A Visual Characteristic Land-Scape Design for EEG Signal Based on LSTM-GAN

Abstract

The ultimate goal of artificial intelligence is to endow machines with human intelligence. Studying and simulating electroencephalogram signals is a way to achieve this goal. The human brain has similar representation abilities for similar visual stimuli. By utilizing this feature, a visual stimulus electroencephalogram signal decoding model based on Long Short-Term Memory Network Bagging was proposed to decode and classify human brain signals. And based on this extracted classification model, a generative adversarial network based on a bi-directional short-term memory network was proposed. It could generate similar visual stimulus images similar to the human brain and represent the visual signals of the human brain. These experiments confirmed that the classification accuracy of the research method in the decoding of electroencephalogram signals reached 91.17%. In terms of extracting visual characteristics and land-scape features from the electroencephalogram, this research model had the highest classification accuracy and recall rates, with 98.38% and 97.94%, respectively. This stimulation image generation model studied had the best actual image generation performance, with an Inception score of 7.27. The study not only improves the accuracy of electroencephalogram signal classification, but also completes the re-construction of brain signals into images. It improves the collaborative representation ability of human-machine collaborative visual cognitive systems and has important significance in brain computer interaction.

Keywords

• Electroencephalogram visual decoding
• Bi-LSTM
• LSTM-B-GAN
• generate adversarial re-construction
• brain computer interaction