DREM: Decoupled region energy model for image segmentation

A novel real-time homogeneous region segmentation method is proposed that decouples appearance and geometric properties of the regional energy model into data-driven external and regularization driven internal energy functions. The initial seed locations are selected using local minimal variance estimated using minimum rectangle converging criteria. Region convergence is obtained through an iterative co-ordinate descent technique. Since this method requires only 14 ms per video frame, it will help compress intelligent transportation system video for transmission over mobile networks for surveillance and cloud-based processing.