FUZZY METHODS FOR OBJECT RECOGNITION

Abstract

The objective of this thesis is to propose fuzzy methods for object recognition. By integrating the pyramid structure and fuzzy set theory, we develop a fuzzy pyramid scheme within which we can achieve automatic object recognition invariant to scale, translation, rotation and distortions. There are five key components in our method. Firstly, we use annular and sector windows for image segmentation. It is shown that it is considerably easier for the annular and sector windows to maintain the invariance properties compared to conventional square windows. Secondly, the extracted features are localized global features that are fuzzy in nature. We choose the zeroth-order moment as a global feature and localize it in the annular and sector windows. By doing this the computational complexity required to extract the features is replaced with spatial segmentation complexity. The localized features are then fuzzified by interpreting them as membership grades in the segmentation windows. Thirdly, extracted features are selected based on the concept of fuzzy entropy. The key step of this algorithm is to segment the annular and sector regions dynamically by calculating the fuzzy entropy of each segmented region. Only one feature is extracted from one such region. The fuzzy entropy of this feature is used as a measure of its fuzziness. The reduction of the total number of the segmented regions based on the idea of dynamic region segmentation results in considerable reduction in feature dimension. Fourthly, the selected features are used to build a fuzzy feature pyramid. This pyramid will be used to make the final recognition decision using a fuzzy decision algorithm. By involving the intra-level and the inter-level weights in the discriminant function1 each feature affects the final decision. Finally, both intra-level and inter-level weights and relevant parameters are obtained by using a new adaptive scheme based on fuzzy entropy. Such an arrangement has resulted in a general purpose technique that is capable of delivering both the desirable recognition invariance and the robustness against noise, distor­ tion and incompleteness of the input image. The new method has been effectively tested for the problem of recognizing a set of tools, maps and characters. In addition, an automated fuzzy segmentation method consisting of crisp seg­ mentation and fuzzy segmentation stages is developed. Multilevel thresholding is used to obtain an initial set of pixels which are then extended by employing fuzzy rules on local properties. The method is tested by applying it to the detection of the glomerular basement membrane in kidney electron micrographs.