In order to improve the stability and discrimination of local feature combination for image representation

two image mediate-level representations

Inter-CSP(inter-class common stable pattern) and ntra-SSP(intra-class special stable pattern) are proposed.The details of processing are given

which can be divided into statistic-filtering

pattern decomposition

pattern summarization

and item-based geometric relation modeling on frequent item_sets mined from image semi-local features.A recognition framework is introduced based on Inter-CSP and Intra-SSP.The experiment results demonstrate that these two kinds of patterns are superior to classical methods.