Blind Single-Image Restoration and Super-Resolution with Inverse PINNs and Trainable PDE Priors
Abstract
Single-image restoration and super-resolution are inherently ill-posed inverse problems in which the degradation operator is rarely known at inference time. Supervised feedforward networks address this ambiguity by learning a prior from large paired datasets, but their performance degrades when the actual degradation deviates from the synthetic corruptions seen during training. This work presents a blind inverse Physics-Informed Neural Network (PINN) that avoids any paired training data by solving one image at a time through a dedicated optimization loop. The proposed system couples three jointly optimized components. First, the unknown high-resolution image is parameterized by a residual SIREN, whose sinusoidal activations provide a spectral bias complementary to the bicubic baseline used as initialization. Second, a differentiable forward model with trainable Gaussian blur and a learnable three-component noise mixture comprising Gaussian, Laplace, and speckle terms adapts the degradation estimate directly to the observed image, without any prior knowledge of the blur kernel or noise statistics. Third, PDE-inspired regularization terms, including Total Variation energy, the Rudin-Osher-Fatemi (ROF) stationary residual, the Perona-Malik anisotropic diffusion residual, and a shock-filter sharpening penalty, are incorporated as additive loss components with exponentially re-parameterized trainable weights, supplemented by a linear warm-up schedule for training stability. Seven inverse PINN variants, differing only in the active PDE prior, are evaluated on the Set5 butterfly image under three scenarios: deblur-denoise $(\times 1), \times 2$ super-resolution, and $\times 4$ super-resolution. All variants consistently outperform bicubic interpolation, with average PSNR gains of +1.09 dB, +1.62 dB, and +3.09 dB, respectively. the spread among PDE variants remains below 0.033 dB in every scenario, demonstrating that the dominant performance driver is the shared blind inverse formulation rather than the specific choice of PDE prior. The estimated blur converges well below the synthetic ground truth, revealing that the solver learns an effective joint decomposition of blur, noise, and image prior rather than an unbiased physical estimate. These results position blind inverse PINNs as a principled and interpretable alternative to supervised superresolution in settings where degradation uncertainty is high and reconstruction quality takes priority over computational throughput.