International Journal of Pattern Recognition and Artificial Intelligence

In pattern recognition, irreducible testors have been used for feature selection. A number of exhaustive algorithms that find irreducible testors have been reported in the literature. One of the latest and more efficient algorithms reported is YYC, an incremental algorithm that finds all the irreducible testors from a training matrix. Its efficiency relies on building a smaller number of feature combinations by finding compatible sets from the top of the matrix to the current row. Nevertheless, as the number of sets currently found grows, YYC execution becomes too slow. This work proposes two improvements of YYC algorithm, incorporated in a pre-processing phase; additionally, a parallel version is implemented. The paper presents some experimental results using synthetic and real data.

Skeletal maturity estimation is an important medical procedure in the early diagnosis of growth disorders. Traditionally, it is performed by an expert physician or radiologist who determines it based on a visual subjective inspection, the approximated bone age of the child. This task is time consuming and is usually dependent on the judgment of each particular physician. Therefore, automated methods are extremely valuable and desirable. In this paper, we propose and describe an automatic method to estimate skeletal maturity through a supervised and incremental learning approach. Our method determines bone age by comparing aligned images with a $K$–$NN$ regression classifier. Here, we have solved the difficult task of image alignment by designing a radiological-hand specific Active Appearance Model, which was developed from a varied set of hand-labeled radiological images. By using this active model, our system constructs its own learned database by increasing a set of shape-aligned training images which are incrementally stored. Thus, when a test image arrives at the system, the alignment process is performed before the classification task takes place. For that purpose, we designed an original layout of landmarks to be located in representative regions of the radiographical image of the hand. Our results show that it is possible to use pixels directly as classification features as long as training and testing images have been previously aligned in shape and pose.

Recently, Human Activity Recognition (HAR) has become an important research area because of its wide range of applications in several domains such as health care, elder care, sports monitoring systems, etc. The use of wearable sensors — specifically the use of inertial sensors such as accelerometers and gyroscopes — has become the most common approach to recognize physical activities because of their unobtrusiveness and ubiquity. Overall, the process of building a HAR system starts with a feature extraction phase and then a classification model is trained. In the work of Siirtola et al. is proposed an intermediate clustering step to find the homogeneous groups of activities. For the recognition step, an instance is assigned to one of the groups and the final classification is performed inside that group. In this work we evaluate the clustering-based approach for activity classification proposed by Siirtola with two additional improvements: automatic selection of the number of groups and an instance reassignment procedure. In the original work, they evaluated their method using decision trees on a sports activities dataset. For our experiments, we evaluated seven different classification models on four public activity recognition datasets. Our results with 10-fold Cross Validation showed that the method proposed by Siirtola with our additional two improvements performed better in the majority of cases as compared to using the single classification model under consideration. When using Leave One User Out Cross Validation (user independent model) we found no differences between the proposed method and the single classification model.

Breast cancer (BC) is one of the leading causes of death in adult women worldwide and the best way to reduce mortality and improve prognosis is through early diagnosis. Thus, it is necessary to optimize diagnostic methods; one option could be the automatic detection of patterns in 1D-II. In that respect, through recent analysis of unidimensional Immunoblot Images (1D-II), it was possible to distinguish between women with and without breast disease using as a discrimination criterion the presence of autoantibodies (bands) in their blood. However, the analysis of 1D-II is a difficult task even for an expert, generating great subjectivity and complexity in the process of interpretation.

In the present study, a semi-automatic methodology for the bands’ analysis contained in the 1D-II’s was implemented and evaluated, the bands were extracted using digital image processing techniques. This was possible through the recognition of banding patterns represented as time series to distinguish between three classes: women with breast cancer (BC), women with benign breast pathology (BBP) and women without breast pathology (H). The classification was performed using the machine learning algorithm k-nearest neighbors (KNN) with different parameters over the time series representation.

The semi-automatic method here presented was able to reduce the time, complexity and subjectivity of the image analysis with the performance metrics compared, obtaining similar percentages for both representations. With the traditional analysis, binary representation [Accuracy 72.8%, Precision 73.42% for three classes (BC, BBP and H) and Accuracy 90.91% Accuracy 92.55% Sensitivity 93.57% and Specificity 92.99% for two classes (BC and H)], versus Time series representation [Accuracy 66.4%, Precision 67.07% for three classes (BC, BBP and H) and Accuracy 86.36% Accuracy 87.31% Sensitivity 95.86% and Specificity 85.56% for two classes (BC and H)].

