Image retrieval methods

The retrieval method we recommend is based on stroke pairs that contain similar information in two images.

Stroke-pair based retrieval : html version / text version(best preformance)

To evaluate the performance of this method, we compared it with retrieval methods that only use image information in the color space.

Global color histogram intersection on original images : html version / text verion

Global color histogram intersection on filtered images : html version / text version

Global color histogram intersection on paintings produced by SPRI : html version / text version

Cluster-based color matching : html version / text version

We tested other methods to use stroke information. These methods seperate strokes into different collections based on color; image-similarity is based on similarity between the statistical data of stroke-collection-pair. These methods differ from how to quantize color and how to find stroke-collection-pair.

Stroke-collection-based retrieval, global color histogram defining quantized colors : html version / text version

Stroke-collection-based retrieval, color cluster defining quantized colors(1) : html version / text version

Stroke-collection-based retrieval, color cluster defining quantized colors(2) : html version / text version

Retrieval results of the above 8 methods are evaluated on the groundtruth(html version) / groundtruth(text version) in human's vision. We consider two factors in evaluation. The first one is how many similar images in groudtruth are included in retrieval results. The second one is the sorting order of the similar images in retrieval results. The evaluation results is as below.