<![CDATA[Lightweight crowd counting is important for real-time surveillance and resource-constrained deployment, where both computational efficiency and effective supervision are required. Although teacher-free self-distillation can improve lightweight density-regression models by guiding intermediate representations without an external teacher, the influence of composite loss weights in such frameworks has not been sufficiently analyzed. This paper presents a focused coefficient-wise loss-weight analysis within the Lightweight Self-Knowledge Distillation framework for single-image crowd counting. Instead of proposing a new architecture, the study investigates how the coefficients α, β, γ, and λ₂ affect optimization behavior and counting accuracy under a fixed experimental setup on ShanghaiTech Part B. Specifically, α controls intermediate feature alignment, β controls consistency supervision, γ controls direct density-regression supervision, and λ₂ controls the structural similarity term in the regression loss. The results show that moderate values of α and β improve performance by providing useful internal regularization, while excessive auxiliary weighting can slightly degrade accuracy. The analysis also indicates that γ should remain dominant because direct density-map regression is the primary learning signal. The best observed configuration is α = 6.0, β = 2.0, γ = 13.0, and λ₂ = 0.2, achieving 8.94 MAE and 11.51 RMSE on ShanghaiTech Part B. These findings highlight the importance of balanced supervision design within the evaluated LSKD framework on ShanghaiTech Part B.]]>
