Invariant image recognition using projections
Many advanced industrial applications for machine vision systems call for vision algorithms that can effectively deal with the numerous dimensions of variability that complicate the interpretation of a 2-D scene. A new technique for recognition of 2-D gray-scale images irrespective of their variable appearance as affected by a priori unknown position, orientation, and size is suggested and investigated in this dissertation. The method uses a combination of Radon, Fourier, and Mellin transforms to produce an "invariant pattern representations" (IPR) of the image, and hence copes with variability in the image formation processes without using contextual knowledge. Using the Radon transform (projections), the 2-D recognition problem is reduced to a set of one-dimensional problems which require only 1-D transformations, and thus the method has the potential of dramatically reducing the computational complexity of the required transformations. Further reduction in computations will be possible by processing only the two orthogonal projections of the image. It is shown that the Fourier-Mellin (FM) transform based on Fourier magnitude for removing delay and scale effects in signals may not be suitable for classification purposes. A novel invariant transformation called Fourier phase-Mellin (F$\Phi$M) transform is developed for the classification of signals degraded by unknown amounts of delay and scale factors. A comparison is made between F$\Phi$M and FM transforms and it is shown that F$\Phi$M performance is superior to FM in producing invariant representations when classification is of concern. Methods for determining the deformation parameters are presented and applied to test images to investigate their effectiveness. Simulations are provided to demonstrate the performance of the recognition schemes developed here. ^
Engineering, Electronics and Electrical
Majid Moztarebi Farahani,
"Invariant image recognition using projections"
Dissertations and Master's Theses (Campus Access).