Media-Text: a Media Industry-Based Dataset for Scene Text Detection

Detection of text in digital images used in the media industry is challenging, mostly due to the lack of specialized datasets and the unknown performance of existing models. Here, a unique Media-Text dataset, comprising 400 images of banners, posters, and covers, annotated with 7,744 text instances is introduced. Two state-of-the-art deep learning models trained on three publicly available datasets were evaluated on the Media-Text dataset to test their generalization capabilities. Additionally, the application of transfer learning techniques with synthetic data and their impact on improving model performance was tested. Regardless of the model and training dataset used, a significant decrease in text detection performance on the media-type dataset was observed, in comparison to hold-out test sets of public data. These results indicate a strong need for media-specialized datasets to enhance the robustness and adaptability of existing text detection models. Future work will explore the use of diverse data sources and advanced model architectures, such as transformers, to potentially improve model generalization. The dataset and code are available at https://github.com/ZAEDPolSl/MediaText.