In order to realize automatic modulation recognition, an algorithm based on geometric shape was put forward. At first, signal samples were processed with discrete wavelet denoise method. After that, signal samples were rearranged from 1-D to 2-D and multi-fractal dimension vectors were obtained. At last, feature vectors projected on the Fisher hyper plane were classified with maximum likelihood method. The simulation result showed that high rate of recognition is achieved at signal-noise rate from -3 dB to 20 dB.
