State-of-the-art emotion recognition models were too large and power-hungry for on-device mobile deployment. Running inference in the cloud introduced latency, privacy concerns, and dependency on connectivity.
Designed a multi-phase pipeline: model selection and benchmarking on FER-2013, fine-tuning on RAF-DB for improved real-world accuracy, then dynamic INT8 quantisation to shrink the model footprint for mobile CPUs. Built the pipeline with PyTorch and OpenCV.
A production-ready pipeline from research dataset to mobile-optimised model. The INT8 quantised variant runs efficiently on-device while retaining competitive accuracy, enabling privacy-preserving, offline emotion recognition.