Design And Implementation Of A Cross-modal Distributed Retrieval Engine For Social Network

With the rapid development of internet technology and intelligent terminals,various social network applications are increasingly closely related to people’s lives.The amount of data generated by them is increasing exponentially,and the types of data are becoming increasingly diverse.Therefore,people’s demand for multimodal data retrieval is also increasing.At present,some mainstream social media platforms in China only have single modal retrieval.Even though the general search engine Baidu in China can search for images through text,its technology for achieving this function mainly uses the matching of search text and tags.Essentially,it is still a single modal retrieval form of text retrieval,and has not truly achieved cross modal retrieval.At present,scholars have made a lot of research on cross modal retrieval technology,but there are still several challenges.Firstly,the heterogeneous gap caused by inconsistent representation of different modal data at the bottom layer leads to low retrieval accuracy;The second issue is the slow retrieval speed caused by the excessively high dimensionality of the extracted data feature vectors in massive multimodal data.Based on this background,this thesis focuses on social network and conducts research on related issues.The main work and contributions include the following aspects:1.A cross modal hash model based on deep learning is proposed.This thesis is based on the concept of lightweight model design.In the feature extraction part of the model,deep separable convolution is used as the design subject of convolution calculation.While ensuring the accuracy of feature extraction,it greatly reduces the calculation of convolution operations and greatly reduces the inference time of the model.At the same time,the channel attention mechanism is specially integrated to enhance the model’s ability to capture subject features,and self-supervised methods are used to guide feature learning,enhancing the effective utilization of data label information.The effectiveness of the model designed in this thesis has been demonstrated through comparative experimental results on the NUS-WIDE dataset.2.A two-level hybrid index structure based on cross modal hash vector was designed.Based on the mapping concept of inverted index and the idea of graph partition,combined with the characteristics of cross modal hash retrieval,this thesis designs a two-level index structure to accelerate hash vector search.The first level inverted index is used to speed up the search of vector set,and the second level graph index is used to speed up the search of internal points of vector set.Finally,through the analysis of test results,the effectiveness of this index structure for cross modal hash retrieval was verified.3.Designed and implemented a cross modal retrieval system for social network.This thesis uses a laboratory self-developed graph database as the storage support at the bottom of the system.Finally,in terms of system function implementation,not only cross modal retrieval is completed,but also related functions such as graph query are implemented,expanding the data retrieval and analysis capabilities of this system.The final system test data in this thesis comes from real social data on Twitter website.From the experimental results,it meets people’s expectations for cross modal retrieval.Therefore,the research content in this thesis has certain application value for exploring how cross modal retrieval technology can play a role in practical applications in the industry.