Crowd Counting Network Based on Multi-Scale Gated Aggregation and Progressive Refinement

In crowd counting, accurately recovering high-resolution density maps remains a critical challenge due to severe scale variations and complex background interference. To address this issue, this paper proposes a progressive density modeling framework, which reformulates conventional single-stage prediction as a multi-stage progressive optimization process. To mitigate information loss during encoding, a Multi-Scale Feature Aggregation (MSFA) module is designed to capture contextual information and suppress background noise via dilated convolutions and adaptive gating mechanisms. In the decoding stage, an enhanced residual module combining spatial-channel dual attention is introduced to strengthen the reconstruction capability of foreground features. Unlike traditional single-step prediction, this paper adopts a three-stage progressive output strategy of “coarse prediction, refinement, and correction” to progressively optimize density map details. Experiments on mainstream datasets such as ShanghaiTech and UCF_QNRF demonstrate that the method maintains high counting accuracy (e.g., a Mean Absolute Error (MAE) of 56.6 on ShanghaiTech Part_A) while exhibiting good robustness. Furthermore, the model achieves an inference speed of 148.1 frames per second (FPS) on an NVIDIA RTX 4090 for inputs with a resolution of 384 × 384, demonstrating strong potential for practical deployment.