Abstract: A crowd counting model based on background suppression and noise monitoring is proposed to solve the problems of large-scale change of crowd, complex background, and label noise. In the coding stage, the first 13 layers of VGG16_bn are used as the backbone, and the initially extracted features are sent to the two-branch feature extraction module and the background information aggregation module respectively, to mitigate the large- scale changes of the population and improve the discriminability of the background. Finally, the information processed by the two modules is fused, and the predictive density map is generated by decoder regression, which is supervised with the ground truth density map to achieve noise suppression. Compared with other algorithms, the counting accuracy of this model has been improved. MAE(Mean Absolute Error) and MSE(Mean Squared Error) on ShanghaiTech PartA are 58. 1 and 95. 9 respectively. Ablation experiments conducted on ShanghaiTech PartA also verified the effectiveness of the modules. Experimental results show that the algorithm can effectively improve the accuracy of crowd counting. 

Key words: crowd counting, density map, convolutional neural network, deep learning, noise supervision