Aiming at the problems of background interference and drastic change of crowd scale in crowd counting， which leads to poor counting effect， a crowd counting method enhanced by dense connected attention and scale perception recombination was proposed. First， a feature extraction network with dense connected attention mechanism was designed to enhance the crowd counting features and suppress the background interference by using the inflated convolutionally improved VGG19 network as the model coarse feature extraction network and embedding the dense connected dual-channel attention mechanism. Then， the scale-aware reorganized upsampling and soft mask feature enhancement and delivery structures were designed to achieve the full utilization of crowd feature information at different scales from shallow to deep， and overcome the problem of poor counting performance due to drastic changes in crowd scales. Secondly， a multi-resolution fusion module was proposed to enhance the interaction between multi-resolution information， reduce the semantic gap between different resolutions， and improve the accuracy of crowd counting. Finally， comparison experiments were conducted on ShanghaiTech， UCF-QNRF， JHU_CROWD++ and other crowd counting datasets， and the results show that the proposed method outperforms the comparison algorithms. For instance， compared with DM-Count crowd counting algorithm， the MAE and MSE error values of proposed method are reduced by 15.98% and 14.52%， respectively， and the proposed method has higher counting performance in crowd counting.