The two-dimensional optical fiber spectral images are the observation results of a spectrometer in an astronomical telescope system, and they are followed by a series of post-processing steps to produce the common one-dimensional spectra. Owing to the optical distortion caused by the spectrometer and CCD, obvious bending can be seen from the two-dimensional optical fiber spectral images, especially at both ends of the fibers. So far for this bending problem, there has been no good solution, nor in any reference to see any relevant work. And this kind of bending can cause great troubles to the subsequent spectral lines extraction and other works, which will affect the wavelength calibration to a great extent, as well as the accuracy of the one-dimensional spectrum. In this paper, a normal mapping method has been used to correct the bending phenomenon of the two-dimensional optical fiber spectral images. The method deals with each optical fiber spectrum in the images separately, and corrects the spectrum into vertical straight. The first step is preprocessing, which is to extract each fiber’s centerline. Then the centerline is taken as a smooth curve to obtain the normal direction at each point. The ideal vertical line is based on a point (usually the most salient point) of the whole centerline, and all the points of the optical fiber centerline are projected onto the ideal vertical line along the relevant normal directions, so as to realize the alignment of the optical fiber centerline points, which is the second step. The third step is to handle the whole fiber spectrum. Because the fiber’s width is 7 pixels in the two-dimensional optical fiber spectral image, the fiber’s centerline moves 3 pixels one by one to the left and right, respectively, realizing the above 2 steps and obtaining the straightening result of the whole fiber spectrum. In the whole process, there are two key problems to be noticed: one is the homogenization problem of the coordinate points, and the other is the accuracy maintenance problem of pixels’ values. The uniformity of the sitting punctuation is due to the use of the normal mapping in this method, which results in uneven density of the points after the formation of the alignment, which is unfavorable to the subsequent processing. The solution is to use the cubic spline interpolation method to achieve the density uniformity of the point on the line, to ensure a series of integer coordinates to facilitate the subsequent processing. To maintain the accuracy of pixels’ gray values, the 64-bit high precision number of pixels’ gray values is always kept in the process of the interpolation calculation. At the end of the implementation of the method, it is necessary to intercept both of the two ends of the spectrum to keep the end-to-end consistency, removing the pixels extending out of the height range of the image, and retaining only the pixel points in the image range. If without this process, the mapped vertical lines are different in length, so that causing difficulties in subsequent processing. This paper deals with the two-dimensional optical fiber spectral images, solves the problem of brightness deviation in the process of extracting optical fiber centerlines by using curve fitting method, obtains the corrected two-dimensional spectra after straightening, and makes a follow-up comparison between the one-dimensional spectra. In the comparison, it can be seen that the spectral line differences before and after correction are more obvious than those of the spectra at both ends of the curve. The experimental results show that this method can completely improve the bending of the spectra. It can be seen that the changes at both ends are large and the changes in the middle are very small, which accords with the basic understanding of the observation. Furthermore, due to the dislocation of pixels and the interpolation of pixels’ values, the effect of the superposition of the pixels’ values changes significantly. Therefore, by observing the movement of the spectral lines after the calibration, it is proved that this method has an important influence on the precise acquisition of the wavelength positions of the characteristic spectral lines. This paper creatively designs the method of automatic bending correction of two-dimensional optical fiber spectra and verifies the validity of the method by experiments.