In order to meet the requirement of the engineering application about the real-time image processing, according to the one billion pixel transient cloud imaging system which has been designed based on the combined structure of a concentric spherical lens and micro camera mosaic array, an adaptive image mosaic algorithm of parallel acceleration based on the compute unified device architecture(CUDA) and prior information has been proposed. First, the imaging overlap region of the adjacent micro camera has been calibrated with high-precision four-axe calibration table, and the speed-up robust features(SURF) method has been used to extract the candidate feature points of the overlap region. Then, the approximate nearest neighbor(ANN) search algorithm based on random K-D tree which has been accelerated by the CUDA basic linear algebra subroutines(CUBLAS) is used to obtain the initial matching points. Finally, the improved parallel progressive sample consensus(IPROSAC) algorithm is used to eliminate the false matching points and estimate the parameters of the space transformation matrix, and the spatial geometry transformation relationship has been obtained about mosaic images. Experimental results indicate that the image mosaic time is 287 ms and the speed is improved about 30 times compared with serial algorithm using CPU.