In real-world scenarios, text embedded in images is frequently subjected to corruption due to various factors such as occlusion, noise, fading, or physical damage. These degradations are prevalent in diverse applications, including street view text recognition and scene text understanding, significantly hindering the utility of text in downstream tasks such as optical character recognition and natural language processing. Existing methods for text image inpainting often rely on additional prompts or external data to guide the restoration process, which not only increases complexity but also limits their applicability in scenarios where such auxiliary information is unavailable. To address these challenges, we propose a novel approach, termed MultiModalTextInpaintingNetwork (MMTINet), which directly extracts and fuses textual semantics from the image itself without requiring any external prompts. By leveraging the inherent textual information within the image, our method achieves a more seamless and context-sensitive restoration, significantly improving the accuracy and fidelity of the reconstructed text. We evaluate MMTINet on a benchmark dataset, demonstrating its effectiveness in recovering missing or occluded text while preserving the original content and style. Experimental results highlight the efficacy and potential of our approach, showcasing its superior performance in advancing text restoration tasks through image-driven text semantics.