Crowd occlusion and background noise in crowd images from different scenes negatively affect the training and prediction of crowd counting network. We design a novel network structure (SABCrowd), based on scale aggregation and background suppression, which uses VGG19 and Transformer as the base framework. To address the crowd occlusion problem, we design a scale aggregation block to capture and fuse feature information from different scales, enabling more comprehensive detection of occluded regions. To tackle background interference, we design a background weakening block that generates a background attention mask to help remove background noise. Additionally, we design a multichannel attention block to enhance the dependency between different channels. By conducting experiments on several current mainstream and complex datasets, namely ShanghaiTech, UCF-QNRF, JHU++, and NWPU, the final result demonstrate that our proposed method achieves Excellent prediction accuracy.