Grain shape is an important feature, which determines the grain quality and yield. Traditional manual determination of grain shape is time-consuming, laborious and subjective. A system for automatically determining the grain shape is presented based on a line-scan camera and a coveyor belt. In order to improve the grain shape determination efficiency, parallel processing technique based on Graphics Processing Unit (GPU) is used. On the basis of Compute Unified Device Architecture (CUDA), the grain shape determination algorithm is optimized. The experiment result shows that the GPU-based accelerated algorithm can achieve a good effect on the measurement efficiency. When the number of grain in an image is approximately 2 000, the optimized algorithm gets a speedup of more than 400 times. Moreover, as the quantity of kernels in the images gets larger, the performance of the grain shape determination algorithm is improved more significantly.