Excel to HTML
MONDAY 2 MARCH 2026
Wave-Based Scientific Imaging (Joint Session with Computational Imaging)
15:30 - 16:30
Grand Peninsula C
15:30MLSI-301
Coherent computational imaging with synthetic waves, [view abstract]
Imaging through scattering scenes or materials severely limits the visual acuity of optical imaging systems. This talk discusses how diversity in illumination wavelength can be utilized to circumvent the problem of phase randomization in scattered light fields. Our technique probes the scene at two closely spaced optical wavelengths and computationally assembles a complex "synthetic field" at a "synthetic wavelength," which is used for further processing. As the synthetic wavelength is the beat wavelength of the two used optical "carrier" wavelengths, it can be picked orders of magnitudes larger, and the computationally assembled synthetic field becomes immune to scatter. Amongst other applications, the introduced method allows for holographic measurements of hidden objects through scattering media or around corners, or for interferometric measurements of macroscopic objects with rough surfaces. During the talk, different flavors of the technique will be introduced, including a method to retrieve the complex synthetic field in single-shot.
15:50MLSI-302
Generative AI and pre-trained foundation models for seismic imaging and interpretation, [view abstract]
Recent advances in Generative AI (GenAI) and pre-trained general-purpose foundation models are redefining how we image and interpret the subsurface. We present an integrated framework that unifies physics-informed generative modeling, diffusion-based synthesis, and transfer-learned foundation models for seismic imaging and interpretation. Generative diffusion transformers simulate broadband seismic wavefields consistent with elastic physics, enabling virtual dense-array simulations and uncertainty-aware data augmentation. Building on these physics-consistent representations, pre-trained vision foundation models are adapted to the seismic domain via parameter-efficient fine-tuning--using low-rank adaptation and prefix prompting--to retain general visual priors while embedding geological constraints through relative-geologic-time and graph-based reasoning. Self-supervised neural operators coupled with partial-differential-equation constraints unfold and unfault complex 3-D structures, allowing geological interpretation of seismic images without significant human expertise. Together, these components establish a scalable GenAI-foundation-model framework that bridges physical modeling, representation learning, and structural reasoning. By unifying pre-trained representations and geological priors, this paradigm moves seismic interpretation beyond task-specific networks toward adaptive, interpretable, and foundation-scale digital twins of the subsurface--marking a new era for data-driven Earth imaging.
16:10MLSI-303
A CNN workflow for stochastic seismic property estimation, [view abstract]
As one of the major tools in resolving the non-uniqueness challenge in subsurface interpretation and reservoir characterization, stochastic inversion from post- and pre-stack seismic data remains challenging, which not only requires heavy computational resources but also relies on intensive manual supervision. Inspired by the recent advances in deep learning particularly convolutional neural networks (CNNs) for interdisciplinary data integration, in this talk, we will first present a deep learning workflow that enables stochastic property estimation by efficiently integrating seismic images with sparse wells. It starts with sampling a set of property prior models (PPMs) from densely-measured actual logs at well locations and corrupting local seismic patterns with Gaussian noise. The core idea is to train a structure-guided CNN by mapping the contaminated seismic with the sampled PPMs while enforcing structural consistency to avoid overfitting in the common scenario of sparse wells. The baseline and uncertainty of target properties are estimated by running multiple realizations of the trained CNN. Next, we will demonstrate how the proposed workflow performs over three use cases, including rock acoustic/elastic property estimation from 3D post-/angle-stack seismic and soil geotechnical properties from 2D ultra-high-resolution seismic. Finally, we end with a summary and discussion on future work.
TUESDAY 3 MARCH 2026
Generative AI for Imaging (Joint Session with Computational Imaging)
08:30 - 09:30
Grand Peninsula C
08:30MLSI-304
Metric 3D human pose estimation via CAD-calibrated monocular depth, [view abstract]
The estimation of 3D human pose from a single 2D image is a fundamentally ill-posed problem due to the inherent ambiguity between an object's scale and its depth. This challenge is particularly pronounced in applications within geometrically constrained environments, where metric-scale (absolute) joint positions are critical for real-world interaction, monitoring, or safety systems. This work presents a novel, multi-model deep learning framework that achieves robust absolute 3D pose estimation using only a single RGB or IR camera. By decoupling the tasks of relative pose estimation and absolute depth inference and introducing environment-specific geometric priors, our system overcomes the limitations of traditional monocular approaches. Our work uses a vehicle cabin as a primary example of such a constrained environment for in-cabin sensing, but the methodology is applicable to robotics, surveillance, factory automation etc.
