動画像分類のための微分積和型特徴抽出演算子

抄録

In this paper, nonlinear spatiotemporal image operators for extracting motion-like brightness changes and emergence-like brightness changes are proposed. First, we show that a null determinant of a 3-D gradient covariance matrix indicates the image brightness variations are caused purely by motion. Then, in order to decide its nullity independently on spatiotemporal contrast and resolution, we derive a dimensionless normalized equation from the determinant based on an inequality between a geometrical and an algebraic average. After adding noise suppressing features to the equation, we obtain three feature extraction operators which respond locally to 1) motion-like variations, 2) emergence-like variations, and 3) total variations. Basicly, they function as region operators to discriminate stationary, moving, and emerging segments in an image. By a theoretical analysis, however, the emergence operator is shown to act as a motion edge operator which respond in selective and shape/orientation-invariant manners to velocity discontinuities at e.g. occluding boundaries. The above mentioned unique characteristics of the operators are tested by several computer simulations and an application to a dynamical 3-D scene.