Hybrid LBPH-inspired CNN for interactive web interface-based multi-face emotion recognition on embedded and IoT platforms

Abstract

Deploying real-time facial emotion recognition on internet of things (IoT) nodes, embedded boards, and reconfigurable-computing platforms is difficult, since conventional deep convolutional neural networks (CNNs) are usually too heavy, computationally for such hardware. This work proposes a hybrid, local binary patterns histograms (LBPH)-inspired CNN pipeline: an LBPH-style illumination-normalization stage is paired with a compact, purpose-built convolutional network, keeping memory and compute demands low enough to later port to embedded and field-programmable gate array (FPGA) targets. A Streamlit-based web interface prototypes the pipeline, supporting multi-face detection with live histogram display. Tested on FER2013, the model reaches 68.5% accuracy ahead of plain LBPH's 58.4%, and close to much deeper CNNs while needing only 11.5 MB of memory (2.9 MB once int8-quantized) and 22 ms of central processing unit (CPU) inference time per face. An input-resolution ablation study identifies 96×96 as the best accuracy-latency trade-off. Beyond a single accuracy number, per-class metrics are reported, along with a discussion of where the model is likely to confuse similar emotions and a layer-by-layer breakdown of parameters and multiply-accumulate operations, intended to be directly useful for later hardware-synthesis work.