The footprint of colour in EEG signal

Abstract

Our perception of the world is inherently colourful, reflecting a tight interaction between colour and object processing. Previous studies have shown that colour information can be decoded from non-invasive electroencephalography (EEG) signals when participants view simple, uniformly coloured stimuli. Here, we test whether colour can also be decoded from EEG responses to complex natural images, without explicit colour cueing. We analysed the THINGS EEG dataset, comprising 64-channel recordings from participants viewing 1,800 object concepts (16,740 images), each presented for 100 ms, yielding over 82,000 trials. We established a behavioural ground truth for perceived colour via a psychophysical experiment in which participants selected all perceived colours from a 13-option palette. To model colour-object perception under brief exposure, we used the Segment Anything Model (SAM) to segment images and replaced each segment with its average colour. We then trained an artificial neural network, CUBE (ColoUr and oBjEct decoding), with two objectives: (1) alignment with a frozen pretrained CLIP vision encoder, and (2) prediction of human-derived colour distributions. Results show that CUBE robustly decodes colour from EEG signals (average F-score = 0.50). Removing the CLIP alignment reduces performance (F-score = 0.46), indicating that while semantic representations enhance colour decoding, substantial colour information remains decodable from EEG signals even without semantic guidance. Finally, we evaluated object recognition across two EEG and MEG datasets in a 200-class task. Incorporating colour-segmented images improved accuracy by approximately 5% across seven vision encoders. Together, these findings demonstrate that EEG signals during natural vision carry rich colour information that interacts with object processing.