Abstract: Aiming at the problems of low efficiency and poor accuracy in the evaluation of roundness errors of shaft parts, a visual evaluation method for roundness errors of shaft parts based on improved Zernike moments is proposed. Firstly, a visual inspection platform for shaft parts is built, and different cross-section images are taken under the control of a displacement platform. The camera is calibrated using a checkerboard calibration method to correct distortions and calculate pixel equivalents. Secondly, based on the Sobel operator, the measurement accuracy is improved through a three-gray edge transition model, and the optimal threshold is determined using the Otsu method to improve the subpixel detection method of Zernike moments. Finally, the key coordinates of the contour are processed using an improved PSO algorithm for roundness error evaluation, and measured with a three-coordinate measuring machine. The experimental results show that the visual evaluation method for roundness errors has an error within 3 μm compared with the measurement results of CMM, and has good stability, making it feasible for practical use in roundness error evaluation.