5. Histogram Analysis

Color

histograms are one of the oldest and best known global feature used in image

retrieval 17. Video is sequence of frames captured by camera. Videos are

sequence of shots & shots are sequence of frames. Distinct type of bodies

can exist between shot. It is impossible for human eyes to detect it openly.

Research has been done on automatic content analysis and segmentation of video.

We can perform a histogram analysis of the RGB features of every frame

analyzed. The histogram analysis is done to find out the number of frames with

the same histogram as compared to others. This is considered to be similarity

of the frames. Cosine techniques measure can be used for detecting histogram

changes in sequence frames.

• Histogram Creation: The procedure used collates consecutive frames based

upon 3 64-bit histograms (1 of luminance, and 2 of chrominance). These three

histograms are then put series of interconnected events to form a single a

N-dimensional vector, where N is the total number of bins in all three

histograms (in our case N=192).

• Cosine Similarity Measure: The dissimilarity

analogue of the cosine measure is used for comparing the histograms of adjacent

frames. The two N-dimensional vectors, A and B, represent the color signatures

of the frames. The distance Dcos(A,B) between vectors A and B is given by,

Dcos

(A,B) = 1-?i=1N(ai.bi) / (?i=1N ai2 .

?i=1N bi2) 1/2

Where ai is one bin in A and bi is the

corresponding bin in B. It can be seen that cosine measure is basically the dot

product of two unit vectors. The result is the cosine of the angle between the

two vectors subtracted from one. Hence a small value for Dcos indicates that

the frames being considered are similar, while a large Dcos value indicates

dissimilarity. A high cosine value can indicate one of two things. Firstly, it

can (and should) signal that a shot boundary has occurred. Secondly, it can be

the result of ‘noise’ in the video sequence, which may be caused by fast camera

motion, a change in lighting conditions, computer-generated effects, or

anything that causes a perceptual change in the video sequence without being an

actual shot boundary. As can be seen, the algorithm used is quite easy compared

to some previously published examples. This is representative of the fact that

the system is currently a working-progress. Also, previous studies have shown

that simpler algorithms often outperform more complex ones on large

heterogeneous video test sets, due to the absence of “hidden” variables and

simpler relationships between threshold settings and results.