MLP-based weakly supervised crowd counting approaches have made significant advancements over the past few years. However, owing to the limited datasets, the current MLP-based methods do not consider the problem of region-to-region dependency in the image. For this, we propose a weakly supervised method termed SR2. SR2 consists of three parts: scale-reasoning module, scale-ranking module, and regression branch. In particular, the scale-reasoning module extracts and fuses the region-to-region dependency in the image and multiple scale feature, then sends the fused features to the regression branch to obtain estimated counts; the scale-ranking module is used to understand the internal information of the image better and expand the datasets efficiently, which will help to improve the accuracy of the estimated counts in the regression branch. We conducted extensive experiments on four benchmark datasets. The final results showed that our approach has better and higher competing counting performance with respect to other weakly supervised counting networks and with respect to some popular fully supervised counting networks.