PLANTEAMIENTOS DE INTERVENCIÓN
CAPÍTULO 4: EL DISTRITO DE SAN MARCOS EN LOS AÑOS 2006 2009 Y LAS EXPERIENCIAS PREVIAS EN AGUA
4.2 El ámbito local
4.2.2 Las localidades seleccionadas y el contexto rural en salud y saneamiento
Features
The comparative gesture recognition performance of MLP-BP-NN, RBF-NN, KLMS- RBFNN-GAF and KLMS-RBFNN-SAF classifiers with proposed DWT-FR features for three hand gesture databases are tabulated in Table4.4. From Table 4.4, it is noticed
Table 4.4: Comparative performance of gesture recognition using MLP-BP-NN, standard RBF-NN, KLMS-RBFNN-GAF and KLMS-RBFNN-SAF classifiers with DWT-FR features for three distinct databases
Database Classifier Acc (%) Sen (%) Ppr (%) Spe (%)
Database I MLP-BP-NN 99.98 99.80 99.81 99.99 RBF-NN 99.98 99.80 99.81 99.99 KLMS-RBFNN-GAF 100.00 100.00 100.00 100.00 KLMS-RBFNN-SAF 100.00 100.00 100.00 100.00 Database II MLP-BP-NN 99.53 94.33 94.49 99.75 RBF-NN 99.77 97.25 97.42 99.88 KLMS-RBFNN-GAF 99.80 97.58 97.64 99.89 KLMS-RBFNN-SAF 99.81 97.75 97.78 99.90 Database III MLP-BP-NN 99.90 98.75 98.78 99.95 RBF-NN 99.97 99.58 99.60 99.98 KLMS-RBFNN-GAF 99.97 99.67 99.68 99.99 KLMS-RBFNN-SAF 99.98 99.75 99.76 99.99
that the average accuracy, sensitivity, positive predictivity and specificity of gesture recognition using DWT-FR features with MLP-BP-NN classifier are 99.98%, 99.80%,
99.81% and 99.99% respectively for Database I, 99.53%, 94.33%, 94.49% and 99.75% respectively for Database II and 99.90%, 98.75%, 98.78% and 99.95% respectively for Database III, and with RBF-NN classifier are 99.98%, 99.80%, 99.81% and 99.99% re- spectively for Database I, 99.77%, 97.25%, 97.42% and 99.88% respectively for Database II and 99.97, 99.58, 99.60 and 99.98% respectively for Database III. However, the av-
0 2 4 6 8 10 x 10−5 0.997 0.9975 0.998 0.9985 0.999 0.9995 1 1.0005 FPR TPR (a) 1 1.5 2 2.5 x 10−3 0.94 0.95 0.96 0.97 0.98 FPR TPR (b) 1 2 3 4 5 6 x 10−4 0.984 0.986 0.988 0.99 0.992 0.994 0.996 0.998 FPR TPR (c) MLP−BP−NN classifier RBF−NN classifier KLMS−RBFNN−GAS classifier KLMS−RBFNN−SAF classifier
Figure 4.6: ROC graphs of gesture recognition using MLP-BP-NN, RBF-NN, KLMS-RBFNN-GAF, and KLMS-RBFNN-SAF classifiers with DWT-FR fea- tures for (a) Database I, (b) Database II and (c) Database III.
