<![CDATA[Deep learning based acne severity classification requires consistent color distribution and image illumination for stable feature extraction. Color imbalance, noise, and lighting variations can affect the accuracy and generalization ability of the model, making image preprocessing optimization a crucial aspect in dermatological classification. This study analyzes the impact of image preprocessing on the performance of EfficientNet-B5 in classifying three levels of acne severity using the Kaggle Acne Grading dataset (999 images; 80% training, 20% testing). The experiment compares the default preprocessing (resize, normalization) with the proposed preprocessing: gray-world white balance for color stabilization, bilateral filtering for edge preservation, and adaptive gamma correction for adaptive illumination. The evaluation uses accuracy and loss curves, confusion matrices, and classification reports, focusing on the macro F1-score to assess the balance between precision and recall. The results show a slight increase in accuracy from 77% to 78%, a macro F1 score of 75%, and more controlled overfitting with smaller differences in accuracy and loss between training and validation. Improving image quality before feature extraction contributes to feature representation and balance in multi-class classification.]]>
Copyrights © 2026