08:50MLSI-305
Self-supervised learning for spatiotemporally scalable computational imaging, Kevin Zhou, University of Michigan (US) [view abstract]
Spatiotemporal throughput can be defined as the number of pixels or voxels per second an imaging system can capture, governing its field of view, spatial resolution, and frame/volume rate. Conventional optical imaging systems face a variety of bottlenecks and tradeoffs that limit their spatiotemporal throughput, such as scanning, asynchronous/sequential acquisitions, aberrations, and limited detector bandwidths. This talk presents multiple large-scale computational imaging systems that alleviate such bottlenecks through a synchronized array of cameras and a corresponding large-scale reconstruction algorithm that computationally fuses the parallel video streams. While such parallelized computational imaging systems can capture data at very high throughput, the associated iterative reconstruction algorithms struggle to keep up, especially as the number of cameras and/or the length of the video increases. To this end, we discuss self-supervised learning-based strategies that establish a mapping between the raw data and the desired reconstruction using a fraction of the acquired data. The trained model can then generalize to the remainder/future unseen frames, enabling scalability to arbitrarily long videos.
Imaging Pipelines I (Joint Session with Computational Imaging)
09:30 - 10:30
Grand Peninsula C
09:30COIMG-127
Computational imaging for aberration correction, Laura Waller, UC Berkeley (US) [view abstract]
Computational imaging is permeating cameras and microscopes across many scientific applications, enabling new high-resolution and multi-dimensional measurement capabilities (e.g. phase, 3D, hyperspectral). Many methods require acquisition of multiple images across a large field-of-view, so motion blur and aberrations can cause severe artifacts. This talk will describe new space-time algorithms that correct aberrations and motion artifacts with imperfect optical systems or approximate forward models. Traditional model-based image reconstruction algorithms work together with neural networks to optimize the inverse problem solver and the data capture strategy.
09:50COIMG-128
Efficient-coding--inspired in-sensor compression, Yuhao Zhu, University of Rochester (US) [view abstract]
Energy-efficient image acquisition on the edge is crucial for enabling remote sensing applications where the sensor node has weak compute capabilities and must transmit data to a remote server/cloud for processing. To reduce the edge energy consumption, this paper proposes a sensor-algorithm co-designed system, which compresses raw pixels in the analog domain inside the sensor. We use coded exposure (CE) as the in-sensor compression strategy as it offers the flexibility to sample, i.e., selectively expose pixels, both spatially and temporally. We make three contributions. First, we propose a task-agnostic strategy to learn the sampling/exposure pattern based on the classic theory of efficient coding. Second, we co-design the downstream vision model with the exposure pattern to address the pixel-level non-uniformity unique to CE-compressed images. Finally, we propose lightweight augmentations to the image sensor hardware to support our in-sensor CE compression. Evaluating on action recognition and video reconstruction, our system outperforms state-of-the-art video-based methods at the same speed while reducing the energy by up to 15.4 times.
Semiconductor Metrology and Computational Lithography (Joint Session with Computational Imaging)
15:30 - 17:30
Grand Peninsula C
15:30COIMG-137
It's just flat: Optimization landscape of inverse lithography, [view abstract]
Many people characterize the optimization in inverse lithography as nonlinear, nonconvex, nonsmooth, hard-to-converge, ill-posed, many local minima, etc, which essentially portraits the subject as one of the evilest combinations of the worst adjectives in optimization textbooks. But is the optimization landscape of inverse lithography really this horrible, aka with peaks and valleys, and with spikes everywhere plus a lot of singularities? We took a closer look at the optimization landscape of the problem. We found that there isn't anything complicated about the optimization landscape. It is just flat. We will explain the source of this flatness, and techniques to overcome the flatness.
