Abstract: Aiming at the problem of residual network super-resolution reconstruction by residual algorithm, an improved deep composite residual network model for residual calculation was proposed. In this research, the original residual block was improved, which could make full use of all the convolutional layer feature information inside the residual block to improve the quality of the generated image; a double-layer composite structure was set to deepen the depth of the model structure, it could enhance the feature extraction of the image by the model and could extract more image features; the image feature information was enhanced using the method of transfer learning through transfer learning in the deep network structure, making the performance of the model more stable. The application experiment of the Tiangong-1 grayscale image show that the improved deep residual dense network proposed in this study performs well in the Tiangong-1 grayscale image super-resolution reconstruction, and has application value and research significance in the field of satellite imagery.