Chai Recognition between a Large Number of Flower Species

The goal of this thesis is to design and build a system which automatically classifies an image of a flower for hundreds of flower species. The classification should be done within a reasonable time so that it is usable for a real-time computer vision application. The classification performance of the system is improved if only the flower region (foreground) of the image is considered, and the background region is ignored. To this end we present a superpixel-based flower foreground segmentation method, which has a processing speed 10 times faster than state-of-art flower segmentation algorithms, while maintaining the segmentation quality as evaluated on the 849 images of the Oxford Flower 17 Data Set. For the actual classification task, we use a standard linear support-vector-machine in combination with the Bag-of-Words representation for the features. We investigate several different descriptors, kernels and heuristics. By combining our new segmentation and improved recognition methods, we achieve a class-average recognition accuracy of 81.4% over the state-of-art 76.3% on the Oxford Flower 102 Data Set. In the final part of the thesis, we propose a simple, scalable and unsupervised co-segmentation method (to simultaneously segment a set of related images into foreground/background) called BiCoS. This is able to automatically produce training segmentations for our superpixel-based foreground segmentation. The algorithm outperforms several recent co-segmentation methods on both Oxford Flower data sets. It also gives competitive results on other benchmark data sets, such as the Weizmann horses and Caltech-101 categories.