A view-based approach to three dimensional object recognition

To improve the performance of three-dimensional object recognition systems, we propose a view-based method in this study. First we extract wavelet moments, texture features and color moments from the 2D view images of 3D objects. Wavelet moments have the multi-resolution properties in addition to the invariant properties under translation, scaling and rotation. Texture features can distinguish objects which have similar shapes and different appearance. Color moments are robust and insensitive to the size and pose of objects. Support Vector Machine (SVM) is chosen as classifier. Then the feature subset selection and SVM parameters optimization are accomplished automatically and simultaneously using Genetic Algorithm (GA) in an evolutionary way. We assessed our method based on the original and noise corrupted 3D object dataset COIL-100. One hundred percent correct rate of recognition was obtained when the number of presented training views for each object was 36 (10 degrees interval) and 18 (20 degrees interval). When the number of training views was reduced, the correct rale of recognition was also satisfied. © 2009 Asian Network for Scientific Information.