Clustered Gauss mixture models for image retrieval

With the rapid advances in multimedia technology, on-line databases of digital images are assuming an increasing role. Effective access to such image archives requires that conventional query techniques based on textually annotated keywords be replaced by queries based upon content such as the visual features of searched images. What is central to this retrieval into some space of features and support indexing and comparison of visual contents. Depending on the characteristics of images in the database, such models can rely on different features: color and texture distribution, shape of appearing objects, spatial arrangement, and Gauss mixture model parameters.; Two novel approaches are proposed and studied experimentally on a small image database of 500 images, consisting of 50 types or classes. Precision and recall are compared with that of other popular features, such as color histograms.; First, a universal codebook containing 64 Gauss components is trained using Lloyd clustering with a minimum discrimination information (MDT) distortion measure. The underlying distribution of an image is estimated as a Gauss mixture and the mixture parameters are calculated by applying the universal codebook to the image. These model parameters (or probability mass functions of codewords) form the feature vector that represents the image. Supervised learning using a classification and regression tree is conducted to classify and index the images.; Second, a set of class codebooks, each of which is trained for a separate class of images, is used to better approximate or estimate the underlying density of images in the same class. Again Lloyd clustering with MDT distortion constructs the codebooks and minimizes the number of Gauss mixture components required for each class. At querying, the codebook that yields the minimal MDT distortion is selected as the basis of similarity comparison. The corresponding Gauss mixture model parameters of images are classified to find similar images that should be retrieved.; Experimental results show Lloyd clustered Gauss mixtures—called Gauss mixture vector quantizers—using either a universal codebook or separate class codebooks yield better precision and recall than features such as color histograms, color layouts and textures.