Digital images play an important role in information transmission, and image super-resolution technology can enrich image details. To address the problems of insufficient effective feature reuse of low-resolution images and excessive parameters in many networks, this study combines convolution kernels of different sizes and attention residual mechanism to construct the image super-resolution network. Three convolution layers of different scales are used to extract the image features, of which the second and third layers replace the large convolution kernels with small ones, and after the three-layer convolution fusion, the attention mechanism is introduced. Finally, the traditional Bicubic interpolation is used to directly provide low-frequency information for the network. By doing this, while reducing the number of parameters and mitigating the disappearance of gradients, the proposed network can make the effective high-frequency information gain greater weights and can enhance the nonlinear expression ability between the networks, which is conducive to the iterative convergence of network training. Experimental results show that the proposed network can enhance the image reconstruction ability to a certain extent.