The Multi-Scale Attention-Based for Crowd Density Estimation

To address the challenges of large scale variations and strong background interference in dense crowd images, this paper proposes a crowd counting model named Multi-scale Selective Density Aggregation Network (MSDA-Net). The model introduces three collaborative modules: the Multi-scale Tuning Module (MTM), the Selective Attention Module (SAM), and the Density Context-aware Module (DCF), which collectively enhance feature representation and density map quality, thereby improving counting accuracy in complex congested scenes. Specifically, the MTM captures multi-scale contextual information through multi-branch atrous convolutions, alleviating the impact of drastic scale variations. The SAM employs spatial and channel attention mechanisms to suppress background noise and highlight foreground crowd regions. The DCF models the contextual dependencies of density maps using convolutional LSTM architecture, enhancing continuity between adjacent regions to generate smoother and more accurate density maps. A joint loss function combining Euclidean distance loss and structural similarity (SSIM) loss is adopted to optimize pixel-level regression and local structural consistency simultaneously. Extensive experiments are conducted on four public benchmarks, including ShanghaiTech Part_A, ShanghaiTech Part_B, UCF_CC_50, and UCF_QNRF. The proposed MSDA-Net achieves superior performance on all datasets, e.g., MAE of 58.1 on Part_A, 6.9 on Part_B, 205.3 on UCF_CC_50, and 84.8 on UCF_QNRF, outperforming many state-of-the-art methods. Ablation studies further validate the individual contribution of each module. Overall, MSDA-Net demonstrates strong generalization ability and practical application value in real-world crowded scenes. 针对密集人群图像中目标尺度变化大、背景干扰强等问题，本文提出一种基于多尺度选择性密度聚合网络（MSDA-Net）的人群计数模型。该模型引入了三个协同模块：多尺度调制模块（MTM）、选择性注意力模块（SAM）和密度上下文感知模块（DCF），三者共同增强特征表达能力和密度图质量，从而提高拥挤复杂场景下的计数精度。具体而言，MTM模块通过多分支空洞卷积捕获不同尺度的上下文信息，缓解目标尺度剧烈变化带来的影响；SAM模块利用空间与通道注意力机制抑制背景噪声，突出前景人群区域；DCF模块采用卷积长短时记忆网络对密度图的上下文依赖关系进行建模，增强相邻区域间的连续性，以生成更平滑、准确的密度图。实验中采用了联合损失函数，结合了欧氏距离损失与结构相似性（SSIM）损失，同时优化像素级回归和局部结构一致性。在四个公开数据集上进行了充分验证，包括ShanghaiTech Part_A、ShanghaiTech Part_B、UCF_CC_50和UCF_QNRF。所提出的MSDA-Net在所有数据集上均取得了优异性能，例如在Part_A上MAE为58.1，Part_B上为6.9，UCF_CC_50上为205.3，UCF_QNRF上为84.8，超过了多种最先进的方法。消融实验进一步验证了每个模块的单独贡献。总体而言，MSDA-Net在真实拥挤场景中表现出强大的泛化能力和实际应用价值。