This paper proposes a supervised learning method for regularization of large-scale inverse problems using primary operators composed of noisy data. This approach is relevant to super-resolution imaging using sampling metrics in inverse scattering theory. This study aims to accelerate the spatiotemporal regularization process for this type of inverse problem to enable real-time imaging. The proposed method uses neural operators to map each pattern on the right-hand side of the scattering equation to a corresponding regularization parameter. The network is trained in two stages: (1) training with a low-resolution regularization map provided by the Morozov discord principle using a non-optimal threshold, and (2) optimizing network predictions by minimizing a Tikhonov loss function conditioned on a validation loss. The second stage allows the approximation map from the first stage to be adjusted for high-quality image generation. This method allows direct learning from test data and does not require prior knowledge of the optimal regularization map. The network trained on low-resolution data rapidly generates a dense regularization map for high-resolution imaging. This paper highlights the importance of the training loss function on the network's generalization performance. In particular, we demonstrate that networks informed by the logic of the disparity principle produce higher-contrast images. In this case, the training process involves multi-objective optimization. In this paper, we propose a novel method that adaptively selects appropriate loss weights during training without additional optimization. The proposed method is synthetically validated for damage evolution imaging of elastic plates. The results demonstrate that the disparity-informed normalization network not only accelerates the imaging process but also significantly improves image quality in complex environments.