16:50COIMG-138
Advancing non-destructive evaluation in electronics: 3D X-ray microscopy and deep learning for failure analysis and quality assurance, Herminso Villarraga-Gomez, Zeiss (US) [view abstract]
X-ray-based technologies, initially developed for medical applications, have become essential tools for non-destructive evaluation (NDE) in the electronics industry. Techniques such as radiography, computed tomography (CT), and 3D X-ray microscopy (XRM) enable detailed assembly analysis, structural characterization, and fault isolation in electronic devices, including printed circuit boards (PCBs), system-in-package technologies, and advanced packaging architectures. While traditional 2D radiography and laminography provide valuable insights, they face limitations in resolving overlapping materials and obstructed components. Advanced 3D XRM systems, featuring resolution-at-a-distance (RaaD) capabilities and deep learning (DL)-based reconstruction algorithms, overcome these challenges by delivering high-resolution imaging of sub-micrometer and nanoscale features. DL-based algorithms, such as DeepRecon and DeepScout, significantly enhance image quality, reduce noise, and improve data acquisition efficiency by up to 10x. These advancements enable faster, more accurate failure analysis (FA) and quality assurance workflows, while preserving the integrity of electronic components. Furthermore, correlative imaging workflows, integrating 3D XRM with techniques like scanning electron microscopy (SEM) and transmission electron microscopy (TEM), provide multiscale insights into complex electronic structures. By addressing the challenges of modern electronics manufacturing, 3D X-ray imaging and DL-based innovations support the development of reliable, high-performance devices, meeting the growing demands of the global electronics industry.
17:10COIMG-166
Fast X-ray micro- and nano-laminography imaging of integrated circuits at the upgraded Advanced Photon Source, [view abstract]
THURSDAY 5 MARCH 2026
Imaging Pipelines II (Joint Session with Computational Imaging)
11:00 - 13:00
Grand Peninsula C
11:00COIMG-160
3D field of junctions: A noise-robust, training-free structural prior for volumetric inverse problems, [view abstract]
Volume denoising is a foundational problem in computational imaging, as many 3D imaging inverse problems face high levels of measurement noise. We propose a fully volumetric version of Field of Junctions, an explicit structural prior that has shown strong denoising properties for 2D images. Like its 2D counterpart, our 3D Field of Junctions (3D FoJ) extracts and preserves sharp boundary structure by optimizing a junction of 3D wedges that best explain each 3D patch of a full volume, while encouraging consistency between overlapping patches. As a denoising prior, 3D FoJ (i) requires no training data, and thus precludes the risk of hallucination, (ii) preserves and enhances sharp edge and corner structures in 3D, even under low signal to noise ratio (SNR), and (iii) can be used as a drop-in denoising representation via projected gradient descent for any volumetric inverse problem with low SNR. We demonstrate successful volume denoising with 3D FoJ across three diverse 3D imaging tasks: low-dose X-ray computed tomography (CT), denoising volumes from cryogenic electron tomography (cryo-ET), and denoising point clouds such as those from lidar with miscalibration or adverse weather. Across these domains, 3D FoJ outperforms a mixture of classical and neural denoisers.
11:20COIMG-161
Near real-time stitching for a heterogenous camera array, Adel Al-Ghazwi, University of Arizona, Wyant College of Optical Sciences (US); Gordon Hageman, University of Arizona, Wyant College of Optical Sciences (US); Kimberly Doty, University of Arizona, Wyant College of Optical Sciences (US); David Brady, University of Arizona, Wyant College of Optical Sciences (US) [view abstract]
We demonstrate a streamlined pipeline for near real-time image stitching from a heterogeneous camera array. Methods using static homography struggle with dynamic scenes and parallax-induced artifacts. Our approach overcomes these limitations by integrating modern lightweight neural networks, ALIKE and LightGlue, for fast robust feature matching. These models are converted into an optimized TensorRT engine, allowing for periodic homography re-estimation during the capture stream. We also address parallax errors by replacing a single global transformation with local mesh-based homographies, which provide more precise, non-rigid alignment. This pipeline delivers a dynamic and accurate stitching pipeline with low latency that is adaptable to different camera array systems.
