Approximating Content-Based Object-Level Image Retrieval

Object-level image retrieval is an active area of research. Given an image, a human observer does not see random dots of colors. Rather, he/she observes familiar objects in the image. Therefore, to make image retrieval more user-friendly and more effective and efficient, object-level image retrieval technique is necessary. Unfortunately, images today are mostly represented as 2D arrays of pixels values. The object-level semantics of the images are not captured. Researchers try to overcome this problem by attempting to deduce the object-level semantics through additional information such as the motion vectors in the case of video clips. Some success stories have been reported. However, deducing object-level semantics from still images is still a difficult problem. In this paper, we propose a “color-spatial” approach to approximate object-level image retrieval. The color and spatial information of the principle components of an object are estimated. The technique involves three steps: the selection of the principle component colors, the analysis of spatial information of the selected colors, and the retrieval process based on the color-spatial information. Two color histograms are used to aid in the process of color selection. After deriving the set of representative colors, spatial knowledge of the selected colors is obtained using a maximum entropy discretization with event covering method. A retrieval process is formulated to make use of the spatial knowledge for retrieving relevant images. A prototype image retrieval tool has been implemented on the Unix system. It is tested on two image database consisting of 260 images and 11,111 images respectively. The results show that the “color-spatial” approach is able to retrieve similar objects with much better precision than the sole color-based retrieval methods.