongly improved. We provide numerical and real examples to reinforce the proposal.
more …
By
Sossa, Humberto; Guevara, Elizabeth
1 Citations
In this paper a modified dendrite morphological neural network (DMNN) is applied for 3D object recognition. For feature extraction, shape and color information were used. The first two Hu’s moment invariants are calculated based on 2D grayscale images, and color attributes were obtained converting the RGB (Red, Green, Blue) image to the HSI (Hue, Saturation, Intensity) color space. For testing, a controlled lab color image database and a real image dataset were considered. The problem with the real image dataset, without controlling light conditions, is that objects are difficult to segment using only color information; for tackling this problem the Depth data provided by the Microsoft Kinect for Windows sensor was used. A comparative analysis of the proposed method with a MLP (Multilayer Perceptron) and SVM (Support Vector Machine) is presented and the results reveal the advantages of the modified DMNN.
more …
By
Ojeda, Leonardo; Vega, Roberto; Falcon, Luis Eduardo; SanchezAnte, Gildardo; Sossa, Humberto; Antelis, Javier M.
Show all (6)
4 Citations
EEGbased BCIs rely on classification methods to recognize the brain patterns that encode user’s intention. However, decoding accuracies have reached a plateau and therefore novel classification techniques should be evaluated. This paper proposes the use of Lattice Neural Networks with Dendritic Processing (LNND) for the classification of hand movements from electroencephalographic (EEG) signals. The performance of this technique was evaluated and compared with classical classifiers using EEG signals recorded form participants performing motor tasks. The result showed that LNND provides: (i) the higher decoding accuracies in experiments using one electrode (
$$DA=80\,\%$$
and
$$DA=80\,\%$$
for classification of motor execution and motor imagery, respectively); (ii) distributions of decoding accuracies significantly different and higher than the chance level (
$$p<0.05$$
, Wilcoxon signedrank test) in experiments using one, two, four and six electrodes. These results shows that LNND could be a powerful technique for the recognition of motor tasks in BCIs.
more …
By
Sossa, Humberto; Barrón, Ricardo; Vázquez, Roberto A.
5 Citations
In this paper we study how the performance of a median associative memory is influenced when the values of its elements are altered by noise. To our knowledge this kind of research has not been reported until know. We give formal conditions under which the memory is still able to correctly recall a pattern of the fundamental set of patterns either from a nonaltered or a noisy version of it. Experiments are also given to show the efficiency of the proposal.
more …
By
Vázquez, Roberto Antonio; Sossa, Humberto; Barrón, Ricardo
1 Citations
Object recognition under occlusions is an important problem in computer vision, not yet completely solved. In this note we describe a simple but effective technique for the recognition objects under occlusions. The proposal uses the most distinctive parts of the objects for their further detection. During training, the proposal, first detects the distinctive parts of each object. For each of these parts an invariant description in terms of invariants features is next computed. With these invariant descriptions a specially designed set of associative memories (AMs) is trained. During object detection, the proposal, first looks for the important parts of the objects by means of the already trained AM. The proposal is tested with a bank of images of real objects and compared with other similar reported techniques.
more …
By
Sossa, Humberto; Cortés, Griselda; Guevara, Elizabeth
1 Citations
This paper presents the initial results concerning a new Radial Basis Function Artificial Neural Network (RBFNN) architecture for pattern classification. Performance of the new architecture is demonstrated with different data sets. Its efficiency is also compared with different classification methods reported in literature: Multilayer Perceptron, Standard Radial Basis Neural Networks, KNN and Minimum Distance classifiers, showing a much better performance. Results are only given for problems using two features
more …
By
Cruz, Benjamin; Barron, Ricardo; Sossa, Humberto
Associative memories (AMs) were proposed as tools usually used in the restoration and classification of distorted patterns. Many interesting models have emerged in the last years with this aim. In this chapter a novel associative memory model (Geometric Associative Memory, GAM) based on Conformal Geometric Algebra (CGA) principles is described. At a low level, CGA provides a new coordinatefree framework for numeric processing in problem solving. The proposed model makes use of CGA and quadratic programming to store associations among patterns and their respective class. To classify an unknown pattern, an inner product is applied between it and the obtained GAM. Numerical and real examples to test the proposal are given. Formal conditions are also provided that assure the correct functioning of the proposal.
