The super-resolution algorithm based on the convolutional neural network has great advantages compared with the traditional super-resolution algorithms. But there are still some problems to be improved, such as long training time, lacking of image texture reconstruction and so on. Owing to this, on the basis of the original convolutional neural network super-resolution reconstruction algorithm, the following optimizations are carried out. The original rectified linear unit function is discarded and the new activation function is used instead. The network structure is changed and image reconstruction is achieved by the final deconvolution upsampling. The original stochastic gradient descent optimization algorithm is replaced by adaptive moment estimation algorithm whose optimizes performance is faster and better. Comparative experiments are carried out on Set5 and Set14 test sets, respectively. The experimental results show that the reconstruction effects of the improved method with less training time are greatly improved on the objective evaluation index, for example, the power signal-to-noise ratio increases up to 2.33 dB, and the texture is clearer, the edges are more complete and the reconstruction effect is better on the subjective visual effects.