Document images often suffer from specular highlights caused by reflective surfaces or uneven lighting conditions, which significantly compromise document readability and reduce optical character recognition (OCR) accuracy in camera-captured document images. However, current document specular highlight datasets face critical limitations such as low resolution, unrealistic synthetic highlights, and insufficient diversity, restricting their applicability to real-world scenarios. In addition, existing highlight removal methods are primarily designed for natural scenarios, which struggle to preserve fine-grained textual details and structural consistency required in real-world documents. To address these challenges, we first introduce DocHighlight, a high-resolution, real-world dataset specifically designed for document specular highlight removal. DocHighlight comprises 2,201 paired images captured under diverse conditions, featuring various document types, illumination settings, and capture devices. Subsequently, we propose Document Specular Highlight Removal Network (DocSHRNet), a new highlight removal method incorporating the Document Structure Attention (DSA) and Adaptive Receptive Field (ARF) modules. These modules facilitate precise structural preservation and adapt to multi-scale highlight patterns, ensuring high-quality restoration. Extensive experiments on the DocHighlight, RD, and SD1 datasets demonstrate that DocSHRNet delivers competitive performance in reconstruction quality and OCR accuracy. These results demonstrate the effectiveness of DocHighlight as a real-world dataset and the robustness of DocSHRNet in addressing document specular highlight removal challenges, providing a solid foundation for real-world applications. The dataset and code are publicly available at https://github.com/shallweiwei/DocSHRNet .