Since the emergence of QR codes, ensuring their reliable scanning and decoding in different environments has been a research hotspot. Relevant studies have mainly aimed at improving the reliability of QR codes from the perspectives of error correction codes, denoising techniques, and image enhancement. In terms of error correction codes, QR codes primarily utilize Reed–Solomon error correction techniques to rectify reading errors, enhancing the data reliability and readability [27]. However, in cases of actual damage, the performance of this technique is not ideal. Due to potential interferences such as stains, blurriness, rotation, and scaling, researchers have developed a range of anti-interference techniques, including filtering, denoising, rotation correction, and scale normalization, to reduce the impact of interference on QR code recognition. In terms of denoising techniques, Tomoyoshi Shimobaba et al. (2018) proposed a holographic image restoration algorithm using autoencoders. Through numerical diffraction calculations or holographic optical reconstruction, they obtained reconstructed images and utilized autoencoders to remove image noise pollution, thus restoring clearer holographic images of QR codes [28]. Furthermore, researchers have employed the Cahn–Hilliard equation to address QR code image restoration, particularly in restoring low-order sets (edges, corners) and enhancing edges [21]. When dealing with severely damaged QR codes, the effectiveness of traditional image processing techniques is relatively limited. As a result, in recent years, researchers have shifted their focus towards deep learning technology. For blurred QR images, researchers have begun exploring from the perspective of deep learning. For instance, Michael John Fanous et al. proposed GANscan, a high-speed imaging method based on generative adversarial networks, which is employed to capture QR codes on rapidly moving scanning devices. This method primarily utilizes GANs to process motion-blurred QR video frames into clear images [12]. The above studies extended QR technology to adapt to various environments from different perspectives. However, these methods mainly addressed QR code blurriness and noise issues, with limited research on addressing damaged QR codes.

Image inpainting refers to the task of reconstructing lost or corrupted parts of images based on the surrounding available pixels. Its core idea is to utilize spatial continuity and texture similarity in natural images to synthesize plausible content for missing regions. Early methods relied on traditional signal processing techniques for image completion, extending based on the inherent image similarity and structure, yielding limited effects. For example, pixel-by-pixel and block-by-block filling, both of which start from the boundaries of the image’s damaged areas, gradually fill unknown regions in images using known information from the surrounding or similar areas in the image based on calculated priorities, aiming to synthesize visually continuous images [25,26]. However, this method is only suitable for small area restoration, and it presents issues of blurriness and unnatural textures in repairing complex backgrounds and large missing areas. In recent years [29], deep learning has seen rapid development, particularly excelling in image inpainting tasks, enabling the learning of intrinsic priors in images and yielding more realistic completion results. For example, the Context Encoders model utilizes end-to-end convolutional neural networks for image completion, representing one of the earliest successful applications of deep learning to this task. Its core innovation lies in the introduction of the encoder–decoder structure of generative models, which can directly process images with holes [30]. Given the development of deep learning in image inpainting tasks in recent years, more people have focused on its applied research, such as object removal and photo restoration. For example, Wan et al. proposed a method using deep learning to restore severely degraded old photos. The method involves training two VAEs to transform old photos and clean photos into latent spaces and learning transformations between the two latent spaces on synthesized image pairs, and then designing global and local branches to handle structural and non-structural defects in old photos, respectively. The two branches are fused in the latent space to enhance the recovery from composite defects [14]. This provides new perspectives for researching QR code restoration. However, unlike ordinary image restoration tasks such as object removal and photo restoration, QR images contain semantic information, requiring a delicate balance between the visual effect and readability. How to employ the image inpainting techniques for high-quality QR code restoration still offers many unexplored possibilities.