Raman spectroscopy is a non-destructive and rapid detection technology that can provide qualitative and quantitative information of the material. Therefore, it has been widely used in many fields such as medicine and chemical industry. However, the Raman spectrum suffers from the baseline drift due to the background fluorescence of the sample. Moreover, it has a serious impact on the identification of characteristic peaks of Raman spectra and the Raman imaging. At present, there are two methods to solve this problem, that is, improve the experimental methods and numerical processing. The improve the experimental methods include polarization modulation method and high frequency modulation method. However, they suffer from the disadvantages of complicated experimental equipment and difficult detection technology. The numerical processing includes polynomial fitting and wavelet transform. However, it is prone to suffer from the over and under-fitting. In order to solve this problem, we propose the baseline correction algorithm for Raman spectroscopy based on adaptive window spline fitting, which based on the existing equipment and the traditional baseline correction algorithm. Firstly, the optimal search interval of the trough value is obtained based on the peak recognition algorithm and the initial search step, and then the trough recognition algorithm is used to complete the fitting of the trough curve. Secondly, the peak position of the trough curve is obtained based on the optimal search interval and the peak recognition algorithm. Then, the adaptive rectangular window is symmetrically added at this position, in order to delete the peak, and fitting the trough curve. Thirdly, the fitting trough curve is compared with the original Raman spectrum, point by point, and taking the smaller value to fit a new trough curve. Finally, the operation above will continue until the width of the adaptive window is lower than the threshold. Afterwards, the baseline fitting of the Raman spectrum is completed. And then the baseline correction of the sample is obtained based on our algorithm and the traditional methods. It can be seen that our algorithm can effectively eliminate the baseline drift, and some weaker Raman characteristic peaks can be better remaining. Simultaneously, the over and under-fitting is avoided, and the result of baseline correction is good. Therefore, it provides reliable information on the further analysis of the Raman spectrum and the realization of the Raman imaging.