This work addresses fine‑grained liquid identification, i.e., distinguishing similar substances from compact mmWave measurements. We introduce StructAttn, which encodes the radar feature tensor into eleven structured tokens derived from antenna, feature, branch, and global complementary views, and fuses them with a three‑layer Transformer. Experiments are conducted on the public FG‑LiquID dataset covering 30 liquids and 3,000 samples. Across five random seeds, the mean best‑epoch accuracy reaches 96.70%, matching the original MMNet baseline and surpassing a scalar‑token attention model by 1.50 percentage points. A five‑seed probability ensemble attains 97% accuracy, corresponding to only 18 errors among 600 test samples. Final‑epoch and five‑fold results further confirm robustness, with an out‑of‑fold accuracy of 95.30%. Token and feature analyses reveal that the structured views contribute significantly to prediction, while residual confusions concentrate in water, soft‑drink, and edible‑oil classes. In summary, structured tokenization proves to be a practical design choice for attention‑based liquid recognition.
Review