SCRSNet : An Efficient Crowd Counting via Lightweight Spatial‐Channel Reconstructive and Scale‐Aware Network

ABSTRACT Accurate crowd counting is of paramount importance for urban management and public security. Currently, to pursue superior counting precision, mainstream methods tend to construct cumbersome models with a massive number of parameters. Although such progress undeniably boosts counting performance, the substantial parameter and heavy computational overhead of these models severely impede their deployment on resource‐limited hardware, such as drones and mobile edge computing nodes. Lightweight architectures have surfaced as tailored solutions for edge computing, aiming to strike a compromise between inference speed and predictive precision. Nevertheless, these lightweight models often falter when confronted with real‐world visual complexities, notably intricate background interference (e.g., architectural occlusions, cross‐modal noise) and severe scale variations. To overcome these limitations, we propose the Spatial‐Channel Reconstructive and Scale‐aware Network for Crowd Counting (SCRSNet). Adopting a divide‐and‐conquer strategy, the proposed framework disentangles the latent representation into spatial and channel dimensions, attentively modelling their inter‐dependencies to filter out crowd‐irrelevant features. Specifically, a Spatial and Channel Reconstructive (SCR) module is constructed to mitigate the adverse effects of background noise. Concurrently, a Dilated Scale‐aware Attention (DSA) module is introduced to capture multi‐scale features by aggregating semantic details across varying scales. Extensive experiments on five crowd counting benchmarks demonstrate that the proposed SCRSNet surpasses existing state‐of‐the‐art methods in both accuracy and efficiency.