Curvelet transform is a new multi-scale transform developed upon wavelet transform. Beside scale and position, its constructive factors still include directions. All these make curvelet transform have a better directional characteristic. Based on these properties, we transform seismic data into curvelet domain, apply a window-shrinking algorithm to attenuate the random noises and improve the quality of seismic data finally. Both model and real data all obtain good results. It is available and necessary to set shrinking window size as 3*3 or 5*5 and the value of σ as 6-7% of maximum amplitude in seismic data denoising.