Fast Emotion Recognition Neural Network for IoT Devices

The last decades have witnessed rapid IoT technologies development, which provided ubiquitous human-computer interactions. Building intelligent systems of various types, among which emotion recognition systems, is important challenge nowadays. Especially pressing problem is to build a real-time portable system which can be embedded in low performance hardware. We propose a high accuracy emotion recognition system, which can be deployed on a single board Raspberry Pi computer to perform real-time recognition of 4 facial expressions: neutral, angry, surprised, and happy. Recognition pipeline is divided into two main stages: human face detection and facial expression classification. Both stages are performed by deep neural networks with simple yet effective design. Several optimization techniques, such as weights quantization and model tracing, were applied after model training, to gain extra execution time reduction. The introduced system is lightweight and fast, is executed locally on a cost-effective single board computer, and requires minimum resources to make and transmit predictions, what makes proposed system an effective IoT device.