Existing CNN-based image super resolution reconstruction methods are usually realized on full-resolution or progressively low-resolution image representations. The former can achieve a spatially accurate but contextually weak super-resolution reconstruction result, while the latter can obtain a semantically reliable but less spatially accurate output. To solve the above-mentioned problems, a new super-resolution reconstruction method based on across-multi-resolution information flow and multiple attention mechanism is proposed in this paper. Multi-scale feature extraction and aggregation are realized by using cross-multi-resolution information flow and information interaction technology. Multiple attention mechanism is used for capturing context information to enhance image high-frequency information. A new weighted loss function is designed and used for optimizing model parameters. The experimental results on five public datasets show that, compared with Bicubic, SRCNN, VDSR, RDN and MuRNet reconstruction algorithms, the PNSR and SSIM of the proposed method are improved by 0.33dB and 0.0048, and the proposed method has better super-resolution reconstruction effect.