more …
By
Vázquez, Roberto A.; Sossa, Humberto
Median associative memories (MEDAMs) are a special type of associative memory based on the median operator. This type of associative model has been applied to the restoration of gray scale images and provides better performance than other models, such as morphological associative memories, when the patterns are altered with mixed noise. Despite of his power, MEDAMs have not been applied in problems involving truecolor patterns. In this paper we describe how a median heteroassociative memory (MEDHAM) could be applied in problems that involve truecolor patterns. A complete study of the behavior of this associative model in the restoration of truecolor images is performed using a benchmark of 14400 images altered by different type of noises. Furthermore, we describe how this model can be applied to an image categorization problem.
more …
By
VillegasCortez, Juan; Olague, Gustavo; Aviles, Carlos; Sossa, Humberto; Ferreyra, Andres
Show all (5)
1 Citations
Associative Memories (AMs) are mathematical structures specially designed to associate input patterns with output patterns within a single stage. Since the last fifty years all reported AMs have been manually designed. The paper describes a Genetic Programming based methodology able to create a process for the automatic synthesis of AMs. It paves a new area of research that permits for the first time to propose new AMs for solving specific problems. In order to test our methodology we study the application of AMs for real value patterns. The results illustrate that it is possible to automatically generate AMs that achieve good recall performance for problems commonly used in pattern recognition research.
more …
By
Ocampo, Ricardo; Luna, Marco A.; Vega, Roberto; SanchezAnte, Gildardo; FalconMorales, Luis E.; Sossa, Humberto
Show all (6)
1 Citations
DNA microarrays is a technology that can be used to diagnose cancer and other diseases. To automate the analysis of such data, pattern recognition and machine learning algorithms can be applied. However, the curse of dimensionality is unavoidable: very few samples to train, and many attributes in each sample. As the predictive accuracy of supervised classifiers decays with irrelevant and redundant features, the necessity of a dimensionality reduction process is essential. In this paper, we propose a new methodology that is based on the application of Principal Component Analysis and other statistical tools to gain insight in the identification of relevant genes. We run the approaches using two benchmark datasets: Leukemia and Lymphoma. The results show that it is possible to reduce considerably the number of genes while increasing the performance of well known classifiers.
more …
By
Vazquez, Roberto A.; Sossa, Humberto; Garro, Beatriz A.
1 Citations
Several associative memories (AM) have been proposed in the last years. These AMs have several constraints that limit their applicability in complex problems such as face recognition. Despite of the power of these models, they cannot reach its full power without applying new mechanisms based on current and future studies on biological neural networks. In this research we show how a network of dynamic associative memories (DAM) combined with some aspects of the infant vision system could be efficiently applied to the face recognition problem. Through several experiments by using a benchmark of faces the accuracy of the proposal is tested.
more …
By
Garibay, Carlos; SanchezAnte, Gildardo; FalconMorales, Luis E.; Sossa, Humberto
Show all (4)
DNA microarrays are being used to characterize the genetic expression of several illnesses, such as cancer. There has been interest in developing automated methods to classify the data generated by those microarrays. The problem is complex due to the availability of just a few samples to train the classifiers, and the fact that each sample may contain several thousands of features. One possibility is to select a reduced set of features (genes). In this work we propose a wrapper method that is a modified version of the Inertial Geometric Particle Swarm Optimization.We name it MIGPSO. We compare MIGPSO with other approaches. The results are promising. MIGPSO obtained an increase in accuracy of about 4 %. The number of genes selected is also competitive.
