Structured-token attention improves mmWave liquid recognition accuracy
Structured-Token Attention for Fine-Grained mmWave Liquid Identification
Artificial Intelligence
Summary
Identifying liquids from radar signals is tricky because many liquids have similar signatures. The authors introduce a new way to organize radar data into multiple pieces called tokens that capture different perspectives, like antenna views and features. They use a transformer model to combine these tokens and distinguish between 30 different liquids with high accuracy. Their approach matches or outperforms previous methods and shows promise for practical liquid recognition tasks.
What this means in practice
- •For security screening teams: Improve identification of liquid substances in security checkpoints using millimeter-wave radar measurements with structured attention models.
- •For industrial quality control engineers: Apply structured-token attention to differentiate liquids during manufacturing processes with compact radar sensors for real-time monitoring.
Authors
Junseo Min, Joongheon Kim
Abstract
Fine-grained liquid identification requires a classifier to distinguish similar substances from compact millimeter-wave measurements. We investigate structured-token attention, which represents a radar feature tensor through complementary antenna, feature, branch, and global views. The proposed StructAttn model embeds these views into eleven tokens and combines them with a three-layer transformer. Experiments on the public FG-LiquID data cover 30 liquids and 3,000 samples. Over five seeds on the released split, StructAttn achieves 96.70\% mean best-epoch accuracy, matching the original-recipe MMNet baseline and exceeding a scalar-token attention model by 1.50 percentage points. A five-seed probability ensemble reaches 97\%, corresponding to 18 errors among 600 test samples. Final-epoch and five-fold results complement the checkpoint-selected comparison, with StructAttn obtaining 95.30\% out-of-fold accuracy. Token and feature analyses show that structured views are useful for prediction, while residual confusions concentrate in water, soft-drink, and edible-oil classes. These results support structured tokenization as a practical design choice for attention-based liquid recognition.