Abstract

Acne is a skin condition that commonly affects young individuals and can cause permanent scars, emotional stress, and physical discomfort. Current dermatologist treatment is still manual, making it time-consuming and labor-intensive. Therefore, a lightweight, artificial intelligence-based acne localization model that can be implemented on devices with limited resources is needed. This implementation can reduce treatment time for dermatologists. Therefore, this study proposes the YOLOv11nMod model, a modification of YOLOv11n that has fewer parameters and layers than the original version. Our proposed model was tested using 5-fold K-Fold Cross Validation. The data used in this study came from Kaggle in the form of acne facial images. Data preparation was carried out and resulted in 915 image data and 915 txt data. The test data for the data split was 183 data points, with the remainder for training on K-Fold Cross Validation. From the test results, which were conducted by averaging the accuracy of 5 times, our proposed model achieved a detection accuracy of 66.64%. 3.34% Superior to YOLOv12n and 0.02% lower than YOLOv11n and 0.42% lower than YOLOv8n. Although the accuracy obtained is 0.42% lower, our model has a smaller size, namely 1.7 MB, compared to the YOLOv8n, YOLOv11n, and YOLOv12n models. Thus, our proposed model is more efficient than the original version with a similar accuracy. Therefore, our proposed model can be implemented on devices with limited resources. However, the accuracy of external data testing is still very low. Therefore, additional data is needed in future research.

Concepts :

Access to Document

Citations by Year