| Summary: | 碩士 === 國立中山大學 === 資訊工程學系研究所 === 102 === In recent years, the image database systems develop quickly and play a significant role in many applications. To discover the useful information from these image databases, it is crucial to develop an efficient image mining algorithm. Lee et al. propose the 9DSPA-Miner algorithm to mine the image database where each image is indexed with the 9D-SPA representation. They use the Apriori strategy to mine the frequent patterns from the image database. In their approach, since there exists an unknown spatial relation in the generated candidate, they propose a reasoning method to reason the unknown relation. Therefore, they can generate the relations which satisfy the spatial consistency. However, they may also generate the impossible relations which can not be found in the 2D space or in the input database. Therefore, a large number of invalid candidates may be generated. Moreover, they recompute the intersections of image sets when counting the support of the pattern. This takes too much time. Therefore, in this project, we propose a Valid-Candidate approach. In our approach, every generated candidate will be valid. The contributions of our approach are as follows. We define the FPL condition and the IntImage condition. If one of the two conditions is satisfied, any generated candidate will not be frequent whatever the unknown relation is. If both two conditions are not satisfied, we use the relations of size-2 frequent patterns to discover the unknown relation and we use a verification step to avoid generating the impossible spatial relations. Therefore, the number of candidates in our approach is less than that of the 9DSPA-Miner. Moreover, by recording the image set with the discovered pattern, we will not scan the index structure to count the support. Hence, we will not recompute the intersections of image sets. From our simulation results, we show that our proposed approach is more efficient than the 9DSPA-Miner.
(Keywords: Frequent patterns, Image database, Iconic index, Image mining, Spatial association rules)
|
|---|
