Research On Online Social Network Image Classification Based On Deep Belief Network

With the rapid development of internet technology and the advent of Web3.0 era, online social network applications has become increasingly popular in daily life. It provides users with a platform to exchange of information and share blogs, videos and images, which product a massive online social network data. As a significant component of online social network data, image has become an important carrier in the big data era. With the increasing of the number of social network images, the abundant information contained in the social network image is not used efficiently. Therefore, how to extract the information contained in the online social network images effectively and accomplish image classification quickly have become one of the research focuses in the field of image processing.Deep belief network(DBN), one of the common deep learning models, is taken as the research object. Aiming at the problem of the selection of the initial learning rate in training stage, this thesis makes a thorough research on them. The main contributions of this thesis are as follows:1. A training method of deep belief network is improved in this thesis, which combine adaptive learning rate based on reconstruction errors and weight updating rule with additional momentum. Then, the improved DBN and SVM classifier are combined to form an image classification method based on improved DBN-SVM. Finally, compared with other classification methods such as ANN, SVM, DBN and DBN-SVM, the validity of the method improved is verified in the experiment, which conducted on the MNIST database of handwritten digits.2. A Sina microblog data retrieval method based on API is proposed in this thesis. By calling the place/nearby_timeline interface of the location service interface group provided by Sina microblog open platform, which can help us to obtain dynamic data around a certain site. The dynamic data not only includes images upload by user, but also the time, location of microblog published and the total number of records location of microblog published and so on. The dynamic data parsed by program can be used as the dataset of Sina microblog.3. Taking Sina microblog as an example, depended on the convenience of feature extracted of online social network image, an online social network image classificationmethod based on improved DBN-SVM is proposed in this thesis. Then, compared the method with other four classification methods in simulation experiment, which conducted on the Sina microblog dataset. Finally, applying this method to estimate the popularity of different sport brands.The thesis ends with the conclusions of the ahceived and future research, which extend our thoughts in the future.