Position, Scale, and Rotation Invariant Optical Pattern Recognition for Target Extraction and Identification

摘要

This thesis investigates the feasibility of optically implementing a real-time, pattern recognition system using correlation techniques in a position, scale, and rotation invariant (PSRI) feature space. Input target templates were optically Fourier transformed using an improved high-resolution, high-pass filter positioned in the focal plane of the transforming lens. A logarithmic-polar coordinate transform of the magnitude-squared of the Fourier transform (/FT/2) was performed with an improved method of focusing the magnified /FT/2 onto a computer generated hologram (CGH), which was shown to scale linearly along the horizontal axis and logarithmically along the vertical axis. Optical, matched-filter correlations on the magnitude-squared Fourier transform logarithmic-polar (FLRT) feature space were performed using thermoplastic, phase-relief holography and Vander Lugt filtering. Correlation results prove that scale and rotation changes of the input can be predicted accurately based on linear shifts of the correlation peak. Also, the FLRT feature space is shown to provide excellent discrimination for multiple-input scenes. The need for a cyclic correlation is verified, and digital simulations prove the validity of the optical results. (RH)