more …
By
Cuevas, Erik; Sossa, Humberto; Osuna, Valentín; Zaldivar, Daniel; PérezCisneros, Marco
Show all (5)
Automatic circle detection in digital images has received considerable attention over the last years. Recently, several robust circle detectors, based on evolutionary algorithms (EA), have been proposed. They have demonstrated to provide better results than those based on the Hough Transform. However, since EAdetectors usually need a large number of computationally expensive fitness evaluations before a satisfying result can be obtained; their use for real time has been questioned. In this work, a new algorithm based on the Harmony Search Optimization (HSO) is proposed to reduce the number of function evaluation in the circle detection process. In order to avoid the computation of the fitness value of several circle candidates, the algorithm estimates their values by considering the fitness values from previously calculated neighboring positions. As a result, the approach can substantially reduce the number of function evaluations preserving the good search capabilities of HSO. Experimental results from several tests on synthetic and natural images with a varying complexity range have been included to validate the efficiency of the proposed technique regarding accuracy, speed and robustness.
more …
By
Sossa, Humberto; Barrón, Ricardo
1 Citations
Median associative memories (MEDMEMs) first described in [1] have proven to be efficient tools for the reconstruction of patterns corrupted with mixed noise. First formal conditions under which these tools are able to reconstruct patterns either from the fundamental set of patterns and from distorted versions of them were given in [1]. In this paper, new more accurate conditions are provided that assure perfect reconstruction. Numerical and real examples are also given.
more …
By
Sossa, Humberto; Barrón, Ricardo; Oropeza, José L.
In this note, we introduce an associative memory useful to recall realvalued patterns altered with mixed noise (additive and subtractive). Numerical and real examples are given to show the effectiveness of the proposal. Conditions under which the proposed memories are able to recall patterns either from the fundamental set of patterns and from distorted versions of them are also given.
more …
By
GudiñoMendoza, Berenice; Sossa, Humberto; SanchezAnte, Gildardo; Antelis, Javier M.
Show all (4)
2 Citations
The identification of each phase in the process of movement arms from brain waves has been studied using classical classification approaches. Identify precisely each movement phase from relaxation to movement execution itself, is still an open challenging task. In the context of BrainComputer Interfaces (BCI) this identification could accurately activate devices, giving more natural control systems. This work presents the use of a novel classification technique Lattice Neural Networks with Dendritic Processing (LNNDP), to identify motor states using electroencephalographic signals recorded from healthy subjects, performing selfpaced reaching movements. To evaluate the performance of this technique 3 biclassification scenarios were followed: (i) relax vs. intention, (ii) relax vs. execution, and (iii) intention vs. execution. The results showed that LNNDP provided an accuracy of (i) 65.26%, (ii) 69.07%, and (iii) 76.71% in each scenario respectively, which were higher than the chance level.
more …
By
Rosa, Erick; Yu, Wen; Sossa, Humberto
Deep learning techniques have been successfully used for pattern classification. These advantage methods are still not applied in fuzzy modeling. In this paper, a novel datadriven fuzzy modeling approach is proposed. The deep learning methods is applied to learn the probability properties of input and output pairs. We propose special unsupervised learning methods for these two deep learning models with input data. The fuzzy rules are extracted from these properties. These deep learning based fuzzy modeling algorithms are validated with three benchmark examples.
more …
By
Peña, Alejandro; Sossa, Humberto; Gutiérrez, Agustin
1 Citations
A powerful and useful approach for modeling knowledge and qualitative reasoning is the Cognitive Map. The background of Cognitive Maps is the research about learning environments carried out by Cognitive Psychology since the nineteenth century. Along the last thirty years, these underlying findings inspired the development of computational models to deal with causal phenomena. So, a Cognitive Map is a structure of concepts of a specific domain that are related through causeeffect relations with the aim to simulate behavior of dynamic systems. In spite of the short life of the causal Cognitive Maps, nowadays there are several branches of development that focus on qualitative, fuzzy and uncertain issues. With this platform wide spectra of applications have been developing in fields like game theory, information analysis and management sciences. Wherefore, with the purpose to promote the use of this kind of tool, in this work is surveyed three branches of Cognitive Maps; and it is outlined one application of the Cognitive Maps for the student modeling that shows a conceptual design of a project in progress.
more …
By
Garro, Beatriz A.; Sossa, Humberto; Vázquez, Roberto A.
