Evaluation of CNN and RNN architectures for multiclass visual content classification from visually evoked EEG signals

Abstract This study evaluates and compares advanced deep learning architectures for classifying electroencephalography (EEG) signals elicited by visual stimuli. The objective is to identify visual content perceived by individuals through brain signal analysis and to rigorously assess the performance of different neural network models. Five architectures—Long Short-Term Memory (LSTM), classical Convolutional Neural Network (CNN), 1D-CNN, 2D-CNN, and a hybrid CNN-RNN—were implemented. EEG data were recorded from six channels across four visual categories and preprocessed using two distinct strategies. All models were trained under identical experimental settings and evaluated based on metrics such as accuracy and sensitivity. The findings reveal that the hybrid CNN-RNN architecture achieved the highest mean accuracy of 92.49%, highlighting its superior ability to analyze EEG signals for multiclass visual content classification. Additionally, 2D-CNN and LSTM demonstrated competitive performance, whereas simpler architectures like 1D-CNN and classical CNN were less effective in capturing the temporal dynamics of EEG data. These results underscore the importance of combining well-designed EEG acquisition protocols with hybrid deep learning models to enhance classification accuracy and robustness. Based on these results, the CNN-RNN model is recommended for future applications in EEG-based visual content recognition.