The footprint of colour in EEG signal

Our perception of the world is inherently colourful, and colour provides well-documented benefits for vision: it helps us see things quicker and remember them better. We hypothesised that colour is not only central to perception but also a rich, decodable source of information in electroencephalography (EEG) signals recorded non-invasively from the scalp. While previous work has shown that brain activity carries colour information for simple, uniform stimuli, it remains unclear whether this extends to natural, complex images with no explicit colour cueing. To investigate this, we analysed the THINGS EEG dataset, which contains 64-channel recordings from participants viewing 1,800 distinct objects (16,740 images) presented for 100 ms each, yielding over 82,000 trials. We established a perceptual colour ground truth through a psychophysical experiment in which participants viewed each image for 100 ms and selected the perceived colours from a 13-option palette. An artificial neural network trained to predict these scene-level colour distributions directly from EEG signals showed that colour information was robustly decodable (average F-score of 0.5). We further examined the effect of colour features on object decoding. Using a contrastive learning framework, we modelled colour-object perception with the Segment Anything Model (SAM), in which all pixels within a segment were replaced with their average colour, followed by standard feature extraction using CLIP vision encoders. We trained an EEG encoder, CUBE (ColoUr and oBjEct decoding), to align features in both object and colour spaces. Across EEG and MEG datasets in a 200-class recognition task, incorporating colour improved decoding accuracy by approximately 5%. Together, these findings demonstrate that EEG signals recorded during natural vision carry substantial colour information that interacts with object perception. Modelling this interaction enhances the power of neural decoding.