Unsupervised and supervised fuzzy neural network architecture, with applications in machine vision fuzzy object recognition and inspection

Scope of study. In this research, an unsupervised fuzzy neural network for fuzzy patterns, termed FUZART, based on Adaptive Resonance Theory networks has been proposed. FUZART employs the two stages of self-organization and decision making. It accepts fuzzy input patterns and provides output decisions in terms of membership values. As an extension of FUZART, a new supervised fuzzy neural network scheme called FUZAMP has been developed that can quickly and efficiently handle hybrid mixtures of fuzzy data and numerical data. This fuzzy neural network can be applied to classification problems with non-linearly separable fuzzy data, and can also be employed as a fuzzy inference engine using linguistic knowledge described by fuzzy rules and numerical data sampled by measurement instruments. In order to implement this work in a real vision system, a multilayer multi-input, multi-output fuzzy logic controller (FLC) has been proposed and implemented to realize automatic adjustment of the camera parameters "gain" and "offset" to compensate for power fluctuation, changes in ambient light, and camera sensitivity drift. The multilayer FLC yields faster response with less overshoot than that of a conventional single layer FLC, and provides excellent camera performance.; Findings and conclusions. The new unsupervised and supervised fuzzy neural networks have been evaluated by simulations and real machine vision applications. FUZART has the ability to learn on-line using only a few training epochs and to provide reasonable clustering decisions for fuzzy patterns. FUZAMP has superior fuzzy classification and fuzzy inference capability and stability with fuzzy data. The advantages of FUZAMP compared with other fuzzy neural networks are that FUZAMP can realize faster and more efficient training for fuzzy data and achieve better performances. FUZAMP has been used to deal with situations where the available training data from a machine vision system includes uncertainty. It performs well when used to recognize different types of fuzzy objects presented at different locations and orientations in the camera Field of View. In addition, FUZAMP has been implemented to correlate human evaluations with machine evaluations of the cleanliness of dishes. Results are compared to those obtained using the so-called fuzzy ARTMAP neural network, with FUZAMP achieving better accuracy than the fuzzy ARTMAP using the same training exemplars.