Pattern Recognition

Feature selection and attribute reduction have been tackled in the Rough Set Theory through fuzzy reducts. Recently, Goldman fuzzy reducts which are fuzzy subsets of attributes were introduced. In this paper, we introduce an algorithm for computing all Goldman fuzzy reducts of a decision system, this algorithm is the first one reported for this purpose. The experiments over standard and synthetic data sets show that the proposed algorithm is useful for datasets with up to twenty attributes.

Polygonal array graphs have been widely investigated, and they represent a relevant area of interest in mathematical chemistry because they have been used to study intrinsic properties of molecular graphs. For example, to determine the Merrifield-Simmons index of a polygonal array \(A_n\) that is the number of independent sets of that graph, denoted as \(i(A_n)\).

In this paper we consider the problem of extending an initial polygonal array \(A_n\) adding a new polygon p to form \(A_{n+1}\), for minimizing or maximizing the Merrifield-Simmons index \(i(A_{n+1}) = i(A_n \cup p)\). Our method does not require to compute \(i(A_n)\) or \(i(A_n \cup p)\), explicitly.

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.

Selecting contrast patterns is an important task for pattern-based classifiers, especially in class imbalance problems. The main reason is that the contrast pattern miners commonly extract several patterns with high support for the majority class and only a few patterns, with low support, for the minority class. This produces a bias of classification results toward the majority class, obtaining a low accuracy for the minority class. In this paper, we introduce a contrast pattern selection method for class imbalance problems. Our proposal selects all the contrast patterns for the minority class and a certain percent of contrast patterns for the majority class. Our experiments performed over several imbalanced databases show that our proposal selects significantly better contrast patterns, obtaining better AUC results, than other approaches reported in the literature.

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.

Similarity searching consists in retrieving from a database the objects, also known as nearest neighbors, that are most similar to a given query, it is a crucial task to several applications of the pattern recognition problem. In this paper we propose a new technique to reduce the number of comparisons needed to locate the nearest neighbors of a query. This new index takes advantage of two known algorithms: FHQT (Fixed Height Queries Tree) and PBA (Permutation-Based Algorithm), one for low dimension and the second for high dimension. Our results show that this combination brings out the best of both algorithms, this winner combination of FHQT and PBA locates nearest neighbors up to four times faster in high dimensions leaving the known well performance of FHQT in low dimensions unaffected.

In this paper we propose a feasible method based on projections using a curvilinear search for solving optimization problems with orthogonality constraints. Our algorithm computes the SVD decomposition in each iteration in order to preserve feasibility. Additionally, we present some convergence results. Finally, we perform numerical experiments with simulated problems; and analyze the performance of the proposed methods compared with state-of-the-art algorithms.

Support vector machines (SVMs) have been found highly helpful in solving numerous pattern recognition tasks. Although it is challenging to train SVMs from large data sets, this obstacle may be mitigated by selecting a small, yet representative, subset of the entire training set. Another crucial and deeply-investigated problem consists in selecting the SVM model. There have been a plethora of methods proposed to effectively deal with these two problems treated independently, however to the best of our knowledge, it was not explored how to effectively combine these two processes. It is a noteworthy observation that depending on the subset selected for training, a different SVM model may be optimal, hence performing these two operations simultaneously is potentially beneficial. In this paper, we propose a new method to select both the training set and the SVM model, using a genetic algorithm which alternately optimizes two different populations. We demonstrate that our approach is competitive with sequential optimization of the hyperparameters followed by selecting the training set. We report the results obtained for several benchmark data sets and we visualize the results elaborated for artificial sets of 2D points.

Reinforcement Learning (RL) algorithms create a mapping from states to actions, in order to maximize an expected reward and derive an optimal policy. However, traditional learning algorithms rarely consider that learning has an associated cost and that the available resources to learn may be limited. Therefore, we can think of learning over a limited budget. If we are developing a learning algorithm for an agent i.e. a robot, we should consider that it may have a limited amount of battery; if we do the same for a finance broker, it will have a limited amount of money. Both examples require planning according to a limited budget. Another important concept, related to budget-aware reinforcement learning, is called risk profile, and it relates to how risk-averse the agent is. The risk profile can be used as an input to the learning algorithm so that different policies can be learned according to how much risk the agent is willing to expose itself to. This paper describes a new strategy to incorporate the agent’s risk profile as an input to the learning framework by using reward shaping. The paper also studies the effect of a constrained budget on RL and shows that, under such restrictions, RL algorithms can be forced to make a more efficient use of the available resources. The experiments show that as the even if it is possible to learn on a constrained budget with low budgets the learning process becomes slow. They also show that the reward shaping process is able to guide the agent to learn a less risky policy.

In this paper we show the advantage of modeling dependencies in supervised classification. The dependencies among variables in a multivariate data set can be linear or non linear. For this reason, it is important to consider flexible tools for modeling such dependencies. Copula functions are able to model different kinds of dependence structures. These copulas were studied and applied in classification of pixels. The results show that the performance of classifiers is improved when using copula functions.

