Real-time interactive object extraction system for high resolution remote sensing images based on parallel computing architecture

Random Walks has less interaction, better accuracy and higher computing independency. We introduce local intensity entropy to modify the weight function in Random Walks, in order to consider not only the intensity change of adjacent pixels, but also the statistical features of regions. Then we put forward a real-time interactive object extraction system for high resolution remote sensing images based on improved Random Walks method, and implement this system on general-purpose GPU with nVidia CUDA platform. Experiment results show that the improved Random Walks we provide could accurately extract the boundaries of residential area, water area, plant area as well as road networks. The whole system is built on NVidia 8800GTX GPU using CUDA platform, and still achieves real-time performance when dealing with high resolution RS images larger than 100M pixels.