Exploring Electroencephalography-Based Visual Imagery Using Convolutional Neural Networks For 2D Shapes Classification

Brain-Computer Interfaces (BCIs) using electroencephalography (EEG) enable intuitive, non-invasive neural communication with external systems. While many other BCI paradigms have been widely studied in EEG-BCI, visual imagery (VI) remains relatively unexplored, particularly for intra-class object differentiation. This study examines the viability of using state-of-the-art deep learning (DL) models for classifying EEG signals generated during imagination of 2D geometric shapes within the same visual category such as circles and triangles. Two leading convolutional neural network architectures, EEGNet and FBCNet, are evaluated for end-to-end learning from raw time-series data. The limitations of classical machine learning approaches such as feature extraction bottlenecks and low signal-to-noise ratio are addressed. Data augmentation with a crop window technique is used to mitigate restrictions on experimental data conditions while maintaining model performance. Results demonstrate that DL models can efficiently classify intra-category VI, predicting 216 testing samples in under 2 seconds. Challenges remain due to limited training data and subtle neural distinctions between simple geometric shapes. Topographic brain maps further illustrate class-specific activation patterns. This work presents, to our knowledge, the first implementation of DL classification on full waveform VI data for primitive 2D shapes in EEG-BCI systems.