General Intelligent Imaging and Uncertainty Quantification by Deterministic Diffusion Model

Weiru Fan,Xiaobin Tang,Yiyi Liao,Da-Wei Wang

2024-08-23

Abstract:Computational imaging is crucial in many disciplines from autonomous driving to life sciences. However, traditional model-driven and iterative methods consume large computational power and lack scalability for imaging. Deep learning (DL) is effective in processing local-to-local patterns, but it struggles with handling universal global-to-local (nonlocal) patterns under current frameworks. To bridge this gap, we propose a novel DL framework that employs a progressive denoising strategy, named the deterministic diffusion model (DDM), to facilitate general computational imaging at a low cost. We experimentally demonstrate the efficient and faithful image reconstruction capabilities of DDM from nonlocal patterns, such as speckles from multimode fiber and intensity patterns of second harmonic generation, surpassing the capability of previous state-of-the-art DL algorithms. By embedding Bayesian inference into DDM, we establish a theoretical framework and provide experimental proof of its uncertainty quantification. This advancement ensures the predictive reliability of DDM, avoiding misjudgment in high-stakes scenarios. This versatile and integrable DDM framework can readily extend and improve the efficacy of existing DL-based imaging applications.

Image and Video Processing,Optics

What problem does this paper attempt to address?

The paper attempts to address the problem of how to recover local structures from non-local data in Computational Imaging (CI) tasks. Specifically, existing Deep Learning (DL) based methods perform poorly when handling global-to-local (i.e., non-local to local) transformations due to their locality bias, which makes them ineffective in dealing with non-local patterns. Moreover, current DL methods lack interpretability and uncertainty assessment capabilities, leading to unreliable predictions in high-risk scenarios.

To overcome these issues, the authors propose a new deep learning framework—Deterministic Diffusion Model (DDM). DDM achieves efficient and accurate image reconstruction from non-local patterns through a step-by-step denoising strategy. Unlike traditional Diffusion Probabilistic Models (DPM), DDM employs a deterministic degradation function, reducing the inference steps, accelerating the inference process, and embedding Bayesian inference to quantify the uncertainty and reliability of predictions.

Experimental results show that DDM can reconstruct rich details and precise image contours when dealing with non-local patterns generated by strong scattering and Second Harmonic Generation (SHG), outperforming existing state-of-the-art DL algorithms. Additionally, DDM provides a theoretical framework and experimental validation, demonstrating its effectiveness in uncertainty quantification, ensuring the reliability of predictions and the accuracy of decisions.