Catálogo de publicaciones - libros

Image mining is more than just an extension of data mining to image domain but an interdisciplinary endeavor. Very few people have systematically investigated this field. Mining association rules in medical images is an important part in domain-specific application image mining because there are several technical aspects which make this problem challenging. In this paper, we extend the concept of association rule based on object and image in medical images, and propose two algorithms to discover frequent item-sets and mine interesting association rules from medical images. We describe how to incorporate the domain knowledge into the algorithms to enhance the interestingness. Some interesting results are obtained by our program and we believe many of the problems we come across are likely to appear in other domains.

Hybrid feature ranking is a feature selection method which combines the quickness of the filter approach and the accuracy of the wrapper approach. The main idea consists in a two steps procedure: building a sequence of feature subsets using an informational criterion, independently of the learning method; selecting the best one with a cross-validation error rate evaluation, using explicitly the learning method. In this paper, we show that in the protein discrimination domain, few examples but numerous descriptors, compared to a traditional approach where each descriptor is evaluated separately in the first step, to take account of their redundancy in the construction of candidate subsets of features reduces the size of the optimal subset and improves, in certain cases, the accuracy.

Prediction of protein subcellular localization is one of the hot research topics in bioinformatics. In this paper, several support vector machines (SVM) with a new presented coding scheme method based on N-terminal amino compositions are first trained to discriminate between proteins destined for the mitochondrion, the chloroplast, the secretory pathway, and ‘other’ localizations. Then a decision unit is used to make the final prediction based on several SVMs’ outputs. Tested on redundancy-reduced sets, the proposed method reached 89.6 % (plant) and 91.9% (non-plant) total accuracies, which, to the best of our knowledge, are the highest ever reported using the same data sets.

There are a number of dirty data in the load database produced by SCADA system. Consequently, the data must be adjusted carefully and reasonably before being used for electric load forecasting or power system analysis. This paper proposes a dynamic and intelligent curve adjusting model based on data mining theory. Firstly the Kohonen neural network is meliorated according to fuzzy soft clustering arithmetic which can realize the collateral calculation of Fuzzy c-means soft clustering arithmetic. The proposed dynamic algorithm can automatically find the new clustering center, namely, the character curve of data, according to the updating of swatch data. Combining an RBF neural network with this dynamic algorithm, the intelligent adjusting model is introduced to identify the dirty data. The rapidness and dynamic performance of model make it suitable for real-time calculation. Test results using actual data of Jiangbei power supply bureau in Chongqing demonstrate the effectiveness and feasibility of the model.

It is an important research topic to improve detection rate and reduce false positive rate of detection model in the field of intrusion detection. This paper adopts an improved boosting method to enhance generalization performance of intrusion detection model based on rule learning algorithm, and presents a boosting intrusion detection rule learning algorithm (BIDRLA). The experiment results on the standard intrusion detection dataset validate the effectiveness of BIDRLA.

This paper states that the RoleOf Relationship can provide a general approach to resolve instantiation problems of design patterns. The problems come from the fact that pattern logic scatters across multiple business classes (classes specific to each application). This causes problems such as decreasing reusability of pattern logic, and losing of the instantiation information of pattern (traceability and overlapping problem) etc. To resolve these problems in design level, an approach for design pattern instantiation based on RoleOf relationship is proposed. In our approach, roles of pattern are treated as the independent modeling elements and RoleOf relationship is used to associate a role with a business class. The meta model of RoleOf relationship for pattern instantiation and its semantics are proposed as well. Examples are used to illustrate this approach. Implementation and behavior description of RoleOf relationship are also presented in the paper.

A novel approach for three-dimensional reconstruction and inspection of industrial parts with image sequence acquired by single non-metric CCD camera is proposed. The purpose of the approach is to reconstruct and thus inspect the producing imprecision (of deformation) of industrial sheetmetal parts. Planar control grid, non-metric image sequence and CAD-designed data are used as information sources. Principles of least squares template matching to extract lines and points from the imagery are presented. Hybrid point-line photogrammetry is adopted to obtain the accurate wire frame model. Circles, connected arcs and lines on the part are reconstructed with direct object space solution. The reconstructed CAD model can be used for inspection or quality control. Experimental results are very satisfying.

In this paper, an advanced distributed control scheme that connects the control networks to the IP networks, based on LonWorks technology, which is one of the control networks, are presented. The proposed approach is implemented by using a simple programmable basic Lon node (BLN) and a IBM PS/2 (PC) compatible computer as an Internet server. BLN is a developed electric board in this research and it physically contains a transceiver, Neuron chip and some memory devices. To perform various functions as an Internet server of PC, control software of Lon on internet system (LOIS) with C-language for GNU/LINUX environment is also developed. Our approach makes system designers to easily implement their various specific applications, only with the download of a control program from serial port (RS-232) of PC.

Structural damage detection is very important for identifying and diagnosing the nature of the damage in an early stage so as to reduce catastrophic failures and prolong the service life of structures. In this paper, a novel approach is presented that integrates independent component analysis (ICA) and support vector machine (SVM). The procedure involves extracting independent components from measured sensor data through ICA and then using these signals as input data for a SVM classifier. The experiment presented employs the benchmark data from the University of British Columbia to examine the effectiveness of the method. Results showed that the accuracy of damage detection using the proposed method is significantly better than the approach by integrating ICA and ANN. Furthermore, the prediction output can be used to identify different types and levels of structure damages.

We have shown [8] that LZ78 parse length can be used effectively for a music classification task. The parse length is used to compute a normalized information distance [6,7] which is then used to drive a simple classifier. In this paper we explore a more subtle use of the LZ78 parsing algorithm. Instead of simply counting the parse length of a string, we use the coding dictionary constructed by LZ78 to derive a valid string kernel for a Support Vector Machine (SVM). The kernel is defined over a feature space indexed by all the phrases identified by our (modified) LZ78 compression algorithm. We report experiments with our kernel approach on two datasets: (i) a collection of MIDI files and (ii) Reuters-21578. We compare our technique with an -gram based kernel. Our results indicate that the LZ78 kernel technique has a performance similar to that obtained with the best -gram performance but with significantly lower computational overhead, and without requiring a search for the optimal value of .