Refereed full papers (journals, book chapters, international conferences)

1997

In image processing, the first step is detection (location of all regions of interest (ROI's) in a scene, where candidate objects may exist) with high detection probability P_D and low false alarm rate P_FA. This must be achieved when multiple objects in different classes in different distortions are present. We present several algorithms to achieve this using morphological and wavelet transform techniques. No one detection algorithm is best and thus we introduce the concept of algorithm fusion to reduce P_FA; i.e. we apply several different detection algorithms to the same scene and we nonlinearly combine the results of different algorithms using fuzzy logic concepts. Case study test results are presented. Distortion-invariant correlation filters represent an attractive technique to recognize different distorted aspect views of one objects and distinguish them from another object. We review algorithms to achieve this and present initial test results. Once ROI's have been located, they are further processed by image enhancement, features of each ROI are then extracted and fed to a classifier. We describe a new neural net (NN) classifier that uses a new feature space trajectory (FST) representation of distorted objects. We show case study distortion-invariant multi-class object recognition using this classifier and we note how it overcomes various problems with standard classifiers. Finally, we consider a product-inspection case study in Agriculture inspection. This involves segmentation of touching input objects, processing and feature extraction of each item, and classification of the class (good, minor or major defect) of each input item. This utilizes various new image processing techniques, feature extraction and use of our neural net classifiers.