Fluorescence is the main part of background within direct Raman spectrum, and a faithful and truthful method is needed to remove it from the Raman detection, and to provide one pure Raman spectrum. There are two paths for setting background of direct spectra, i.e., baseline fitting and fluorescence finding. The baseline fitting is a currently path whose significant advantage reflects in “visual experience” without additional hardware, nevertheless, “visual” means surface and not essential, so that its results probably may be wrong or unreal. The other way attends to find out the real fluorescence within spectra, but it needs additional design and cost, such as two and more wavelengths in laser sources. Furthermore, some handle methods, reagent for suppress fluorescence or long time photo-bleaching, are insufficient operation and low efficiency. This research focus on the time difference of Raman and fluorescence to find out fluorescence from the direct spectra of a stable system. For the stable system, and within a tiny time cell, a few milliseconds or so, the excitation light does not make the system to be changed significantly and a part of fluorescence has been out of the life cycle at the same time, that is, fluorescence is fading and Raman keep stable within the time cell. So that this difference of the time cell can be treated as a fluorescence infinitesimal and used to distinguish fluorescence and Raman from mixed signal. Based on this principle, the fluorescence photo-bleaching difference approach (FBDA) has been presented to eliminate the background under direct Raman spectra. The steps of FBDA is carried out as follows: Firstly, a series of Raman spectra between tiny time cell is directly measured, and then a series differences for each time cell is computed, these differences are de-noised by low pass filtering, to get the infinitesimals of fluorescence. Secondly, these infinitesimals are mean and normalized to be a fluorescence unit. Thirdly, the total fluorescence can be summed by inverse difference of the unit, in addition, the silence area of Raman, 2 000～2 500 cm-1, which usually do not appear Raman signal, can be used as a benchmark for deciding the gross of fluorescence. Finally, the fluorescence gross deducted from the original spectrum, and the background deduction or baseline correction is completed. The paper takes an example, the Raman spectrum measuring of metformin hydrochloride tablets, to illustrate the FBDA method and its validity. The FBDA is more objective and true than others baseline correction methods, which are considered to work well, such as, ALS and airPLS. Further advances of FDBA are more convenient and less costly than the current fluorescence-finding path, because all of the data for FDBA are collected by existing instruments without any change or addition. It should be noted that the tiny time cell is a key requirement for FDBA, and tiny time cell can ensure the real-time performance of spectrum, the difference of long time would affect the accuracy of the results. In addition, the applicability of FBDA needs further development under the complex background from photochemical reaction and other non-fluorescence.