Catálogo de publicaciones - libros

Emotion is fundamental to human experience influencing cognition, perception and everyday tasks such as learning, communication and even rational decision-making. This aspect must be considered in human-computer interaction. In this paper, we compare four different weighting functions in weighted KNN-based classifiers to recognize five emotions, including anger, happiness, sadness, neutral and boredom, from Mandarin emotional speech. The classifiers studied include weighted KNN, weighted CAP, and weighted D-KNN. To give a baseline performance measure, we also adopt traditional KNN classifier. The experimental results show that the used Fibonacci weighting function outperforms than others in all weighted classifiers. The highest accuracy achieves 81.4% with weighted D-KNN classifier.

This paper introduces the idea that recognition and interpretation of engineering drawings should be two interwoven phases, with each providing feedback to another, and applies this idea to a dynamic-rule-based method. Recognition rules with attributes are obtained by an automatic object feature extraction procedure, and stored in rule database. During the recognition phase, rules are firstly selected according to two attributes, domain and priority. Then the thresholds of the rules are adjusted automatically to obtain better match results and their priorities are modified dynamically to improve recognition efficiency. Especially, the interpretation phase based on the recognition is also valued in validating and rectifying the recognition result automatically and efficiently. This approach was implemented in a system for recognizing and interpreting architectural structure drawings, and has shown to embody good self-adaptability to various drafting conventions.

A novel method is developed for the dichromat’s visual correction. This scheme includes three steps. Firstly, two similarity matrices are established respectively for the normal three-dimensional color space and the color plane in which dichromats can distinguish all the colors. Then a 3D-2D mapping relationship is searched based on these similarity matrices. Finally, the original color image is transformed to a new one which is suitable for dichromats. The experiments on both color test images and real images demonstrate the ability of the scheme for color blindness correction. With the fixed transformation of color space, this scheme may have capability to help training dichromats to recognize most colors.

A new algorithm for decomposition of mixed pixels based on orthogonal bases of data space is proposed in this paper. It is a simplex-based method which extracts endmembers sequentially using computations of largest simplex volumes. At each searching step of this extraction algorithm, searching for the simplex with the largest volume is equivalent to searching for a new orthogonal basis which has the largest norm. The new endmember corresponds to the new basis with the largest volume. This algorithm runs very fast and can also avoid the dilemma in traditional simplex-based endmember extraction algorithms such as N-FINDR which generally produces different sets of final endmembers if different initial conditions are used. Moreover, with this set of orthogonal bases, the proposed algorithm can also determine the proper number of endmembers and finish the unmixing of the original images which cannot be done by the traditional simplex-based algorithms. Experimental results of both artificial simulated images and practical remote sensing images demonstrate that the algorithm proposed in this paper is a fast and accurate algorithm for the decomposition of mixed pixels.

Two-dimension road following is one of the crucial tasks of vision navigation. For the reasons of environment complexity and the discrepancy between motion images, the robust outdoor road following for two-dimension image sequence is still a challenging task. This paper proposes a novel road following method, which firstly uses the Mean Shift algorithm with embedded edge confidence to partition the images into homogenous regions with precise boundary. Then, according to the color statistic information of the road/non-road model obtained from previous frames, the Graph Cuts (GC) algorithm is used to achieve the final binary images and update the road/non-road model simultaneously. This algorithm combines the advantages of Graph Cuts algorithm and Mean Shift algorithm, and effectively solves some difficult problems of conventional methods, such as the adaptive selection of road model under complex environments, and the choice of effective criteria for the region merging. Experiment results indicate our method possesses excellent performance under complicated environment, and meets the requirements of fast computing.

This paper presents a digital shot system based on three dimensions integrated accelerometer. The digital shot with almost the same size and weight as the standard shot for open female has been designed, fabricated and tested. The data collecting and processing system, human-machine interface and athlete train guiding system are illuminated in detail. By using wavelet transformation, the characteristics of acceleration signals during the shot-putting period can be extracted. In this manner, the force sensing system serves as a powerful tool for coaches and sports scientists to make scientific researches on shot-put techniques. It also provides an intuitive and reliable guidance for the shot-put athletes to improve their skills.

From the Bayesian Ying-Yang (BYY) harmony learning theory, a harmony function has been developed for finite mixtures with a favorite property that its maximization can make model selection automatically during parameters learning. In this paper, we propose an adaptive gradient learning rule for maximizing the harmony function on Poisson mixtures which are applied more and more extensively in practice, especially for overdispersed data. It is demonstrated by simulation experiments that this adaptive gradient learning rule can determine the number of Poisson distributions during the parameters learning on a sample data set.

In order to recognize the acoustic emission source with different characteristics, the parameter-ratio method was put forward to analyze the characteristic parameters of acoustic emission from different source further. According to the peak amplitude, counts, energy and rise-time, the three ratios of the amplitude to the energy difference, the amplitude to the counts difference and the amplitude to the rise-time difference were used as the parameter-ratios. Based on the matter-element of extension theory, a matter-element model was built to describe the characteristics of the acoustic emission. The dependent function and degree of the characteristics of the acoustic sources were introduced to evaluate the possibility of the acoustic sources. The acoustic sources can be recognized, putting forward the recognition rules of parameter-ratio method. The recognition example was taken to validate the parameter-ratio method. It is shown that the parameter-ratio method can recognize the acoustic emission source well.

In this paper we present the experiments and results obtained in the classification of infant cry using a variety of classifiers, ensembles among them. Three kinds of cry were classified: normal (without detected pathology), hypo acoustic (deaf), and asphyxia. The feature vectors were formed by the extraction of Mel Frequency Cepstral Coefficients (MFCC); these were then processed and reduced through the application of five statistics operations, namely: minimum, maximum, average, standard deviation and variance. For the classification there were used supervised machine learning methods as Support Vector Machines, Neural Networks, J48, Random Forest and Naive Bayes. The ensembles used were combinations of these under different approaches like Majority Vote, Staking, Bagging and Boosting. The 10-fold cross validationtechnique was used to evaluate precision in all classifiers.

The aim of this study is to identify the author of an unauthorized document. Ten different feature vectors are obtained from authorship attributes, n-grams and various combinations of these feature vectors that are extracted from documents, which the authors are intended to be identified. Comparative performance of every feature vector is analyzed by applying Naïve Bayes, SVM, k-NN, RF and MLP classification methods. The most successful classifiers are MLP and SVM. In document classification process, it is observed that n-grams give higher accuracy rates than authorship attributes. Nevertheless, using n-gram and authorship attributes together, gives better results than when each is used alone.