Abstract
Recognizing acne disease and evaluating its type is vital for the efficacy of the medical treatment. This report collects a dataset of 420 images and then labels them into seven different classes by a well-experienced dermatologist. After a pre-processing step, including local and global contrast enhancement and noise removal by a smoothing filter, the dataset size is enhanced using augmentation. The images of the dataset and the augmented ones are all fed into a novel integrated dual deep convolutional neural network (CNN) model to recognize acne disease and its type by classifying it into seven groups. First, two CNN-based units are designed to extract deep feature maps, later combined in a feature aggregation module. The aggregated features provide rich input information and classify the acne by a softmax. The proposed architecture's optimizer, loss function, and activation functions are all tuned so that both CNN units are trained with minimum kernel size and fewer training parameters. Thus, the computational cost is minimized. Compared with three machine learning-based classifiers and five pre-trained models, our model achieves competitive state-of-the-art performance with an accuracy of 97.53% on the developed dataset.
