Research On Image Retrieval Algorithm Based On Semi-supervised Hashing Algorithm

With the rapid development of Internet technology and multimedia technology, the number of digital images which people have access to get increases exponentially. Because of being content-rich and having direct visual performance, digital images are used in increasingly wide range of application areas, including national defense, medical, digital libraries, media, industrial technology, entertainment and so on. So how to find the picture which user needs quickly and efficiently in massive image library becomes extremely important. The early 1990s, content-based image retrieval (CBIR) emerged. The main idea of CBIR is to represent an image using color, texture, shape and other features. And then calculate the similarity distance between the image characteristics. At last return retrieval results according to the value of similarity distance. In CBIR, the "dimension disaster" and the high time complexity of similarity measure function has been plagued researchers. In recent years, researchers have proposed map high dimensional content feature to lower dimensional Hamming space using hashing algorithm, using a short binary sequence to represent an image. So the Hamming distance between binary codes can represent the distance between images. With this method, we can effectively alleviate the "dimension disaster" problem and improve the retrieval speed.Due to "dimension disaster" of traditional content-based image retrieval method and urgent needs for fast image retrieval at present, this paper focuses on image retrieval methods based on supervised hashing algorithm, unsupervised hashing method and semi-supervised hashing method.In addition, this paper proposed a balanced semi-supervised hashing method by dividing image into several blocks. With the help of improved semi-supervised hashing, we obtain a short hash code of each block, which jointed together forms a hash code of an integrated image. In the improved semi-supervised hashing, the supervised information is completed by combining the similarity of image pairs and label information. Extensive experiments results show that the balanced semi-supervised hashing Algorithm can overcome "precision rate bottleneck" problem, which existing hashing algorithms usually run into, when hash codes reach a certain length. And the balanced semi-supervised hashing algorithm can achieve higher precision when the entire lengths of the hash codes of picture are the same. In order to further improve the retrieval accuracy, we add pre-retrieval process to narrow the target image library into the image retrieval system based on balanced semi-supervised hashing algorithm. The pre-retrieval process is achieved by combination of wavelet decomposition and semi-supervised hashing algorithm. The experiments results show that we can get better retrieval accuracy after adding the pre-retrieval process.