the training. All slide frames are collected from our lecture
video database. To create a non-slide frame set with varying
image genres, we have collected additional 3,000 images
from flickr.3 Moreover, about 2,000 non-slide video frames
have been collected from the lecture video database. The
test set consists of 240 slide frames and 233 non-slide
frames, which differ from the training set.
In the classifier training the SVM-parameters were determined
by using the grid-search function. To calculate the
image intensity histogram, 256 histogram bins were initially
created corresponding to the 256 image grayscale values.
Then the normalized histogram values were applied to train
the SVM classifier. We have evaluated the normalization
factor (cf. Fig. 5), which has proven to serve best when set to
1,000.4
The comparison results of two features are illustrated in
Table 1. Both features achieved a good recall rate for recognizing
slide frames. However, compared with the HOG feature
the intensity histogram feature showed a considerable
improvement in precision and F1 measure.