A Fast Image Segmentation Algorithm for Interactive Video Hotspot Retrieval

This paper presents a fast image segmentation algorithm for extracting areas-of-interest (”hotspots”) from video sequences. Because video hotspots, which are the visual analogues of HTML hyperlinks, highly depend on a user’s interest, it is very difficult to automatically extract the exact outlines of objects of interest in videos. Our approach of interactive video hotspot retrieval is to provide candidate hotspots with segmented regions of video frames when requested, and web designers using the tool for web site creation may select the area (or areas) he is interested in as a video hotspot. This segmentation algorithm follows a three step coarse-to-fine scheme. The original image is first divided into blocks and a block-merging process is applied to get a coarse segmentation. Then we represent the presegmented image as a graph where the value of edges in this graph are evaluated based on multi-features, and by using a minimum spanning tree searching algorithm, we merge the vertices in the graph to generate the segmented regions of the image. Finally, we refine the boundary of segmented regions on the pixel level using an improved K-means clustering algorithm. This segmentation algorithm is fast and efficient and our final segmentation results compare favorably with the results obtained using manual outlining. We also provide a convenient way for the web page designer to select multiple regions as a video hotspot. Our method allows the designer to click on an image during video playback and to get a segmented image almost instantly. The extracted video hotspots can then be used for video hyperlinks, an extension of the general concept of hyperlink to interactive videos. This work was done when the first author was an intern in Siemens Corporate Research.