In this episode of Visions: A Machine Vision and Automation Solutions Podcast, VSD Head of Content Sharon Spielman speaks with Dr. Michael Fanous of Fanous Photonics about the concept of a "blur budget" — deliberately allowing and engineering motion blur as a tool rather than an error. They discuss continuous motion imaging approaches such as GANSCAN and BlurryScope, balancing stage velocity, exposure, optics, and computation, and why training models on real acquisition conditions is critical.

The conversation highlights research insights, potential industrial impacts, and guidance for teams exploring computational imaging strategies.

Related: Designing the Blur Budget

Visions: A Machine Vision and Automation Solutions Podcast, is the podcast for engineers, designers, integrators, and end users who want to keep an informed eye on the imaging and machine vision industry. Every Tuesday we will explore the latest in imaging trends, developments and solutions. Here you will find interesting, useful insights and observations from expert interviews, solo episodes, even the occasional panel discussion, all of which aim to expand your knowledge on imaging and machine vision.

Transcript

Well, hello and welcome to Visions: A Machine Vision and Automation Solutions Podcast. I'm your host, Jim Tatum, senior editor of Vision Systems design and Visions is an Endeavor Business Media production from your friends at Vision Systems Design. Here you'll find the latest on everything from end user machine vision solutions to trends, developments, and perspectives on all things machine vision and imaging. Whether you've been working in the industry for a while or you're just starting to take a closer look at it, this podcast is designed to grow your knowledge and bring greater focus to your understanding of the imaging and machine vision industry. And now on to our show.

Welcome. Today we are excited to feature a pioneering voice in imaging system design doctor Michael John Furniss from Furniss Photonics. He's redefining how we think about motion blur not as a flaw to be eradicated, but as a powerful design tool that can unlock new possibilities and speed, precision and computational imaging. If you haven't heard of designing the blur budget before, don't worry, we are going to unpack what it means and why. Embracing a little imperfection could lead to bolder, more efficient imaging architectures. So whether you're an engineer, research integrator, or just fascinated by how AI is transforming machine vision, this episode promises some fresh perspectives and practical insights. Doctor Fanis Michael, thank you so much for joining us today.

Thanks so much for having me. Thank you. I really appreciate being here. You bet. So would you mind sharing just a little bit about your journey and how you became interested in continuous motion imaging and this unique approach to managing motion blur?

Absolutely. So this all began during my doctoral program. We were flirting a lot with different image to image ideas for biomedical purposes primarily. This is around twenty twenty and it was commonplace to use um, devices like detection, classification segmentation on histological data, but converting one image to another still seemed to offer a lot of exciting opportunities. So that's something that we were pursuing actively. We eventually turned our attention to motion blur by considering the degrees of freedom of the average digital microscope in our lab. If you actually count the number of ways a modern microscope can be fiddled with physically, it comes up to over a dozen. And so each knob or switch can be thought of as a potential opportunity for AI enhancement. This particular concept, the motion blurring, is inherently an operational scheme where the method of acquisition revolves around an AI premise. This is very different from, say, inferring one kind of a stain or image type or modality from another. Here you're seeking to improve the way an image is taken. Other translations may involve the way a specimen is processed or the way the specimen is illuminated. And there are still actually, I believe, a lot of untapped opportunities in that domain. So that's how it began.

Well, thank you very much for that. So for our listeners who might be new to this topic and for me, who is very new to this topic, um, could you explain the conventional approach to motion blur and imaging systems and why? Historically it's been something engineers just try to avoid.

Absolutely. So most microscopes are made of very strong materials that are very steady. They seek to avoid any kind of vibrations because naturally you don't want imperfections in the image. And this is very hard to obtain an actual crisp, clear in-focus image. That's the gold standard. Um, and it's considered what in microscopy, you're trying very hard to maintain the highest level at the image plane before you take any acquisition. So, uh, up until I'd say ten years ago, it would have been preposterous to sort of court and invite the idea of a deliberate defect like motion blur in order to then compensate it using an algorithm. Um, so what most scanners and what most systems will use is what's called stop and stare. The specimen needs to be physically halted before you're taking an acquisition. Anything that does continuous scanning will use what's called a time domain integration or line scanners, in which case it's not interactive and you need that. In that case, you need very precise synchronization between the mechanics of the acquisition and the camera. So that complicates the situation. It's also quite expensive. So that's what standard now with the advent of all of this AI sophistication, um we realized that we can undermine the traditional orthodox methods and potentially gain a lot of value, um, through, you know, specific algorithms and tailored AI models.

Okay. That's pretty insightful. So can, can you give us a sense of some industries or applications where continuous motion, uh, imaging approach guided by this blur budget might be a game changer?

