Due to the over-fitting problem caused by imbalance samples, there is still room to improve the performance of data-driven automatic modulation classification (AMC) in noisy scenarios. By fully considering the signal characteristics, an AMC scheme based on adaptive fusion network (AFNet) is proposed in this work. The AFNet can extract and aggregate multi-scale spatial features of in-phase and quadrature (I/Q) signals intelligently, thus improving the feature representation capability. Moreover, a novel confidence weighted loss function is proposed to address the imbalance issue and it is implemented by a two-stage learning scheme.Through the two-stage learning, AFNet can focus on high-confidence samples with more valid information and extract effective representations, so as to improve the overall classification performance. In the simulations, the proposed scheme reaches an average accuracy of 62.66% on a wide range of SNRs, which outperforms other AMC models. The effects of the loss function on classification accuracy are further studied.