Video Retrieval based on Patterns of Oriented Edge Magnitude

In this work, a video retrieval system is proposed based on POEM (Patterns of Oriented Edge Magnitudes) descriptor. In the first stage, the input video is partitioned into shots based on Gabor moments and keyframes are selected from each shot based on Temporally Maximum Occurrence Frame (TMOF). In the next stage, the POEM descriptor is computed from each keyframe for robust image/frame representation. Given a query frame, the descriptor is obtained from it in a similar manner, and this descriptor is compared with the descriptors of the video keyframes using nearest neighbour matching technique to find the matching keyframe. We have conducted experiments on the TRECVID video segments to exhibit the superiority of the proposed approach for video retrieval applications.