Research On Crowd Counting Based On Fully Convolutional Networks

In recent years,Convolutional Neural Network(CNN)have been the network of choice for various types of research in deep learning.Benefiting from the powerful feature representation capabilities of CNN,earlier models conducted crowd counting studies with base CNN and achieved significant performance improvements compared to traditional handcrafted features.At present,more effective models based on fully convolutional network(FCN)have become the mainstream models in crowd counting research.However,the current deep model is only roughly divided into several levels when dealing with the scale change problem.When acquiring the characteristics of the crowd,VGG16 is directly used to extract the features.This will lead to the problem that the extracted features are not sufficiently effective in dealing with complex backgrounds and scenes with a wide range of scales.To effectively deal with the above problems,this paper designs two crowd counting network models based on the fully convolutional network architecture.In order to effectively deal with the problem of scale variation,this paper proposes a continuous scale feature extractor(CSE)based on residual U＿blocks,which can obtain a wider range of scale features.Using CSE,a novel crowd counting network(ERUNet)is designed,which can capture more scale information from different-level features and fuse different-level features.This network is a U-shaped structure design,mainly composed of three parts: encoding,decoding and attention density map generator(AMG).The encoding provides multi-level semantic feature information.Decoding extracts wider scale features,fuses semantic feature information at different levels,and outputs feature maps for generating attention maps and density maps.AMG is used to enhance the identification of crowded areas.In order to verify the effectiveness of residual U＿blocks,the effectiveness is analyzed through comparative experiments,and performance evaluation experiments are conducted on four mainstream datasets.The experimental results show that ERUNet has strong robustness to deal with scale changes,density changes and other problems,and has high counting accuracy.In addition,this paper also designs a counting model(FCN-CBAM)based on Convolutional Attention Mechanism Module(CBAM),which uses CBAM to process the features extracted by VGG16 to generate more effective features,thereby generating high-quality densities picture.The structural design of the entire network framework is also a fully convolutional structure with a U-shaped structure,so the input image can be of any size.In order to verify the effectiveness of CBAM,the comparison model was verified,and the performance evaluation experiments were also carried out on 4 mainstream datasets.Experiments show that the effect of CBAM on the network is significantly improved,and FCN-CBAM has higher computational accuracy and better generalization performance.The above two networks can effectively alleviate the problems caused by scale changes,and provide a new idea for the design of crowd counting models.