Multi-scale dilated convolution of feature Fusion Network for Crowd counting

Crowd counting has long been a challenging task due to the perspective distortion and variability in head size. The previous methods ignore the multi-scale information in images or simply use convolutions with different kernel sizes to extract multi-scale features, resulting in incomplete multi-scale features extracted. In this paper, we propose a crowd counting model called Multi-scale Dilated Convolution of Feature Fusion Network (MsDFNet) based on a CNN (convolutional neural network). Our MsDFNet is based on the regression method of the density map. The density map is predicted by the parameters learned by CNN to obtain better prediction results. The proposed network mainly includes three components, a CNN to extract low-level features, a multi-scale dilated convolution module and multi-column feature fusion blocks, a density map regression module. Multi-scale dilated convolutions are employed to extract multi-scale high-level features, and the features extracted from different columns are fused. The combination of the multi-scale dilated convolution module and the multi-column feature fusion block can effectively extract more complete multi-scale features and boost the performance of counting small-sized targets. Experiments show that the problem of various head sizes in images can be effectively solved by fusing multi-scale context feature information. We prove the effectiveness of our method on two public datasets (The ShanghaiTech dataset and the UCF_CC_50 dataset). We compare our method with the previous state-of-the-art crowd counting algorithms in terms of MAE (Mean Absolute Error) and MSE (Mean Square Error) and significantly improves the performance, especially in case of various head sizes. On the UCF_CC_50 dataset, our method reduces the MAE index by 28.6 compared with the previous state-of-the-art method. (The lower the MAE, the better the performance).