Rotation and scale invariant texture segmentation applied to SAR sea ice images

The navigation of ships in cold ocean regions such as Voiseys Bay, Labrador, Canada, is influenced by the types of ice and the properties of any such ice that surrounds the ship. Ship navigation can be aided with the use of synthetic aperture radar (SAR) imagery. SAR sea ice images for the most part have distinct regular textures for each type of ice when viewed by the human eye or a machine vision system. These regular textures can be employed in sea ice description and classification. It is for this reason that approaches to identifying ice types and classifying ice properties often are dependent on texture discrimination techniques.; This thesis reinforces the ice labeling problem mainly by investigating the extraction of features vectors for unsupervised texture segmentation. Several approaches of texture segmentation are investigated. The segmentation approaches that are reviewed are the well known methods such as Gabor filtering and the Gray Level Co-occurrence Matrix. Additionally, the issue of separating texture from form information in an image based on a grating cell operator is explored. The foundation of much of the work presented here is based on Gabor filtering.; Texture orientation issues are addressed in this thesis since certain ice types may have orientation information due to environmental conditions such as the neighboring coastline, winds and sea currents. A new method based on a circular Meyer wavelet is proposed and applied to the texture segmentation problem. This method extracts rotationally invariant features since the kernel of the circular Meyer wavelet is rotationally symmetric. This new method is compared to an established method based on circular Gabor functions. Several experiments are performed to compare the performance of the methods.; Finally, the issue of scale invariance is also investigated in this thesis since certain ice types may appear in an image at different scales and the segmentation system will be degraded if it is not scale invariant. The scale invariance is achieved by Discrete Fourier Transform (DFT) encoding of rotationally invariant feature vectors. The DFT encoding of feature vectors is investigated with a circular Gabor function producing the rotationally invariant features and with the new method using the circular Meyer wavelet producing the rotationally invariant, features. The performances of the circular Gabor and the circular Meyer wavelet DFT encoded feature vectors are evaluated.