Images of different origin contain textures, and textural features in such regions are frequently employed in pattern recognition, image classification, information extraction, etc. Noise often present in analyzed images might prevent a proper solution of basic tasks in the aforementioned applications and is worth suppressing. This is not an easy task since even the most advanced denoising methods destroy texture in a more or less degree while removing noise. Thus, it is desirable to predict the filtering behavior before any denoising is applied. This paper studies the efficiency of texture image denoising for different noise intensities and several filter types under different visual quality criteria (quality metrics). It is demonstrated that the most efficient existing filters provide very similar results. From the obtained results, it is possible to generalize and employ the prediction strategy earlier proposed for denoising techniques based on the discrete cosine transform. Accuracy of such a prediction is studied and the ways to improve it are considered. Some practical recommendations concerning a decision to undertake whether it is worth applying a filter are given.

A crucial step in plagiarism detection is text alignment. This task consists in finding similar text fragments between two given documents. We introduce an optimization methodology based on genetic algorithms to improve the performance of a plagiarism detection model by optimizing its input parameters. The implementation of the genetic algorithm is based on nonbinary representation of individuals, elitism selection, uniform crossover, and high mutation rate. The obtained parameter settings allow the plagiarism detection model to achieve better results than the state-of-the-art approaches.

This paper presents a system for offline recognition of cursive Arabic handwritten text based on Hidden Markov Models (HMMs). The proposed work reports an effective method taking into account the context of character by applying an embedded training-based HMMs to perform and enhance the character models. The system is analytical without explicit segmentation; extracted features preceded by baseline estimation are statistical and structural to integrate both the peculiarities of the text and the pixel distribution characteristics of the word image. The experiments are done on benchmark IFN/ENIT database. The proposed work shows the effectiveness of using embedded training-based HMMs for enhancing the recognition rate, and the obtained results are promising and encouraging.

Recent developments in speech synthesis have produced systems capable of producing speech which closely resembles natural speech, and researchers now strive to create models that more accurately mimic human voices. One such development is the incorporation of multiple linguistic styles in various languages and accents. Speech synthesis based on Hidden Markov Models (HMM) is of great interest to researchers, due to its ability to produce sophisticated features with a small footprint. Despite some progress, its quality has not yet reached the level of the current predominant unit-selection approaches, which select and concatenate recordings of real speech, and work has been conducted to try to improve HMM-based systems. In this paper, we present an application of long short-term memory (LSTM) deep neural networks as a postfiltering step in HMM-based speech synthesis. Our motivation stems from a similar desire to obtain characteristics which are closer to those of natural speech. The paper analyzes four types of postfilters obtained using five voices, which range from a single postfilter to enhance all the parameters, to a multi-stream proposal which separately enhances groups of parameters. The different proposals are evaluated using three objective measures and are statistically compared to determine any significance between them. The results described in the paper indicate that HMM-based voices can be enhanced using this approach, specially for the multi-stream postfilters on the considered objective measures.

This paper introduces the Gamma Rough Sets for management information systems where the universe objects are represented by continuous attributes and are connected by similarity relations. Some properties of such sets are demonstrated in this paper. In addition, Gamma Rough Sets are used to improve the Gamma associative classifier, by selecting instances. The results indicate that the selection of instances significantly reduces the computational cost of the Gamma classifier without affecting its effectiveness. The results also suggest that the selection of instances using Gamma Rough Sets favors other lazy learners, such as Nearest Neighbor and ALVOT.

In Digital Image Steganography, Pixel-Value Differencing (PVD) methods use the difference between neighboring pixel values to determine the amount of data bits to be inserted. The main advantage of these methods is the size of input data that an image can hold. However, the fall-off boundary problem and the fall in error problem are persistent in many PVD steganographic methods. This results in an incorrect output image. To fix these issues, usually the pixel values are either somehow adjusted or simply not considered to carry part of the input data. In this paper, we enhance the Tri-way Pixel-Value Differencing method by finding an optimal pixel value for each pixel pair such that it carries the maximum input data possible without ignoring any pair and without yielding incorrect pixel values.

