Video Summarization using Deep Learning

Abstract

Today, the technology is used in some or the other way in almost all areas newlineand the advancement in the technology have brought a revolution in the field of newlineresearch. It has forced the people to adopt the updates in technology to smoothen newlinetheir day-to-day life. Images and Videos have increased in a large extent and more newlinealgorithms are being designed to classify such images and videos. One of the major newlineroles has been played by Artificial Intelligence that has changed the thinking power newlineof a human being. The revolution in Artificial Intelligence has made human life so newlineeasy such that humans are forced to adopt it. newlineToday, the datasets are generated in the form of Video. The videos are too newlinelong in length and due to this, there occur 2 main complexities: 1) Very Difficult to newlinefind the required information from Video and 2) Time Consuming with respect to newlinelength of the video. Owing to these complexities, there is a huge demand of video newlinesummary. The main aim of the research work was to minimize the mentioned newlineresearch gap using various CNN models for training various datasets. newlineThe focus of the research was on summarizing any video with the newlineappearance of any object in the video along with their time frames. The search for newlinethe required datasets was fulfilled by KAGGLE repository. The idea was to newlinesummarize the videos based on Animals Detection, Birds Detection, Flowers newlineDetection, Bollywood Actors Detection as well as Vehicles and Number Detection. newlineThe research work presented in this thesis provides the video summary on newlineAnimals Detection and Birds Detection from a large video. The summary is in the newlineform of table which contains the presence of the object with its start time and the newlineend time. The work has achieved about 97% accuracy level for both the datasets. newlineFurthermore, we have presented the comparison of various CNN models namely newlineGoogleNet, Squeezenet and Alexnet on the flower dataset to test the accuracy level newlineof Object Detection; although Alexnet has least number of layers,