Sure. Absolutely. So it's much more than just biomedical or pathology. Um, in terms of real world applications, it's important to understand that motion blur is actually comprised of more information in some ways than its crisp counterpart, because it's really an integration of a few lateral frames slightly shifted. And so that's what is producing the smearing effect. So this is essentially a throughput for resolution kind of trade off before any, you know, deep learning involvement. So any kind of fast, continuous large image measurement could drastically benefit from this sort of scanning tactic, whether for speed, cost or quality. For instance, wafer inspections, the chips of manufacturing lines may use expensive cameras or strobe lighting techniques to deal with the throughput demands. But again, scan like approach could potentially prove very convenient to alleviate costs or improve speeds here. Uh, microfluidics is another area high speed movement of cells, and scenarios like that may have phototoxicity issues that could be improved with a similar approach to this. Um, even satellite like terrain mapping can possibly benefit from a continuous scanning approach, uh, using models that are trained on slower, um, combing techniques and then building a registration pipeline and making a model that is specifically tailored for that. So there's a number of different areas where this sounds like it's, it's, it's a, it's a sensing technique.

I know that AI and computational imaging are transforming very broadly. Um, machine vision today. So how does your blurb budget philosophy. Mhm. Say that five times fast fit within the wider trends in the field.

Okay. So, uh, I'll, I'll just talk about a general trend that's happening that I see more and more. And there's a wave of innovations that is doing this. So mostly it involves AI very, very fundamentally, most innovations with AI involve manipulating the software or the algorithm to fit whatever physical hardware or optic and photonic configuration that you already have. The reason for this is obvious. It's much easier to manipulate the algorithm and play around with that around long standing and traditional data than it is to reassemble or manufacture an entirely new physical configuration. And that system, that AI for hardware, AI for optics is starting to shift now from to optics for AI or hardware for AI, in which case you start with the AI, you start with with whatever algorithm that you like. And with AI, there's a great deal more flexibility, but it's also at the same time, more daunting in that sense, where sky is the limit because now you're just dealing with pure abstract ideas. The AI is really take one set of data and transform it into another in any way that you like. And so that really causes you to reevaluate the whole system and the whole optical configuration, and to design the physical aspect around the AI, which is very counterintuitive and feels very foreign and alien and uncomfortable. But it's what will give rise, I believe, to a lot of really exciting new technologies.

Yeah, that's really good context. So why don't we dig into the technical part of your work and your recent research? Um, so designing a system that balances stage velocity, exposure time, optics and computation, it sounds pretty complex. So how do you approach engineering that precise equilibrium in practice?

Okay. That's a great question. Uh, for equilibrium, balancing all of the different variables whenever you're playing and making compromises with certain factors, you need to be very careful. And we approach this initially extremely conservatively. I literally used factors and variables that I knew would fall comfortably within what a basic model this is, you know, four or five years ago. So less robust could reliably reconstruct. This is before the mayhem with the LMS, and we were painstakingly hand-picking hyperparameters. And, you know, we had to wait nervously for weeks just to see if the model was even remotely successful. And so there was very little room to be bold and wild and aggressive with what we were trying to compromise. There's always this push and pull when you're dealing with this kind of a tactic. And so initially we want to be very conservative. Now that we have the whole system set up, and this is what I would recommend to others who are interested in pursuing such a strategy, is to then let a kind of automated self-training improvement system run free and push the boundaries of what you think is possible. In our particular case, we are always interested in challenging the bottleneck of speed and doing whatever we can, and trying to play around with the different variables and to keep that equilibrium, but accelerate the system as much as possible. That's a really dazzling prospect for us to push forward.

Okay. You know, you wrote a paper on this very topic and we just published it very recently. So in the paper, you highlight how critical it is for AI models to be trained on data matching the real acquisition conditions, not just the synthetic blur. So what are some of the biggest challenges in building and validating those real world training data sets?

Okay. Yeah. So the training data sets are really important because we played around with synthetic data sets. And synthetic data sets are very useful because you can manipulate them every which way, and you can introduce whatever factors you like to enhance the training as much as possible. The problem is, there are so many little micro factors in an actual system that are not being captured for our specific situation. You have stage vibrations, the acceleration of the stage at the edges. So that's changing the length of the blur. You also have rolling shutter effects, illumination changes, straying away from the focal plane. This is not something you can always control. So different tissue thicknesses as well. Even within pathology and even with the same processing situation you're going to have variations. Stain types also differ so that the importance of gathering enough of those real instances in the real data is something that we're very conscious about. And we want to have representative in a kind of comprehensive way for our data sets. So the training models can gather all of those micro features, which you won't get with synthetic data. But synthetic data is very important to supplement that process, if that makes sense.

