Estimating the population in a given scene is a process known as crowd counting. The field has recently garnered significant attention, and many innovative methods have emerged. However, intense scale variations and background interference make crowd counting in realistic scenes always challenging. To address these in this paper, a multi-level attention network with multi-scale feature fusion named MLANet is proposed. The network consists of three sections: a multi-level base feature extraction front-end network, a centralized dilated multi-scale feature fusion mid-end network with a global attention module, and a back-end network for the generation of density maps. By incorporating a flexible attention module and multi-scale features, the method can accurately capture crowd information at different scales and achieve accurate counting results. We evaluated the method on four public datasets (UCF_CC_50, ShanghaiTech, WorldExpo’10, and Beijing BRT), and the experimental results demonstrate a significant reduction in counting error when compared with existing methods.