Superpixel Based Inter-Frame Prediction for Video Coding

Traditional video codec uses the block-based approach for motion estimation and compensation, which is unable to capture the true underlying motion. Blocks consisting the boundary of moving object need more bit for better prediction of the current frame because real-world objects are not block-based and their motion is not translational. One of the main goal of the current standard (HEVC) is to minimize the prediction error energy by using quad-tree splitting method. But if motion discontinuity exits, quad-tree based splitting method uses a large amount of motion bit-rate and also introduces a potential coding overhead in the video encoder and decoder. In this paper, we propose a segmentation-based method of using arbitrary shape superpixel and affine motion registration technique for reducing the prediction error energy between the current and reference frame. This method uses the already encoded reference frame for predicting the current frame. The experimental result shows that good subjective quality with PSNR up to 27.71dB can be obtained using this approach.