Vol.4 No.2 2011
11/66

Research paper : ARGUS: Adaptive Recognition for General Use System (N. Otsu et al.)−82−Synthesiology - English edition Vol.4 No.2 (2011) K=NC3 two-dimensional (3×M) matrices, and taking HLAC features from each 3 × 3, a feature vector with K × HLAC dimensions is obtained. By performing multivariate analysis on this (PCA, DA), analysis of such time series can be conducted (abnormality detection and discrimination). This method has been applied to abnormality detection in electrocardiograms[19] and in the analysis of the movement of a multi-fingered robot hand with multiple degrees of freedom[20].Moreover, causal relationships, which can be interpreted as the asymmetric interrelationship (correlation) among time series data, are important in many fields. Granger Causality[21] has been proposed as an analysis index using a linear auto-regression model, but the present paper expands this model to a polynomial auto-regression model[22] (which therefore involves Higher-order Local Auto-Correlation features). Furthermore, by introducing a weighting function w(t) (a Causality Marker) to indicate the existence of a causal relationship, we have proposed a method that automatically extracts where a causal relationship exists[23].5.8 Correspondence learningCorrespondence learning is connected to a wide range of common applications. Retrieval through impressions and interactive searching as well as automatic evaluation (prediction) become possible by approximating (canonical correlation analysis (CCA) or multiple regression analysis (MRA)) through learning the correspondence between the expression of a person’s judgment or evaluation (external criteria) toward a pattern (still or moving images), such as the qualitative expression in the form of keywords or impression (sensitivity) vocabulary yNote7 or rating y, and the feature vector expression (HLAC/CHLAC) x of a pattern. Figure 19 shows applications to retrieving of family crests by impressions (CCA)[24], and to the automatic evaluation of exercise (MRA)[25].The former has been further applied to general image annotation and retrieval[26], and the latter to the automatic indexing of sports video images[27], as well as to the judgment of beef meat quality (BMS) based on ultrasound video images[28].6 Effectiveness of the theoretical approachThis paper has thus far given an outline of an Adaptive Recognition for General-Use System (ARGUS) constructed to fulfill basic required conditions, based on feature extraction theory in pattern recognition. This paper has also discussed the system’s application, focusing on a variety of practical applications in visual systems.Unlike the scientific approach in physics and chemistry (elucidation of phenomena), in engineering applications, and in particular information technology, construction methods designed for realizing functionality are highly flexible and tend to become ad hoc and arbitrary. Thus, it is important to design proper and novel solutions from a theoretical perspective based on the fundamental requirement conditions of application demand.By considering the fundamental framework of pattern recognition based on a theoretical standpoint, the method in this paper comprises a twostage method of Higher-order Local Auto-Correlation (HLAC/CHLAC), which is a geometrical invariant feature extraction, and multivariate data analysis, which is a statistical discriminant feature extraction. By using the latter, it is possible to learn from examples appropriate to the task. The method requires neither any model of the object nor prior knowledge, and the shape and movement of the object pattern are distinguished as points in discriminant feature space. Since segmentation of the object is also unnecessary and the computation is small with a fixed quantity of the sum of products, even moving images can be processed at far greater speeds than real-time operation on a normal PC. The features of this method are asfollows:• non-model base methodology high versatility.• basic initial features (HLAC/CHLAC) applicable to a wide range of data formats.• statistical learning (MDA) task adaptability and increased accuracy.• parallel sum of products operations possible to process large amounts of data at high speed.Almost as expected, through a variety of applications, it has outperformed the schemes that have been developed thus far. This can be attributed greatly to the method that is substantiated by theory, in particular the predominance of the Higher-order Local Auto-Correlation features and its essence. In contrast to being restricted to a two-point relationship of usual autocorrelation, by increasing to higher orders of threepoint relationship, the features obtained have become specific, e.g., curvature (convexity/concavity) rather than local straight-line direction for a contour in still images, and acceleration rather than velocity in moving images. These basic and essential initial features do not use an arbitrary iterative procedure or logical decisions (such as threshold processing and conditional branching, etc.). Rather, they use a multivariate data analysis technique and are integrated into new effective features in a parallel and comprehensive manner, forming a robust system with low information loss.HLAC/CHLAC are fundamental general-purpose features, i.e., statistics (correlation and frequency) of spatio-temporally localized “patterns”. In that sense, this constitutes a precedent for such trends as “from a model collation base to local feature

元のページ 

page 11

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