Creating fake pictures becomes more accessible than ever, but tampered images are more harmful because the Internet propagates misleading information so rapidly. Reliable digital forensic tools are therefore strongly needed. Traditional methods based on hand-crafted features are only useful when tampered images meet specific requirements, and the low detection accuracy prevents them from using in realistic scenes. Recently proposed learning-based methods improve the accuracy, but neural networks usually require to be trained on large labeled databases. This is because commonly used deep and narrow neural networks extract high-level visual features and neglect low-level features where there are abundant forensic cues. To solve the problem, we propose a novel neural network which concentrates on learning low-level forensic features and consequently can detect splicing forgery although the network is trained on a small automatically generated splicing dataset. Furthermore, our fusion network can be easily extended to support new forensic hypotheses without any changes in the network structure. The experimental results show that our method achieves state-of-the-art performance on several benchmark datasets and shows superior generalization capability: our fusion network can work very well even it never sees any pictures in test databases. Therefore, our method can detect splicing forgery in realistic scenes.