For more than a decade, progress in imaging has been defined by a familiar set of metrics: smaller pixels, higher resolution, improved dynamic range. These advances have driven remarkable improvements in image quality, particularly in consumer devices, where the race for better pixels has shaped both roadmaps and marketing narratives.
But that model is beginning to change.
Across the imaging industry, a new pattern is emerging. Performance is no longer defined solely by the sensor itself, but by the intelligence surrounding it. From AI-driven image pipelines to the increasing demands of automotive and industrial applications, imaging is evolving from a component-level discipline into a system-level one.
This shift was a recurring theme in discussions at Image Sensors Europe earlier this year.
As Marie-Charlotte Leclerc of STMicroelectronics put it succinctly:
“The next big step is not just better sensors, it’s better systems.”
This shift is subtle, but significant. It does not replace the importance of sensor innovation, but it reframes it. Metrics such as quantum efficiency, pixel size, and dynamic range still matter. Increasingly, however, they are becoming baseline expectations rather than points of differentiation. What now matters more is how the sensor interacts with the wider imaging pipeline: the image signal processor, the system-on-chip, and, increasingly, AI models.
One of the clearest signs of this transition is the growing overlap between image capture and image processing. Technologies such as hybrid shutter architectures, in-sensor HDR, and sensor–AI co-design are dissolving the traditional boundary between hardware and software. The imaging pipeline is no longer a linear sequence of steps. It is becoming a tightly integrated system where capture and interpretation are interdependent.
Neural ISP approaches illustrate this shift particularly well. Rather than relying on fixed-function processing blocks, these systems use machine learning models to reconstruct images directly from raw sensor data. As discussed by Tom Bishop of Glass AI, this approach has the potential to replace large portions of the conventional pipeline.
If that transition continues, the implications are far-reaching. Imaging performance becomes less about hardware specifications and more about algorithms. The centre of gravity shifts from semiconductor differentiation to software and model development. This is not just a technical shift, it has consequences for how value is created across the ecosystem.
At the same time, the long-running trend toward ever-smaller pixels is beginning to encounter fundamental constraints. Over the past decade, pixel pitch has shrunk dramatically, from around two microns to as little as half a micron in leading smartphone sensors. This scaling has delivered higher resolution and enabled new computational imaging techniques, but it is no longer a straightforward path forward.
As Joren Hoet of eyeo observed:
“We’re trying to fit a one-micron photon into a half-micron pixel — it’s not a winning game.”
The physics has always been understood. What is changing now is the willingness to acknowledge it more openly. Pixel scaling is no longer a simple roadmap. It is a compromise between optics, signal-to-noise ratio, and diminishing returns.
As a result, innovation is shifting into adjacent areas. Pixel binning has become standard practice. Optical design is evolving, with larger apertures and folded optics compensating for reduced light capture. And increasingly, software is stepping in where physics imposes limits, using AI to reconstruct detail that the sensor alone cannot resolve.
While the pressure of pixel scaling is most visible in mobile imaging, the strongest growth drivers are emerging elsewhere. Automotive, in particular, is reshaping expectations across the industry. Modern vehicles are rapidly increasing their use of cameras, moving from a handful of sensors to potentially a dozen or more, supporting everything from parking assistance to driver monitoring and in-cabin sensing.
As Alastair Attard of UTAC Group noted:
“As automation increases, camera counts go up dramatically.”
But the significance of automotive lies not only in volume, but in requirements. Unlike consumer applications, where failure is inconvenient, in automotive failure is unacceptable. This drives stricter standards for reliability, redundancy, and validation. In effect, automotive is forcing the imaging industry to mature, quietly reshaping how sensors and systems are developed.
A similar shift is underway in industrial vision, although less visibly. Applications such as food sorting, logistics, battery inspection, and robotics are increasingly reliant on machine vision, yet deploying these systems remains slow and resource-intensive. Installation, tuning, and validation in real-world environments can take months, sometimes years, creating a bottleneck that is now attracting increasing attention.
This is where simulation and digital twin technologies are beginning to play a role. Platforms such as NVIDIA Omniverse allow engineers to design and validate vision systems virtually before deployment. If these approaches scale effectively, they could significantly reduce development cycles, an outcome that, for many users, may be more valuable than incremental improvements in sensor performance.
At the same time, the definition of “vision” itself is expanding. Traditional RGB imaging is no longer sufficient for many applications. Multispectral sensing, depth imaging, and event-based vision are gaining traction across robotics, automotive, and security.
The driver behind this shift is artificial intelligence. As Harish Venkataraman of Meta observed:
“Vision is becoming the eyes of the system.”
And yet, the more profound change is not simply that vision feeds AI, but that vision systems are being redesigned for AI consumption. The goal is no longer to produce images optimised for human viewing, but data optimised for machine interpretation. This distinction has significant implications for sensor architecture, processing pipelines, and performance metrics.
Beyond core imaging technologies, there is growing recognition of the importance of packaging, manufacturing, and supply chain considerations. Packaging is evolving rapidly, enabling smaller and more cost-effective systems, but also introducing new challenges, particularly around reliability in demanding environments such as automotive.
At the same time, global supply chains remain a point of vulnerability.
“You cannot build an image sensor if one component is missing.”
The response is a gradual shift toward resilience. Dual sourcing, regionalisation, and diversification are becoming strategic priorities rather than short-term adjustments.
At the other end of the spectrum, scientific imaging continues to push the boundaries of what sensors can achieve. While these applications represent a smaller portion of the market, they often act as proving grounds for technologies that later reach mainstream use.
Stepping back, the overall direction is clear. Imaging is no longer defined solely by the performance of individual components. Sensors are becoming part of broader, intelligent systems in which hardware, software, and data are tightly integrated.
As Harish Venkataraman put it:
“It’s not about if sensors become more powerful — it’s about how they work with AI to understand what they see.”
For those working in machine vision, that shift has practical consequences. It changes how systems are designed, how performance is measured, and ultimately, how value is created.
Better pixels still matter. But they are no longer enough. The next phase of imaging will be defined by smarter systems, and by the intelligence that connects them.
Editor’s Takeaways
- The battleground has shifted: differentiation is moving from sensor performance to system integration and software
- AI is redefining imaging: pipelines are being rebuilt around machine interpretation, not human viewing
- Pixel scaling is reaching limits: physics is forcing a pivot toward computational and optical innovation
- Automotive is raising the bar: reliability and validation are reshaping expectations across industries
- Deployment is the real challenge: simulation could transform time-to-market
- Resilience is now strategic: supply chain decisions are becoming central to system design
