In image acquisition process, the image blur caused by the moving subject or the camera itself will have a negative impact on the subsequent high-level vision tasks. Aiming at the problem that the current deep learning image deblurring method cannot balance the deblurring effect and efficiency, a multi-scale recurrent attention network was proposed, which used separable convolution to reduce the amount of parameters, and improved the attention module to allocate computing resources reasonably. Layers were used for dense connection to improve parameter utilization efficiency, and edge loss was introduced to improve the edge detail information in the generated image. Experiments prove that the proposed method has good generalization performance and robustness. Compared with the typical methods in recent years, the SSIM and PSNR have increased by about 1.15%, 0.86% and 0.91%, 1.04% on the Lai dataset and Köhler dataset, respectively. The average single frame running speed on the GoPro dataset is nearly 2.5 times faster than similar methods.