Image-to-image registration by matching area features using Fourier descriptors and neural networks

An automatic method of image-to-image registration, which may be applied to register overlapped images of a scene from different views and dates, is presented. The proposed approach is a feature-based matching with constraints of orientation consistency and one-to-one match. Area features of homogeneous regions in gray level are extracted from images as matching entities. The boundaries of area features are matched in the frequency domain, i.e., matching their Fourier descriptors. The spatial meaning of the matching is transforming a feature to fit the other optimally. After the matching process, this scheme provides not only a quantitative evaluation of the remaining lack of fitness as an objective index of shape similarity, but also the solved transformation parameters to represent the relative orientation between features. The evaluation of shape similarity is used as the key information in recognizing conjugate features for overlapped images. Furthermore, the consistency of relative orientation between matched pairs is considered as the principal constraint to dissolve improperly registered features. The registration procedure is implemented by combining the factor of shape similarity as well as the constraints of orientation consistency and one-to-one match into a cost function and driving the cost function to reach its lowest value by using an artificial neural network system. The lowest cost represents the optimal solution of matching conjugate features. The planar registration of images then can be solved by using marched conjugate features. Two application examples, registering a stereo pair of aerial images and mosaicking overlapped images for automatic aerial triangulation, presented to show the success of this method.