Key points are not available for this paper at this time.
The field of skin disease categorization has seen substantial changes since the introduction of MobileNet, which is well-known for its adaptable architecture and high computational efficiency. To increase flexibility, this study investigates the use of MobileNet in the field of dermatological diagnosis. Transfer learning using models that have been trained is utilized. An extensive collection of skin disease images is included in the experimental design, and the efficacy of the model is thoroughly assessed in terms of training accuracy, testing accuracy, training loss, and testing loss. The outcomes show how well MobileNet works to strike a compromise between precise classification and computational effectiveness. The model exhibits impressive performance in a range of skin conditions, providing a promising path toward the readily available and real-time classification of skin diseases. The proposed model has been fine-tuned and implemented by changing the learning rates to 0.1, 0.01, 0.001, and 0.0001. Further, the accuracy has been identified as the highest at 96.33% with a learning rate value set to 0.01. Further, the optimizers have been also changed from SGD to Adam, which resulted in Adam showing the highest accuracy of 97.78% at epoch value 33. The dataset for the implementation of the proposed model has been collected from Kaggle, which contained 1159 images. However, these images have been augmented by performing two different zooming operations resulting in an image count to 3477. The model has depicted good accuracy, however, this accuracy can be further enhanced by performing more augmentation operations such as shearing, rotation etc.
Kaur et al. (Thu,) studied this question.