Colour-based image retrieval algorithms based on compact colour descriptors and dominant colour-based indexing methods

Content based image retrieval (CBIR) is reported as one of the most active research
areas in the last two decades, but it is still young. Three CBIR’s performance problem in this study is inaccuracy of image retrieval, high complexity of feature extraction, and degradation of image retrieval after database indexing. This situation led to discrepancies to be applied on limited-resources devices (such as mobile devices). Therefore, the main objective of this thesis is to improve performance of CBIR. Images’ Dominant Colours (DCs) is selected as the key contributor for this purpose due to its compact property and its compatibility with the human visual system. Semantic image retrieval is proposed to solve retrieval inaccuracy problem by concentrating on the images’ objects. The effect of image background is reduced to provide more focus on the object by setting weights to the object and the background DCs. The accuracy improvement ratio is raised up to 50% over the compared methods. Weighting DCs framework is proposed to generalize this technique where it is demonstrated by applying it on many colour descriptors. For reducing high complexity of colour Correlogram in terms of computations and
memory space, compact representation of Correlogram is proposed. Additionally, similarity measure of an existing DC-based Correlogram is adapted to improve its accuracy. Both methods are incorporated to produce promising colour descriptor in terms of time and memory space complexity. As a result, the accuracy is increased up to 30% over the existing methods and the memory space is decreased to less than 10% of its original space. Converting the abundance of colours into a few DCs framework is proposed to generalize DCs concept. In addition, two DC-based
indexing techniques are proposed to overcome time problem, by using RGB and perceptual LUV colour spaces. Both methods reduce the search space to less than 25% of the database size with preserving the same accuracy.