Conclusion

Patch-based image classification with image epitomes showed some success at correctly classifying test images. It performed well on positive example identification on the face dataset and negative example identification on the beach dataset. But there are clearly problems with the technique, as it performed quite poorly on positive example identification on the beach dataset and somewhat poorly on negative example identification on the face dataset. Many of the problems with the approach seem to stem from the somewhat ``brittle'' nature of the patch-based approach. Making the epitome model more flexible by allowing varying patch sizes and transformational mappings may improve the performance considerably. The epitome representation captures textural and shape regularities present in image data, and the results of my experiments suggest that it may be possible to successfully exploit this property to find features that can be effectively used for image classification.