Automatically Categorizing Commercial Segments Using Multiple Computer Vision Techniques

Lake Brantley High School
Eduardo Salazar Salinas
Joseph Patrick Wilson
Computer Vision , key frames


The purpose of Computer Vision is to understand the methods by which humans\r
process visual information and likewise to create computer algorithms similar to these\r
processes. Through careful observation, a computer algorithm was developed to mimic\r
how humans recognize logos in television commercials. After visual analysis of\r
numerous commercial sequences, it was hypothesized that the key frames (frames in\r
which the logo resides) could be found using the intersection of color histograms; the\r
logo region could be found using the edge density within the key frames; and the logo\r
could be identified utilizing a correlation method with a database of stored logos, scaled\r
to different levels using Bilinear Interpolation.\r
Color histograms were implemented using one-dimensional arrays with 24 bins;\r
key frames were determined by calculating the intersection between consecutive frames’\r
color histograms. The edge density was calculated by convolving the key frame with\r
the number of edge pixels within a 21X21 area. The identification of the logo was\r
determined by computing the Sum of Square Differences between the logo region and\r
the database of logos on different scales; SSD values were normalized for different\r
The algorithm was tested on 14 different sequences and determined the key frame\r
with 80% accuracy. By segmenting the sequence into two key frames, the algorithm\r
generated 93% accuracy. The algorithm also identified the logo region with 93%\r
accuracy. The identification of the logo yielded anomalous results. These data suggest\r
that motion between consecutive frames in commercial segments decreases around the\r
display of the logo. They also suggest that the logo region has the most visible edges\r
within the key frame.\r
Future study includes a complete overhaul of the logo recognition algorithm. The\r
correlation algorithm (SSD) does not work accurately enough to be used. Therefore, the\r
next step is possibly to look at the edge information about the key frames. As the Canny\r
algorithm determines the edges of an image, it has to determine the direction (or\r
orientation) of the edges. Therefore, a proposed study includes utilizing an edge\r
orientation histogram of the database of the logos and the key frames. This would mean\r
that the algorithm would identify the logo in the key frames by matching edge\r
orientation histograms.

「為配合國家發展委員會「推動ODF-CNS15251為政府為文件標準格式實施計畫」,以及 提供使用者有文書軟體選擇的權利,本館檔案下載部分文件將公布ODF開放文件格式, 免費開源軟體可至LibreOffice 下載安裝使用,或依貴慣用的軟體開啟文件。」

檔案名稱 檔案大小 格式
Automatically Categorizing Commercial Segments Using Multiple Computer Vision Techniques 130 KB Adobe Reader(Pdf)檔案