In order to reduce the influence of absorption and scattering on tissue fluorescence spectra, the tissue fluorescence and diffuse reflection are simulated under different optical parameters with the Monte Carlo (MC) method, and a fluorescence recovery algorithm based on the tissue diffuse reflection spectrum is proposed. The empirical parameters in the proposed algorithm are coded as a particle in the solution domain, the classification performance is defined as fitness, and then a particle swarm optimization (PSO) algorithm is established to optimize empirical parameters. Skin fluorescence and diffuse reflection spectra of 327 subjects are collected with a tissue detection system for noninvasive screening of diabetes. The fluorescence spectra are recovered by the empirical approach, and the fluorescence intensity before and after recovery is selected as the input variable for the receiver operating characteristic (ROC) curve analysis, which is applied to evaluating the classification performance in diabetes screening. The sensitivity and specificity are 32% and 76% respectively, and the area under the ROC curve is 0.54 when the spectra before recovery are used, while the sensitivity and specificity are 72% and 86% respectively, and the area under the ROC curve is 0.86 when the spectra after recovery are used. The results indicate that using the tissue fluorescence spectrum recovery algorithm based on PSO can improve the application of tissue fluorescence spectroscopy effectively.