J, Sreekanth
(2014)
Scene Segmentation and Classification.
Masters thesis, Indian Institute of Technology, Hyderabad.
Abstract
In this thesis work we propose a novel method for video segmentation and classification, which are
important tasks in indexing and retrieval of videos. Video indexing techniques requires the video
to be segmented effectively into smaller meaningful units shots. Because of huge volumes of digital
data and their dimensionality, indexing the data in shot level is a tough task. Scene classification
has become a challenging and important problem in recent years because of its efficiency in video
indexing. The main issue in video segmentation is the selection of features that are robust to false
illuminations and object motion. Shot boundary detection algorithm is proposed which detects both
the abrupt and gradual transitions simultaneously. Each shot is represented using a key-frame(s).
The key-frame is a still image of a shot or it is a cumulative histogram representation that best
represents the content of a shot. From each shot one or multiple key frame(s) are extracted. This
research work presents a new method for segmenting videos into scenes. Scene is defined as a sequence
of shots that are semantically co-related.
Shots from a scene will have similar color content, background information. The similarity
between a pair of shots is the color histogram intersection of the key frames of the two shots. Histogram
intersection outputs the count of pixels with similar color in the two frames. Shot similarity
matrix with 0
′
s and 1
′
s is computed, that outputs the similarity between any two shots. Shots are
from the same scene if the similarity between the two shots is 1, else they are from different scenes.
Spectral clustering algorithm is used to identify scene boundaries. Shots belonging to scene will
form a cluster. A new method is proposed to detect scenes, sequence of shots that are similar will
have an edge between them and forms a node. Edge represents the similarity value 1 between shots.
SVM classifier is used for scene classification. The experimental results on different data-sets shows
that the proposed algorithms can effectively segment and classify digital videos.
Key words: Content based video retrieval, video content analysis, video indexing, shot boundary
detection, key-frames, scene segmentation, and video classification.
Actions (login required)
 |
View Item |
![[img]](http://raiithold.iith.ac.in/style/images/fileicons/text.png)
Altmetric
Altmetric