Aiming at the defects of limited receptive field in the deblurring of remote sensing images based on convolutional neural network, which leads to the loss of details and incomplete deblurring of images in the recovery process, a multi-scale remote sensing image deblurring algorithm integrating self-attention is proposed. A multi-input, multi-output U-Net is utilized to simulate a single U-Net into a multi-level joint multi-scale convolution operation to achieve effective extraction of features. A Transformer-based multi-head self-attention module is proposed, which enhances the network's spatial feature extraction and global information capture capabilities by embedding it in the middle position of encoder and decoder. A multi-scale edge loss function is introduced to improve the restoration of image edge details. A fuzzy remote sensing image dataset is constructed for the experiment, and quantitative and qualitative analysis of the experimental results shows that the proposed algorithm is superior to the contrast algorithms. To demonstrate the generalization ability of the algorithm, it is validated on the public dataset GOPRO. The results show that the algorithm is of certain practical significance for effective processing of blurred remote sensing images.