21 trained sklearn/CatBoost classifiers for acne severity classification (Grade I / II / III). Part of the Acne CV Playground — an interactive web tool that walks through every pipeline stage from raw pixels to prediction, with live parameter sliders and a 42-dim feature inspector.

Model Description

Each model is a sklearn.pipeline.Pipeline of [StandardScaler → Classifier], trained on a 42-dimensional handcrafted feature vector extracted from face photos:

Group	Dims	Description
Structural	8	Lesion count, total area, intensity mean/std, area max/std, density, circularity
Multi-scale LBP	27	Uniform LBP histograms at radii R=1,2,3 (9 bins × 3 scales, last bin dropped)
GLCM Texture	3	Contrast, homogeneity, energy (4 angles, 64 levels, dissimilarity removed)
Global Redness	4	Mean + std of LAB a* and YCrCb Cr over all skin pixels

Preprocessing pipeline (OpenCV)

Resize to 512×512
CLAHE (clipLimit=3.0, tileGridSize=8×8) on grayscale
Haar cascade face detection — 4-attempt fallback chain (frontal default + alt2)
ROI mask — eyes, nose, lips blacked out via sub-cascades
YCrCb skin segmentation + morphological close/open
Adaptive lesion thresholding — requires both Cr > thr_cr AND a* > thr_a simultaneously
Connected-component shape filtering — aspect ratio, fill ratio, local a* contrast

Training Data

Subset of the ACNE04 dataset — 3-class balanced split:

Split	acne1 (mild)	acne2 (moderate)	acne3 (severe)
Train	300	300	300
Test	218	61	34

Results (42-dim feature set, 3-class)

Rank	Model	Accuracy	Precision	Recall	F1
🥇 1	SGD Classifier	0.7542	0.7530	0.7542	0.7504
2	CatBoost	0.7318	0.7796	0.7318	0.7494
3	Calibrated Linear SVM	0.7486	0.7503	0.7486	0.7466
4	SVM (RBF)	0.7263	0.7605	0.7263	0.7399
5	LDA	0.7207	0.7615	0.7207	0.7371
6	Logistic Regression	0.7207	0.7588	0.7207	0.7360
7	MLP	0.7151	0.7558	0.7151	0.7301
8	Stacking Ensemble (top-5)	0.7151	0.7511	0.7151	0.7297
9	Ridge Classifier	0.7207	0.7309	0.7207	0.7233
10	KNN	0.7095	0.7394	0.7095	0.7222
11–21	(see Evaluation_Summary(42dim).txt)

Per-class note: acne2 (moderate) is the hardest class — best F1 only 0.456 (SVM RBF). Severe acne3 best F1 0.553 (CatBoost). Full per-class rankings in the summary file.

Files

model_42dim/
  acne_sgd_classifier_model.pkl         ← overall champion  (F1=0.7504)
  acne_catboost_model.pkl               ← best precision     (0.7796)
  acne_svm_rbf_model.pkl                ← best acne2 F1      (0.456)
  acne_stacking_ensemble_model.pkl      ← top-5 stacking
  acne_voting_ensemble_model.pkl        ← top-3 soft voting
  acne_extra_trees_model.pkl
  acne_random_forest_model.pkl
  acne_bagging_svm_model.pkl
  acne_gradient_boosting_model.pkl
  acne_histgradientboosting_model.pkl
  acne_adaboost_model.pkl
  acne_knn_model.pkl
  acne_lda_model.pkl
  acne_logistic_regression_model.pkl
  acne_mlp_model.pkl
  acne_qda_model.pkl
  acne_gaussian_nb_model.pkl
  acne_ridge_classifier_model.pkl
  acne_calibrated_linear_svm_model.pkl
  acne_svm_linear_model.pkl
  acne_svm_polynomial_model.pkl
Evaluation_Summary(42dim).txt           ← full per-class metrics table

Usage

python

import joblib
import numpy as np

# Load champion model
model = joblib.load("model_42dim/acne_sgd_classifier_model.pkl")

# feature_vec: 42-dim numpy array from the extraction pipeline (see preprocessing above)
label = model.predict(feature_vec.reshape(1, -1))[0]
# Returns: 'acne1_1024' (mild) | 'acne2_1024' (moderate) | 'acne3_1024' (severe)

# Confidence (available on most models)
proba = model.predict_proba(feature_vec.reshape(1, -1))[0]

Download via huggingface_hub

python

from huggingface_hub import snapshot_download

path = snapshot_download(repo_id="will702/acne-cv-models")
# models at: path/model_42dim/*.pkl

Requirements

scikit-learn==1.7.2   # must match training version exactly
catboost
joblib
numpy
opencv-python-headless
scikit-image

Limitations

Trained on ACNE04 (Asian skin tones, studio lighting). May underperform on other demographics or lighting conditions.
acne2 (moderate) classification is substantially weaker than acne1/acne3 — class imbalance in test set.
No CNN/deep features — intentionally classical for interpretability and the interactive playground.

Citation

If you use these models, please cite the ACNE04 dataset:

bibtex

@inproceedings{wu2019joint,
  title={Joint Acne Image Grading and Counting via Label Distribution Learning},
  author={Wu, Xiaoping and Liang, Wen and Yu, Kezhou and Xu, Fei and Liang, Weiwei and others},
  booktitle={ICCV},
  year={2019}
}