11:40COIMG-162
Trace2SCAD: Converting images Into parametric OpenSCAD models, Henry Dietz, University of Kentucky (US) [view abstract]
Ironically, it is often useful to manipulate 2D graphics for 3D printing. Trace2SCAD was originally written to convert 2D diagrams into printed structures that could be understood by a blind student feeling the surface. However, by its initial open source release in January 2015, it had already been applied to convert pixel-mapped images into parametric 3D models and 3D-printable image renderings (e.g., lithophanes). The original version was a shell script using a variety of helpers, especially potrace. The new version is pure C++ code using the OpenCV library and incorporating many new features, including the ability to render color images as 3D models to be printed for viewing by reflected light (similar to the filament paintings produced by HueForge). It is also significant that the output generated is not mere 3D shapes, but parametric 3D models coded in the OpenSCAD programming language. This paper describes the functionality and algorithms used in the new Trace2SCAD.
12:00COIMG-163
NUTIK: A testbed for functional post-capture manipulation of time and gain, Paul Eberhart, University of Kentucky (US); Henry Dietz, University of Kentucky (US) [view abstract]
Conventional photographic imaging technologies fix the time, interval, and gain of an exposure at the instant of capture, spoiling images if these parameters are not correctly anticipated. Techniques, such as TDCI (Time-Domain Continuous Imaging) allow the integration of an image to be decoupled from the capture of scene data, creating an opportunity to not only adjust the timing and gain of an image after the fact, but manipulate those parameters in previously impossible ways.To facilitate exploration of these new dimensional freedoms, we have created NUTIK, a tool which allows scene data to be captured and computationally post-processed to expose images with user control over the time interval being sampled and the gain of integration, not just for each image rendered but for every site in each rendered image. This paper documents the design and operation of NUTIK, and makes an initial exploration of useful and interesting new photographic techniques enabled by such a tool.
12:20COIMG-164
Arbitrary spatial sampling with the B-FFT for efficient exoplanet imaging simulations, Jamila Taaki, University of Michigan (US); Farzad Kamalabadi, Department of Electrical and Computer Engineering, University of Illinois at Urbana-Champaign (US); Athol Kemball, Department of Astronomy, University of Illinois at Urbana-Champaign (US); Lia Corrales, Department of Astronomy, University of Michigan (US); Alfred Hero III, Department of Electrical Engineering and Computer Science, University of Michigan (US) [view abstract]
Direct imaging of Earth-like exoplanets in the habitable zone of a Sun-like star requires extreme starlight suppression, typically on the order of 10-8 to 10-10. Starshades are petal-shaped external occulters, some 10 s of meters across, flown in formation with space telescopes to suppress on-axis starlight while preserving off-axis planetary signals. Accurate computational simulations of starshade Fresnel diffraction are essential for mission design studies and quantifying performance metrics for future mission concepts such as NASA's Habitable Worlds Observatory (HWO). However, these simulations demand moving between fine spatial resolutions over multiple planes of propagation and across broad wavelength bands. Traditional fast Fourier transforms (FFTs) require large zero-padding factors to accomplish this, increasing computational cost. In this work, we present a general optical model for starshades and demonstrate the adaptation of the Bluestein Fast Fourier Transform (B-FFT) to enable efficient, high-fidelity imaging simulations of starshades at arbitrary spatial scales without zero-padding, thereby achieving greater efficiency than the FFTs or discrete Fourier transforms (DFTs).
12:40COIMG-165
Parallelizable iterative CT reconstruction, Sri Ragha Sai Sowmya Seeram, Elmore Family School of Electrical and Computer Engineering, Purdue University, West Lafayette, Indiana (US); Charles Bouman, Elmore Family School of Electrical and Computer Engineering, Purdue University, West Lafayette, Indiana (US); Gregery Buzzard, Department of Mathematics, Purdue University, West Lafayette, Indiana (US) [view abstract]