The analysis of measured spectra, when it reduces to finding the positions and heights of spectral lines, becomes an extremely difficult problem when the lines are closely spaced and broadened to the point where overlapping occurs, and explicit models for the shape of the peaks are unknown. For the widely used nonlinear fitting methods, good starting values cannot be found by direct inspection of the spectra. Sometimes, it is even impossible to determine the number of overlapping spectral lines. Resolution enhancement procedures aim to undo the broadening in order to go back towards the original situation. Traditional resolution rules-of-thumb, such as the Rayleigh, Houston and Sparrow criteria, must be replaced by computational procedures which invoke appropriate assumptions about the broadening of the spectral lines. Measured spectra can be viewed as convolutions between known smooth kernels and discrete line spectra. In this way, the determination of positions and heights of the spectral lines can be recast as a deconvolution problem. Unfortunately, the theory which guarantees the effectiveness of deconvolution requires smoothness of the solution and, as a consequence, does not apply to discrete line spectra. However, discrete line spectra only need to be approximated by spectral peaks which satisfy the smoothness criteria required by the regularization theory in order to formulate a suitable deconvolution framework. In this paper, the corresponding deconvolution problem is solved effectively using a numerical differentiation procedure which plays a derivative spectroscopy enhancement role. When compared with similar deconvolution methods, its enhanced performance is confirmed by theoretical error analysis which identifies a sense in which this method has best possible convergence rates. In practice, the method provides good starting values for the positions of the lines for the nonlinear fitting methods. Sometimes, the ‘starting values’ are so good that a subsequent nonlinear fitting step is not necessary. By modelling the spectra as probability measures in an appropriate Hilbert space framework, estimates are derived that characterize the nature of the smoothing performed by the regularization recovery methodology proposed. The paper concludes with an examination, using the Hilbert space framework, of the trade-off between resolution and the various types of error that occur in a measured spectrum.