<![CDATA[In the digital content era, movie posters serve as rich visual media conveying crucial semantic information rather than mere promotional tools. This research aims to implement EfficientNet, a modern convolutional neural network (CNN) architecture, to automate multi-label movie genre classification based solely on poster images. The study utilized the Kaggle Movie Poster Dataset and evaluated the performance of EfficientNet-B3 against classical CNN baselines, namely VGG16, ResNet50, and InceptionV3. To address the inherent challenge of class imbalance, the methodology incorporated comprehensive data preprocessing, class-weighting, and dynamic threshold-tuning. Experimental results demonstrated that EfficientNet-B3 significantly outperformed the baseline models, achieving the highest F1 Macro score (0.3200) and AUC-ROC (0.7676), while maintaining the lowest Hamming Loss (0.1269). In conclusion, EfficientNet's compound scaling approach provides a robust and highly effective feature extraction mechanism that successfully mitigates majority bias in imbalanced datasets. This study contributes to the advancement of intelligent visual classification systems, particularly within creative industries where supporting metadata is limited. ]]>
Copyrights © 2026
