Abstract
Commercial facade service classification from street-level imagery benefits from multi-branch pipelines that fuse global facade appearance with signboard-centric cues; however, this design increases inference cost and hinders large-scale deployment. We study pruning for a multi-branch facade recognition pipeline with a trainable signboard descriptor and compare unstructured magnitude pruning with structured baselines, including channel pruning and latency-aware structured pruning. We further propose a customized, architecture-aware pruning strategy that allocates sparsity across modules according to their sensitivity and computational profile. Experiments on a curated Google Street View subset and the Signboard Classification Dataset show that unstructured pruning reduces parameters, but yields limited wall-clock speedup, whereas structured pruning provides substantial acceleration. At$\mathbf{5 0 \%}$sparsity, the proposed method reduces latency from 14.3 ms to 7.9 ms and FLOPs from 2.83 G to 1.26 G while preserving macro-averaged F 1 -score close to the baseline ($\mathbf{0. 8 3 6}$vs.$\mathbf{0. 8 4 4}$on GSV;$\mathbf{0. 8 7 8}$vs.$\mathbf{0. 8 8 7}$on SCD). At 60 % sparsity, it further reduces latency to 6.9 ms and FLOPs to 1.04 G with competitive performance$(0.826$on GSV; 0.868 on SCD). Overall, the proposed strategy improves the accuracy-efficiency trade-off of facade service recognition under realistic domain variability.