Research Of Content-based Commercial Product Image Retrieval And Recommendation System

As is known to all, the merging of computer has exerted a profound influence on oursociety, which dramatically changes our life. With the development of Internet, commoditytrading way a lot of changes, purchasing methods transfer to the online commodity trading.Due to the wide variety of goods and the increasing of the number of goods, people take alarge amount of time to search and select the product they need. It becomes an urgent problemto retrieve and recommend the product with high precision.For the low precision of the retrieval performance of traditional retrieval methods, amulti-feature based multi-stage retrieval system is designed. We determine the catalog of theproduct image firstly, and then use the high precision retrieval methods to extract similarproducts according to the catalog of the product image. In addition, we also start the productrecommendation problem. We use LDA (Latent Dirichlet Allocation) model to learn theinformation between visual features of different commercial product image, and thenrecommend different products according to this information we learned.In the product image catalog determine layer, firstly, we propose an incrementalvocabulary tree to handle the dynamic increasing of product image. Based on this model, weextract local visual features to build visual vocabulary layer, and then use weighted votingmethod to determine the catalog of the product image.In the high precision retrieval layer, firstly, we propose a novel algorithm of clothingsegmentation, combining the saliency detection human, skin area detection and graph cutmethod, which can handle variations in configuration. And then, based on the segmentation,we extract the color, textual and shape feature to retrieve fashion products, achieving a highprecision performance.In respect to product recommendation, we use LDA model to learn the informationbetween visual features of different commercial product image. For the difficulty to fusedifferent visual feature in LDA model, we propose to use multiple LDA model to learn theinformation of the multiple visual feature between different commercial products. Finally webuild a recommendation system for clothing coordinates based on the multiple LDA model.