According to the book "Introductory Raman Spectroscopy" by John R. Ferraro, Kazuo Nakamoto and Chris W. Brown, straight line subtraction is not the only viable option. The following text describes an example of spectrum flattening :

"The five-point smooth is better than the two-point smooth, but the spectrum still has curved sections. In the last case, a four-point polynomial (quartic equation) fit was

applied to the spectrum. Again, the user selects the four points, that are then

fitted with a fourth-degree polynomial, which is subtracted from the spectrum.

The last method produced the flattest baseline for the spectrum."

Therefore different curves may be subtracted. However, the same book has the following notice:

"It should be emphasized that any baseline alterations to correct sloping

backgrounds should be avoided, especially by one new to the field. Important

information may be lost in the flattening process, and important spectral

bands may be distorted."

This is because not only position, but also shape of a peak gives us information about sample.

Generally speaking I think it is much preferred to enhance spectra using better data collection rather than manipulating the spectra after the data is collected.  If the correction is needed because you are using a Raman spectrometer that gives you the ability to adjust the slit width, then it will be better to reduce the slit width in order to enhance the signal to background and the signal to noise ratio.  Furthermore, one can examine the shape of the central peak due to the laser line by examining its characteristics using Raman spectra of samples without Raman peaks below 500 cm-1, such as Si and normalize the shape of the central peak in view of the sample of interest, and then subtract that from the spectra of the sample under investigation to get a better feeling for any possible weak signals that may be washed away in the presence of the tail of the central elastic peak. 

Hi Marcin

It would appear that the Raman spectrum of your sample (i.e. the weak peaks) is superimposed on a broad fluorescence background. In this case, the subtraction of a non-linear background would be appropriate as a means of removing the fluorescence contribution.

Have you tried changing the wavelength of the excitation laser used in the Raman experiment? By careful selection of excitation laser, you may be able to minimize or completely eliminate the fluorescence background. Accordingly, you would not then need to perform a background subtraction on your Raman data.

Hi Marcin, Geoffrey is right. It is not always possible to change instrumentation, so  the removal of broad fluorescence background is a valid option. You might want to try the "adaptive baseline" feature in my free http://Spectragryph.com optical spectroscopy software. This will do it without smoothing or polynomial fitting of the spectrum, just removes the broad background "by magic" :-))