A distorted QR code correction method based on heatmap regression

Abstract To address the issue of Quick Response Code (QR code) decoding failure caused by complex geometric distortions, such as surface attachment and folding deformation, this paper proposes a deep learning correction network based on heat map regression. The method dynamically adjusts the sensory field to adapt the distortion intensity through adaptive dilation convolution (ADConv) to preserve the geometric accuracy of the position detection pattern; designs the grouped CBAM enhancement module (GCE) to group the feature channels for parallel processing to strengthen the significant features of the boundary and key regions; transforms the coordinate prediction into probabilistic heatmap regression and combines with the sub pixel offset prediction to achieve the control point of the high precision The final reconstruction of the regular QR code image is performed by thin-plate spline transformation (TPS). Validation on a dataset containing 12,000 warped QR codes demonstrates that the scheme is successful in terms of PSNR (10.35 dB), SSIM (68.40%), and decode rate (91.4%). The correction scheme proposed in this paper is robust to the correction of severely deformed QR codes, such as large curvature distortion or folding, and provides reliable technical support for logistics traceability, mobile payment, industrial code reading, and other scenarios.