A quick scene classification method based on compact encoding of video feature sequence
Akio Nagasaka, Takafumi Miyatake · Systems and Computers in Japan · 2000
This article proposes a method of real-time scene classification of motion images by detecting the features of an input image that are identical with already stored images. In this method, the feature of an image is represented by an average color of each frame, and continuous input images are compressed and memorized in a time series. Then an image section (no limit in length) that is identical with the latest input image is found from the memorized images. The proposed method is installed in a typical personal computer, and a classification of commercial TV programs has been investigated experimentally. The results show that features of the images were memorized at a rate of 20 bytes/s (average) with no error. The processing time was 10 ms/frame (average) for 8 hours of TV programs. This suggests that TV images for several days can be processed in real time by using this method. © 1999 Scripta Technica, Syst Comp Jpn, 31(1): 102–108, 2000