Experimental results towards content-based sub-image retrieval

We are interested in the problem of sub-image retrieval (CBSIR), i.e., given a query image one must find the best candidate images that contain that query image. We use two kinds of image feature vectors: global color histograms and autocorrelograms and experimented with several distance measures for both feature vectors in our experimental system. After extensive experimentation we found that using autocorrelograms with the so-called S₁ distance measure yielded excellent results for sub-image retrieval with an acceptable processing overhead.