Tianjin Key Laboratory of Wireless Mobile Communications and Power Transmission, Tianjin Normal University, Tianjin 300387, China.
College of Artificial Intelligence, Tianjin Normal University, Tianjin 300387, China.
Rev Sci Instrum. 2024 Aug 1;95(8). doi: 10.1063/5.0221637.
Electroencephalogram (EEG) signals, serving as a tool to objectively reflect real emotional states, hold a crucial position in emotion recognition research. In recent years, deep learning approaches have been widely applied in emotion recognition research, and the results have demonstrated their effectiveness in this field. Nevertheless, the challenge remains in selecting effective features, ensuring their retention as the network depth increases, and preventing the loss of crucial information. In order to address the issues, a novel emotion recognition method is proposed, which is named Res-CRANN. In the proposed method, the raw EEG signals are transformed into four dimensional spatial-frequency-temporal information, which can provide a more enriched and complex feature representation. First, the residual block is incorporated into the convolutional layers to extract spatial and frequency domain information. Subsequently, gated recurrent unit (GRU) is employed to capture temporal information from the convolutional neural network outputs. Following GRU, attention mechanisms are applied to enhance awareness of key information and diminish interference from irrelevant details. By reducing attention to irrelevant or noisy temporal steps, it ultimately improves the accuracy and robustness of the classification process. The Res-CRANN method exhibits excellent performance on the DEAP dataset, with an accuracy of 96.63% for valence and 96.87% for arousal, confirming its effectiveness.
脑电信号（EEG）作为客观反映真实情绪状态的工具，在情绪识别研究中具有重要地位。近年来，深度学习方法已广泛应用于情绪识别研究，其在该领域的有效性已得到证实。然而，选择有效特征、确保随着网络深度的增加保留这些特征以及防止关键信息丢失仍然是一个挑战。为了解决这些问题，提出了一种新的情绪识别方法，称为 Res-CRANN。在提出的方法中，原始 EEG 信号被转换为四维空间-频率-时间信息，这可以提供更丰富和复杂的特征表示。首先，将残差块引入卷积层，以提取空间和频域信息。然后，使用门控循环单元（GRU）从卷积神经网络输出中捕获时间信息。在 GRU 之后，应用注意力机制来增强对关键信息的意识，并减少对不相关细节的干扰。通过减少对不相关或嘈杂的时间步骤的注意力，最终提高了分类过程的准确性和鲁棒性。Res-CRANN 方法在 DEAP 数据集上表现出优异的性能，其效价的准确率为 96.63%，唤醒度的准确率为 96.87%，证明了其有效性。