10 Citations
The design of an Artificial Neural Network (ANN) is a difficult task for it depends on the human experience. Moreover it needs a process of testing and error to select which kind of a transfer function and which algorithm should be used to adjusting the synaptic weights in order to solve a specific problem. In the last years, bioinspired algorithms have shown their power in different nonlinear optimization problems. Due to their efficiency and adaptability, in this paper we explore a new methodology to automatically design an ANN based on the Differential Evolution (DE) algorithm. The proposed method is capable to find the topology, the synaptic weights and the transfer functions to solve a given pattern classification problems.
more …
By
Sossa, Humberto; Flusser, Jan
Abstract
Geometric moments have been proven to be a very efficient tool for description and recognition of binary shapes. Numerous methods for effective calculation of image moments have been presented up to now. Recently, Sossa, Yañez and Díaz [Pattern Recognition, 34(2):271276, 2001] proposed a new algorithm based on a morphologic decomposition of the image into a set of closed disks. Their algorithm yields approximative results. In this paper we propose a refinement of their method that performs as fast as the original one but gives exact results.
more …
By
Vazquez, Roberto A.; Sossa, Humberto
2 Citations
An associative memory (AM) is a special kind of neural network that allows associating an output pattern with an input pattern. In the last years, several associative models have been proposed by different authors. However, they have several constraints which limit their applicability in complex pattern recognition problems. In this paper we gather different results provided by a dynamic associative model and present new results in order to describe how this model can be applied to solve different complex problems in pattern recognition such as object recognition, image restoration, occluded object recognition and voice recognition.
more …
By
Virgilio G, Carlos D.; Sossa, Humberto; Antelis, Javier M.; Falcón, Luis E.
Show all (4)
We report the development and evaluation of brain signal classifiers, specifically Spiking Neuron based classifiers. The proposal consists of two main stages: feature extraction and pattern classification. The EEG signals used represent four motor imagery tasks: Left Hand, Right Hand, Foot and Tongue movements. In addition, one more class was added: Rest. These EEG signals were obtained from a database provided by the Technological University of Graz. Feature extraction stage was carried out by applying two algorithms: Power Spectral Density and Wavelet Decomposition. The tested algorithms were: KNearest Neighbors, Multilayer Perceptron, Single Spiking Neuron and Spiking Neural Network. All of them were evaluated in the classification between two Motor Imagery tasks; all possible pairings were made with the 5 mental tasks (Rest, Left Hand, Right Hand, Tongue and Foot). In the end, a performance comparison was made between a Multilayer Perceptron and Spiking Neural Network.
more …
By
Hernández, Gerardo; Zamora, Erik; Sossa, Humberto
2 Citations
In this paper, a comparative study between two different neural network models is performed for a very simple type of classificaction problem in 2D. The first model is a deep neural network and the second is a dendrite morphological neuron. The metrics to be compared are: training time, classification accuracies and number of learning parameters. We also compare the decision boundaries generated by both models. The experiments show that the dendrite morphological neurons surpass the deep neural networks by a wide margin in terms of higher accuracies and a lesser number of parameters. From this, we raise the hypothesis that deep learning networks can be improved adding morphological neurons.
more …
By
Sossa, Humberto; Sánchez, Hermilo
We present two formulations and two procedures that can be used for computing the number of bubbles and tunnels of a 3D binary object. The first formulation is useful to determine the number of bubbles of an object, while the second one can be used to calculate the number of tunnels of an object. Both formulations are formally demonstrated. Examples are provided to numerically validate the functioning of both formulations. On the other hand, the first procedure allows obtaining the number of bubbles and tunnels of a 3D object while the second procedure allows computing the number of bubbles and tunnels of several 3D objects. Examples with 3D images are provided to illustrate the utility and validity of the second procedure.
more …
By
Furlán, Federico; Rubio, Elsa; Sossa, Humberto; Ponce, Víctor
Show all (4)
In this paper we study the problem of rock detection in a Marslike environment. We propose a convolutional neural network (CNN) to obtain a segmented image. The CNN is a modified version of the Unet architecture with a smaller number of parameters to improve the inference time. The performance of the methodology is proved in a dataset that contains several images of a Marslike environment, achieving an Fscore of 78.5%.
more …
By
Sossa, Humberto; Barrón, Ricardo; Vázquez, Roberto A.
