Abstract: To address the degradation of infrared images caused by translation motion, radial motion, and turbulence blurs, a transformer-based infrared image deblurring neural network using deformable sampling and multidirectional shifting windows (DMST) is proposed. With deformable sampling and a multi-direction window shifting mechanism, the proposed method allows self-attention windows to better conform to the irregular distribution of informative features in images while enhancing the network's perception of interactions with global features. On three infrared datasets featuring translation motion, radial motion, and turbulence blurs, the proposed method achieves peak signal-to-noise ratios (PSNR) of 35.69, 30.19, and 33.55 dB, respectively, and structure similarity index measure (SSIM) values of 0.9566, 0.8882, and 0.9230, respectively. The experimental results indicate that compared to the advanced deblurring methods reported in recent years, the proposed method exhibits superior deblurring performance for infrared images.