The modulation format is a key parameter that influences the monitoring of the intercepted signals. Automatic modulation classification (AMC) is utilized to recognize the modulation format of the intercepted signals. However, most recent AMC methods neglect the complementarity accross different features. In this paper, we propose a novel feature fusion based AMC scheme using the convolutional neural network (FFCNN). Fused feature is generated by concatenating the two-dimensional spectrum correlation function (SCF) images and the graphic constellation (GC) images. Moreover, the FFCNN classifier is adopted to obtain more discriminative representations, leading to improved final modulation classification performance. Extensive simulations demonstrate that the proposed FFCNN scheme outperforms other recent methods.