EEG Emotion Recognition via a Lightweight 1DCNN-BiLSTM Model in Resource-Limited Environments

In the application of wearable medical monitoring devices, electroencephalogram (EEG) emotion recognition tasks need to be implemented in resource-constrained environments. Therefore, the proposed lightweight 1DCNN-BiLSTM network aims to achieve comparable emotion recognition accuracy to existing models while significantly reducing computational costs and memory usage on resource-constrained devices. First, a low computational cost preprocessing method is used to eliminate the interference of baseline signals in the raw EEG signal. Second, a shallow hybrid network of 1DCNN-BiLSTM is proposed to extract spatial features between different channels and temporal forward–backward features in EEG signals. Finally, quantize the trained model to reduce memory consumption and replace floating-point operations with fixed-point operations. Experiments on DEAP and DREAMER datasets achieve more than 90% recognition accuracy. The memory usage of the quantized network is 17.4 KB, and the computation for a single classification is 1.7 MFLOPs. The model is ultimately deployed on an embedded processor, attaining an inference speed of 352.51 mss, thereby enabling emotion recognition in resource-constrained environments.