Abstract
In this note we describe a new set of associative memories able to recall patterns in the presence of mixed noise. Conditions are given under which the proposed memories are able to recall patterns either from the fundamental set of patterns and from distorted versions of them. Numerical and real examples are also provided to show the efficiency of the proposal.
more …
By
Arce, Fernando; MendozaMontoya, Omar; Zamora, Erik; Antelis, Javier M.; Sossa, Humberto; CantilloNegrete, Jessica; CarinoEscobar, Ruben I.; Hernández, Luis G.; Falcón, Luis Eduardo
Show all (9)
Dendrite ellipsoidal neurons are a novel and different alternative for classification tasks, giving competitive results compared with typical classification methods. Based on kmeans++ algorithm, the network allows each dendrite to build a hyperellipsoidal in order to assign each incoming pattern
$$x_{i}=(x_{1},x_{2},\ldots ,x_{n})^{T}$$
to its respective C class. The main disadvantage of this training algorithm is the lack of accuracy in high dimensional datasets. In this research, we solved this problem by training the dendrite ellipsoidal neuron using stochastic gradient descent. Furthermore, electroencephalography data were acquired during two mental conditions (imaginary movements of the left and right hand) in order to test the new training algorithm. The proposed algorithm outperformed the accuracy acquired by a dendrite ellipsoidal neuron based on kmeans++ obtaining 76.02% and 62.77%, respectively. Also, the algorithm was compared with multilayer perceptrons and support vector machines which are some of the most common classifiers used to detect motorrelated information in brain signals. These achieved an accuracy of 72.38% and 65.81%, respectively.
more …
By
Cruz, Benjamín; Barrón, Ricardo; Sossa, Humberto
2 Citations
Associative memories (AMs) have been extensively used during the last 40 years for pattern classification and pattern restoration. A new type of AMs have been developed recently, the socalled Geometric Associative Memories (GAMs), these make use of Conformal Geometric Algebra (CGA) operators and operations for their working. GAM’s, at the beginning, were developed for supervised classification, getting good results. In this work an algorithm for unsupervised learning with GAMs will be introduced. This new idea is a variation of the kmeans algorithm that takes into account the patterns of the a specific cluster and the patterns of another clusters to generate a separation surface. Numerical examples are presented to show the functioning of the new algorithm.
more …
By
Sossa, Humberto; Carreón, Ángel; Santiago, Raúl
1 Citations
In this short communication, we explain how a Multilayered Perceptron (MLP) can be used to compute the Euler number or Genus of a 2D binary image. We take as basis the results provided by a mathematical formulation that is known providing exact results in the computation of this important topological image feature to derive two MLPbased architectures, one useful for the 4connected case and one useful for 8connected case. We present results with a set of realistic images and compare our proposals in terms of processing with other approaches reported in literature.
more …
By
Sossa, Humberto; Barrón, Ricardo; Cuevas, Francisco; Aguilar, Carlos; Cortés, Héctor
Show all (5)
Abstract
We show how the binary αβ associative memories recently proposed by Yáñez in [1] can be extended to work now in the graylevel case. To get the desired extension we take the operators α and β, foundation of the αβ memories, and propose a more general family of operators among them the original operators α and β are a subset. For this we formulate a set of functional equations, solve this system and find a family of solutions. We show that the α and β originally proposed in [1] are just a particular case of this new family. We give the properties of the new operators. We then use these operators to build the extended memories. We provide the conditions under which the proposed extended memories are able to recall a pattern either from the pattern’s fundamental set or from altered versions of them. We provide real examples with images where the proposed memories show their efficiency.
