Study On Intelligent TV Program Classification

The intelligent TV program classification is used to classify the scene of TV program.Through the time information of the scene,the broadcast time of the advertisement program and the normal program in the integrated video can be obtained.It helps to speed up the cataloging of TV programs.The classification of TV programs is not good.The research work of this thesis mainly includes the following three aspects.1.TV program video preprocessing includes scene detection,feature extraction and scene segmentation.Scene detection is performed on the video using the Py Scene Detect tool.Video feature extraction includes image feature extraction and Audio feature extraction.The image feature extractor is an Inception v3 network model trained on the Image Net dataset,and the Audio feature extractor is a VGGish model trained on the Audio Set dataset.The extracted features are segmented using the result of scene detection.2.Training and optimization of video classification model includes the training of the baseline classification model,the training of the feature adjustment model and the training of the competing classification model.The baseline classification model is trained by DBo F model and Mo E classifier.The feature adjustment model is trained with a modified frame feature extraction model Ne Xt NEW and classifier Mo E,and is applied in the modification of the samples in training.The modified samples are used to train the competing classification model.3.The test and verification of the classification models,include the baseline classification model test,the competing classification model test and the actual video program classification test.Comparing recall ratio and accuracy of advertising video and other video in baseline classification model and competing classification model,then come to the following conclusions: when the clustering center number of DBo F model is 1024 or2048,the baseline classification model is better than competing classification model;when the clustering center number is 4096,the competing classification model is better than the baseline classification model;The actual classification accuracy of the two models needs to be further improved.