An inverted index for image retrieval using colour pair feature terms

An algorithm is presented for indexing and retrieving images using an inverted index that stores colour features. The features are extracted using a modification of the Multi-modal Neighbourhood Signature (MNS). Images are divided into regular patches and modes of the colour distribution are derived using the meanshift algorithm. The colour values of patches with one, two or three dominant modes are recorded, and quantised into bins that form the colour feature terms. The terms and their frequencies are stored in an inverted index implemented in a relational database. Retrieval is performed using four different techniques, including a variation of the Term Frequency, Inverse Document Frequency (TF/IDF) algorithm used in text retrieval, that weight the query image features against those in the index. This new approach is compared to our previous work with indexed features and more traditional colour retrieval algorithms. The comparison is performed against a database of photographic images containing a wide variety of scenes. Two types of retrieval are tested - full image and sub-image queries. The performance of the algorithms are presented both in terms of computational speed and retrieval accuracy.