Finger multimodal recognition methods, which combine fingerprint, finger-knuckle-print, and finger-vein pattern, offer superior accuracy and robustness. However, current research faces significant challenges due to the instability of image acquisition caused by individual differences and feature space mismatches between modalities. This paper proposes a finger trimodal feature recognition model called FCNet. First, three CNNs are used to extract features from the trimodal images, separately. Then, nodes and node-features are generated based on the CNN feature maps. Next, all the nodes are fused into a crystal-shaped graph. Finally, the graph is fed into the GCN model for further optimisation to obtain a feature graph that accurately describes the finger information. Additionally, to improve the robustness of FCNet, we build a contaminated image dataset to train and test the model. The experimental results demonstrate that FCNet exhibits superior recognition performance and robustness in comparison to other models.