In this study, we propose a lightweight causal Mamba network (LCM-QRIR) for blurred QR code image restoration. Its core architecture comprises a HOGMamba module that integrates HOG features with Mamba, which is employed to process long-range spatial dependencies in degraded QR code images with linear complexity. To further adapt to degraded image restoration, we introduce a dynamic interactive feed-forward (DIFF) module to promote channel-space interaction, a wavelet downsampling enhancement (WDE) module to mitigate information loss during encoder downsampling, and a sparse artifact similarity-weighted loss combined with a distortion invariant learning (DIL) strategy to guide the model toward learning more invariant features. By evaluating on the blurred QR code image dataset (BQRCI) and the publicly available GoPro dataset for natural image deblurring tasks, experimental results demonstrate that LCM-QRIR exhibits competitive advantages in both performance metrics and computational complexity. Furthermore, the model displays robust capabilities in interpretability and causal reasoning.
Zhang et al. (Fri,) studied this question.