In complex environments， traditional infrared detection methods are severely limited， necessitating the fusion of polarization technology with infrared technology. This fusion is aimed at existing convolutional neural network methods， which lack the capacity to extract multi-scale information and fuse subnetworks adequately. The objective is to enhance the quality of infrared polarization fusion images and improve the target recognition capability of infrared imaging technology in complex backgrounds. This paper proposes a progressive infrared polarization image fusion method （PIPFuse） based on multi-scale features. Firstly， the feature extraction component employs a semantic extraction module and a multiscale dense block， which are utilized to extract semantically enhanced multiscale depth features. Secondly， to reduce the information loss and enhance the salient information， the fusion subnetwork incorporates a progressive difference information enhancement fusion module for the feature fusion. Finally， the final fused image is obtained by decoding the fused features. In comparison to nine classical image fusion methods， this method demonstrates superior performance in six evaluation indexes. Furthermore， the subjective visualization of the target texture is more distinct and exhibits higher contrast.