Nonlinear correlation filter and morphology neural networks for image pattern and automatic target recognition
Yonggwan Won, Paul Gader · 1995
The most important property of a pattern recognition system is it's generalization capability. Previous research shows that neural networks generalize well and approximate arbitrarily complex functions. Feature extraction and decision-making are both important components. Unfortunately, designing effective feature extraction procedures is a very difficult task. For this reason, a heterogeneous neural network that can learn feature extraction and classification simultaneously is very attractive. The nonlinear correlation filter neural networks (NCFNN) learn correlation filters in the frequency domain and simultaneously learn a nonlinear combination of the outputs of the correlation filters for object detection. Two morphological neural networks, the generalized-mean morphology neural network (GMMNN) and the ordinary morphology neural network (OMNN), learn morphological structuring elements as feature extractors simultaneously with classification. They provide a general problem-independent methodology for designing morphological structuring elements. They perform feature extraction using a novel gray-scale Hit-Miss transform. The OMNN is invariant to shifts in gray-scale. The GMMNN and the OMNN were applied to pattern recognition problems. For binary handwritten digits, they produced performance comparable to that obtained using the ordinary linear shared-weight neural network (LSNN and NBLSNN) that has been used by others previously. However, the OMNN trained faster. For gray-scale patterns, the morphological neural networks produced superior performance. The LSNN, the OMNN and the NCFNN were applied to automatic target recognition (ATR) problems. Two data sets were used: Forward Looking Infrared image scenes of tanks and visual image scenes of a parking lot containing occluded vehicles. Several performance measurements and target-aim-point selection algorithms were defined. The OMNN performed significantly better than the other; especially at detecting occluded vehicles and reducing false alarm rates. All the networks performed significantly better than a Minimum-Average-Correlation-Energy filter technique. The new neural networks provide a general methodology for designing nonlinear filters without specific knowledge of the problem domain. They performed better than existing shared-weight network and matched filter approaches. In particular, the OMNN produced the best performance. It trained relatively quickly and is independent of shifts in gray-level. The NCFNN, the LSNN and the NBLSNN all produced similar performance rates. These networks are not problem-specific and can be widely used for other pattern recognition and ATR problems.