Low Level Operations and Learning in Computer Vision

This thesis presents some concepts and methods for low level computer vision and learning, with object recognition as the primary application. An efficient method for detection of local rotational symmetries in images is presented. Rotational symmetries include circle patterns, star patterns, and certain high curvature patterns. The method for detection of these patterns is based on local moments computed on a local orientation description in double angle representation, which makes the detection invariant to the sign of the local direction vectors. Some methods are also suggested to increase the selectivity of the detection method. The symmetries can serve as feature descriptors and interest points for use in hierarchical matching structures for object recognition and related problems. A view-based method for 3D object recognition and estimation of object pose from a single image is also presented. The method is based on simple feature vector matching and clustering. Local orientation regions computed at interest points are used as features for matching. The regions are computed such that they are invariant to translation, rotation, and locally invariant to scale. Each match casts a vote on a certain object pose, rotation, scale, and position, and a joint estimate is found by a clustering procedure. The method is demonstrated on a number of real images and the region features are compared with the SIFT descriptor, which is another standard region feature for the same application. Finally, a new associative network is presented which applies the channel representation for both input and output data. This representation is sparse and monopolar, and is a simple yet powerful representation of scalars and vectors. It is especially suited for representation of several values simultaneously, a property that is inherited by the network and something which is useful in many computer vision problems. The chosen representation enables us to use a simple linear model for non-linear mappings. The linear model parameters are found by solving a least squares problem with a non-negative constraint, which gives a sparse regularized solution.