Multistream Dilated Convolutional Feature Fusion Neural Network for SSVEP Classification

• Published in: 2023 5th International Conference on Frontiers Technology of Information and Computer (ICFTIC) pp. 1044 - 1047
• Main Authors: Li, Xiujun, Zhang, Yongzheng, Wu, Yan, Li, Qi
• Format: Conference Proceeding
• Language: English
• Published: IEEE 17.11.2023
• Subjects: Adaptation models; attention mechanism; Brain modeling; Convolution; Feature extraction; feature fusion; multi-stream atrous convolution; ssvep classification; Target recognition; Vectors; Visualization
• DOI: 10.1109/ICFTIC59930.2023.10456289

Based on steady-state visual evoked potentials (SSVEP), a neuroelectric phenomenon where the brain's electrical signals respond to specific frequency stimuli, which holds significant application value. Due to signal noise and individual differences, achieving accurate SSVEP classification remains highly challenging. To address these challenges, we propose a multi-stream atrous convolutional feature fusion convolutional neural network (MACNN) model. The model adopts a parallel structure with multiple streams of atrous convolution for feature fusion. Each parallel convolution stream utilizes different dilation rates, sharing weights across various streams, and incorporates a feature fusion module, allowing the model to leverage information from multiple feature maps. Finally, an attention mechanism is introduced to adaptively emphasize critical feature channels, thereby enhancing the discriminative power of classification. Results from data involving 35 subjects indicate that, with a 1 -second data length, the average accuracy and information transfer rate increase to 79.94% and 141.57 bits/min, respectively. Consequently, the proposed method holds significant importance in the research of SSVEP signal classification.