VCA-Net: Visual Creativity Analysis using Deep Neural Network✱

Abstract

Electrical signals generated in the brain, known as Electroencephalographic signals (EEG) are used to measure electrical activities in the brain. EEG methods being non-invasive, affordable, and portable, are used in this work. Analysis of Visual Creativity has clinical application in monitoring the progress of patients suffering from neurodegenerative diseases. In this work, EEG is used to analyze Visual Creativity by classifying the two brain states (1) Creative and (2) Non-creative using the in-house collected data. Data is collected from the participants performing the following tasks: (1) Planning to draw, (2) Drawing, and (3) Monotonous Tasks. Drawing and Monotonous Tasks are considered for the classification in the proposed work. Two types of analysis are carried out (1) converting EEG time-series data into a sequence of topology preserving feature maps and (2) feature maps generated using Spectral Entropy. The EEG data is first pre-processed to down-sample and band-limit the signals followed by ICA to remove the artifact. The artifact-removed EEG data is decomposed into three sub-bands. The decomposed time-series data is converted to a sequence of feature maps that preserves spatial, spectral, and temporal information. Another set of feature maps is also generated using Spectral Entropy from artifact-removed EEG data. Two types of Deep Learning frameworks are proposed (1) VCA-Net video - a deep learning framework consisting of temporal convolution that extracts spatial information and preserves temporal information which is extracted by a combination of LSTM and 1D convolution and (2) VCA-Net image - a deep learning framework which extracts the spatial information using Feature Pyramid Network (FPN) and Atrous Spatial Pyramid Pooling (ASPP). Drawing and Monotonous tasks are classified with a mean accuracy of 70.87% and 65.62% using VCA-Net image and VCA-Net video respectively.