Since the existing image enhancement model based on the deep network directly learns the mapping relationship between the degraded image and the clear image,and ignores the constraint of the fidelity term in the observation model,the reconstructed image has false texture and loss of details.This paper proposes an improved deep network for infrared image enhancement,which embeds the denoising network module into an iterative-based enhancement task,and achieves data consistency constraint by interleaving and optimizing the denoising auto-encoder module and the back projection module.Our proposed deep network can not only use the deep feature to describe the priori information,but also use the consistency priori of the observation model.The experimental results show that the proposed algorithm can achieve very competitive reconstruction results in image denoising and deblurring tasks,and can also achieve clear reconstruction results in low-contrast areas such as point object.