Image registration using artificial neural networks

Image registration is a procedure that transforms different sets of data that are multiple photographs or data from different sensors, from different times, or from different viewpoints into one coordinate system. In this thesis affine transform is chosen as the transform model.In this thesis image registration process is done using neural networks in the presence of noise. In the applications, we had three images and affine transform was applied all of the images. At first, features were extracted from the images and these features were given to the network as inputs, then estimated parameters were obtained at the output. These features were extracted by the methods of discrete cosine transform (DCT) and two dimensional principal component analysis (2DPCA). In a pre-registration phase, extracted features from a set of translated, rotated and scaled images of the same scene are employed to train both a Radial Basis Function Neural Network (RBF NN) and a Feed-forward Neural Network (FNN). In the registration phase, the features are extracted from the test image and these features are given to the network. By this way, registration parameters are obtained.The results were compared both according to the different feature extraction methods and different type of neural networks.