Research And Implementation Of Multi-feature Based Two-stage Product Image Retrieval Technology

With the rapid development of E-commerce website, to search the needed product byusers, with high speed and accuracy, becomes a research hotspot. Currently, E-commercemostly uses text-based image retrieval method, which could not satisfy of the demand of usersfor product image research. Therefore, people wish to apply Content-Based Image Retrieval(CBIR) technology in the area of E-commerce. However, it becomes a great challenge torapidly and accurately find out the needed product from the massive product image database.With fully consideration of the characteristics of product images, we use a hierarchy imageretrieval method, and improve and optimize the image features in product image retrieval. Amulti-feature based two-stage image retrieval system is designed and implemented.Firstly, this paper introduces the key techniques related to CBIR. Emphatically improvedthe Edge Histogram Descriptor (EHD), we propose an adaptive threshold based EHD to applyin shape feature retrieval method. The optimized EHD could filter the noise edge and theweak edge effectively. This descriptor improves the robustness.Secondly, based on a model of hierarchy image retrieval, considering the individualfeatures of product image, a two-stage product image retrieval model is designed. In the filterstage, single feature of image is extracted, which extremely reduce the retrieval space andimprove the efficiency. In the matching stage, the combined features are extracted to fullysearch and improve the search precision. To better improve the search accuracy, the weightsof combined features in matching stage are optimized, depending on the characteristic oftwo-stage image retrieval model. And an optimization method that adjusts the weights byretrieval performance is proposed. The experiments show that the ratio of precision and recallof our two-stage product image retrieval method with weight optimization is greatlyimproved. Finally, combined with the improved method of feature extraction, a multi-feature basedtwo-stage product image retrieval system is designed. In the filter stage, this system applywith the adaptive threshold based EHD to sketchy matching. The candidate images withsimilar shape are preserved to the next stage; in the matching stage, the combined featureswith optimized weights are applied with the precise matching to get the retrieval results. Theexperiments show that our system better improves the efficiency and accuracy. This methodprovides a useful solution for the product image retrieval in E-commerce.