Abstract: The image super-resolution reconstruction algorithm based on residual model can increase the number of network layers and improve network performance, but it has the problem of large network scale. To solve this problem, and effectively extract features from different network layers, as well as distinguish low-frequency and high-frequency information of input images, a lightweight dual-branch super-resolution reconstruction model based on skip residual connections and attention mechanism is proposed. Firstly, the attention residual branch is composed of skip residual connection and spatial attention mechanism, and the skip residual connection is composed of convolution layer and small residual structure. Secondly, the image is directly upsampled in the image upsampling branch. Finally, the characteristics of the two branches are fused. The algorithm is tested on five base data sets, Set5, Set14, BSD100, Urban100, and Manga109. Experimental results show that the proposed algorithm has higher peak SNR and structural similarity in some datasets of ×2 and ×3 amplification factor models. In the ×4 magnification factor model, the evaluation indexes of all data sets are optimal, and the reconstructed images have more texture details and fewer artifacts. Compared with other lightweight super-resolution reconstruction algorithms (the number of parameters is less than 1.5M), the proposed algorithm contains fewer parameters and higher peak signal-to-noise ratio.