Parts-based model for image search and retrieval

Abstract

The project focuses on developing a pattern recognition model that would be useful in the area of image search and retrieval as well as object recognition. The model is based on object or image parts and their spatial relationships, and the parts are abstracted as ellipses. The queries for this model are therefore visual keywords or sketches composed of ellipses. For the purpose of this study the decomposition of the image into ellipses is carried out manually and a dataset of various objects is created. A number of images are each manually decomposed into an elliptical structure, describing the original image leading to the creation of a dataset. The model developed is based on graph operations that are useful in modelling the similarity between images. Gaussian noise is added to the decompositions to model errors inherent in machine decompositions. The model gave satisfactory results, where some classes work effi ciently well, such as cars, which have a general recognition rate of over 95%, while giving signi cantly low results for houses, due to the fact that buildings are rectangular in shape. Intra-class similarity is observed among intermediary cases; images of the cats and rabbits, which are easily misinterpreted as one another, under the considered decomposition. It is not uncommon to have a high level of intra-class similarity. However in general decomposition into ellipses yields a low complexity model, but results in decreased discrimination between some classes. The used of superquadrics or the addition of other shapes, such as rectangles and triangles should improve discrimination at the expense of added complexity, while a large dataset for a particular class could improve the result by a series of similar hits.