Accelerating and Characterizing Seam Carving Using a Heterogeneous CPU-GPU System

Seam carving has been widely used for contentaware resizing of images and videos with little to no perceptible distortion. Unfortunately, for high-resolution videos and large images it becomes computationally unfeasible to do the resizing in real-time using small-scale CPU systems. In this paper, we exploit the highly parallel computational capabilities of CUDA-enabled Graphics Processing Units (GPUs) for accelerating the content-aware resizing of videos and images. The performance results show that our implementation of the seam carving algorithm achieves up to 100x and 14x speed-ups on the computationally-intensive part of the algorithm compared to the faster single-threaded and the faster multithreaded CPU implementations, respectively, on the systems tested. The overall resizing operation is over 6x and 2x faster than the best single-threaded and multithreaded CPU implementations, respectively, which demonstrates the potential to resize videos and large images in real-time.