Scene change detection using Multi-class Support Vector Machine with MPEG encoding information

As the amount of digital video is increasing, efficient w ay s of searching and annotating it according to its content are req uired. T he fi rst step tow ard video index ing is to detect scene changes. A scene is usually defi ned as a seq uence of video frames w ith no significant changes b etw een frames in terms of their visual content. T he simp lest scene cut is rep resented b y a camera b reak , that is, an ab rup t transition due to an editing cut. M ore sop histicated changes are gradual transitions such as dissolves, w ip es, fade-ins, fade-outs, resulting from chromatic, sp atial, and comb ined edits. In this p ap er w e address the p rob lem of ab rup t transition recognition or, b riefl y , video-cut detection. M ost techniq ues of video segmentation w ork s on uncomp ressed data and rely on features such as color histograms ([1]), track ing of feature p oints ([2 ]), sub seq uent frame diff erences ([3 ]), and motion features ([4 ]). S ome algorithms have also b een develop ed to w ork directly on M P E G -encoded video seq uences ([5 ][6 ]), and have imp roved the comp utational efficiency , video comp ression is also generally done w ith signal-p rocessing techniq ues cap ab le of deriving useful features, for ex amp le, motion vectors in M P E G . W hile techniq ues that w ork in the uncomp ressed domain usually achieve high rob ustness, techniq ues that w ork directly on M P E G -encoded video are usually faster. W e p rop ose an original method comb ining the rob ustness of the uncomp ressed domain to the sp eed of M P E G -encoded video and p resent the fi rst ex p erimental results. Almost all these methods rely on a threshold selected b y a human op erator. B ecause it is not alw ay s easy to manually fi nd a good threshold w hen several features are used, the op erator uses a near op timal threshold. S ometimes, neural netw ork s ([7 ]) are used to determine these thresholds. W hile they can b etter fi nd relations b etw een the features, they can b e slow to train. W e are p rop osing an algorithm that can ex tract seven features from M P E G -encoded information, three features from D C comp onents and four features from B -frame macrob lock s. W e use these features to train a M ulti-class S up p ort V ector M achine ([8 ]) to automatically design a classifi er for detecting video cuts. T he organization of the p ap er is as follow s: S ection 2 p rovides a b rief descrip tion of the M P E G videocomp ression standard, w hile S ection 3 describ es details on the p rop osed algorithm to detect cuts. S ection 4 p resents the results of an ex p erimental study.