Okay. Yeah, I appreciate that. All right. So you've demonstrated dramatic throughput gains with projects like Gan Scan and Blurry Scope. So how do those gains translate into practical benefits for research labs or industrial settings?

Yeah. So what we're proposing and what we're doing with these throughput strategies is going to feel to users very unfamiliar at first. And this is why we're always coupling what we have with what we think approximates the most conventional and traditional way. So whenever we have a disruptive alien feature, we have its counterpart familiar and comfortable one. And so for the most obvious is continuous scanning versus stop and stare scanning. We have two big buttons, continuous and stop and stare. So the idea is to make the segue as smooth and seamless as possible for users, because some of these features are quite disruptive. They will feel uncomfortable and unusual. And so we're not naive in the sense we don't think it's going to be very easy for people to implement this right away. So we think the adoption speeds will differ. Some people right away, they'll pick it up and some people may be more reluctant. Maybe they'll use continuous in the high throughput scanning style one out of a dozen times. But the important thing is introduction and just seeing, you know, putting it into practice, even if it's just for now, a very modest amount. That's our goal and that's what we hope to achieve.

Yeah, it's probably important that we say it's right now it's a research only.

Yes it is a research only. Yeah. Assistive review and supplemental.

Absolutely. So you know, a fascinating feature that you had described is the quality control gate that can, um, it can reject data outside validated conditions rather than producing misleading outputs. So can this concept redefine trust and reliability in AI driven machine vision? Do you think?

Absolutely. So our end final absolute metric, because this is ultimately a med tech device, will have to rely on expert pathology professionals and domain pathologists. So domain. So what I mean is a trained pathologist. So it can go through every quantitative or qualitative test and metric. The images the outputs, the reconstructions that you like, if it fails or if it doesn't meet the standards of the user. In our case, a pathologist, if it seems odd or slightly bizarre, or is, is just, uh, there's something that is errant about it, then ultimately it's a failure. So that is the final assessment for us. And so we're working very closely with a lot of consultants in the field because I am not the biological expert. So I can only do the optics, the mechanics, the AI. Ultimately, the user has to flag the final situation and the final reconstruction. So that is the ultimate test because you can have every beautiful technical metric and be as rigorous as you like. But if it's not suitable or if it's not, uh, convincing or compelling to the end user, in our case pathologist, then it, it did not succeed.

So I know we're talking about mainly, uh, the, the end user being pathologists. So looking forward, how might embracing a deliberate blur budget influence machine vision applications beyond microscopy? Um, maybe in drone imaging or automated inspection or other fields?

Absolutely. So that's a great question. And so yeah, again, it's a little funny to see where we've come that I'm talking about having not only encouraging the defect of motion blur, but actually talking about a budget for how much you can spend, and maybe even going above that and playing with it. Um, so we wrote a very extensive review at UCLA in Professor Oscar's lab on compromising certain just imaging metrics in general terms. So, uh, point spread function, the signal to noise ratio, the sampling density, doing that for enhancements of field of view, depth of field space, bandwidth product. All of this can be applied to any kind of imaging situation or scenario, drones included. And so we covered over a dozen papers and projects using this. It's a very powerful underlying principle. And that is you start with the final objective that you want to improve, whether it's duration, form factor, maybe you want to change the dimensions, make it smaller, make it larger, make it oversee a larger field of view, or just reduce the cost. So there may be certain components that are pain points. And then you consider all of the variables around that system, and then you weigh them against what can be compromised against something else. Now you're dealing with calculated intentions for training on superior data to recover whatever that you have decided to compromise or remove. And so this is the gist of the trick, which may appear initially quite crude, but if it's executed just right, I think that the effects in various areas and domains can be really far reaching. So that's what that's what I would say for that's what I would recommend.

All right. So before we wrap up, is there anything else you want to say to our audience?

Uh, I just like to say I'm really grateful for being able to talk with you and your platform. I'm really grateful for everything that you do for this community. I think that this is a really exciting time for vision systems and for design in general. So that's what I'd like to say. Just express my gratitude. Uh, we, we really appreciate that.

And we appreciate the great minds like yours. Michael, it was a pleasure to get to speak with you today and thank all of you in the VSD audience for tuning in. Until next time, this is Sharon Spielman with Vision Systems Design. Stay curious.

Well, that's a wrap for this episode of visions, produced by Endeavor Business Media, a division of endeavor B2B. Thanks very much for tuning in. If you enjoyed today's show, be sure to subscribe to the podcast and share this episode with a colleague who would find it helpful. Until our next episode, you can find us at vision dash systems dot com or on LinkedIn, Facebook or X. For more insights, updates, and breaking news to keep you in the know. Thanks for tuning in. Until next time, stay focused on your visions.