Orientation intégrale multi-échelle
Asim Imdad Wagan · HAL (Le Centre pour la Communication Scientifique Directe) · 2010
This thesis examines and proposes a new set of image region descriptors based on the multiscale localized orientation histograms. The orientation histograms have been proven useful in many image analysis tasks and they have been quite extensively studied in the literature. In this these we have proposed multiscale integral orientations which are a fast multiscale version of the localized orientation histograms. We have also tested the MSIO descriptor on many different image analysis problems and shown the accuracy of this descriptor. In the case of the document image analysis we have tackled with problem of word spotting and writer classification as the test problems for the MSIO descriptor. The results have shown that the feature provide good results and warrant an excellent study on other problems related in the domain of the document image analysis. In the domain of natural image analysis we have tackled the problems of face detection and recognition using the MSIO descriptor. The results there also show the effectiveness of this descriptor. lastly, we have tested the MSIO descriptor to match 3D models based on their shape similarity with an accuracy of 95%. The results have shown the promising aspect of the MSIO descriptor and we hope to use these descriptors for more work in the image analysis domain.