Cognitive Digital-Twin-Guided Image Restoration: A Difference-Map-Driven Deep Learning Framework for Compound Degradation Correction

Abstract

Image restoration under compound, realistic degradation simultaneous sensor noise, motion or defocus blur, low-illumination attenuation, and quantization remains a persistent challenge because most restoration networks are trained to invert a single degradation type and therefore generalize poorly to overlapping distortions encountered in real imaging pipelines. This paper proposes a Cognitive Digital-Twin-Guided Restoration (CDT-Restore) framework that decomposes restoration into two cooperating stages. First, a Digital Twin Predictor network learns a virtual expectation of the clean scene directly from the degraded observation, functioning as a data-driven digital twin of the undistorted image. Second, a Cognitive Restoration Network consumes the degraded image, the digital twin's expected image, and an explicit pixel-wise difference map and its absolute value as a four-way fused input, allowing the network to reason about where and how the observation deviates from its expected clean state before applying a learned residual correction. The framework is trained end-to-end with a composite Mean Absolute Error and Structural Similarity (SSIM) loss and evaluated using PSNR, SSIM, and MAE rather than classification accuracy, since restoration is a continuous-valued regression task rather than a discrete labelling task. On a genuine CIFAR-10-derived pilot benchmark subjected to a compound degradation operator (Gaussian noise, average-pool blur, stochastic low-light attenuation, and 5-bit quantization), the proposed Cognitive Restorer achieves 20.31 dB PSNR and 0.8023 SSIM, improving on the degraded input (18.82 dB, 0.7555) and a matched-capacity baseline U-Net restorer without digital-twin guidance (17.97 dB, 0.7624) by 2.34 dB and 0.0399 SSIM respectively, while reducing MAE from 0.1137 to 0.0829. We report these as genuine measured results from an executed, reduced-scale training run (900 training / 200 test images, 16–22 epochs per network due to CPU-only compute availability) and provide the exact architecture, hyperparameters, and a full-scale reproduction protocol so that the reported gains can be validated at production scale (standard CIFAR-10 splits, GPU training, extended epochs) prior to journal submission. The central contribution is architectural and mathematical: an explicit difference-map fusion mechanism that converts an implicit digital twin prediction into an explainable deviation signal usable by a downstream correction network, together with a justification of why this decomposition should outperform single-pass restoration as degradation compounding increases.