Testor Theory allows performing feature selection in supervised classification problems through typical testors. Typical testors are irreducible subsets of features preserving the object discernibility ability of the original set of features. However, finding the complete set of typical testors for a dataset requires a high computational effort. In this paper, we make an empirical study about the performance of two of the most recent and fastest algorithms of the state of the art for computing typical testors, regarding the density of the basic matrix. For our study we use synthetic basic matrices to control their characteristics, but we also include public standard datasets taken from the UCI machine learning repository. Finally, we discuss our conclusions drawn from this study.

This work extends the use of the non-parametric dimensionality reduction method t-SNE [11] to unseen data. Specifically, we use retrieval experiments to assess quantitatively the performance of several existing methods that enable out-of-sample t-SNE. We also propose the use of deep learning to construct a multilayer network that approximates the t-SNE mapping function, such that once trained, it can be applied to unseen data. We conducted experiments on a set of images showing Maya hieroglyphs. This dataset is specially challenging as it contains multi-label weakly annotated instances. Our results show that deep learning is suitable for this task in comparison with previous methods.

An algorithm for filtering images contaminated by additive white Gaussian noise in discrete cosine transform domain is proposed. The algorithm uses a clustering stage to obtain mean power spectrum of each cluster. The groups of clusters are found by the proposed fast algorithm based on 2D histograms and watershed transform. In addition to the mean spectrum of each cluster, the local groups of similar patches are found to obtain the local spectrum, and therefore, derive the local Wiener filter frequency response better and perform the collaborative filtering over the groups of patches. The obtained filtering results are compared to the state-of-the-art filters in terms of peak signal-to-noise ratio and structural similarity index. It is shown that the proposed algorithm is competitive in terms of signal-to-noise ratio and in almost all cases is superior to the state-of-the art filters in terms of structural similarity.

The defects in tiles are directly related with changes in the structure or color components producing spots or stains in the final product. Usually, a visual inspection is carried out in order to detect one of such common defects in tiles; however this process depends on the expertise and abilities of the operator on duty. In this paper, we present the automation of defect detection in tiles using vision algorithms and Artificial Neural Networks (ANN). Color and texture information extracted from real tile images are used as input to a classifier based on neural networks. Setting parameters for extracting the texture attributes are obtained performing detailed tests of different distances, orientations and window sizes. An initial architecture of the ANN is obtained using texture features extracted from Brodatz images. Next, the neural network parameters are computed using real images from the tile database. The experimental tests validate the global performance, accuracy and feasibility of our approach.

In computer vision, facial images have several applications such as Facial Expression Recognition and Face Recognition. The segmentation of Regions Of Interest (ROIs) in face images are relevant, because those provide information about facial expressions. In this paper a method to segment mouth and eyebrows in face images based on edge detection and pixel density is proposed. According to the experimental results, our approach extracts the ROIs in face images taken from different public datasets.

Background subtraction is an important task in video processing and many algorithms are developed for solving this task. The vast majority uses the static behavior of the scene or texture information for separating foreground and background. In this paper we present a novel approach based on the integration of the unsteady vector field embedded in the video. Our method does not learn from the background and neither uses static behavior or texture for detecting the background. This solution is based on motion extraction from the scene by plane-curve intersection. The set of blobs generated by the algorithm are equipped with local motion information which can be used for further image analysis tasks. The proposed approach has been evaluated with a standard benchmark with competitive results against state of the art methods.

Mobile robots are more and more used on diverse environments to provide useful services. One of these environments are supermarkets, where a robot can help to find and carry products, maintain the account of them and to mark out from a list, the products already in the shopping car (maybe the same robot). However, a common problem on these environments is the autonomous localization, due to the fact that supermarkets are a set of aisles, and most of them look the same for laser range finders; sensors commonly used for localization. On this paper, we present an approach to localize autonomous mobile robots on supermarket by using a perspective reconstruction of the shelves and then an statistical comparison of the products present in them. In order to detect the shelves, the vanishing points are estimated to provide a fast and efficient way to segment products on them. To avoid multiple vanishing points on this kind of environments, result of the variety of products present, a variation of a RANSAC approach is proposed. Once a vanishing point has been determined, an homography process is applied to the shelves in order to rectify images. And finally, by horizontal histograms the robot is able to segment individual products to be compared to the data base. Then the robot will be able to detect by a probability function the correct aisle where it is.

In this study we tested some methods for detecting clandestine graves using hyperspectral remote sensing technology. Specifically, we addressed three research questions: What is the true potential of hyperspectral images for detecting buried remains? What is the useful information in hyperspectral images for detecting buried remains? When they should be acquired following a burial? For this matter, we simulated seven graves with varying number of carcasses of domestic pigs and monitored the spectral reflectance of the surface during a period of six months. A total of twelve hyperspectral images were formed and analyzed using standard pattern recognition methods. Results indicated that hyperspectral data can indeed have a true potential for detecting buried remains, but the detection can succeed only after three months from burial, and the useful wavelength intervals are mainly distributed along the spectral range of 700–1800 nm and with several narrow intervals that could not have been discovered using multispectral sensors.

