The paper introduces a novel method for signature detection and restoration in document images, which is robust to remove noise like stamps, seals and dates. The signature detection involves, object detection deep learning models such as YOLOv8 as single-stage detection model and Detectron2 as two-stage detection model with pre-trained weights ResNet-50 C4, ResNet-50 FPN, ResNet-101 C4, ResNet-101 FPN, and ResNeXt-101 32 × 8d FPN. After detection, the signature regions are extracted and restored using an advanced cleaning model like CycleGAN, Morphological operations, and Pix2Pix to eliminate noise and to preserve the sign information. After the cleaning process, the restored signatures are compared with its ground-truth signatures to ensure distinctive qualities. The proposed approach achieves 94% detection accuracy and restoration with an effectiveness of 98% accuracy. From the advanced cleaning methods, Pix2Pix method outperforms best in removing noise. This novel result, highlight its potential for real-world applications involving noisy document images and document authenticity.