Automatic analysis of Arabic text in news videos presents a major challenge due to visual noise, low image quality, and the difficulty of reading Arabic script. In this study, we present a model called Hybrid TransUNet, which integrates the advantages of Vision Transformers (ViTs), capable of understanding the image in its entirety, and of U-Net, known for its accuracy in perceiving visual subtleties. Our approach is based on two complementary steps. The first uses a ViT-based encoder to obtain the overall context of the video image. This allows us to identify areas likely to contain text, even if the color or brightness changes. The second method uses a U-Net-type decoder to refine this information and make detection more precise. It also has a recognition module that transforms detected areas into readable Arabic text. Our tests show that our hybrid model is better than other traditional methods.