The design and use of unconstrained image filters and features for SAR detection and recognition

We consider the problem of detection and recognition of objects in Synthetic Aperture Radar (SAR) image scenes. Detection is the location of candidate regions of interest (ROIs) that contain targets in a scene; this must be done for multiple classes of objects with distortions present and in the presence of clutter. Recognition is determining the class identity of any object in an ROI. Fast and efficient algorithms (few operations per pixel) are needed for detection, since these will be applied to large image scenes. The recognition algorithms could be more computationally intensive (more operations per pixel) since they are applied to only a small number of detected regions of interest (ROIs). In this thesis, we present new algorithms for object detection and recognition in synthetic aperture radar (SAR) data.; Correlation is an attractive parallel and shift-invariant approach to detect known objects (ROIs) in cluttered scenes. However, as the number of objects and object distortions possible in a scene increases, the number of correlation operations grow if simple matched filters are used for each object distortion. Our work addresses new distortion-invariant filters that can locate a wide variety of object classes with aspect distortions present.; For detection, we consider distortion-invariant filters which also give a controlled spatially-localized correlation response for objects of interest. We design two such new eigenimage-based filters for object detection. Both filters are unconstrained (they do not have fixed correlation peak constraints) and eigen-based (they are linear combination of preprocessed eigendata). In the first approach, we separate the object classes into two macro-classes (half the objects in each class) and design pairs of filters that give large average differences between the outputs for the two macro-classes. A filter that can discriminate between two similar object classes is also expected to reject clutter. We refer to these filters as eigen-detection filter pairs. In the second approach addressed, the combination coefficients are chosen to reduce the variance of the output correlation peaks for the training set about a specified value. Rejection of clutter is achieved by selecting proper frequency domain preprocessing. We refer to these as eigen-detection filters.; For recognition, we use features which are versions of the eigen-detection filter pairs without the pre-processing (for one class versus all other classes); they are still unconstrained and eigen-based. We analyze the detection filter outputs to determine the central point to use for each training image and for the test object in an ROI. The combination coefficients for these recognition eigenimages are chosen to separate each class from all other classes. These eigen-recognition features are the set of features we use to identify each class. All such features are combined in a feature space trajectory classifier.; We present initial results and conclusions based on our study of six objects in a synthetic aperture radar military object database.