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

Vol.4 No.2 2011

8/66
Research paper : ARGUS: Adaptive Recognition for General Use System (N. Otsu et al.)−79−Synthesiology - English edition Vol.4 No.2 (2011) Fig. 9 Simultaneous recognition (/enumeration) of multiple objects[5][6]Fig. 11 Example of the JAFFE facial expression dataset (3 people)[14]Fig. 10 Recognition (/enumeration) of topological characteristics[5][6]Fig. 8 Multivariate data analysis method (by objective)Whether an external criterion exists corresponds to being supervised. A quantification method is for the case of qualitative data (Yes/No, 1/0). the masks over the solid frame XYT. The dimension is 279 for a gray-scale moving image and 251 for a binary moving image.The feature extraction methodology using the integral features of HLAC (CHLAC) is a fundamental and general-purpose feature extraction method for “object shape (and movement)” and satisfies the required conditions R1 (shift-invariance) and R2 (frame-additivity). Employing these methods, the recognition object can always be captured and represented in a unified manner as a single point (a vector) x in an invariant feature space.4.2 Discriminant feature extraction (MDA)In the next stage of adaptive learning (satisfying R3), statistical discriminant feature extraction, various multivariate data analysis (MDA) techniques are applied as linear mappings (Fig.8). This refers to adaptively deriving a new feature y optimized for the given recognition task, as the weighted linear sum of the elements of the HLAC or CHLAC feature vector x (R3: adaptive trainability) (Fig.5) ; since the mappings are linear, this secures the required condition R2 (frame-additivity).A similarity can be found in neural networks, but owing to its nonlinearity R2 is not preserved. In addition, it requires an iterative solution for optimization and takes much computation time. On the other hand, MDA has the advantage that by learning through examples, the weights optimal for the tasks are easily obtained in an analytically explicit formNote3. 4.3 Characteristics of the ARGUS recognition systemThis formulation comprising these two stages of feature extraction does not require segmentation or positioning of the object, and is unique in not requiring any knowledge or model of the object. Thus, the formulation has a versatility that makes it applicable for various recognitions, measurements or enumerations of still and moving images. Moreover, since it basically performs only the multiply-accumulate operation, even CHLAC can run on a normal PC with an extremely high processing speed (2 msec/frame).5 Application examples5.1 Simultaneous recognition (/enumeration) of multiple objectsAs an example of recognition of still images, an application for the enumeration task of simultaneously recognizing multiple objects is presented (Fig.9). This can be easily realized by utilizing factor analysis (FA), based on the shift-invariance Discriminant Analysis (QuantificationⅡ)Canonical Correlation AnalysisRegression Analysis(QuantificationⅠ)Cluster AnalysisMultidimensional Scaling(Quantification Ⅳ)Quantification ⅢPrincipal Component AnalysisFactor Analysisdirectlyspatial configuration of similaritiescontingency tablevariable compositionlatent factordiscriminationcorrelationpredictionNoexternal criterionYesprediction discrimination correlationfactor analysisclusteringlatent structure analysisS4S1V2H2D2U2INDEXS1 =19.00S4 = 4.00D2 = 3.00U2 = 5.00H2 = 6.00V2 = 7.002319105a) Number of objectsb) Number of holes

元のページ