Synthetic discriminant function filter employing nonlinear space-domain preprocessing on bandpass-filtered images

Area of Science:

• Optics and Photonics
• Image Processing
• Pattern Recognition

Background:

• Previous work established the difference-of-Gaussians (DOG) filter for pattern recognition.
• Nonlinear preprocessing of the DOG filter demonstrated enhanced tolerance to in-class variations and noise.

Purpose of the Study:

• To synthesize a modified synthetic discriminant function (SDF) filter incorporating nonlinear preprocessing.
• To evaluate the performance of the new filter regarding discrimination, robustness, and invariance.

Main Methods:

• Incorporation of a nonlinear preprocessed difference-of-Gaussians operation into SDF filter synthesis.
• Testing the filter's performance against in-class variations, noise, background clutter, and out-of-plane rotations.

Main Results:

• The modified SDF filter produced sharp correlation peaks.
• Excellent discrimination was achieved without requiring out-of-class objects during training.
• The filter demonstrated good invariance to out-of-plane rotations up to 90 degrees.
• Nonlinear preprocessing significantly improved robustness to background clutter.

Conclusions:

• The nonlinear preprocessed SDF filter offers superior performance compared to linear filters.
• This approach enhances tolerance to variations and robustness to challenging input scenes.
• The filter is effective for pattern recognition tasks requiring high discrimination and invariance.