<![CDATA[High interest in eyelash extension services at Le'Goub Beauty is accompanied by challenges in visual perception between clients and therapists, which affect the consistency of final results. This study compares three deep learning models: a basic CNN, MobileNetV2, and EfficientNet-B0. The comparison is used to classify three types of eyelash extensions (anime, natural, and volume) from 662 internal salon images, using a stratified k-fold cross-validation scheme (K=5) along with a two-phase training approach (freezing and fine-tuning) and the Adam optimizer. The basic CNN, without transfer learning, produces low and unstable performance (36.02% accuracy, SD 11.98%), with a tendency toward bias for the anime class. Applying transfer learning improves performance considerably: MobileNetV2 achieves a higher average accuracy of 82.59% (SD 2.65%), slightly above EfficientNet-B0's 80.95%, though EfficientNet-B0 shows greater stability (SD 1.76%). Both models perform consistently in the anime class but continue to have difficulty distinguishing between the natural and volume classes due to their visual similarity. MobileNetV2 is recommended as the primary model for Le'Goub Beauty, with EfficientNet-B0 as an alternative in scenarios where consistency is a priority. Keywords: Deep learning; EfficientNet-B0; MobileNetV2; image classification; eyelash extension.]]>
Copyrights © 2026
