Flower classification: Training augmentation using manifold images

We investigate the classification problem of visually similar objects, like flowers. We propose to substitute the lower number of original training images with their large number of counterparts, artificially generated by manifold mapping. Since, the inter-class similarity is very high in visually similar object classification problems; we try to imitate the test images in the low dimensional space with this large number of manifold images. Until now, we have compared our scheme with other methods only for the bag-of-words color features on Oxford 17 class flower dataset. For this single feature, with very low resolution (64×64) manifold images, we almost reach the same accuracy as that obtained with original resolution images (both row and column have dimensions more than 500) in the state-of-the-art literatures.