※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

7/66
Research paper : ARGUS: Adaptive Recognition for General Use System (N. Otsu et al.)−78−Synthesiology - English edition Vol.4 No.2 (2011) Fig. 6 Local 3 × 3 masks up to the 2nd order [3][4]Fig. 5 Adaptive Recognition for General-Use System (ARGUS)4 Adaptive Recognition for General-Use System(ARGUS)An “adaptive image recognition system for general use”Note 1 has been devised that meets these basic demands and is implemented using the simplest form for the aforementioned pattern recognition, especially for the framework of feature extraction theory[3][4]. This system comprises a two-stage feature extraction following the theoretical framework of feature extraction (Fig.5).4.1 Invariant feature extraction (HLAC/CHLAC)The most basic features were considered to be parallel shift-invariant (position-invariant), for an initial feature in the first stage, namely, feature extraction that is invariant from a geometrical aspect. This is because recognition is essentially independent of the position r of the spatio-temporal pattern f(r). As position-invariant features, the auto-correlation function r() = f (t) f( t + ) dt in the field of time series analysis of audio, has been known for a long time. This extracts the relative relationship (correlation) in a wave profile pattern that does not depend on time position. The higher-order expansion of this, Nth order auto-correlation function is known mathematically, x(a1, , , aN) = f (r) f (r+a1) ・ ・ ・ f(r+aN) dr (3)and the 2nd order (generally even-order) autocorrelation functions form a complete system. As such, several interesting properties pertaining to pattern recognition applications, have been discussed[9]. For an image, f(r) is the gray-scale value at the reference point (image pixel) r, and ai is the relative displacement around the reference point r. However, the number of feature values becomes exponentially large according to the combination of N displacements, and their computation is almost impossible. Thus, combinations of limited orders and displacements are used in a practical application.In fact, patterns in the real world are spatio-temporally localized and the local relative relationships are essencial. Moreover, this localization also satisfies frame-additivity (R2). Therefore, as nonlinear features that satisfy both R1 and R2, Higher-order (Nth order) Local Auto-Correlation (HLAC) features, obtained from the higher-order auto-correlation function in Eq. (3) limited to a localized displacement, were devised and adopted[3][4].HLAC: For an actual 2D image (a still image) f(x, y), restricting the order to 2nd order and displacements to a local 3×3 region, there are 25 patterns of local masks for taking inequivalent and independent sum of products when considering the shift variance (Fig.6). For the full screen (or a sub-region) XY, scanning each of the local masks shown in Fig.6 and finding the sum of products of the pixel values corresponding to black dots gives the HLAC feature vector x. Its dimension is 35 for a gray-scale image (e.g., for mask No. 1, distinguishes f , f 2, f 3), and for a binary image (0/1, white/black), it degenerates to 25 dimensions (e.g., for mask No. 1 yields f = f 2 = f 3 as idempotent)Note2.CHLAC: In the case of a moving image (3D) f(x, y, t), since there are three-dimensional (solid) numerical data over XYT formed from the two-dimensional still images lined up along the time axis, features are extracted for CHLAC (Cubic HLAC), which naturally expanded HLAC by including the time axis[11]. Fig.7 shows an example of a local 3×3×3 mask for CHLAC. There are 251 independent local mask patterns. As with HLAC, the CHLAC feature vector x(t) is obained by finding the sum of products using Input imageWeight ARecog.Enum.Learning from examplesVarious MDAmethods(PCA,DA,RA,ARA,CCA,...)(nonlinear)(linear is enough)AnswerHLAC/CHLACMDA=1m1mn111nm1nInitial featureNewfeaturexxxxyyyyaaaaANo.5No.4No.3No.2No.1No.9No.8No.7No.6No.13No.12No.11No.10No.17No.16No.15No.14No.21No.20No.19No.18No.25No.24No.23No.22Fig. 7 Example of CHLAC mask (hr’b”) [5][11]YXTttt+2t+1

元のページ