RDC-SAL: Refine distance compensating with quantum scale-aware learning for crowd counting and localization

As one of the most meaningful research topics in computer vision, crowd counting and localization problems have been applied in many applications such as Video surveillance and Dense object detection. The most recent works solved the crowd counting and localization problems as a regression task via convolutional neural networks (CNNs). However, it is relatively hard for a basic CNN framework to extract adequate features of the crowd scenes. In this work, a refine distance compensating with quantum scale-aware learning framework (RDC-SAL) is proposed to solve crowd counting and localization task based on the Front-end quantum feature extraction, Multi-scale and Refine distance compensating modules. First, the Front-end quantum feature extraction module is adopted with qubit rotation and Pauli operators to calculate the crowd feature using classical CNN architecture. Then the Multi-scale feature extraction module is used to handle the quantum feature with different feature extraction branches by branching procedure. Finally, the Refine distance compensating module is proposed to estimate the density map, which uses the Refine distance compensating factor to fuse several feature extraction branches with different Upsample layers. To the best of our knowledge, it's the first time to introduce the hybrid classical-quantum network to model the crowd counting and localization problem. Experimental results on some benchmark datasets show that the proposed RDC-SAL can restore the predicted density maps with the high spatial resolution for crowd scenes and achieve improved performance to deal with the localization task compared with state-of-the-art works.

Automated summary

This paper proposes a hybrid classical-quantum network based on quantum feature extraction and distance compensation to solve crowd counting and localization problems. By using quantum feature extraction module, multi-scale feature extraction module, and refine distance compensating module to fuse feature extraction branches, this method achieves better performance in handling crowd scenes and localization tasks compared to existing methods.