Fig. 1. Example of a query image (top left) and correctly retrieved images.

Fig. 2. System overview.

Fig. 3. Effect of interfering background on histogram peak location: (A) original image, (B) global hue histogram, (C) hue histogram of cell marked in white, and (D) final peaks in discretized HSV space.

Fig. 4. Example of spatial proximity graph (SPG) construction: (A^′) synthetic image divided into cells, (B^′) cells marked with nodes (peaks) contained in them. The intermediate graph is shown in broken lines, (C^′) SPG constructed from (B^′).

Fig. 5. Example of scale and rotation invariance obtained by collapsing the intermediate graph to produce a spatial proximity graph (SPG) construction. The first row shows the intermediate graph, the second row shows the steps in collapsing the graph with the final nodes remaining shown boxed and the final links in curved lines, and the last row shows the SPG obtained for (A^′) query, (B^′) target image containing queried object at one-quarter the original size, (C^′) target image with smaller queried object rotated by 90°.

Fig. 6. SPG filtering on the synthetic example in Fig. 4: (A) query image and graph, (B) correspondence between query and candidate peaks obtained from first phase of matching, (C) construction of reduced SPG from the SPG shown in Fig. 4C^′ by deleting unmatched peaks and relabeling nodes.

Fig. 7. Example of SPG filtering: (A) “Blueberry Morning” query image with SPG superimposed, (B) a false match with reduced SPG superimposed.

Fig. 8. Example of reduction of SPGs after phase I: (top) query and query graph; a correctly retrieved image; SPG stored offline; reduced SPG in which a match was detected. (bottom) false match retrieved after phase I; SPG for the image; reduced SPG which did not match the query graph (hence this image is deleted by phase II).

Fig. 9. Steps in query processing: (A) query image labeled with the peak color labels, (B) mask defining neighbors—the cross marks the center pixel and the shaded pixels are its neighbors, (C) pixel pairs counted supporting each adjacency, (D) query color adjacency matrix obtained by thresholding (C).

Fig. 10. Refinement of retrieval by second phase of processing: the query is marked by a white box: (top two rows) results after the first phase of retrieval, (last row) results after completion of second phase.

Fig. 12. Example of query selection and result: (top) portion of image (from original image shown in Fig. 6) with query marked by a box and the query image generated; (bottom) retrieved images—the first three images have the query object embedded in the lower right corner.

Fig. 11. Recall–precision graph after phase II for a set of 25 randomly selected queries and 15 queries with more than three colors each.

Fig. 14. (Left) Recall–precision graph after phase II with cell sizes of 100 × 100 (Default), 200 × 200 (Double), and 50 × 50 (Half). (Right) Default cell locations and cell locations shifted by half cell width (50 pixels).

Fig. 15. Online user interface to FOCUS showing a query box being selected and the results after first phase of processing (where the first, third and fifth images contain the query object).

Fig. 16. First five retrieved images for three different queries (in the advertisement images domain) in order of rank. The query is marked by a white box. (First row) First, second, and fourth images are correct matches. (Second row) First, second, and fifth images are correct matches. (Third row) First, second, third, and fifth images are correct matches.

Fig. 17. First five retrieved images for queries in the natural objects domain, in order of rank with the query marked by a white box.

Fig. 13. Comparison of recall–precision graphs obtained with FOCUS and whole image color histogram-based retrieval on a set of 20 random queries.

Table 1. Retrieval results for 10 queries: (Recall) images retrieved/No. of correct images in database; (Prec 1) precision after phase I; (Prec 2) precision after phase II

Table 2. Comparative retrieval results for 25 queries on the advertisements only