Conformal Prediction for Reliable Image Super-Resolution

Tapabrata Chakraborti, Samiran Dey
Proceedings of the Fourteenth Symposium on Conformal and Probabilistic Prediction with Applications, PMLR 266:756-757, 2025.

Abstract

Single image super-resolution (SISR) has been employed over a wide range of applications to enhance the visual quality and details of images. For training super-resolution (SR) models, low-resolution (LR) images are synthesized from the high-resolution (HR) images. However, these artificial intelligence (AI) methods for SR (like diffusion based or generative adversarial models) have stochastic elements that are inherent to the learning process that can be mitigated but not avoided. Effectively this means that for the same input LR image, different instantiations of the generative process is expected to produce slightly different HR output images. While this might not be an issue for certain applications, and in fact might provide interesting variations in tasks like AI art generation, in certain other high stakes applications like medical image super-resolution, such variations need to be tightly controlled and rigorously quantified. After all, when superresolving a biomedical image (say radiology) one would ideally expect the output to be invariant for a patient if it is a static time-independent image. In fact, though the point of super-resolving a biomedical image is to provide the human expert (or indeed an equivalent AI system) the visual clarity to make a better evaluation, having degradation of clinical features or introduction of spurious morphological features would defeat the purpose, and potentially increase the chances of a false inference. Thus it is important in such high risk applications to predict uncertainty bounds for the generated images using a conformal prediction inspired estimate of maximum calibrated coverage.

Cite this Paper

BibTeX
@InProceedings{pmlr-v266-chakraborti25a, title = {Conformal Prediction for Reliable Image Super-Resolution}, author = {Chakraborti, Tapabrata and Dey, Samiran}, booktitle = {Proceedings of the Fourteenth Symposium on Conformal and Probabilistic Prediction with Applications}, pages = {756--757}, year = {2025}, editor = {Nguyen, Khuong An and Luo, Zhiyuan and Papadopoulos, Harris and Löfström, Tuwe and Carlsson, Lars and Boström, Henrik}, volume = {266}, series = {Proceedings of Machine Learning Research}, month = {10--12 Sep}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v266/main/assets/chakraborti25a/chakraborti25a.pdf}, url = {https://proceedings.mlr.press/v266/chakraborti25a.html}, abstract = {Single image super-resolution (SISR) has been employed over a wide range of applications to enhance the visual quality and details of images. For training super-resolution (SR) models, low-resolution (LR) images are synthesized from the high-resolution (HR) images. However, these artificial intelligence (AI) methods for SR (like diffusion based or generative adversarial models) have stochastic elements that are inherent to the learning process that can be mitigated but not avoided. Effectively this means that for the same input LR image, different instantiations of the generative process is expected to produce slightly different HR output images. While this might not be an issue for certain applications, and in fact might provide interesting variations in tasks like AI art generation, in certain other high stakes applications like medical image super-resolution, such variations need to be tightly controlled and rigorously quantified. After all, when superresolving a biomedical image (say radiology) one would ideally expect the output to be invariant for a patient if it is a static time-independent image. In fact, though the point of super-resolving a biomedical image is to provide the human expert (or indeed an equivalent AI system) the visual clarity to make a better evaluation, having degradation of clinical features or introduction of spurious morphological features would defeat the purpose, and potentially increase the chances of a false inference. Thus it is important in such high risk applications to predict uncertainty bounds for the generated images using a conformal prediction inspired estimate of maximum calibrated coverage.} }
Endnote
%0 Conference Paper %T Conformal Prediction for Reliable Image Super-Resolution %A Tapabrata Chakraborti %A Samiran Dey %B Proceedings of the Fourteenth Symposium on Conformal and Probabilistic Prediction with Applications %C Proceedings of Machine Learning Research %D 2025 %E Khuong An Nguyen %E Zhiyuan Luo %E Harris Papadopoulos %E Tuwe Löfström %E Lars Carlsson %E Henrik Boström %F pmlr-v266-chakraborti25a %I PMLR %P 756--757 %U https://proceedings.mlr.press/v266/chakraborti25a.html %V 266 %X Single image super-resolution (SISR) has been employed over a wide range of applications to enhance the visual quality and details of images. For training super-resolution (SR) models, low-resolution (LR) images are synthesized from the high-resolution (HR) images. However, these artificial intelligence (AI) methods for SR (like diffusion based or generative adversarial models) have stochastic elements that are inherent to the learning process that can be mitigated but not avoided. Effectively this means that for the same input LR image, different instantiations of the generative process is expected to produce slightly different HR output images. While this might not be an issue for certain applications, and in fact might provide interesting variations in tasks like AI art generation, in certain other high stakes applications like medical image super-resolution, such variations need to be tightly controlled and rigorously quantified. After all, when superresolving a biomedical image (say radiology) one would ideally expect the output to be invariant for a patient if it is a static time-independent image. In fact, though the point of super-resolving a biomedical image is to provide the human expert (or indeed an equivalent AI system) the visual clarity to make a better evaluation, having degradation of clinical features or introduction of spurious morphological features would defeat the purpose, and potentially increase the chances of a false inference. Thus it is important in such high risk applications to predict uncertainty bounds for the generated images using a conformal prediction inspired estimate of maximum calibrated coverage.
APA
Chakraborti, T. & Dey, S.. (2025). Conformal Prediction for Reliable Image Super-Resolution. Proceedings of the Fourteenth Symposium on Conformal and Probabilistic Prediction with Applications, in Proceedings of Machine Learning Research 266:756-757 Available from https://proceedings.mlr.press/v266/chakraborti25a.html.

Related Material