ADAC System Design and Its Application to Mine Hunting Using SAS Imagery

This PhD thesis considers the problem of automatic detection and classification. On the one hand, a set of application independent design issues for classification is tackled. On the other hand, the specific application of underwater mine hunting using synthetic aperture sonar imagery is considered. A novel resampling method is proposed in order to solve two fundamental issues involved in the design of classification systems, namely, the selection of the classifier and the estimation of the optimal feature set dimensionality. The method estimates both the probability distribution of the misclassification rate (or any other figure of merit of the classification system) subject to the size of the feature set and the probability distribution of the optimal dimensionality given a misclassification rate. The latter allows for the estimation of confidence intervals for the optimal feature set size. Unlike previous methods, no assumption for the features distribution is required. Based on the probability distribution of the figure of merit, a quality assessment for classifier performance is proposed. By contrast with previous works, the proposed algorithm allows to compare different classifiers without bonds to a specific feature set. In addition, the problem of determining the optimal feature subset is considered. In this respect, novel extensions of the Sequential Forward Selection and Sequential Forward Floating Selection methods are proposed. It alleviates the limitations of the methods, yielding a better performance. A system for automatic detection and classification of underwater objects using synthetic aperture sonar imagery is developed within this design framework. It consists of three steps: detection, feature extraction and classification. In order to detect the objects in the sonar images, three segmentation algorithms are compared: iterative conditional modes, min-cut/max-flow and active contours. Novel initialization schemes addressing the application at hand are proposed, since they significantly influences the final result. An extensive set of features is extracted for each object, both geometrical and statistical. They are designed to remain invariant to changes in the object position and also in poor segmentation scenarios. The selection of the optimal feature subset is accomplished by the extended feature selection algorithms, only after the resampling method has determined the best out of four classifier candidates (k-nearest neighbor, Mahalanobis’, linear discriminant analysis and support vector machines). The proposed methods have been applied to two databases of real sonar images containing over 57,000 m2 of underwater images and 600 mines of different types