Intelligent Engineering Systems through Artificial Neural Networks
56 Performance Analysis of Moments in Invariant Object Classification
Download citation file:
- Ris (Zotero)
- Reference Manager
Moments have been widely used for creating invariant features for image classification and object recognition problems. In this study, different moments that are extracted from a database of 1200 images are used for object classification. The images are obtained from 10 different objects, each of which had 120 images that are rotated in different angles with different lighting conditions. For performance analysis, geometric invariant moments, Zernike moments, Pseudo- Zernike moments, Tchebichef moments and statistical features of the objects are extracted from each image. 800 of the images (80 images from each object) were used in training and the remaining 400 images (40 images from each object) were used for performance testing. In the classification step Nearest Neighbor (NN) and Bayesian classifiers are used. The results from each moment are compared. Overall, Pseudo- Zernike moments showed the best classification results even though other moments were also not far behind. Bayesian classifier outperformed NN classifier independent of the selected features. The preliminary results indicate these moments can be used in real world applications where object recognition with 2-D invariance and different lighting conditions are needed.