In order to avoid the problem that the reconstructed image is easy to cause under-deblurred or over-deblurred due to inaccurate blur kernels estimation,amulti-scale encoder-decoder network deblurring model is proposed.Firstly ,the skip connection and multi-scale recurrent connection are added to the tradi tional encoder-decoder network,and combine the image feature of each layer to make the network training more stable.Then,the proposed encoder-decoder is combined with the nested residual network.The image information of different scales is further extracted by the method of “coarse-to-fine” to enhance the texture details of the image.Finally,the image is directly blind deblurred by using the end-to-end mode to output a clear image.The multi-scale network model has fewer parameters,the network structure is simpler,and does not need blur kernels estimation.The experimental results on the GoPro and Khler datasets show that the PSNR of the proposed method is improved by about 6.61 dB and 2.13dB compared with the traditional SFD method.Compared with MRF CNN、Multi-scale CNN、DeblurGAN and other methods based on CNN,the maximum increase in PSNR is about 5.61dB and 1.59dB,respectively.The experimental results are better than the other fou r methods.The restored image is closer to the real image,and the texture details are clearer.It is an effective motion image deblurring method.