Computer-aided detection and classification of minelike objects using template-based features

In this paper the classification of small sidescan sonar images is considered. The features of the image that are used for classification are the outputs from various convolution masks based on ray-generated images of targets and their corresponding edge masks at various aspects and ranges. This procedure has the advantage that a detailed segmentation of the image into highlight and shadow is avoided. The procedure is efficient to implement and robust. In this paper we concentrate on the identification of cylindrical objects. This case is more complicated than some other targets because of the aspect dependence. A set of sidescan sonar images of a specific type of cylinder, as well as images from a large variety of clutter events is considered. A subset of the cylinder images is used for training the algorithm.