Özet:

Classification of image groups has different applications in remote sensing, photogrammetry, digital painting catalogues and security related areas. Researchers from remote sensing, computer vision and photogrammetry have used different approaches based on image features (color, textures, object shapes) in order to develop robust classification methods. In this study, classification of viewpoint independent image groups is developed based on the principle of invariant properties of conic sections under projective transformation. Since invariant signatures are high dimensional data, Support Vector Machines which is a stochastic pattern recognition algorithm is used instead of classical deterministic methods. Performance of classification is evaluated using ROC (Receiver Operating Characteristics) analysis.At the beginning of this study a data set is created for testing the method. This dataset consists of nine categories of images, for each category ten images used that are taken from different viewpoints. Edge detection is applied on each image to detect boundary of objects in images. Detected edges are used for conic fitting so that each conic will be represented as a set of conic sections. Under projective transformation conic sections remain as conic section even if their shapes change. For each image in the dataset an invariant signature is computed using set of conic sections. It is assumed that there is similarity between invariant signatures of images belong to the same image category. These invariant signatures are used in histogram form for visual representation and computations.

Anahtar Kelimeler: