Similarity of Hybrid Object Representations With Applications in Object Recognition and Classification

Abstract: Similarity measures between images that are robust to noise and other kinds ofdistortion, while sensitive to transformations in a smooth and stable way, are of great importance in many image analysis problems. In this thesis a family of measures basedon fuzzy set theory which combine shape and intensity, is extended to vector-valued fuzzy sets for hybrid object representations such as intensity and gradient magnitudeas well as multi-spectral images such as color images. Several novel distance measures are proposed, discussed with regards to theoretical and practical properties, and evaluated empirically on both synthetic images and real-life object recognition and classification tasks. Performance metrics, such as number of local minima and size of catchment basin, which are important for distance-based local search techniques are evaluated for varying degrees of distortion by additive noise and number of discrete membership levels. The proposed distance measures are shown to enable utilizationof information-rich object representations and to outperform distance measures between scalar-valued fuzzy sets on various object detection and classification tasks.