Research On The Technology And Key Problems Of Automatic Video Clip And Mixing Based On Natural Language Processing

With the popularization of computer network and the rapid development of Artificial Intelligence,people put forward higher and higher requirements for Artificial Intelligence.One of the main objectives of Artificial Intelligence research is to enable the machine to be competent for some complex tasks that usually require human intelligence.This thesis is aimed at the task of automatic video clip and mixing.Human editors need to accumulate a large number of video materials to make a material library and then select the appropriate video materials from the material library.Generally,more than 120 videos are needed to be selected for a 3-minute video editing.It's a lot of effort.Therefore,this thesis proposes a model which can complete the task of automatic video clip and mixing based on theme sentences from the perspective of Natural Language Processing,and designs experiments to verify its effectiveness.In this thesis,we design an automatic video clip and mixing model,which can replace human to complete the task of video mixed cutting according to the theme sentence.This thesis also proposes a theme filtering model,which is the core part of the automatic video clip and mixing model,and is responsible for the semantic filtering of video clips.In addition,this thesis proposes a semantic sentence matching model(MFF)which combines multiple features.The feature interaction layer in the MFF includes five feature fusion interaction processes.This thesis also designs an implementation model for the task of blind movie automatic generation,which includes an original sequential correspondence algorithm.For the above models,this thesis designs experiments to verify their effectiveness.Among them,the test accuracy of the multi feature interactive semantic sentence matching model on the natural language reasoning standard dataset "SNLI" and sentence rewriting standard dataset "Quora Question Pairs" is much higher than that of the standard model,which shows the superiority of the model for semantic feature extraction.In the experiment of theme filtering model,we compare the BLEU evaluation value between the theme filtering model based on MFF and the theme filtering model based on the semantic sentence matching model based on siamese structure.The results show that the theme filtering model based on MFF proposed in this thesis has excellent effect.The experiment also shows the video clips automatically mixed and cut according to the given theme sentences,and the results show the effectiveness of the proposed model.In the part of blind movie,this thesis explores the methods to improve the matching accuracy in the experiment of script and movie subtitle preprocessing design,and explores the effective preprocessing methods in the experiment,and the sequence corresponding algorithm proposed in this thesis also improves the accuracy.The final blind movie result has proved the effectiveness of the model.