Sancar Yasin
Open Education Faculity, Atatürk University, Erzurum, Yakutiye, Turkey.
PeerJ Comput Sci. 2025 Apr 15;11:e2841. doi: 10.7717/peerj-cs.2841. eCollection 2025.
Quick-response (QR) codes have become an integral component of the digital transformation process, facilitating fast and secure information sharing across various sectors. However, factors such as low resolution, misalignment, panning and rotation, often caused by the limitations of scanning devices, can significantly impact their readability. These distortions prevent reliable extraction of embedded data, increase processing times and pose potential security risks. In this study, four super-resolution models Enhanced Deep Super Resolution (ESDR) network, Very Deep Super Resolution (VDSR) network, Efficient Sub-Pixel Convolutional Network (ESPCN) and Super Resolution Convolutional Neural Network (SRCNN) are used to mitigate resolution loss, rotation errors and misalignment issues. To simulate scanner-induced distortions, a dataset of 16,000 computer-generated QR codes with various filters was used. In addition, super-resolution models were applied to 4,593 QR codes that OpenCV's QRCodeDetector function could not decode in real-world scans. The results showed that EDSR, VDSR, ESPCN and SRCNN successfully read 4,261, 4,229, 4,255 and 4,042 of these QR codes, respectively. Furthermore, the EDSR, VDSR, ESPCN and SRCNN models trained by OpenCV's deep learning-based WeChat QR Code Detector function to read 2,899 QR codes that were initially unreadable and simulated on the computer were able to successfully read 2,891, 2,884, 2,433 and 2,560 of them, respectively. These findings show that super-resolution models can effectively improve the readability of degraded or low-resolution QR codes.
快速响应（QR）码已成为数字转型过程中不可或缺的一部分，有助于在各个部门实现快速、安全的信息共享。然而，诸如低分辨率、未对齐、平移和旋转等因素，通常是由扫描设备的局限性导致的，会显著影响其可读性。这些失真会妨碍可靠地提取嵌入数据，增加处理时间，并带来潜在的安全风险。在本研究中，使用了四种超分辨率模型，即增强深度超分辨率（ESDR）网络、非常深度超分辨率（VDSR）网络、高效亚像素卷积网络（ESPCN）和超分辨率卷积神经网络（SRCNN），以减轻分辨率损失、旋转误差和未对齐问题。为了模拟扫描仪引起的失真，使用了一个包含16000个带有各种滤镜的计算机生成的QR码数据集。此外，将超分辨率模型应用于4593个QR码，这些QR码在实际扫描中OpenCV的QRCodeDetector函数无法解码。结果表明，EDSR、VDSR、ESPCN和SRCNN分别成功读取了其中的4261个、4229个、4255个和4042个QR码。此外，通过OpenCV基于深度学习的微信QR码检测器函数训练的EDSR、VDSR、ESPCN和SRCNN模型，用于读取2899个最初不可读且在计算机上模拟的QR码，分别成功读取了其中的2891个、2884个、2433个和2560个。这些发现表明，超分辨率模型可以有效地提高退化或低分辨率QR码的可读性。
