Article Preview
Top1. Introduction
The enormous development of multimedia technologies and availability of computing resources have led to the explosion of video data on the internet. However, the surge of video data has not been induced by an increase in its accessibility. Thus, there is a need for techniques that can efficiently access, browse, index and retrieve the video data. Shot Boundary Detection (SBD) is a preliminary step for video abstraction, video segmentation and video retrieval approaches (Hanjalic, 2002). A video shot represents a sequence of interrelated frames captured in a single take with one camera (Pal et al., 2015). Detection of shot boundaries in a video is mainly based on identifying the editing effects that are used to combine shots into video sequence. The hierarchal structure of video is represented in Figure 1.
Figure 1.
Hierarchal structure of video
Generally, transition of shots in a video are categorized as: Abrupt and Gradual transition (Sengupta et al., 2015) and it is depicted in Figure 2. Abrupt transition occurs when there is a rapid change between consecutive frames/images whereas, gradual transition occurs when change of the boundary is over multiple frames. Fade-in, Fade-out, dissolve and wipes are the frequently used editing effects of gradual transition.
Figure 2.
Types of video shot transition
Extraction of essential features from video frames that efficiently represents the visual information plays an important role in SBD (Jadon et al., 2001). The frame features can be extracted either globally or locally (Thounaojam et al., 2016) and it provides different information of frame at computational level. The global features describes the visual content of whole image and it is represented by a single vector. Contrastingly from global features, local features describes the visual content of the frame in patches or considering pixels of small group and it is represented by a set of vectors. However, global features have certain limitations such as scaling, sensitivity to noise, illumination variation and it often fails to identify the essential features of the image (Kabbai et al., 2018). Thus, global features are not suitable for few applications. Their flaws are fixed by the use of local features which encodes the local information to get the finest details of the image such as interest points (Kabbai et al., 2018). Hence, several approaches available in the literature have combined both local and global features in various domains such as image retrieval (Chaudhary & Upadhyay, 2014), SBD (Rashmi & Nagendraswamy, 2020), object detection (Muralidharan & Chandrasekar, 2012) etc., and obtained good results. This has motivated us to carry out the proposed SBD approach.