The traditional mismatching eliminating method RANSAC has poor performance when used for the mismatching eliminating of the cylindrical images. A mismatching eliminating algorithm based on curve fitting was proposed in this paper in order to solve this problem. First, SURF features were extracted from two images and were matched using the nearest neighbor method. Then, the matching point set was sorted by ordinal position. At last, the slope set and the Euclidean distance set of the lines build by every two matched feature points were done curve fitting respectively to remove the mismatching points. Experiments show that the proposed method has improved the matching point set number close to the edge of the cylinder after removing, and improved the precision of cylindrical matching obviously