Combined Holistic Feature And Local Feature For Commodity Image Retrieval

In this constant changing tide of Internet age, people can obtain all kinds of information through more convenient and quick accesses, the products which are supplied by the internet have been changing people’s life in every aspect. Therefore the applying of electrical commercial website has become very popular under this circumstance. How to improve the accuracy and efficiency of merchandise image retrieval which happens in mobile terminal, I did a deep study in this field and has put forward some major points about the study, the major contents of my paper are presented below:First, this paper described and analyzed the typical content-based image retrieval system. In the traditional content-based image retrieval system, the retrieval part is mainly using either a holistic feature or a local feature to retrieval. If the system uses a holistic feature, it will extract the information about color and edges information. The image retrieval system with using the holistic feature has a better retrieval quality and a higher precision than the system with using local features to the pictures which have an obvious difference to the prospects background. The holistic feature image retrieval indexed by the color and edge directivity descriptor has a good efficiency. The accuracy of the retrieval results and integral rate will differ a lot in different situation. Thus, we must find a measure to reduce the error rates and the fluctuations.In order to eliminate the impact of the fluctuation derived from using the different measure, this paper uses a weighted undirected graph method to combine the global feature and local features. We will execute the preliminary process of merchandise images in the database and save all the commercial information before we construct the graphs. In the graph, the node represents a picture and the edge between two nodes represents the pictures which have similarities. The weight in the graph represents the similarity values between the query image and database image. We construct a graph using candidate get from the holistic feature retrieval and construct another graph using the candidates with local feature retrieval results. Then we put the two graphs together and fuse them to a new weighted undirected graph. The weight in the graph will be updated. Finally we use the updated data to find out the new retrieval results.This system implemented a client-server model with using the system which combined holistic features and local features. The system will transfer the image to the service system. Then extract the images’ arithmetic features and descriptions. The query image will be processed via the new image retrieval system and get back the finally retrieval results as well as the corresponding commercial messages to the client. At last we conduct the experiments and get the retrieval statistic data and the accuracy. The experimental results proved that the proposed retrieval system which combined the feature retrieval method has higher retrieval accuracy. Besides, this method can prevent the wide fluctuation in accuracy when processing different categories images.