Content-based image retrieval with the normalized information distance

The main idea of content-based image retrieval (CBIR) is to search on an image’s visual content directly. Typically, features (e.g., color, shape, texture) are extracted from each image and organized into a feature vector. Retrieval is performed by image example where a query image is given as input by the user and an appropriate metric is used to find the best matches in the corresponding feature space. We attempt to bypass the feature selection step (and the metric in the corresponding feature space) by following what we believe is the logical continuation of the CBIR idea of searching visual content directly. It is based on the observation that, since ultimately, the entire visual content of an image is encoded into its raw data (i.e., the raw pixel values), in theory, it should be possible to determine image similarity based on the raw data alone. The main advantage of this approach is its simplicity in that explicit selection, extraction, and weighting of features is not needed. This work is an investigation into an image dissimilarity measure following from the theoretical foundation of the recently proposed normalized information distance (NID) [M. Li, X. Chen, X. Li, B. Ma, P. Vitányi, The similarity metric, in: Proceedings of the 14th ACM-SIAM Symposium on Discrete Algorithms, 2003, pp. 863–872]. Approximations of the Kolmogorov complexity of an image are created by using different compression methods. Using those approximations, the NID between images is calculated and used as a metric for CBIR. The compression-based approximations to Kolmogorov complexity are shown to be valid by proving that they create statistically significant dissimilarity measures by testing them against a null hypothesis of random retrieval. Furthermore, when compared against several feature-based methods, the NID approach performed surprisingly well.