vincent lee
Menu

02Computer vision research

Facial skin classification

Testing when better contrast helps a lightweight CNN.

A controlled comparison of three CNN architectures and CLAHE preprocessing across five facial skin condition categories.

View source dataset (opens in a new tab)
My role
Research developer
Team
Two student researchers and two academic supervisors
Project status
Research manuscript
Controlled comparison3 architectures · 6 configurations
Best test accuracy89.47%

MobileNetV3-Large
with CLAHE preprocessing

  1. Dataset3,033 imagesFive facial skin categories
  2. PreprocessingGamma + optional CLAHEImages resized to 224 × 224 pixels
  3. Comparison3 CNN architecturesEach with and without CLAHE
  4. Evaluation456 held-out imagesA common test split

Accuracy comparison · Bars use a 0-100% scale

ArchitectureWithoutWith CLAHE
MobileNetV3-LargeMacro F1: 0.8929 → 0.8952
89.25%89.47%+0.22 pp
MobileNetV2Macro F1: 0.8876 → 0.8628
88.82%86.40%−2.42 pp
EfficientNetB0Macro F1: 0.8421 → 0.8746
84.21%87.50%+3.29 pp
456 held-out images. CLAHE improved MobileNetV3-Large by 0.22 percentage points. Improvements varied by architecture. Research evaluation, without clinical validation.

Context

Facial photographs vary in brightness, shadows, and contrast. For image classification, improving those properties may help a model recognize relevant features, but enhancement can also amplify noise or obscure useful detail.

This research investigated that trade-off using lightweight CNNs for five facial skin condition categories: redness, dark spots, inflammatory acne, non-inflammatory acne, and pigmentation.

My contribution

I led engineering and dataset preparation, developed the preprocessing workflow, and conducted model experiments and performance analysis. The research involved another student researcher and two academic supervisors.

The manuscript is titled MobileNetV3-Based Facial Skin Condition Classification with CLAHE-Based Data-Centric Optimization for Teledermatology.

Approach

We organized 3,033 images into a 70% training, 15% validation, and 15% test split. Images were resized to 224 × 224 pixels and processed with adaptive gamma correction. For the enhanced configurations, CLAHE was applied to the luminance channel in LAB color space to preserve color information.

MobileNetV3-Large, MobileNetV2, and EfficientNetB0 were each evaluated with and without CLAHE. Training first fitted a custom classification head with the backbone frozen, then fine-tuned the top 30% of backbone layers. Augmentation was applied to training images.

Results

Architecture CLAHE Accuracy Macro F1
MobileNetV3-Large No 89.25% 0.8929
MobileNetV3-Large Yes 89.47% 0.8952
MobileNetV2 No 88.82% 0.8876
MobileNetV2 Yes 86.40% 0.8628
EfficientNetB0 No 84.21% 0.8421
EfficientNetB0 Yes 87.50% 0.8746

CLAHE improved MobileNetV3-Large accuracy by 0.22 percentage points and EfficientNetB0 by 3.29 points, while MobileNetV2 fell by 2.42 points. The useful finding was the variation across architectures: preprocessing needed to be tested as part of each model configuration.

Limitations and lessons

These are research-dataset results. Visually similar categories remained difficult to separate, and the evaluation does not establish clinical performance or smartphone inference efficiency.

Further work includes broader skin-tone coverage, diverse real-world photographs, repeated runs, explainability, and evaluation on mobile hardware. The central lesson was to measure the effect of an enhancement method rather than assume that clearer-looking inputs always produce better predictions.

MY CONTRIBUTION AT A GLANCE

  • Led engineering and dataset preparation for the classification workflow.
  • Developed adaptive gamma correction and luminance-channel CLAHE preprocessing.
  • Conducted model comparisons, evaluation, and analysis with academic supervision.