Discovering Robustness Amongst CBIR Features

Digital photography faces the challenges of image storage, retrieval and provenance at the consumer and commercial level. One major obstacle is in the computational cost of image processing. Solutions range from using high-throughput computing systems to automatic image annotation. Consumers can not dedicate computing systems to image processing and handling nor do consumers have large-scale image repositories to make automatic image annotation effective. Nevertheless, we consider an alternative approach: reducing computational cost in image processing. Using a 25,000 image collection, we consider using a subset of image features to evaluate image similarity. We discover several robust features displaying comparable relevancy performance with the additional benefit of reduced processing cost.