PVGGCount: A Lightweight VGG Framework for Efficient Crowd Counting via Partial Convolutions

This study proposes an efficient crowd counting model based on VGG, called Partial Convolution VGG Crowd Counting Model (PVGGCount), to address the issues of high redundancy in feature extraction and low computational efficiency in existing crowd counting methods. The performance optimization is achieved through three stages: First, partial convolution and dilated convolution modules are introduced into the VGG network architecture to build an efficient backbone network, Partial Convolution VGG (PVGG), suitable for crowd counting. Second, a multi-scale feature enhancement structure is adopted to design the crowd counting model, PVGGCount, with PVGG as the backbone network. Finally, an end-to-end coordinate regression network is established to achieve accurate head position prediction. Cross-dataset experiments show that, compared with the baseline model, our method reduces the number of model parameters and floating-point operations (FLOPs) by more than 60%, while maintaining the same level of accuracy. Systematic experiments comparing PVGGCount with state-of-the-art methods using ResNet, YOLOv11, and FasterNet as backbones further validate its effectiveness and applicability in crowd counting tasks.