FGEFNet: Fine-Grained Extraction and Flow Network for Crowd Counting

Abstract Crowd counting is an important application of artificial intelligence in computer graphics and one of the most challenging research areas in the field of computer vision. Most existing methods lack consideration for fine-grained information and the role of advanced features, making it challenging to simultaneously focus on fine-grained information while extracting global information In this paper, we propose a fine-grained extraction and flow network (FGEFNet) for crowd counting. Firstly, we propose a Feature Selection Fusion Pyramid structure, which fully exploits important information in high-level feature maps and ensure the flow and fusion of fine-grained features. Secondly, we propose an Adaptive Channel Focus Module (ACFM),whcih can make the model focus on global features while also paying attention to fine-grained features. We innovatively introduce ACFM at the backend of the network in a fine-grained manner, providing fine-grained channel perception capability to detect subtle features in crowd images. Finally, we conducted extensive experiments on four widely used datasets, achieving State-Of-The-Art (SOTA) results on the SHHA datasets. It is worth noting that FGEFNet achieves a remarkable improvement by reducing Mean Absolute Error (MAE) by 2.56 and Mean Squared Error (MSE) by 7.4 on the SHHA dataset compared to the existing best model. The code and models are available at https://github.com/lele-progammer/FGEFNet