Classification of Live Video Stream from Pakistani News Channels (Urdu) using Deep Learning Latest Techniques /
Muhammad Afzal
- 134p. Soft Copy 30cm
In our contemporary era, information is of prime importance and its dominant use by social media and TV channels for making public opinion and cultural influence is quite evident. Videos form the major portion of media and contain more elaborate information than a single image. Today, videos are piling up in millions every day and their segregation, classification and analysis are upheaval tasks. Live TV video stream contains voice, metadata and image frames full of multiple information including written scripts etc. which can contribute to video classification. But utilization of each type of data we need to do a separate study. However, we have focused on classification of video stream using deep learning (DL) neural networks which are well established solutions for images and small videos classification and gesture recognition. In our study, we have suggested a mechanism for classification of big or live video streams obtained from Pakistani TV News Channels into 5 classes (Advertisement, News, Talk Show, Sports & Entertainment Program) using supervised DL pretrained neural networks. Due to non-availability of authentic dataset on this subject, we have created a customized data of videos recorded (approximately 335 hours videos) from various sources like different TV channels‘ websites and YouTube. Videos were processed to extract image frames to prepare a trainable dataset. For our experimentation, we have mainly used pretrained ResNet variants (ResNet18, ResNet34, ResNet50, ResNet101 & ResNet152) on ImageNet dataset and few other models like AlexNet, ConvNeXt_Tiny, DenseNet121, SqueezeNet and VGG11 for comparison purposes. Then modified the last classification layer of the network as per number of target classes and finetuned all weights of neural network on the subject dataset. We carried out various experiments on these neural networks and achieved quite encouraging results having accuracies ranging from 95% to 99%. For testing of videos on trained models, dynamic averaging time domain window was applied to diminish the jitters in the output results. This can be useful in many other applications as well including social media & advertisements analysis, classification of small videos, industrial and business automation etc.