Efficient and Robust Scene Text Classification: Distinguishing Natural and Artificial Text Using Compact Deep Learning Models

This paper addresses the task of classifying detected scene text in images as either natural or artificial, a key capability for enabling computer vision systems to interpret text contextually. We extended the original MediaText dataset by adding detailed annotations and constructed a dedicated training set. In the original dataset, we annotated 6,775 legible instances (5,304 artificial, 1,471 natural), and added 9,037 new (5,350 artificial, 3,687 natural). We evaluated several most popular backbones combined with a lightweight classification module. Hyperparameters were optimized using a Random Search strategy. EfficientNet-based models achieved the best results on the independent test set (balanced accuracy equal to 69.7% and AUC equal to 0.76). Despite fewer parameters than ResNet50, it outperformed them, confirming the effectiveness of compact architectures with minimal classification heads. These findings suggest that lightweight models can achieve competitive performance in scene text classification tasks, offering a favorable trade-off between accuracy and computational efficiency. The updated dataset is available at GitHub repository: https://github.com/ZAEDPolSl/MediaText.