This paper presents a model-based, filter response algorithm for Automatic Target Recognition (ATR) in Sidescan Sonar (SSS). The filters are created from the projection of a large set of generating boxes. The first and second-order statistics over the highlight and shadow regions of the projected boxes form a feature vector for each pixel in the image. The algorithm selects the features which best describe the target object, during training. When the algorithm is applied to an image, the projection of the generating boxes under rotation and translation is used to approximate invariant regions of the object under the equivalent transformation. The performance of the algorithm is compared to the standard Haar cascade. It is shown that the algorithm presented in this paper has a reduced dependence on the image formation model, requires a lower number of features to train and matches the performance of the Haar cascade. Operationally this brings two key advantages. A single classifier can be trained on data from several different models of sonar without a significant loss in performance. Additionally, the algorithm uses a smaller feature vector, reducing the time required to train the algorithm from several days to under one hour. This makes it possible to train the algorithm in-situ, increasing robustness with respect to new sea-floor types and environments.