more …
By
ValadezGodínez, Sergio; González, Javier; Sossa, Humberto
In previous works, a successful scheme using a single Spiking Neuron (SN) to solve complex problems in pattern recognition has been proposed. This consists in using the firing frequency response to classify a given input pattern, which is multiplied by a weight vector to produce a constant stimulation current. The weight vector is adjusted by an evolutionary strategy where the objective is to obtain an optimal frequency separation. The problem is that the SN has to be numerically simulated several times when the weight vector is being adjusted. In this work, we propose fitting the SN frequency response curve to a piecewise linear function to be used instead of the costly SN simulation. A high fitting degree was found, but, more importantly, the computational cost of the training and testing phases was drastically reduced.
more …
By
Barrón, Ricardo; Sossa, Humberto; Cortés, Héctor
2 Citations
Morphological neural networks consider that the information entering a neuron is affected additively by a conductivity factor called synaptic weight. They also suppose that the input channels account with a saturation level mathematically modeled by a MAX or MIN operator. This, from a physiological point of view, appears closer to reality than the classical neural model, where the synaptic weight interacts with the input signal by means of a product; the input channel forms an average of the input signals. In this work we introduce some geometrical aspects of dendrite processing that easily allow visualizing the classification regions, providing also an intuitive perspective of the production and training of the net.
more …
By
Vázquez, Roberto A.; Sossa, Humberto
2 Citations
An associative memory AM is a type of neural network commonly used for recalling output patterns from input patterns that might be altered by noise. Most of these models have several constraints that limit their applicability in complex problems. Recently, in [13] a new AM based on some aspects of human brain was introduced, however the authors only test its accuracy using image patterns. In this paper we show that this model is also robust with other type of patterns such as voice signal patterns. The AM is trained with associations composed by voice signals and their corresponding images. Once trained, when a voice signal is used to stimulate the AM we expect the memory recall the image associated to the voice signal. In order to test the accuracy of the proposal, a benchmark of sounds and images was used.
more …
By
Peña, Alejandro; Sossa, Humberto; Gutierrez, Francisco
4 Citations
This work proposes a framework for the design and development of Ontology Agents oriented to manage Rule Base Fuzzy Cognitive Maps (RBFCM). The approach takes into account the foundations of the Ontology Agents and the baseline of the Fuzzy Cognitive Maps depicted by Rule Bases. With these underlying elements, a specification of a conceptualization about the modeled domain is outcome. Moreover, a knowledge structure, composed by concepts and causal relationships that fit a Fuzzy Rule Base, is grown from. As a result, a semantic repository is stated by means of the Ontology Web Language (OWL). The management of the ontology is fulfilled by an Ontology Agent. This kind of agent takes over the services required to define and update the Ontology items. Also, the Ontology Agent achieves the tasks for answering the queries sent by a community of agents. This set of agents recreates a MultiAgent System (MAS) that is deployed on the Internet by means of Web Services, where the system carries out causal inferences based on RBFCM.
more …
By
Cruz, Benjamín; Sossa, Humberto; Barrón, Ricardo
8 Citations
Pattern reconstruction or pattern restoration in the presence of noise is a main problem in pattern recognition. An essential feature of the noise acting on a pattern is its local nature. If a pattern is split into enough subpatterns, a few of them will be less or more affected by noise, others will remain intact. In this paper, we propose a simple but effective methodology that exploits this fact for the efficient restoration of a pattern. A pattern is restored if enough of its subpatterns are also restored. Since several patterns can share the same subpatterns, the final decision is accomplished by means of a voting mechanism. Before deciding if a subpattern belongs to a pattern, subpattern restoration in the presence of noise is done by an associative memory. Numerical and real examples are given to show the effectiveness of the proposal. Formal conditions under which the proposal guaranties perfect restoration of a pattern from an unaltered or and altered version of it are also given.
more …
By
Cruz, Benjamín; Barrón, Ricardo; Sossa, Humberto
1 Citations
Conformal Geometric Algebra (CGA) is a high level language commonly used in mathematical, physics and engineering problems. At a top level, CGA is a free coordinate tool for designing and modeling geometric problems; at a low level CGA provides a new coordinate framework for numeric processing in problem solving. In this paper we show how to use quadratic programming and CGA for, given two sets p and q of points in ℝ^{n}, construct an optimal separation sphere S such that, all points of p are contained inside of it, and all points of q are outside. To classify an unknown pattern x, an inner product must be applied between x and S. Some numerical and real examples to test the proposal are given.
more …
By
Vázquez, Roberto A.; Sossa, Humberto; Garro, Beatriz A.
