Vol.4 No.2 2011
5/66

Research paper : ARGUS: Adaptive Recognition for General Use System (N. Otsu et al.)−76−Synthesiology - English edition Vol.4 No.2 (2011) and body type among other features are compared (or quantified) as partial images in order to reach a final decision. However, overall recognition is dependent on the recognition of parts, which implies that the overall recognition will be incorrect if the part recognition is erroneous.Thus, most ordinary approaches are “serial and procedural types” that first segment each individual object in this manner from the image, and then perform recognition according to a pre-prepared model. However, because the pattern generally has different variations, the model must also be made proportionately more complex in various ways. Moreover, accumulation of errors at each stage of processing in serial procedures results in overall vulnerability; a large amount of calculation is involved and it is difficult in practice to obtain the required recognition performance. The problem lies in the tendency to consider this as a logical procedure in an ad hoc manner at the image level. In a way, it is an approach dominated by the Neumann-type computer programming paradigm.As the antithesis of this method, from the late 1980s, the “parallel and adaptive (learning) type” method was proposed using neural networks[7]; in addition to the study of the theoretical aspects, various applications have been attempted, especially in pattern recognition and control. However, because of the constraints that the elements are nonlinear and have bounded values [0, 1], information representation and feature extraction usually tend to be ambiguous. In recent years, additional problems such as the arbitrary nature of the model and the learning speed and convergence have been indicating the need for a change toward the nonlinear multivariate analysis, such as the kernel method[8].To examine a new methodology for visual systems, recognition systems in general, it is necessary to theoretically reconsider the general framework of the underlying pattern recognition mechanism, especially information representation and feature extraction.2.1 General framework for pattern recognitionIn pattern recognition, recognition is accomplished by multiple extraction (thus represented by the vector x) of some feature values effective for recognition (generally functionals xi =φi[f] defined as functions of the function f) from the pattern, a signal expressed by a function f localized in space-time. Typically, as shown in Fig.2, the framework comprises a two-stage process of “feature extraction” and “recognition”. Recognition can be divided into classification and clustering. Classification is the determination of whether the input pattern corresponds to one of the several known categories, and is called supervised learning because the answer is given in the learning stage. Clustering is called unsupervised learning,which discriminates the input pattern into several clustersFig. 3 General framework for pattern recognition (detail)Fig. 2 General framework for pattern recognition (typical)(categories). Many techniques have been proposed regarding classification, and it is already theoretically known that the minimum error rate classification method is the Bayesian decision rule, which decides on the category Cj with a maximum posterior probability P(Cj |x). This implies that the feature extraction at the first stage is important as the requirement which dominates the recognition system efficiency, however various ad hoc or heuristic techniques have been suggested until date.2.2 Feature extraction theoryThe author has conducted a theoretical study of these feature extractions[2]. General framework for feature extraction comprises “invariant feature extraction” as the geometrical aspect and “discriminant feature extraction” as the statistical aspect. In principle, it is important that feature extraction comprises these two stages in this order. Fig.3 thus demonstrates the general framework for pattern recognition, as a natural consequence of this theory.2.2.1 Invariant feature extraction (geometrical aspect)The observed image f as a pattern is subject to various continuous geometrical transformations (generally, projective transformations) such as translation, scaling, and rotation due to the relative position and movement of the observer and the object. However, recognition results are independent of these and remain invariant. In Pattern space(function space)Feature space(vector space)Category set(discrete set)Feature extraction ΨRecognition (clustering/classification)CjfxPattern space(function space)Invariant feature space(vector space)Discriminant feature space(vector space)Category set(discrete set)Discriminant feature extractionInvariant feature extraction Recognition(clustering/classification)xy Ψ(x)x=Φ[ ]=y( )(λ)( | )( | )xyiCiCiCppfTfr

元のページ 

page 5

※このページを正しく表示するにはFlashPlayer10.2以上が必要です