Prediction of enhancer-promoter interactions (EPIs) is key to regulating gene expression and diagnosing genetic diseases. Due to limited resolution, biological experiments perform not as well as expected while precisely identifying specific interactions, giving rise to computational biology approaches. Many EPI predictors have been developed, but their prediction accuracy still needs to be enhanced. Here, we design a new model named EPIMR to identify enhancer-promoter interactions. First, Hilbert Curve is utilized to represent sequences to images to preserve the position and spatial information. Second, a multi-scale residual neural network (ResNet) is used to learn the distinguishing features of different abstraction levels. Finally, matching heuristics are adopted to concatenate the learned features of enhancers and promoters, which pays attention to their potential interaction information. Experimental results on six cell lines indicate that EPIMR performs better than existing methods, with higher area under the precision-recall curve (AUPR) and area under the receiver operating characteristic (AUROC) results on benchmark and under-sampling datasets. Furthermore, our model is pre-trained on all cell lines, which improves not only the transferability of cross-cell line prediction, but also cell line-specific prediction ability. In conclusion, our method serves as a valuable technical tool for predicting enhancer-promoter interactions, contributing to the understanding of gene transcription mechanisms. Our code and results are available at https://github.com/guofei-tju/EPIMR.