Scholarlink

<p>Electroencephalogram (EEG) signals are generally captured from the central nervous system, leading to their popularity in recognizing distinct emotion tasks. Current research on emotion recognition approaches mainly focuses on the usage of different types of neuron-based network architectures by exploiting the spectral, temporal, or spatial characteristics from EEG for classifying the emotional state. However, most of the human emotion recognition approaches fail to retrieve the spatiotemporal information from EEG signals, thus degrading the performance of the model. Understanding the emotional insights in the traditional EEG-based model is quite challenging. Due to dynamic and non-stationary EEG signals, the traditional model is not effective for capturing the emotional states in the EEG data. In this research, a novel approach is developed using an optimization-based advanced deep learning technique for an emotion recognition process. This process has evolved into three different phases of data collection, deep feature extraction, and emotion classification. Initially, a standard database-related EEG signal is considered for the process. From the EEG signal, three sets of features are attained. Utilizing Recurrent Convolutional Autoencoder, the deep features is attained. Moreover, the Short-Time Fourier Transform (STFT) model is applied for converting EEG signals into the spectrogram image. Then, the spectrogram images are extracted by the Vision Transformer (ViT) technique, which is considered as a second set of features and the third set of wave features are extracted directly from the EEG signal. For classifying distinct emotion states, the extracted features are then given into newly designed Multiscale Adaptive MobilenetV2 (MA-MNet) model. A novel heuristic algorithm, namely Fitness Weighted Greylag Goose Optimization (FW-GGO), is developed for optimizing the hyperparameters in the MobilenetV2. Further, the empirical findings of the developed model are compared with various standard classification metrics. Throughout the validation, the performance outcome of five-fold cross-validation shows a more accurate result of 94.99%, 92.29% and 93.62, whereas four-fold cross-validation shows the specificity outcome of 95.73%, 98.39% and 98.84% in terms of EEG Brain wave, BATH and DEAP datasets, respectively.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep Feature Analysis and Heuristic Enabled Multiscale Adaptive Deep Learning-Based Emotion Recognition Model with EEG Signals

  • 期刊论文

  • 发表日期：

    2026年5月22日

查看全文
  • Sanjay Singh,
  • Prakriti Trivedi

摘要

Electroencephalogram (EEG) signals are generally captured from the central nervous system, leading to their popularity in recognizing distinct emotion tasks. Current research on emotion recognition approaches mainly focuses on the usage of different types of neuron-based network architectures by exploiting the spectral, temporal, or spatial characteristics from EEG for classifying the emotional state. However, most of the human emotion recognition approaches fail to retrieve the spatiotemporal information from EEG signals, thus degrading the performance of the model. Understanding the emotional insights in the traditional EEG-based model is quite challenging. Due to dynamic and non-stationary EEG signals, the traditional model is not effective for capturing the emotional states in the EEG data. In this research, a novel approach is developed using an optimization-based advanced deep learning technique for an emotion recognition process. This process has evolved into three different phases of data collection, deep feature extraction, and emotion classification. Initially, a standard database-related EEG signal is considered for the process. From the EEG signal, three sets of features are attained. Utilizing Recurrent Convolutional Autoencoder, the deep features is attained. Moreover, the Short-Time Fourier Transform (STFT) model is applied for converting EEG signals into the spectrogram image. Then, the spectrogram images are extracted by the Vision Transformer (ViT) technique, which is considered as a second set of features and the third set of wave features are extracted directly from the EEG signal. For classifying distinct emotion states, the extracted features are then given into newly designed Multiscale Adaptive MobilenetV2 (MA-MNet) model. A novel heuristic algorithm, namely Fitness Weighted Greylag Goose Optimization (FW-GGO), is developed for optimizing the hyperparameters in the MobilenetV2. Further, the empirical findings of the developed model are compared with various standard classification metrics. Throughout the validation, the performance outcome of five-fold cross-validation shows a more accurate result of 94.99%, 92.29% and 93.62, whereas four-fold cross-validation shows the specificity outcome of 95.73%, 98.39% and 98.84% in terms of EEG Brain wave, BATH and DEAP datasets, respectively.