Seam Carving on CUDA architectures

The objective of this thesis is to optimize the Seam Carving method in CUDA (Compute Unified Device Architecture) [5]. We want to determine, if a general method can be effectively transferred into a parallel method, and whether such a step means a better (faster) performance. The method of Seam Carving is a content aware image scaling method. Changing the size is paired with the desire to keep as much relative information in the picture as possible. This is achieved by carving out the optimal seams, which contain the least information. We find that the suitability of the method is dependent on the content of the image, and suggest optimal use. CUDA is a highly parallel architecture that runs on NVIDIA graphic cards. It exploits the ability of graphic cards that can simultaneously execute a large amount of threads over a certain set of instructions and data. We present its limitations and weaknesses but also major strengths in implementing the algorithms that are highly parallel in nature and have a high density of arithmetic operations. We determine that the CUDA implementation gives much better results when used on larger images. In addition, the implementation of the algorithm relieves the load on the host CPU (central processing unit)