Binary fringe projection image is widely used in high speed and high precision 3D surface measurement, and improving the sinusoidal properties of binary coded fringe is of positive significance for improving the accuracy of 3D surface measurement. The traditional and improved error diffusion kernel mostly uses universal diffusion to check fringe image for binary coding, while the influence of image features and projective defocus degree on phase extraction accuracy is less considered. First, the genetic algorithm is used to find the better error diffusion kernel coefficient. Second, the optimization objective function related to defocus degree and sinusoidal fringe period is constructed by linear fitting. Finally, the sinusoidal error diffusion kernel of the optimized binary coded fringe is obtained. Simulation and experimental analysis show that the error diffusion nuclei with minimum phase error are different in different periods under different fuzzy degrees, which confirms the binary value of diffusion check image coding quality is related to image features. Experiment further proves that the phase error of the proposed algorithm can be reduced by 43.86%, 64.37% and 50.10%, respectively, compared with the universal Floyd-Steinberg diffusion method, under three defocus degrees. Compared with the improved Floyd-Steinberg diffusion method, the phase error of the proposed algorithm can be reduced by 13.51%, 18.48% and 17.65%, respectively.