This paper aims to propose a multi-scale dense convolution network (SRMD) to solve the problem of low utilization of convolution feature map and low reconstruction quality of high-power image. In this paper, the dense connection module of SRDenseNet was improved, and the batch normalization layer was removed. Referring to the existing network, the information integration layer of multi-scale feature extraction layer and 1×1 was designed to form a multi-scale dense convolution module. SRMD stacked 64 low-level feature images through a multi-scale feature extraction layer, and then stacked 1024 feature images through 8 multi-scale dense convolution modules after dense connection. Finally, SRMD output super-resolution reconstruction images through information integration and sub-pixel convolution modules. In this paper, the test is carried out on Set5, Set14, B100 and U100. The peak signal-to-noise ratio of SRMD reconstructed image is 30.1570, 26.9952, 25.7860 and 23.4821 dB, respectively, and the structural similarity is 0.8813, 0.7758, 0.7243 and 0.7452. Compared with the existing networks, SRMD, DRCN and VDSR have the same performance, superior to SRDenseNet and BiCubic methods.