erage accuracy, sensitivity, positive predictivity and specificity of gesture recognition using DWT-FR features with KLMS-RBFNN-GAF classifier are 100.00%, 100.00%, 100.00% and 100.00% respectively for Database I, 99.80%, 97.58%, 97.64% and 99.89% respectively for Database II and 99.97%, 99.67%, 99.68% and 99.99% respectively for Database III, and with KLMS-RBFNN-SAF classifier are 100.00%, 100.00%, 100.00% and 100.00% respectively for Database I, 99.81%, 97.75%, 97.78% and 99.90% respec-
4.4 Performance Evaluation
tively for Database II and 99.98%, 99.75%, 99.76% and 99.99% respectively for Database III. For proposed DWT-FR features, the ROC graphs of gesture recognition using MLP- BP-NN, RBF-NN, KLMS-RBFNN-GAF and KLMS-RBFNN-SAF classifiers for three distinct databases are shown in Figure 4.6. Table 4.4 and Figure 4.6demonstrate that the gesture recognition performances of modified RBF-NN classifiers (KLMS-RBFNN- GAF and KLMS-RBFNN-SAF) are better compared to MLP-BP-NN and RBF-NN classifiers for all three databases when DWT-FR features are used as input feature set. The average sensitivity of hand gesture recognition using MLP-BP-NN, RBF-NN, KLMS-RBFNN-GAF and KLMS-RBFNN-SAF classifiers with KM, DCT, proposed combined and proposed DWT-FR features for three hand gesture databases are shown in Figure 4.7. From Figure 4.7, it is observed that the performance of hand gesture recognition using proposed DWT-FR features is better compared to using KM, DCT, proposed combined features for all four different classifiers that are MLP-BP-NN, RBF- NN, KLMS-RBFNN-GAF and KLMS-RBFNN-SAF classifiers and all three different databases used except for two cases where the performances of DWT-FR features are equal with the performances of DCT features.
From the above experimental study it is demonstrated that in overall, gesture recog- nition performances of the proposed modified RBF-NN classifiers (KLMS-RBFNN-GAF and KLMS-RBFNN-SAF) are better compared to existing classifiers that are MLP- BP-NN and RBF-NN classifiers for all different feature sets like KM, DCT, proposed combined and proposed DWT-FR feature sets. The modified RBF-NN classifiers of- fer higher recognition performance because (i) they incorporate all the advantages of standard RBF-NN; (ii) they use a k-mean based center-selection algorithm which se- lect the centers of RBF units in automatic manner; and (iii) they use an LMS based weights update technique to update the estimated weight matrix at training phase with the goal to minimize the MSE and this in turn helps to improve the gesture recogni- tion performance. From the above experimental study, it is also demonstrated that the KLMS-RBFNN-SAF classifier offers better recognition performance compared to KLMS- RBFNN-GAF for all cases except for Database II using KM features and Database I. For
MLP−BP−NN RBF−NN KLMS−RBFNN−GAF KLMS−RBFNN−SAF 99 99.5 100 Sensitivity (%) (a) MLP−BP−NN RBF−NN KLMS−RBFNN−GAF KLMS−RBFNN−SAF 86 88 90 92 94 96 98 Sensitivity (%) (b) MLP−BP−NN RBF−NN KLMS−RBFNN−GAF KLMS−RBFNN−SAF 94 96 98 100 Sensitivity (%) (c)
KM features DCT features Combined features DWT−FR features
Figure 4.7: The average sensitivity of hand gesture recognition using MLP- BP-NN, RBF-NN, KLMS-RBFNN-GAF and KLMS-RBFNN-SAF classifiers with KM, DCT, proposed combined and proposed DWT-FR features for (a) Database I, (b) Database II and (c) Database III.
these cases, KLMS-RBFNN-SAF classifier provides an equal recognition performance with KLMS-RBFNN-GAF classifier. In most of the cases, the KLMS-RBFNN-SAF classifier provides better recognition performance over KLMS-RBFNN-GAF classifier because KLMS-RBFNN-SAF uses a set of composite sigmoidal functions as activation function which estimate the similar training patterns better compared to Gaussian ac- tivation functions.
4.5 Conclusions
4.5
Conclusions
The salient points covered in this chapter are:
• This chapter proposes two modified RBF-NN classifiers that are KLMS-RBFNN- GAF and KLMS-RBFNN-SAF classifiers for better recognition of hand gesture images.
• In KLMS-RBFNN-GAF classifier, the Gaussian function is used as activation func- tion, the centers are automatically selected through k-means based center-selection algorithm and the initial trained weight matrix is further updated using LMS al- gorithm whereas in KLMS-RBFNN-SAF, activation function is formed using a set of composite sigmoidal functions, centers are automatically selected using k-mean algorithm and initial trained weight matrix is further updated using LMS algo- rithm. For both cases, selected centers and updated weight matrix are stored at the end of training phase and these are used during testing phase to recognize hand gesture images.
• The experiments are separately conducted on three hand gesture databases us- ing KM, combined, DCT and DWT-FR features to evaluate gesture recognition performance of the proposed KLMS-RBFNN-GAF and KLMS-RBFNN-SAF clas- sifiers and to compare them with other reported classifiers such as MLP-BP-NN and RBF-NN.
