The single image dehazing is a challenging problem because of the acute degradation of the hazing image. Traditional CNNs-based dehazing models suffer from two fundamental issues: one is that the convolutional layers of dehazing networks are often content-independent and cannot effectively learn remotely dependent feature information. The other is that transparent images are used to training the dehazing networks without utilizing hazing information. Therefore, we introduces a multi-layer fusion network based on contrast learning that integrates multiple models into a single layer centrally and improves the image dehazing performance. First, we utilize the MSBDN-DFF model as the original feature extraction of initial dehazed images; then, with the obtained features, we further predict the hazy-free results of the atmospheric scattering module and the four hazy layer separation modules; Finally, we add a contrast learning mechanism to make our model effective in generating highquality hazy-free images to fully use the information from fuzzy and clear images. An extensive evaluation shows that our proposed method is effective in the RESIDE dataset and outperforms current techniques. And our method improves the best-published PSNR value by 0.55 dB on the SOTS test dataset and the original model by 1.12dB/0.83dB on the SOTS indoor/outdoor test dataset, respectively. A comprehensive analysis of our method is also performed through additional experiments and detailed ablation studies.