| A challenging pattern recognition problem is to develop a rotation, scale, and translation invariant recognition system without high computational expense. This dissertation attempts to develop an integrated software program to recognize monochrome images of traffic signs. The test images were recorded using a video camcorder at different times of day, season, weather, angles, and range.;The result is a unique software capability for image pattern recognition on a Personal Computer which allows Windows interaction and real-time results. New understanding of visual pattern recognition and its limitations and usage was achieved. The images were not segmented. An attempt was made to show that the features of the sign would dominate the non-stationary clutter. A truncated Gabor wavelet feature was developed to demonstrate rotation, scale, and translation invariance with bandpass wavelets using linear Fourier techniques.;An integrated workstation in a Windows environment has been used to perform image processing and pattern recognition functions. The image processing functions considered herein include the Fourier Transform, noise addition, linear and morphological filters, and Gabor feature analysis while the pattern recognition functions are template matching, statistical pattern recognition, and the multi-layer perceptron. The software design objective was to demonstrate the advantage of using self-similar Gabor wavelet features in statistical pattern matching and a multi-layer perceptron over whole image template matching. A database of 540 training images and 50 test images of signs and backgrounds supports the software. |
