Molecular features play a crucial role in molecular property prediction. Although molecular fingerprints (FPs) and graph neural networks (GNNs) like GAT are widely used, both have limitations. FPs lack structural context, while GATs struggle to capture cooperative interactions between adjacent atoms, making it difficult to model complex molecular structures. To overcome these challenges, we propose a new three-branch hybrid deep learning framework, FPG2DNet, for molecular property prediction. Firstly, four types of molecular fingerprints capture chemical substructures, while a GCN-GAT hybrid extracts local structural features. Simultaneously, a pre-trained ResNet-50 learns global information from molecular images. These multi-source features are fused into a unified representation and fed into a prediction layer for molecular property prediction. Experiments conducted on 10 public datasets and 14 cell-based phenotypic screening datasets show that FPG2DNet achieves state-of-the-art performance on 6 public and 8 screening datasets. Further analysis of the contribution of the module proves the complementarity and effectiveness of different feature extraction strategies. Overall, FPG2DNet provides a new method for accurately predicting molecular properties and provides valuable support for drug design and optimization.