In order to change the control approach that commercially available Unmanned Aerial Vehicles (UAVs) use to execute a flight plan, which is based on the Global Positioning System (GPS) and assuming an obstacle-free environment, we propose a hierarchical multi-layered control system that permits to a UAV to define the flight plan during flight and locate itself by other means than GPS. The work presented on this article aims to set the foundation towards an autonomous airborne agent, capable of locating itself with the aid of computer vision, model its environment and plan and execute a three dimensional trajectory. On the current stage of development we locate the vehicle using a board of artificial markers, the flight plan to execute was defined as either a cubic spline or a Lemniscate. As results, we present the resultant flight data when the proposed control architecture drives the vehicle autonomously.

In this paper we present an approach for the automatic evaluation of relations in ontologies of restricted domain. We use the evidence found in a corpus associated to the same domain of the ontology for determining the validity of the ontological relations. Our approach employs Latent Semantic Analysis, a technique based on the principle that the words in a same context tend to have semantic relationships. The approach uses two variants for evaluating the semantic relations and concepts of the target ontologies. The performance obtained was about 70% for class-inclusion relations and 78% for non-taxonomic relations.

People prefer to share and express opinions in their own language. Internet is a biggest repository for sharing opinions. Opinion mining uses Natural Language Processing (NLP), text analysis and computational linguistics to identify and extract subjective information in data. Opinion mining for Urdu language is not a well explored area. Therefore, an approach has been proposed which identifies and extracts adji-units and decisions from the given text using lexicon-based approach focusing on Urdu language. Adji-units are the expressions which contain subjective text in a sentence. Our proposed approach uses two-step lexicon to extract opinions from text chunks. Moreover, for Urdu language no such lexicons exist. The main aim is to develop a diverse two-step lexicon and highlight the linguistic as well as technical aspects of this multidimensional research problem. The performance of the proposed system is evaluated on multiple texts and the achieved results are quite satisfactory.

The screening of chromosomal defects, as trisomy 13, 18 and 21, can be obtained by the measurement of the nuchal translucency thickness scanning during the end of the first trimester of pregnancy. This contribution proposes an automatic methodology to detect mid-sagittal sections to identify the correct measurement of nuchal translucency. Wavelet analysis and neural network classifiers are the main strategies of the proposed methodology to detect the frontal components of the skull and the choroid plexus with the support of radial symmetry analysis. Real clinical ultrasound images were adopted to measure the performance and the robustness of the methodology, thus it can be highlighted an error of at most 0.3 mm in 97.4% of the cases.

This paper presents the Breast Ultrasound Analysis Toolbox (BUSAT) for MATLAB, which contains 62 functions to perform image preprocessing, lesion segmentation, feature extraction, and lesion classification. BUSAT is useful to codify programs for computer-aided diagnosis (CAD) purposes in reduced time; hence, to replicate several approaches proposed in literature is feasible. We provide the implementation of a CAD system to classify breast lesions into benign and malignant classes and an example to evaluate the classification performance. BUSAT could be downloaded from the following permanent link: http://​www.​tamps.​cinvestav.​mx/​~wgomez/​downloads.​html.

Quality assurance of dry cleaned industrial textiles is still a mostly manually operated task. In this paper, we present how computer vision and machine learning can be used for the purpose of automating defect detection in this application. Most existing systems require textiles to be spread flat, in order to detect defects. In contrast, we present a novel classification method that can be used when textiles are in inhomogeneous, voluminous shape. Normalization and classification methods are combined in a decision-tree model, in order to detect different kinds of textile defects. We evaluate the performance of our system in real-world settings with images of piles of textiles, taken using stereo vision. Our results show, that our novel classification method using key point pre-selection and convolutional neural networks outperform competitive methods in classification accuracy.

The answering of any test represents a challenge for students; however, foreign students whose first language is not English have to deal with the difficulty of the understanding of a series of questions written on a different language in addition of the effort required to solve the problem. In this study, we recorded the behavior of the brain signals of 16 students, 10 whom first language was English and 6 who were English learners, and used two supervised classification algorithms in order to identify the students’ language proficiency. The results shown that in both approaches, harder problems which required longer time to be responded had a higher accuracy rate; however, more tests are needed in order to understand the physical processing of written math text problem and the difference among both groups.

In steel production the finishing process on hot rolling mill includes a set of essential operations managed by complex control mechanical, electrical and hydraulic equipment. However, accuracy of mill automating mechanisms and sensors is still low due to hot hostile environment with strong vibration and shock. The proposed solution is a computer vision application that exploits morphological filtering and discontinuity masks for detection and separation of rods on rolling mill and provides fast recognition and tracking rod front ends during their deceleration on cooler. The proposed algorithm has been implemented and evaluated in real time conditions achieving precision of rod front end recognition in range of 90–98% on artificial and daylight illumination, respectively.