MTDNet: A crowd counting network based on a multiscale transformer and dilated convolution

Crowd counting methods based on Convolutional Neural Networks (CNNs) often struggle to model contextual features effectively due to their limited feature extraction scope. In contrast, the Transformer model, through its global self-attention mechanism, can consider information from all positions simultaneously, thereby modeling contextual relationships more comprehensively and enhancing the model's ability to understand sequences. Building on this, we propose a crowd counting network that leverages multi-scale features extracted by Transformer and dilated convolutions. Specifically, we utilize Vision Transformer to capture global information and integrate low-level detail features with high-level semantic features, enabling the model to effectively extract semantic features with global contextual information. Additionally, we design a multi-branch dilated convolution regression module, composed of dilated convolutions, to process feature maps generated during the encoding phase, which can effectively produce predicted density maps and yield accurate regression results. Furthermore, we introduce a novel loss function for our pipeline, which enhances the model's fitting capability and results in smoother density maps. Extensive experiments on four public crowd counting datasets demonstrate that our method achieves state-of-the-art results on several benchmarks. Particularly on the ShanghaiTech dataset, MTDNet achieves the most advanced results, with the MSE improving by 2.6% compared to the next best model.