Research On Fabric Image Retrieval Algorithm Based On Pyramidal Multi-Scale LBP And Self-taught Hashing

With the rapid development of Internet and multimedia information technology, content-based image retrieval has become a hot research topic in the information field, which has been widely used in the public security system, the health care system and the intellectual property maintenance system, etc. In this thesis, we focus on the application of image retrieval in the textile industry and explore the suitable method for the fabric image retrieval in order to meet users' increasing fashionable demands. After analyzing the characteristics of fabric image thoroughly, we study the retrieval algorithm from two aspects of feature extraction and index construction, and finally propose fabric image retrieval algorithm based on pyramidal multi-scale LBP and self-taught hashing. The main work and innovations of this research are:1. We introduce several typical LBP patterns and propose a feature extraction method named pyramidal multi-scale LBP, according to the characteristics of fabric image, which overcomes the limitation of some existing multi-scale LBP methods. The method firstly constructs an image pyramid, and then LBP operator with equal sampling points under multiple radiuses is applied to each image in the pyramid. The proposed method can extract feature in a wider region and describe texture information of fabric image effectively than other existing multi-scale LBP methods.2. We analyze several existing hashing methods and propose to apply self-taught hashing method into fabric image retrieval. Meanwhile we make an improvement on self-taught hashing method according to the differences between image and text data. As utilizing the improved self-taught hashing method to construct index for the fabric image, we can greatly save the data storage space and reduce the computation complexity.3. On the basis of the two kinds of algorithm, we propose a fabric image retrieval algorithm based on pyramidal multi-scale LBP and self-taught hashing. The feature extraction method of pyramidal multi-scale LBP can combine local information and global information of fabric image, in this way, the texture structure of fabric image can be described effectively. In addition, it can preserve the similar structure of image data in original space as utilizing the improved self-taught hashing method in index construction stage. The experiments prove that the proposed method is superior to other related methods.