Scale-Aligned Capacity Allocation: A Lightweight Face Detection Framework for Fixed-View Unmanned Restaurant Scenarios

In fixed-view interaction scenarios of unmanned restaurants, face detection models face two core bottlenecks: the mismatch between training data distribution and real deployment scenarios, and the misalignment between model feature capacity allocation and business priority. To address these problems, this paper takes YOLOv8n (You Only Look Once version 8n) as the baseline, proposes a unified Scale-Aligned Capacity Allocation (SACA) theoretical framework, and constructs an end-to-end Scale Distribution Reconstruction Network (SDRNet) for lightweight face detection. First, we define the SACA loss with KL (Kullback-Leibler) divergence as the core optimization objective, which mathematically characterizes the matching degree between model capacity allocation and real scene face scale distribution. Second, a two-stage scene-aware scale distribution reconstruction strategy is designed based on the SACA framework, which derives the core face scale interval of the unmanned restaurant scene through a monocular imaging model, and constructs a scene-adaptive training dataset based on the public WIDER FACE benchmark, which is highly consistent with the real scale distribution of unmanned restaurant scenarios. Third, three scale-aligned lightweight modules, including LFEM (Lightweight Feature Extraction Module), LDown (Feature Segmentation and Sparse Optimization Module), and MSCH (Multi-Feature Shared Convolution Module), are proposed to realize the priority allocation of model capacity to core interaction scales, achieving collaborative optimization of data distribution and model structure. Fourth, a 2 × 2 controlled experiment is designed to separate the independent contributions of the data strategy and architectural improvements, and the robustness of the proposed model is verified on the standard WIDER FACE benchmark. Finally, a scale-specific validation mechanism is established to conduct fine-grained evaluation of the model’s detection performance on faces of different scales, avoiding the overall indicator masking the accuracy fluctuation of core scenarios. Experimental results show that the parameters of the proposed model are reduced to 1.76 M (a decrease of 41%), and the computational complexity is reduced to 5.5 GFLOPs (Giga Floating-point Operations Per Second) (a decrease of 32%). The mAP@0.5 (mean Average Precision) of the core medium-scale face reaches 0.684, with the performance loss controlled within 2% compared with the baseline. On the standard WIDER FACE benchmark, the model maintains competitive detection accuracy under extreme lightweight compression, which fully verifies its robustness. On the NVIDIA Jetson Orin NX embedded platform, the inference frame rate of TensorRT-FP16 reaches 79.9 FPS (Frames Per Second), which fully meets the real-time deployment requirements of resource-constrained unmanned restaurant scenarios.