A new sub-segmentation method has been proposed in 2009 which, in digital images, help us to identify the typical pixels, as well as the less representative pixels or atypical of each segmented region. This method is based on the Possibilistic Fuzzy c-Means (PFCM) clustering algorithm, as it integrates absolute and relative memberships. Now, the segmentation problem is related to isolate each one of the objects present in an image. However, and considering only one segmented object or region represented by gray levels as its only feature, the totality of pixels is divided in two basic groups, the group of pixels representing the object, and the others that do not represent it. In the former group, there is a sub-group of pixels near the most representative element of the object, the prototype, and identified here as the typical pixels, and a sub-group corresponding to the less representative pixels of the object, which are the atypical pixels, and generally located at the borders of the pixels representing the object. Besides, the sub-group of atypical pixels presents greater tones (brighter or towards the white color) or smaller tones (darker or towards black color). So, the sub-segmentation method offers the capability to identify the sub-region of atypical pixels, although without performing a differentiation between the brighter and the darker ones. Hence, the proposal of this work contributes to the problem of image segmentation with the improvement on the detection of the atypical sub-regions, and clearly recognizing between both kind of atypical pixels, because in many cases only the brighter or the darker atypical pixels are the ones that represent the object of interest in an image, depending on the problem to be solved. In this study, two real cases are used to show the contribution of this proposal; the first case serves to demonstrate the pores detection in soil images represented by the darker atypical pixels, and the second one to demonstrate the detection of microcalcifications in mammograms, represented in this case by the brighter atypical pixels.

Most algorithms to mine graph patterns, during the searching process, require a pattern to be identical to its occurrences, relying on the graph isomorphism problem. However, in recent years, there has been interest in the case in which it is acceptable to have some differences between a pattern and its occurrences, whether these differences are in labels or in structure. Allowing some differences and using inexact matching to measure the similarity between graphs lead to the discovery of new patterns, but some important challenges, such as the increment on the number of found patterns, make the post-mining analysis harder. In this work we focus on two extensions of the AGraP algorithm, which mines inexact patterns, addressing the issue of reducing the output pattern set while trying to retain the useful information gained through the use of inexact matching. First, exploring a traditional approach, we propose the CloseAFG algorithm that focuses on closed patterns. Then, we propose the IntAFG algorithm to find a subset of patterns covering the original pattern set, while lessening redundancy among selected patterns. We show the performance of our approaches through some experiments on synthetic databases; additionally, we also show the usefulness of the reduced pattern sets for image classification.

Bayesian learning is applied on two class systems. Partitioning a big sample made up of many elements of two classes of indistinguishable objects, we indistinctly pursue from 2 to 5 training sets called hypotheses in the probability field, with a plausible rate of object from each hypothesis. Objects are taken one by one from the sample. The basic aim faced is to predict one type of objects in the following occasion in which an agent takes one of them from the original sample to test it. We obtain the graph of a posteriori probability for each hypothesis of one of the objects. A prediction that the following object is specifically one of them is acquired in one probability curve by means of training previously accomplished. This methodology is applied on manufacture of glass bottles of two classes: good or crash. The main interest is to predict which machine produced one detected crash bottle because bottles turn to be indistinguishable when they are reviewed. This is solved by fixing a priori probabilities and taking into account all possible probability distribution combinations in the classes.

Automatic image annotation and image segmentation are two prominent research fields of Computer Vision, that are getting higher attention these days to accomplish image analysis and scene understanding. In this work, we present an annotation algorithm based on a hierarchical image partition, that makes use of Markov Random Fields (MRFs) to model spatial and hierarchical relations among regions in the image. In this way, we can capture local, global and contextual information. Also, we combine the processes of annotation and segmentation in an iterative way so that each process benefits from the other. Furthermore, we investigate the selection of the starting segmentation level for the hierarchical annotation process, to show its relevance for the final results. We experimentally validate our approach in three well-known datasets: CorelA, Stanford Background and MSRC-21 datasets. In these datasets, we achieved better or comparable results to other state-of-the-art algorithms, improving our base classifier in all cases.

Partially observable Markov decision processes (POMDPs) are mathematical models for the planning of action sequences under conditions of uncertainty. Uncertainty in POMDPs is manifested in two ways: uncertainty in the perception of model states and uncertainty in the effects of actions on states. The diagnosis and treatment of cerebral vascular diseases (CVD) present this double condition of uncertainty, so we think that POMDP is the most suitable method to model them. In this paper, we propose a model of CVD that is based on observations obtained from neuroimaging studies such as computed tomography, magnetic resonance and ultrasound. The model is designed as a POMDP because the health status of the patient is not directly observable, and only can be deduced, with some probability, from the observations in the cerebral images. The components of the model (states, observations, actions, etc.) were defined based on specialized literature. A diagnosis of the patient’s health status is made and the most appropriate action for the recovery of health is recommended after introducing the observations when operating the model. Consultation of the probable state of health of the patient and alternative actions is also allowed.