Brain-inspired deep learning model for EEG-based low-quality video target detection with phased encoding and aligned fusion

Brain-computer interface (BCI) technologies for video target detection hold great promise across various applications. However, existing algorithms exhibit limited performance in electroencephalogram (EEG) decoding for target detection in low-quality videos. In this paper, to address the limitation, we propose a novel brain-inspired deep learning model that incorporates EEG phased encoding and feature-aligned fusion. We first divide the EEG segments into pre-phase and post-phase, and extract the corresponding compressed temporal features using a novel phased encoder, which is based on multi-scale convolution and attention mechanisms. Subsequently, to capture the full-phase brain response, we align and integrate the features from both phases and extract global temporal features for classification. The proposed model is grounded in our time- and frequency-domain neural analysis, which identifies three critical phases of the brain's response during low-quality video target detection: early target recognition, later target spatial tracking, and sustained attention throughout the entire phase. EEG datasets, with and without ICA-based artifact removal, were used for cross-subject training and evaluation, with the proposed model consistently outperforming baselines. Pseudo-online tests confirmed real-time performance, and additional experiments with cognitively distracted participants further demonstrated the model's robustness. This work addresses a significant gap in low-quality video target detection algorithms and advances brain-inspired EEG classification by combining principles of neuroscience with artificial intelligence techniques. Our code is available at: https://github.com/Wonder-How/PSAFNet.