Vol.4 No.2 2011
10/66

Research paper : ARGUS: Adaptive Recognition for General Use System (N. Otsu et al.)−81−Synthesiology - English edition Vol.4 No.2 (2011) Fig. 18 Example of application to cancer detection (The upper figure shows abnormal values for each specimen.)[18]Fig. 19 Correspondence learning[24][25]distributed in a linear subspace (usual motion subspace) SN in the feature space (with 251 dimensions). Accordingly, once SN has been derived using principal component analysis (PCA) while learning on a regular basis (unsupervised), abnormal behavior does not require prior definitions, and can be detected and recognized immediately with high speed and accuracy, as a deviation (in distance or angle) from SN [17] (Fig.15). Because of its additivity, capability for detecting abnormalities remains constant even for multiple persons (Fig.16).This abnormality detection system is already put into practice with surveillance cameras in elevatorsNote6.This system, where usual cases are learned as a statistical distribution in the CHLAC feature space and abnormalities are detected as deviations from such distribution (unusual), does not necessitate any model or knowledge of the objects. Accordingly, the system can be applied not only for abnormality detection from footage taken by surveillance and car-mounted cameras, but also for various other abnormality detection scenarios. For example, using the HLAC feature space for still images, it can be applied to various appearanceinspections such as in the field of manufacturing semiconductor substrates (Fig.17).Moreover, abnormality detection using HLAC is equally applicable in the medical field for various kinds of tissue examinations, especially in the pathological diagnosis of cancer. Cancer is a cell abnormality. The pathological diagnosis of cancer is conducted under a microscope by a pathologist who determines the degree of change found in the structure and the cells of organ tissues. However, this requires a wealth of experience and knowledge, and experienced pathologists are in short supply with their ever-increasing workload. Thus, there is great demand for systemdevelopment to support pathologists, in the form of alleviating the burden of screening tests through automation,and preventing oversights through crosschecking.When this method was applied to actual lymph node metastasis in stomach cancer, it was possible to obtain analysis results that were close to those obtained by a experienced pathologist[18](Fig.18). Currently, we are collaborating with university hospitals and cancer centers with the goal of setting up a support system for pathological diagnosis.5.7 Time series data analysisIn general, sensing data, not limited to images, is represented by N-dimensional (Ch.) time series data, {si (t)}Ni=1, t = 1, . . . , M. Although it is possible to consider these as an N × M two-dimensional matrix(image) and extract HLAC features, the order of the dimension (Ch.) subscript i is generally optional. Accordingly,if for example a combination of any three is taken arbitrarily, this gives Normal 1Normal 2Cancer 1Cancer 2201510501101201301401Fig. 16 Deviation from the subspace of usual motionsOrtho-complementsubspaceAbnormalityvalueSubspace of normal movementsFig. 17 Example of application to substrate inspectionInspection windowIC substrate(Original image)Result of a conventional method using distance from the average sub-image×Result of the proposed method using HLAC features

元のページ 

page 10

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