Research On The Method Of Video Character Relation Mining Based On Association Rules

Character relationship is an important clue to help users understand video stories.Especially in storytelling videos,it can reveal the high-level semantic information of video stories.At present,people-based video analysis pays more attention to regard each character as an independent individual,ignoring the objective relationship between them.With the increasing demand of movie viewers,people are more concerned about the relationship network that can quickly reflect the development of the whole story.The purpose of mining the relationship between people in video is to take the relationship between people as the research subject according to the content of video,and to obtain the high-level semantic information of the relationship between people in video through video content analysis technology.Mining the relationship between video characters can be considered from the perspective of character symbiosis,that is,the probability of common occurrence of characters under the lens unit.This reduces the complexity of mining and can efficiently mine useful information.Association rule is a mode of conditional probability that responses to the relativity between transactions.In the analysis of video character relationship,we can quantify the correlation of character co-occurrence,obtain the relationship network of characters,and then analyze the high-level semantics of video content.Firstly,the original video was preprocessed to get the key frame sequence.This thesis combined several Perceptual Hashing methods to detect the boundary of video shots,and took the shot boundary frame as the key frame of each group of shots.Secondly,image frames were repaired and face recognition was performed.Aiming at the overexposed blurred image in the extracted key frame,an overexposed blurred image restoration algorithm based on dynamic convolution template was proposed to get the repaired key frame image.Face detection was carried out by combining directional gradient histogram with hierarchical associative classifier for key frame sequence.Face feature points were extracted by ERT algorithm.ResNet network was used to generate feature vectors for face feature points and face frames,and CW clustering algorithm was used to cluster feature vectors to obtain face data items.Finally,the common face item sets under the lens unit were regarded as the same transaction,and the face transaction database was obtained,and the positive and negative association rules were mined.This thesis put forward three concepts of relationship refinement:relationship direction,relationship influence and relationship weight.Using MapReduce framework,we first mined frequent itemsets,then calculated the correlation and confidence of the obtained frequent itemsets,and judged rules such as relationship impact,then got a video character relationship network based on positive and negative association rules.The experimental results show that,compared with the objective character relationship graph of video content,the result graph of this method has obtained the basic structure of the objective character relationship,and the average accuracy rate of the nodes is above 62%.Compared with other methods of character relationship mining,this thesis finishes the refinement of the character relationship,and has a stronger reference value on video content analysis such as video recommendation and video summary.