Mobile Visual Search Based On Hash Methods

In recent years, including smart phones and tablet computers, mobile intelligent terminal developing rapidly, especially in the multimedia processing ability is greatly improved. These devices provide a good platform for mobile image retrieval. The so-called mobile image retrieval is that image photographed by the mobile terminal, the image information transmitted to the server side through wireless network, then retrieved by the server in the database retrieval, finally the information related with the photos returned to the mobile terminal. In this process, due to the limitation of network bandwidth, how to reduce the transmission delay is still a tremendous challenge in the computer industry. This study will be carried out in order to solve this problem.To this end, this paper presents a method of mobile visual search based on hash methods to realize the transmission delay. Firstly, the query image hashed into a set of binary hash code on the mobile terminal; then according to the hash code and progressive transmission strategy, the IDF weight of the hash code calculated, determined the priorities by the weight values, and hash codes with high weight value preferentially transmitted to the server side; when some priority hash codes transmitted to the server, the server begins to match and stop to match the right image; finally the retrieval result related with the query photos returned to the mobile terminal through wireless network. As far as I know, the progressive transmission by the form of the hash code in the mobile image retrieval is the first attempt.In order to improve the retrieval accuracy, the server used the proposed BoHb (Bag of Hash bits) model to match the image, combined with the multi-hash tables as this paper mentioned to weight the dimensions of the hash code in turn, and the dimension of weights table selected as hash subset to comprise multi-hash tables for retrieval, lastly the retrieval result returned to the mobile terminal. Extensive experiments have been done on the public Stanford MVS database, demonstrating that the proposed method achieves higher accuracy and recall compared to other strategies.