Mohammadreza Mostafavi

Mohammadreza works on biomedical signal processing, asking how much can be read from as little recording as possible, and on vision-based monitoring of livestock.

Publications

EEG-based biometric authentication through motor imagery

M. Mostafavi, A. Ayatollahi, S. Rajagopal, S.-B. Ko. Signal, Image and Video Processing, 19:1252, 2025.

• EEG is attractive for authentication because it resists spoofing and cannot be taken under coercion
• On the BCI IV 2a dataset, participants imagine moving the right hand, left hand, both feet or the tongue
• A 2D convolutional network with six layers plus self-distillation; channel-wise kernels pick up localised spatial features and the relationships between channels
• Best results came from combining frequency bands with the full channel set and a four-second window

Detecting schizophrenia from a single EEG channel

M. Mostafavi, S.-B. Ko, S. B. Shokouhi, A. Ayatollahi. Physical and Engineering Sciences in Medicine, 48:3-18, 2025.

• People with schizophrenia often lack awareness of the condition, which makes timely diagnosis both harder and more valuable
• The goal is automatic diagnosis from minimal input: one EEG channel, converted to scalogram images
• Transfer learning combined with knowledge distillation keeps accuracy despite the reduced input

Instance-aware coordinate attention for efficient pig pose estimation

F. Ferri, M. Mostafavi, B. Predicala, S.-B. Ko. Preprint, 2026.

• Existing multi-instance methods lean on heavy backbones such as HRNet and underuse the features they produce
• A lightweight instance-aware head uses Coordinate Attention for spatial-channel dependencies and Conditional Channel Weighting for instance-specific modulation
• Accuracy comparable to CiD with 68.09% fewer FLOPs