Image Hash Search Based On Deep Learning

With the arrival of the era of big data information,coupled with the rapid development of digital storage devices and multimedia technology,digital image data in the Internet presents a blowout growth trend.Compared with the text information,the image information contains more useful information,the expression form is more image and intuitive is an important means for people to obtain information,but with the deepening of the digital society,in the face of massive digital image resources,how to quickly and accurately find the resources needed by the users of digital image data has become a urgent problem.Content-based image retrieval has two problems:semantic gap and curse of dimensionality.For the first problem,deep learning has been proved to learn features strongly.Deep learning can sue the underlying data characteristics to find hidden internal relations features and abstract high-level image features.For the second problem,the hash algorithm is the best retrieval method.The hash algorithm maps the original high dimensional data into a series of binary codes by hash function.The feature representation is simplified and it can use Hamming algoritlhm at the similarity measure,which greatly improve the retrieval rate and strengthen the real-time retrieval.This thesis presents an image hashing retrieval method based on deep learning.In the process of feature extraction,training on stacked auto-encoder using image data without annotation,learning to strong the expression of image feature.The unlabeled learning does not require labeling of image database,reducing the standard of image library,at the same time,also played a strong advantage of deep learning network,get a better image feature representation than the other algorithms.In the index structure,hash algorithm is the best retrieval method in the all of the existing algorithms,which improves the retrieval efficiency and strengthen the real-time retrieval.On this basis,second-level retrieval is proposed,which increases the retrieval accuracy at the little expense of retrieval time.Through a series of experiments,the validity of the proposed algorithm is verified and the performance of retrieval is good.