Parallel processing for accelerated mean shift algorithm with GPU

In order to overcome the shortcomings of the mean shift method for its intensive computational requirement, an improved GPU-based mean shift algorithm is presented. By the novel algorithm, first k-means algorithm is used to pre-classify the source image with a re-sampling, then mean shift runs on the narrowed re-sampled data sets. As a result, the algorithm complexity can be effectively reduced. In addition, through the further study of k-means and mean shift, and with general purpose computation of GPU, k-means and mean shift are respectively parallel processed. Experimental results show that by preprocessing the images, the accurate recognition rate of Geometric Model Finder for intensive noises, low SNR images is effectively improved. At the same time, the efficiency of the modified mean shift algorithm is greatly improved, with the average processing time nearly 40 times faster. It meets the requirement of high-speed machine vision inspection.