A Multi-Operator Retargeting Scheme for Compressed Videos

This research presents a multi-operator retargeting scheme, in which content-based cropping, seam carving/insertion and scaling are applied sequentially to adjust the video frames to the target resolution. Assuming that the video to be processed is encoded with H.264/AVC, compressed-domain data in the bitstream are utilized to classify video shots into different types for further processing. SLIC superpixels are formed to identify boundaries of objects and a saliency map will determine the visual significance of frame areas for appropriate cropping. A motion feature map is constructed to locate moving objects or contents so that possible distortions on them can be avoided. For static scenes, a local-significance-aware seam carving scheme based on one-dimensional gradients is applied. A SSIM-based blur detection method is also developed to extract sharp foreground objects. The experimental results show that the proposed method performs well in various kinds of video shots.