This work is supported by the Government of India, Defence Research & Development Organization, Grant No. ERIP-ER- 0703643-01-1025. ICETECT 2011 Evidence Generation for Dempster-Shafer Fusion Using Feature Extraction Multiplicity and Radial Basis Network Prabha Verma Sensors & Signal Processing Laboratory, Department of Physics, Banaras Hindu University, Varanasi 221005, India. e-mail: pverma.bhu@gmail.com R.D.S. Yadava Sensors & Signal Processing Laboratory, Department of Physics, Banaras Hindu University, Varanasi 221005, India e-mail: ardius@gmail.com , ardius@bhu.ac.in Abstract—Feature extraction methods in pattern recognition tasks seek to transform data variables to abstract mathematical variables such that their scores (called features) reveal hidden data structure of high cognitive value. Various feature extraction methods process raw data from different perspectives. Some depend on statistical correlation or independence such as principal component analysis (PCA), independent component analysis (ICA), singular value decomposition (SVD) and linear discriminant analysis (LDA), and some others aim to model parametric representations such as partial-least-square regression (PLSR). These methods can be viewed as independent observers who generate different feature sets for describing the same data. In supervised pattern recognition problems, this viewpoint can be combined with a classifier function to generate independent sets of class likelihood. The latter can be interpreted as evidences for class identities assigned by independent expert systems consisted of feature extraction method and classifier function combinations. Having created such set of experts, one can employ an information fusion system that could predict class identities. Following this paradigm, we used above mentioned feature extraction methods paired with a radial basis network to generate evidences, and applied Dempster-Shafer (D-S) fusion for pattern classification in a number of benchmark data sets. It is found that DS fusion results in enhanced classification rates compared to results from individual expert systems. Keywords—Evidence generation, Dempster-Shafer fusion, pattern recognition I. INTRODUCTION Information fusion techniques combine multiple sources of information to create basis for making decision. The basic aim is to improve observability of target and boost confidence level of decision making process. The baseline information is collected or created from diverse perspectives so that in combination they provide rich information pool for decision taking [1]. Creation of diversified information can proceed at several levels. It can be done directly from measurements by using different methods such as object images collected by active, passive and spectral methods; by pooling measurements, data bases and expert opinion such as collection of biometric data for disease diagnostics; and by processing measured data from different perspectives such as analyzing time-series data by different integral transforms and statistical correlation methods. In this work, we extend the last paradigm for evidence generation for object recognition. We employ five feature extraction methods to process a given data so that each method generates a different set of features. This is creating diversity of information by processors of varied perspectives from a given set of measurements. For evidence generation, the feature sets are fed to a class predictor which labels class’s identities according to maximum values of a likelihood function. The likelihood values can be interpreted as class evidences. Thus, a feature extractor and class predictor combination can be taken to be an evidence generator. Different extractor-predictor combinations then can then be treated as independent evidence generators. In order to verify the veracity of this approach, we undertook a case study of using five feature extractors in combination with radial basis network to create five different evidence generators. The processors are: principal component analysis (PCA), independent component analysis (ICA), singular value decomposition (SVD), linear discriminant analysis (LDA) and partial-least-square regression (PLSR). Then, combine these evidences by using Dempster-Shafer (D-S) rule to generate class beliefs for identity declaration. II. EVIDENCE GENERATION BASED ON MULTIPLE FEATURE EXTRACTION AND RADIAL BASIS NETWORK Figure 1 shows schematically the information fusion strategy outlined above. In this section we summarize the basis for present approach for evidence generation for Dempster- Shafer fusion. The ground for D-S fusion rule lies in independence of evidences being fused. Feature extraction methods pull out concrete information content from data by applying certain criteria which data structure most probably conforms. These criteria are formulated in objective functions of respective extraction methods. By applying different feature extraction methods on the same data set should therefore describe hidden structure of data from different perspectives. Here, we have chosen five most common feature extraction methods. These methods can be considered as five observers of the same event producing their own measurements based on the criteria they use. In the following we briefly summarize the PROCEEDINGS OF ICETECT 2011 978-1-4244-7926-9/11/$26.00 ©2011 IEEE 542