9 Citations
A novel method for face recognition based on some biological aspects of infant vision is proposed in this paper. The biological hypotheses of this method are based on the role of the response to low frequencies at early stages, and some conjectures concerning how an infant detects subtle features (stimulating points) from a face. In order to recognize a face from different images of it we make use of a bank of dynamic associative memories (DAM). As the infant vision responds to low frequencies of the signal, a lowfilter is first used to remove high frequency components from the image. We then detect subtle features in the image by means of a random feature selection detector. At last, the network of DAMs is fed with this information for training and recognition. To test the accuracy of the proposal a benchmark of faces is used.
more …
By
García, Joel; Zamora, Erik; Sossa, Humberto
Households are responsible for more than 40% of the global electricity consumption [7]. The analysis of this consumption to find unexpected behaviours could have a great impact on saving electricity. This research presents an experimental study of supervised and unsupervised neural networks for anomaly detection in electrical consumption. Multilayer perceptrons and autoencoders are used for each approach, respectively. In order to select the most suitable neural model in each case, there is a comparison of various architectures. The proposed methods are evaluated using realworld data from an individual home electric power usage dataset. The performance is compared with a traditional statistical procedure. Experiments show that the supervised approach has a significant improvement in anomaly detection rate. We evaluate different possible feature sets. The results demonstrate that temporal data and measures of consumption patterns such as mean, standard deviation and percentiles are necessary to achieve higher accuracy.
more …
By
TapiaDueñas, Osvaldo A.; SánchezCruz, Hermilo; López, Hiram H.; Sossa, Humberto
Show all (4)
There are many applications in different fields, as diverse as computer graphics, medical imaging or pattern recognition for industries, where the use of three dimensional objects is needed. By the nature of these objects, it is very important to develop thrifty methods to represent, study and store them. In this paper, a new method to encode surfaces of threedimensional objects that are not isomorphic to the plane is developed. In the proposed method, a helical path that covers the contour is obtained and then, the Freeman F26 chain code is used to encode the helical path. In order to solve geometric problems to find optimal paths between adjacent slices, a modification of the A star algorithm was carried out. Finally, our proposed method is applied to threedimensional objects obtained from real data.
more …
By
OchoaMontiel, R.; FloresCastillo, O.; Sossa, Humberto; Olague, Gustavo
Show all (4)
Image enhancement techniques are needed to decrease the negative effects of blur or unwanted noise in image processing. In biomedical images, the quality of images is very important to achieve an adequate identification to detection or diagnosis purposes. This paper addresses the use of contrast enhancement to facilitate the identification of leukemia in blood cell images. Differential evolution algorithm is used to get parameters required to apply contrast enhancement specifically in the interest region in the image, which facilites the posterior identification of leukemic cells. Identification of leukemic cells is accomplished applying an edges extraction and dilatation. From this image, two types of neural networks are used to classify the cells like healthy or leukemic cells. In first experiment, a multilayer perceptron is trained with the backpropagation algorithm using geometric features extracted from image. While in the second, convolutional networks are used. A public dataset of 260 healthy and leukemic cell images, 130 for each type, is used. The proposed contrast enhancement technique shows satisfactory results when obtaining the interest region, facilitating the identification of leukemic cells without additional processing, like image segmentation.
This way, computational resources are decreased. On the other hand, to identify the cell type, images are classified using neural networks achieving an average classification accuracy of
$$99.83\%$$
.
more …
By
Hernández, Gerardo; Hernández, Luis G.; Zamora, Erik; Sossa, Humberto; Antelis, Javier M.; MendozaMontoya, Omar; Falcón, Luis E.
