Deep Learning Approaches for Hair Disease Classification: A Comparative Analysis of MobileNetV2 and VGG19 Architectures

This work primarily looks into comparing two CNN architectures, namely MobileNetV2 and VGG19, in the light of hair disease classification. Models will be trained and tested with respect to their performances on a balanced set of 12,000 images of ten different types of hair diseases obtained from Kaggle. The dataset will be preprocessed based on resizing into 224×224 pixel images, normalizing, and augmentation through rotation, flipping, zooming, and change of brightness. All the data was divided into an 80:10:10 split for training, validation, and testing datasets. Both models were trained with the Adam optimizer using a learning rate of 0.001 over 20 epochs. Training on callbacks such as ReduceLROnPlateau and ModelCheckpoint was used to improve the performance on models. According to the evaluation metrics, VGG19 did perform slightly better compared to MobileNetV2 by giving an accuracy of 98% compared to 97%. VGG19 also provided superior precision and recall for the specific conditions, such as Alopecia Areata and Male Pattern Baldness. In contrast, MobileNetV2 was much faster in terms of inference time with fewer computational resources due to the lightweight network structure. Results are hence indicative that while VGG19 yields slightly better accuracy, MobileNetV2 offers an efficient trade-off between efficiency and performance. It is thus indicated in resource-constrained environments. This comparative study focuses on the detection of hair disease using automated tools and underlines the choice of the model based on requirements.