Abstract: It is a very challenging problem to extract the amplitude and occurring time of weak nuclear pulse signals in the existence of intensive background noise. To solve this problem, this paper proposes a pulse signal estimation method based on Gabor transform and sparse representation. Firstly, it builds a pulse signal representation dictionary through the Gabor decomposition of mononuclear pulse signal samples. Then it eliminates the fluctuation of the Gabor bases, which is caused by the detector variation and the measurement noise, by using K-SVD algorithm, and learns a self-consistent over-complete dictionary which is used to represent the useful signal being overwhelmed in the background noise. Finally, it reconstructs the desired signal by an improved OMP algorithm, greatly attenuates the noise and achieves the goal of extracting the weak nuclear pulse signal. The effectiveness and efficiency of the proposed method are verified through simulations and experiments on a CsI(Tl) detector. Results confirm that the proposed method outperforms the traditional Salley-Keys smoothing and Kalman filtering methods with smaller estimation errors of the amplitude and peak occurring time of the concerned nuclear pulse signal.