Aşırı öğrenme makinası topluluğuna dayalı görüntü çakıştırma

Image registration is the process that aligns a set of images obtained from various sensors, different views of points or different time in one coordinate system. Image registration with neural network using image features is a scheme that used to estimate the geometrical transformation parameters namely (scaling, rotation and translation). In this thesis, the image registration procedure was carried out using ensembled extreme learning machines to overcome the disturbance effect of the noise, improve the accuracy and robustness of the estimation of the geometrical transformation parameters under noisy conditions. In performed experiments, three different images, which are transformed using affine transformation, were used. The features were extracted from the noisy images and noise-free images by discrete cosine transform. These features were given as input to neural networks and the registration parameters were calculated at the output of the network. Image registration based on the neural network can be realized in two stages. In the pre-registration stage, extreme learning machines, radial basis function and feed-forward type of neural networks were trained by features extracted from images, which are translated, scaled and rotated with various parameters. In the registration stage, the trained networks were fed by features extracted from test images and the registration parameters were observed at the output. The obtained results were compared in terms of different used neural networks. The results showed that extreme learning machines give better result in the present of the noise than the other two neural networks.