Abstract: In view of the tunnel environment with poor lighting conditions, the single-modal tunnel leakage image can no longer meet the needs of disease perception. Therefore, based on the cross-modal heterogeneous image fusion technology of visible and infrared, a heterogeneous image fusion algorithm D-DDcGAN based on dual-path and dual-discrimination Generative Adversarial Network was proposed. Firstly, the generator structure and the double discriminator structure of the two-stream input architecture were designed by using the Generative Adversarial Network (GAN) model to enhance the resilience of the Generative Adversarial Network. Secondly, Structure Similarity Index Measure (SSIM) loss function is introduced to improve Joint Low-Rank Sparse Decomposition. (JLRSD) is used to decompose the features of multi-source tunnel leakage images. Finally, the low-rank graph and sparse graph generated by the JLRSD were input into the GAN to obtain the fused tunnel leakage image. In the experiment, 5000 tunnel leakage images of infrared and visible are selected for heterogeneous fusion verification. The results show that: Compared with the classical Non-Subsampled Shearlet Transform (NSST), Low-Rank Sparse Decomposition (LRSD), JLRSD and FusionGAN models. The D-DDcGAN algorithm has the six quantitative evaluation indicators of Information Entropy, Mutual Information, Average Gradient, Spatial Frequency, Peak Signal-to-Noise Ratio and Root Mean Square Error increased by 1.887%, 0.987%, 9.547%, 12.732%, 1.659% and 2.299%, respectively. The six indicators have obvious advantages. The experimental results show that the proposed method enriches the tunnel leakage image information, increases the target identification, and helps to improve the tunnel leakage perception ability.