Traditional Chinese Medicine · Vision Intelligence

Identify Medicinal
Plants Instantly

A two-stage hybrid intelligence pipeline — Swin Transformer backbone fused with an ensemble of 8 specialist classifiers — delivering 92.76% accuracy across 300 TCM herb species.

Analyze a Plant How It Works
52K+
Training Images
300
Herb Species
92.76%
Val Accuracy
0.998
AUC Score
Scroll
Analysis Engine

Plant Recognition System

Upload a clear photograph or capture one live. Our two-stage pipeline classifies the specimen and retrieves full TCM medicinal details.

Image Input
Awaiting input
Drop herb photo here
Upload a clear image — root, leaf, flower, bark, or processed form
Plant preview
Initializing…
Analysis Results
Ready

Awaiting Analysis

Upload a plant photo to receive classification results

1
Upload or capture a plant photo
2
Click "Unmask This Plant"
3
View species & full TCM medicinal details
--
confidence
—
Pinyin
Scientific
Common

Species Not Confidently Recognised

Confidence is below threshold. Showing closest candidates — use a well-lit, focused photograph for best results.

Top Candidate Species
Research Architecture

Two-Stage Hybrid Pipeline

A novel system that surpasses what any single model achieves — combining deep visual features from a transformer backbone with a diverse ensemble of specialist classifiers.

Stage 01 · Backbone
Swin Transformer
A Swin-B vision transformer pretrained on ImageNet-22K is fine-tuned using a two-phase strategy — four warmup epochs with frozen early stages, then full parameter unfreezing under cosine annealing with layer-wise learning rate differentiation.
Swin-B BackboneImageNet-22KHybridMixCosine Annealing
91.95%
best validation accuracy
Stage 02 · Ensemble
Specialist Ensemble
Ten-view Test-Time Augmentation averages into a 1024-dimensional descriptor, compressed via PCA to 768 dimensions. Eight classifiers trained under 5-fold stratified cross-validation form a weighted soft-vote ensemble and stacking meta-learner.
SVM-RBFLightGBMXGBoostLogRegMLPExtraTrees
8
specialist classifiers in ensemble
Stage 03 · Fusion
Optimal Blend
An exhaustive sweep across 91 mixing ratios identifies the optimal blend at α=0.41. The fusion strategy is validated via cross-validation with improvements concentrated in historically confusable herb classes.
α=0.41 BlendROC AUC 0.998t-SNE
92.76%
final accuracy · +0.81% over backbone
Capabilities

Built for Real-World Demands

Handles the hardest challenges of TCM herb identification: extreme inter-class similarity, varied processing states, and diverse imaging conditions.

Unknown Detection
Automatically flags specimens outside the 300 trained classes using confidence thresholding — no false certainty.
Full TCM Database
Complete TCM pharmacology — nature, taste, meridian affinity, parts used, and clinical indications for all 300 species.
Live Camera Capture
Real-time camera input for on-site identification. Works with mobile rear cameras and desktop webcams.
Top-5 Candidates
Returns ranked candidates with confidence scores — essential for ambiguous specimens with overlapping morphology.
Inference Flow

From Image to Prediction

Each uploaded image passes through a precisely engineered inference chain before a classification is returned.

Image Input
224×224 resize
→
Swin-B Backbone
1024-dim features
→
10-View TTA
averaged pooling
→
PCA · 768-d
≥95% variance
→
Ensemble
8 classifiers
→
α-Blend · 0.41
Swin + Ensemble
→
Classification
300 classes