• Experimental results demonstrate that the proposed modified RBF-NN classifiers provides higher recognition performance compared to other reported classifiers such as MLP-BP-NN and RBF-NN. Experimental results also indicate that the KLMS-RBFNN-SAF classifier offers better or equal recognition performance com- pared to KLMS-RBFNN-GAF classifier.
• From experimental results, it is also observed that the performance of hand ges- ture recognition using DWT-FR features is better compared to using KM, DCT
and proposed combined features for all different classifiers including MLP-BP-NN, RBF-NN, KLMS-RBFNN-GAF and KLMS-RBFNN-SAF classifiers.
C H A P T E R
5
Feature Vector Optimization
using Genetic Algorithm for
Hand Gesture Recognition
5.1
Introduction
I
n general, the feature subset selection (FSS) has been used in the area of pattern recognition where large datasets are involved [50]. A major problem associated with pattern recognition is the “curse of dimensionality” in which too many features are used to represent the patterns [117]. This problem led to a large number of classifier pa- rameters (e.g., weights of MLP-BP-NN) [118] which in turn increases computation for pattern classification. Therefore, the reduction of feature dimensionality has become an important task in the areas of pattern recognition [119], classification [120], data min- ing [121], bioinformatics [122] etc. The reduction of feature dimensionality contributes several advantages [50] for a recognition system employed for a specific application: (i) a reduction in the cost of acquisition of the data, (ii) improvement of the understand- ability of the final classification model, (iii) a faster induction of the final classification mode, (iv) an improvement in recognition performance. Feature Selection (FS) can be defined as a process to select a subset of features from original feature set. This can be carried out through eliminating the redundant, uninformative, and noisy features [50]. By reducing feature vector dimensionality, feature selection process not only reduce the cost of recognition method but also in some cases it can provide a better classification performance due to finite sample size effects [123].In this chapter, a genetic algorithm (GA) based feature vector optimization tech- nique is adopted to optimize the original feature set for hand gesture recognition [9]. The objective of the feature vector optimization is to reduce the dimension of feature vector by selecting an optimal subset of features that contains the most informative and discriminative information of patterns. The length of extracted feature set is re- duced by eliminating the unnecessary, redundant and irrelevant features. This work uses GA for feature vector optimization because it has following advantages: (i) it is a randomized search and optimization technique guided by the principles of evolution and natural genetics; (iii) For an optimization problem, GA performs search in com- plex, large and multimodal landscapes, and provide near-optimal solutions [124]; (iv)
5.1 Introduction
it is considered to be a robust method because no restrictions on the solution space are made during the process; (v) it exploits historical information structures from the previous solution to future solution which in turn to increase the performance of future solution structures [125]; (vi) it is an inherently parallel optimization method where the calculations of the fitness function on all search points of a population are completely independent and can be carried out in several operations. Therefore, rather than be- ing trapped in a suboptimal local maximum or minimum, it can move to near global optimum solution [126,127]. In this chapter, three hand alphabets databases of static gesture are used for recognition. As described in chapter 2, a common training and testing dataset is used for performance evaluation. The extracted sets of KM, DCT, proposed combined and proposed DWT-FR features are separately optimized using GA with proposed KLMS-RBFNN-GAF and KLMS-RBFNN-SAF classifiers. The KLMS- RBFNN-GAF and KLMS-RBFNN-SAF classifiers are discussed in detail in chapter 4. The final optimized feature subsets of KM, DCT, proposed combined and proposed DWT-FR features are applied to the input of the proposed classifiers. The experimen- tal results indicate that the proposed technique significantly reduces the dimension of feature vectors which in turn helps to reduce feature space complexity and for most of the cases, it also provides a better performance for hand gesture recognition.
5.1.1 Organization of the Chapter
The rest of the chapter is organized as follows. The theoretical background of the genetic algorithm (GA) is discussed in Section 5.2. The framework of optimum feature selection technique is thoroughly described in Section 5.3. The experimental results and performance comparisons are presented in Section 5.4. Finally, the chapter is concluded with some final remarks in Section 5.5.