Comparing keyframe extraction for video summarization in CPU and GPU

Most of the information is captured through multimedia techniques. Videos contain many frames which might be redundant. Since processing of many frames is involved, these redundant frames must be removed for better and efficient results. Summarizing these frames by removing similar frames can speed up processing. In this paper video summarization is achieved by generating key frames. Key frames are generated using discrete wavelet transforms (DWT) technique and we subtract background from the keyframes to get region of interest. A video of 920×720 resolution and length 120 second was used as test video and the run-time was 111 second in CPU and 60 second in GPU. The speed up is nearly 100%. A HD video which took 23 minutes in serial implementation to extract foreground object from key frames generated was reduced to 7 minutes using GPU acceleration.