Affine Invariant Object Recognition Based On Wavelet Transform

For the same 2-D object, the geometry shape of image obtained is different when the camera is located different orientation and angle. The geometry distortion of arbitrary two images can be described with perspective transform. The perspective transform can be replaced approximately with affine transform, when the distance between the camera and the object is larger than the dimension of the object. Therefore the affine object recognition is one of the important parts of computer vision, and has been widely applied in the fields such as military affairs and automatization.The imaging process of object can be approximated to affine transform in some condition, therefore in order to recognize object, we have to find affine invariants of object, it is the core of this paper. Firstly in the stage of image segmentation, on the basis of comparing three edge of detector in common use, we propose a new method of edge detection, which is combing Robert detector and contour tracing, and obtain slick and continuous single contour. After getting chain code in the stage of image segmentation, we get the coordinates of contour. Secondly two levels invariant, three levels invariant and six levels invariant have been calculate using different numbers of dyadic levels. Both experimental results and analysis show that the invariants can be easily recognize object accurately. And then the stability of these affine invariants has been tested for a large perspective transformation. Finally in order to overcome the dependence the wavelet transforms to starting-point of contour, a new algorithm of starting-point matching in recognition of 2D object contours is studied. The polynomial approximation technique is adopted in order to reduce the computation complexity. Experimental results show that the algorithm can estimate the misalignment between the starting points for the object contours effectively.