<p>Image dehazing aims to reconstruct a clear image from its hazy version, playing a crucial role in object detection and road surveillance. While current dehazing methods excel on homogeneous thin haze, their performance degrades significantly on non-homogeneous hazy images. Therefore, a model selection and image fusion based non-homogeneous image dehazing algorithm (MSIF-Net) is proposed in this paper. First, we deem that a non-homogeneous hazy image consists of multiple image patches with varying haze concentrations and scenarios. Then, a novel iterative training scheme is devised to get multiple image enhancement network models tailored to different hazy scene modes. Through the image enhancement models, multiple initial enhancement results can be obtained. Finally, an unsupervised image fusion network is designed to integrate the advantageous regions of all initial enhancement results and generate the final dehazing result. Compared to the state-of-the-art methods, the proposed MSIF-Net achieves better performance metrics and visual results on both synthetic and real-world hazy datasets, which can prevent over-enhancement in thin hazy areas and effectively boost structural contrast in dense hazy regions.</p>