| Colour is the most important low-level feature which is used to build image indices, for retrieval of images from a database. Specifically, the colour histogram remains the most popular method for building such indices, due primarily to its simplicity and fast computation. It has many drawbacks and limitations, however, which degrade its effectiveness.;In this thesis, a new framework is proposed for indexing and retrieving images by colour, using representative RGB colour vectors and a perceptually-tuned vector angular-based measure of colour similarity. The feature extraction and indexing stage incorporates a new recursive HSV-based segmentation scheme which identifies pixels based on perceptual prominence and classifies them as bright chromatic, chromatic, black and white. For each of these classes, hue histogram thresholding is performed, while taking into consideration the multi-modal nature of the saturation histogram. Post-processing operations follow the segmentation to arrive at an accurate low-level representation of the colour image. The average colour of all the extracted regions we then stored in an index, along with the added information of colour categorization, number of regions, and colour amount.;For the retrieval process, a new technique is proposed using a multidimensional query distance space, whose dimension is determined by the number of query colours. All images which exhibit similarity to the query colours occupy a position in this space and it is this position, and its relation to the origin and the equidistant line, that determines the overall similarity and ranking of a given image to a colour query. The concept proves to be quite flexible in how queries can be structured, allowing any number of query colours, query-by-example, and also effectively allows the exclusion of a specific colour in a query. An exclusion colour adds an additional component to the multidimensional query distance space which affects the overall position of a given image in the space.;Through extensive testing, it is shown that the proposed colour image retrieval scheme and measure exhibit very high performance in terms of retrieval rate, precision, and recall, especially over common colour histogram techniques and IBM's QBIC system. This was established through Human Query Sets obtained from 25 human volunteers. These sets of images represent the retrieval results containing images judged by humans to best fit a specified query and are used to analyze and compare the retrieval results of the system. Furthermore, the proposed scheme and measure exhibit high resistance to gamma nonlinearity. Retrieval results show that the retrieval rate is the highest over a wider range of gamma nonlinearity values than all other vector measures investigated, and also colour histogram techniques. |
