Analisis Klasifikasi Gambar Deteksi Merokok dengan Metode CNN yang Ditingkatkan Menggunakan Model Fine Tuning pada Arsitektur MobileNetV3L, EfficientNetV2M, dan Vision Transformer

Abstract

Smoke detection faces challenges in detecting small and common events such as smoking, using deep learning techniques. Issues like these have resulted in unsatisfactory privacy and accuracy models. In the EfficientNetV2M model, the author first uses data augmentation to increase the amount and diversity of training data by carrying out transformations of existing data. A lower learning rate allows smoother parameter updates and can improve the final performance of the model, fine-tune the EfficientNetV2M Layer, and Find the Ideal Learning Rate with the LearningRateScheduler Callback. The improved performance in terms of accuracy and robustness shows that this method can be used in related fields and represents significant progress in the field of burn detection with an accuracy rate of up to 97%. In the MobileNetV3L model, the author obtained lower resource usage results, namely with an accuracy rate of 87%. In the Vision Transformer model, the author uses a custom ViT (Vision Transformer) model for the feature extraction stage, then applies PCA for dimensional problems, and finally uses the XGBoost model for the classification stage and gets very satisfying results, namely with an accuracy level of 96 %. Future efforts will focus on improving this technology and finding ways to use it in broader contexts.