Automatic Cricket Highlight Generation Using Event Driven and Exctitement Based Features Using Deep Learning

Abstract

Video summarization plays a crucial role in diverse domains where the major application is sports video summarization. Sports broadcasting channels have shown their keen interest in the summarization of sports videos based on the viewer s interest as they hold massive viewership worldwide. So, to gain transmission benefits and reduce storage, extraction of the exciting clips from the lengthy cricket video has been widely used. Sports video summarization is a tiring task as it includes enormous variations in the camera movements, background noise, lighting conditions, editing effects, etc. As a solution to such a problem, this thesis work presents reliable hybrid methods to effectively identify the key events from the video for highlight generation. The computational complexity of this arduous task has been reduced by proposing key event recognition systems that detect and classify only the important events from the video. This step reduced the video length, making it suitable for summarization. newlineA hybrid deep neural network with emperor penguin optimization (HDNN-EPO) is proposed in the research work to generate cricket video highlights. In this work, the key events of the cricket video are identified, and then summarization is done for the obtained key events. Initially, the exciting clips are extracted using audio features such as shouting, spectators cheering, and applause of the audience. Then the keyframes are generated by determining the shot boundaries in the videos, and these frames are identified using the hue histogram differences of neighborhood frames. The key events such as replay, players gathering, real view, umpire, batsman, fielder, spectators, and field view are then extracted to determine the importance of each clip in summarization. After this process, the concept annotation process is carried out using the proposed HDNN-EPO algorithm for the obtained exciting clips. The EPO algorithm reduced the weight values of the DNN to reduce the error rate in the annotation process. A voting classifier is