Abstract: To solve the problems of fuzzy details and low contrast of low-quality infrared images, a parallel multifeature extraction network for infrared image enhancement is proposed, and a structural feature mapping network and a two-scale feature extraction network are designed. The structural feature mapping network is used to establish the global structural feature weight to maintain the spatial structure information of the original images. The two-scale feature extraction network using multiscale convolutional layers and the attention mechanism fused dilated convolutions is applied to enhance the attention on contextual information, improve the feature extraction capability for regions of interest, and simultaneously learn feature information of different scales, complete the exchange of information of the two scales, and then generate a target enhancement map to achieve adaptive enhancement of detailed texture of target areas. Experiments have proven that the proposed method can effectively improve contrast, avoid overenhancement, enrich image details and textures, and reduce artifacts and halos. Compared with typical traditional methods and deep learning methods, the PSNR and SSIM on the BSD200 dataset are increased by approximately 37.35%, 2.1% and 25.94%, 3.15%, and increased by approximately 30.62%, 1.04% and 24.83%, 2.08% on real infrared images. The proposed method also has good generalization performance on low-quality images with different contrast factors as well.