An improved algorithm of 2-D empirical mode decomposition (EMD) in image processing has been presented. It contains selecting extrema of the pixels and interpolation of them in the course of EMD, in which a variance phenomenon of boundary pixels has been discovered. Delaunay triangulation has been used to partition the selected extrema, then replaces the pixels that not contained in the Delaunay polygon through symmetry principle, which can restrain the variance phenomenon that appeared in the cubic spline interpolation. An image has been processed with the improved algorithm, and the result indicates that the standard deviation between the original image and the reconstructed image is 6.667×10-6 . The reconstructed image is in good agreement with the original image. It demonstrates that the improved algorithm presented is accurate and feasible. The application of method of EMD is more and more popular in image compression and de-noising, therefore, the improved algorithm will increase the calculation speed of image processing based on EMD.