Research And Implementation Of Image Retrieval On Shopping

Along with the development of Internet and e-commerce,online shopping has gradually become the mainstream of consumption,for the countless orders have completed everyday in Jingdong,Taobao,Amazon and the other shopping platforms,the way to make users find the favorite goods quickly and accurately become the focus of the study.Traditional method of commodity retrieval is based on keywords inputted by users;However,keywords are difficult to describe the goods completely and precisely,which cause customers need spending time to choose the target commodity.In order to facilitate users to shop online,content-based image retrieval technology is commonly applied to shopping platforms;same or similar commodities will be returned when users input commodity images of their demand.However,describing commodities using underlying visual features simply may cause the semantic gap and details loss of commodities.This paper introduces the basic theory of content-based image retrieval,mainly study the region based image retrieval.It proposed two new and effective image retrieval methods based on representative region and category respectively.Experimental results show that the proposed methods are prominent than some of the existing retrieval methods,and finally we design a real-time shopping platform using category based image retrieval method.The main work and contribution of this paper are as follows:1.A image retrieval method based on representative region is proposed.Firstly,image is segmented to several categories using K-means and Affinity Propagation(AP)clustering method,the largest region of each category is found as the representative region.Then,the color and the texture features of representative regions are obtained by the HSV color histogram and Local Binary Pattern(LBP).Integrated Region Matching(IRM)algorithm is used for calculating the distance between query images and database images.In representative region method,features of each category are extracted and combined as image feature,which improve the composition similarity between query images and retrieved images;Assigning weights for regions according to their probability as interest object improve the accuracy of retrieval.Experimental results show the proposed algorithm has efficient and reliable performance.2.A novel image retrieval method based on category is proposed.In this method,feature extraction and matching are processed at the category level.Categories in images are obtained by a statistical region merging(SRM)and affinity propagation(AP)methods.HSV color histogram,Local Binary Pattern(LBP),quantity and distribution of regions are extracted to represent the characteristics of categories.Moreover,integrated category matching is used to calculate the images distance.SRM and AP ensure the integrity of regions in segmentation and prevent the over-segmentation;combined feature descriptions contains rich visual information and helps to reduce the semantic gap;integrated category matching is more in accordance to human visual perception due to the significance index are obtained based on centroid.Experimental results show that the proposed method is more effective than most of existing RBIR methods.3.A real-time shopping platform is designed.Firstly,users input commodity image of their demand,the keywords of commodity are obtained by support vector machine(SVM)algorithm and if there are descriptions of commodity input by users they will be considered as the keywords of commodity;then relevant commodities are fetched from Jingdong and Taobao using web crawler through the keywords;finally,the most similar commodities are retrieved from them by category based image retrieval method,and the information of them are feedback to users as well.The platform uses a retrieval method based on text combining content,which can recommend interested goods to users quickly and accurately,and provide users effective product analysis to improve the shopping experience.