Место издания:Saint-Petersburg University Publ Saint-Petersburg
Первая страница:41
Последняя страница:41
Аннотация:Methods, algorithms and computational programs of natural and anthropogenic objects recognition using airborne images of high spectral and spatial resolution are elaborated within the created apparatus and programmatic system. The apparatus part is based on the registration of the hyper-spectral images which are formed by hundreds of spectral channels in visible and near infrared region. The recognition is based on the joint use of spectral and texture features of the observed objects while employing the supervising procedures of a selected classifier with test sampling. A step up method is applied for the channels optimization using data under processing for the analysis of the objective function characterizing the accuracy of the objects recognition by the spectral features for the particular classifier. The high spatial resolution of the hyper-spectral images demands understanding contextual information (the context is characterized by the influence of neighboring pixels, for example, the forest vegetation object is related to a particular class unless a boundary with any other object appears, the account of the context leads to an enhancement of the accuracy of such recognition). A new approach is accounted for while considering sites and labels in the objects pattern recognition procedures. A mathematical formalism is used of finding the maximum posteriori probability for Markov random fields to have in mind the influence of the neighboring pixels for a given class of the objects on the processed images. Standard procedures of smoothness of any distorted image due to the instrument noise and other effects are considered together with the segmentation procedure of separate classes, i.e. the smoothness only inside the selected contours (“spatial regions”) of the objects. Examples are also demonstrated of the perceptual grouping appearance where the test sites with the segmenting features (points, lines, regions) are irregular distributed and the matching problem of these features has emerged. The neighborhood system has just become essential for the pixels of a particular class and the cliques between these pixels as a measure of their inter-relations. Results are shown of the related applications using the proposed original system of hyper-spectral data processing.