Learning Under Multiplicative Noise: Principled Image Enhancement Frameworks for Coherent Imaging 35 views

Computer vision has made a significant impact in numerous scientific fields, enabling specialized algorithms to analyze complex visual data accurately and efficiently. These algorithms have substantially enhanced the performance of traditional image processing techniques and have become indispensable in almost all scientific fields. However, in many of these applications, the acquired data is noisy and spatially degraded due to inherent limitations of the imaging systems and acquisition processes. Such degradations often obscure fine structural details present in the underlying signal and limit the performance of downstream analysis. Despite the remarkable success of deep learning algorithms, these algorithms are data-driven and require large volumes of labeled training data for optimal performance. In several scientific applications, access to large, high-quality datasets remains a challenge. To address these challenges, researchers have explored image enhancement algorithms to improve visual quality and improve downstream performance. Diffusion models have emerged as the state-of-the-art methods for image enhancement and augmentation tasks in recent years. However, these models underperform in real-world scientific applications where noise is signal-dependent. In coherent imaging modalities such as ultrasound, Synthetic Aperture Radar (SAR), sonar, electron microscopy, and laser imaging, the acquired data are often degraded by signal-dependent speckle and modality-specific spatial distortions. This dissertation develops principled image enhancement algorithms tailored to coherent imaging by integrating core image processing concepts, such as speckle and point spread function (PSF), with cutting-edge deep learning architectures. In particular, this dissertation introduces the first diffusion model derived from the first principles of multiplicative noise to reduce speckle. This framework is then extended to propose a PSF-aware diffusion model to enhance images that are corrupted with speckle and further distorted by the PSF of the imaging device. To address applications where images are corrupted with speckle and signal-independent noise and further degraded by the PSF-induced spatial distortions, a diffusion-inspired denoising framework is proposed. Finally, this dissertation introduces a novel self-supervised algorithm for reducing multiplicative noise. Since acquiring clean ground truth images is often infeasible, supervised restoration algorithms face inherent limitations. The proposed model-agnostic framework can efficiently reduce speckle following gamma and Rayleigh distributions without requiring clean, ground truth images. In general, the dissertation advances principled, physics-informed data-driven algorithms for coherent imaging. The proposed methods significantly impact the enhancement of speckle-corrupted images and benefit a wide range of scientific and medical applications where the image quality is distorted by the modality-specific degradations. Furthermore, this dissertation introduces approaches that eliminate the need for iterative refinement for image enhancement algorithms, thus significantly reducing the inference time for enhancing images with diffusion models.

PHD (Doctor of Philosophy)

Speckle; Image denoising; Deep learning; Image analysis; Diffusion models; Signal processing; Point Spread Function; Signal-dependent noise; Physics-informed deep learning; Coherent imaging; Image enhancement; Multiplicative noise

English

2026-04-27