Show all (7)
This study aims to compare classical and MorphologicalLinear Neural Network (MLNN) algorithms for the intention recognition to perform different movements from electroencephalographic (EEG) signals. Three classification models were implemented and assessed to decode EEG motor imagery signals: (i) MorphologicalLinear Neural Network (MLNN) (ii) Support Vector Machine (SVM) and (iii) Multilayer perceptron (MLP). Real EEG signals recorded during robotassisted rehabilitation therapy were used to evaluate the performance of the proposed algorithms in the classification of three classes (relax, movement intention A Int A and movement intention B Int B) using multiCSP based features extracted from EEG signals. The results of a tenfold cross validation show similar results in terms of classification accuracy for the SVM and MLNN models. However, the number of parameters used in each model varies considerably (the MLNN model use less parameters than the SVM). This study indicates potential application of MLNNs for decoding movement intentions and its use to develop more natural and intuitive robot assisted neurorehabilitation therapies.
more …
By
Sossa, Humberto; Barrón, Ricardo; Cuevas, Francisco; Aguilar, Carlos
Show all (4)
5 Citations
In this note we show how a binary memory can be used to recall graylevel patterns. We take as example the α β associative memories recently proposed in Yáñez, Associative Memories based on order Relations and Binary Operators(In Spanish), PhD Thesis, Center for computing Research, February of 2002, only useful in the binary case. Basically, the idea consists on that given a set of graylevel patterns to be first memorized: (1) Decompose them into their corresponding binary patterns, and (2) Build the corresponding binary associative memory (one memory for each binary layer) with each training pattern set (by layers). A given pattern or a distorted version of it, it is recalled in three steps: (1) Decomposition of the pattern by layers into its binary patterns, (2) Recalling of each one of its binary components, layer by layer also, and (3) Reconstruction of the pattern from the binary patterns already recalled in step 2. The proposed methodology operates at two phases: training and recalling. Conditions for perfect recall of a pattern either from the fundamental set or from a distorted version of one them are also given. Experiments where the efficiency of the proposal is tested are also given.
more …
By
Cuevas, Erik; Sención, Felipe; Zaldivar, Daniel; PérezCisneros, Marco; Sossa, Humberto
Show all (5)
57 Citations
This paper explores the use of the Artificial Bee Colony (ABC) algorithm to compute threshold selection for image segmentation. ABC is an evolutionary algorithm inspired by the intelligent behavior of honeybees which has been successfully employed to solve complex optimization problems. In this approach, an image 1D histogram is approximated through a Gaussian mixture model whose parameters are calculated by the ABC algorithm. In the model, each Gaussian function represents a pixel class and therefore a threshold point. Unlike the ExpectationMaximization (EM) algorithm, the ABC method shows fast convergence and low sensitivity to initial conditions. Remarkably, it also improves complex timeconsuming computations commonly required by gradientbased methods. Experimental results over multiple images with different range of complexity validate the efficiency of the proposed technique with regard to segmentation accuracy, speed, and robustness. The paper also includes an experimental comparison to the EM and to one gradientbased method which ultimately demonstrates a better performance from the proposed algorithm.
more …
By
Campos, Yerania; Sossa, Humberto; Pajares, Gonzalo
2 Citations
Precision Agriculture aims to apply selective treatments and tasks at localized areas concerning crop fields. Robotized and autonomous tractors, equipped with perception, decisionmaking and actuation systems, can apply specific treatments as may be required. Correct plant identification through the perception system, including crops and weeds, is an important issue. Additionally, it is well known that, in autonomous vehicles, safety is a major challenge, where unexpected obstacles in the working area must be conveniently addressed in order to guarantee the security and the continuity of the process. The objective of this study was to design a triclass Support Vector Machine classifier for identifying plants (crops and weeds), soil and objects in maize fields based on unsupervised learning. For this, a strategy for automatic sample selection was designed to obtain elements of the three involved classes for the training process. In this context, the identification of obstacles for safe navigation makes an important contribution. A comparative analysis of different texture descriptors and local patterns was carried out with the aim of determining the best for characterizing the classes under study; results have shown that the SpeededUp Robust Features descriptor is the most appropriate to discriminate between plants, soil and objects. The development of an object detection algorithm for agricultural images proved the effectiveness of the triclass classifier with an accuracy of 94.3%.
more …
