Comparative Analysis of Deep Learning Models for Alcohol Scene Detection: A Comprehensive Benchmark Study

The rapid expansion of video content on digital platforms has underscored the importance of automated detection of alcohol-related scenes, a task pivotal for content moderation, public health surveillance, and regulatory compliance. This study rigorously evaluates eight deep learning models—Convolutional Neural Networks (CNN), DenseNet, ResNet, Faster R-CNN (FRCNN), Inception V4, EfficientNet B4, AlexNet, and GoogLeNet—for binary classification of alcohol versus non-alcohol scenes. Utilizing a dataset of 3,000 images sourced from Kaggle and Google Images, processed via Python with TensorFlow/Keras on Google Colab, the models were trained over 10 epochs with a batch size of 80. Performance metrics—accuracy, precision, recall, and F1-score—were derived from confusion matrices, complemented by training and testing accuracy/loss curves. GoogLeNet emerged as the top performer with 86.5% accuracy and 87.5% F1-score, followed by FRCNN (82.7% accuracy, 83.1% F1-score) and Inception V4 (80.7% accuracy, 81.9% F1-score). In contrast, EfficientNet B4 lagged at 56.3% accuracy, highlighting significant variability across architectures. These findings affirm GoogLeNet’s suitability for video frame analysis and propose future directions for temporal modeling and scalability.