It is difficult to fully recover the image details using the existing image super-resolution reconstruction methods. Furthermore, the reconstructed images lack a hierarchy. To address these problems, an image super-resolution reconstruction method based on self-attention deep networks is proposed herein. This method, which is based on deep neural networks, reconstructs a high-resolution image using the features extracted from a corresponding low-resolution image. It nonlinearly maps the features of a low-resolution image to those of a high-resolution image. In the process of nonlinear mapping, the self-attention mechanism is utilized to obtain the dependence among all the pixels in the images, and the global features of the images are used to reconstruct the corresponding high-resolution image, which promotes image hierarchy. During the deep neural network training, a loss function comprising a pixel-wise loss and a perceptual loss is utilized to improve the image-detail reconstruction ability of the neural network. Experiments on three open datasets show that the proposed method outperforms the existing methods in terms of image-detail reconstruction. Furthermore, the visual impression of the reconstructed image is better than that of the images reconstructed using other existing methods.