Automatic UCLl Indexing And Intelligent Recommendation Based On Deep Learning For Video

With the rapid growth of video resources in the Internet,the problem of information overload has become increasingly prominent.Although major video portals use video features and user's data characteristics to recommend videos to users,there still have two problems: First,there is a lack of complete and standardized video semantic feature indexing standards and the commonly used feature extraction algorithm is not sufficient to extract the semantic features that makes poor recommendation of the video.Second,the traditional single video recommendation algorithms cannot simultaneously utilize video features and user features.Although the deep learning algorithm can partially solve the problem,there are still insufficient analysis of the implicit relationship between the two features.Considering the above problems,this dissertation studies video UCL automatic indexing and intelligent recommendation technology through the Uniform Content Label(UCL)which contains rich semantic features,and proposes a video UCL automatic indexing method based on characteristic flexible sampling and attention mechanism(AUIMV-FFSA)and the video recommendation algorithm based on the self-attention mechanism and model fusion(RAVSAMF).On this basis,the automatic video UCL indexing and intelligent recommendation system is designed to verify the AUIMV-FSSA method and the RAV-SAMF algorithm proposed in this dissertation.The main research work of this dissertation is as follows:1)In the view of the problem that the video recommendation system has insufficient video semantic feature extraction and lack of standardized indexing,a FSSA-based video UCL automatic indexing method AUIMV-FSSA is proposed.Firstly,the attention mechanism is introduced into multiple modules of the SV2T(Sequence to Sequence-Video to Text),and the input of the SV2 T model is optimized by sampling video features flexibly using cosine similarity between the lower layer semantic features,generating a SV2T-FFSA(Sequence to Sequence-Video to Text based on FFSA)model.Then combining the speech natural language description features with the video natural language description features generated by SV2 TFFSA model to generate high-leavel semantic features of video keywords.Finally,UCL is used to index the video visual features and the video semantic features extracted by the above methods.2)Aiming at solving the problem that the current video recommendation algorithm can not discover the implicit relationship between the video semantic features and the user features effectively,this dissertation proposes a video recommendation algorithm RAV-SAMF based on SAMF.Firstly,the video semantic features in the video UCL are extracted,and the video semantic features and user features are preprocessed.Then,the self-attention mechanism is introduced into the traditional MLP(Multi-Layer Perception)to generate the new model MLPSA(MLP based on Self-Attention),and using this model to analyze the implicit relationship between features.Finally,the MLPSA model and the xgboost model are merged to generate a new model XGB-MLPSA,which is used to predict the user's score and likes and dislikes of the video,and improve the precision and recall rate of the video recommendation algorithm.3)Based on the above method,this dissertation designs the video UCL automatic indexing and intelligent recommendation prototype system,and verifies the AIUMV-FSSA method and RAV-SAMF algorithm proposed in this dissertation through experiments.The experimental results show that the AUMMV-FSSA method has a higher improvement in the indicators of the video language description than the traditional SV2 T method.The RAV-SAMF algorithm has higher precision and racall rate than other commonly used video recommendation algorithms.At the same time,verifying the effectiveness of UCL in the video recommendation algorithm.