In machine vision, a scene may contain deep shadows and intense highlights at the same time. A conventional pixel with a linear response can capture only a limited range before dark details become difficult to distinguish or bright pixels saturate. HDR machine vision addresses this challenge by combining multiple exposure strategies or changing how pixels respond to light. The right approach depends on the scene, especially its motion and illumination, as well as the camera’s processing requirements.
High dynamic range (HDR) is not a single sensor feature. It is a group of techniques with different compromises in capture timing, spatial resolution, image quality and implementation. Understanding those differences helps engineers and integrators assess whether a camera can deliver useful detail under the conditions of a particular inspection or monitoring task.
Why a linear pixel can struggle
A linear-response pixel accumulates charge in relation to the light it receives during an exposure. Increasing exposure can reveal detail in darker regions, but bright regions may reach saturation. Reducing exposure protects highlights while leaving less signal in shadowed areas. When a scene spans a wide range of brightness, one exposure may not preserve useful information across the full image.
HDR methods seek to capture or represent that broader range. Some take multiple images at different exposures and combine them; others alter the pixel response or readout so that bright and dark information can be represented in a single capture. These approaches are not interchangeable: a method that performs well on a stationary scene may produce artifacts when objects, the camera or the lighting move.
Traditional HDR methods and their trade-offs
Exposure sequencing
Sequencing captures a series of images at different exposure times or gains, then combines them in an image signal processor (ISP) or software. The approach places relatively low demands on the image sensor and can use large differences between exposure times. However, the effective frame rate is reduced because multiple images are needed for an HDR result. Motion between exposures—whether from a moving object, camera movement or changing illumination—can create artifacts.
Interleaved exposure
Interleave HDR assigns different exposures to alternating rows or columns, producing multiple exposure views within a frame. This can provide single-capture HDR, but each interleaved view has reduced spatial resolution. The method can also introduce spatial blur and motion artifacts. In the described implementations, the short exposure begins after a longer pixel-reset phase, limiting its ability to capture fast motion cleanly.
Knee-point and logarithmic response
Knee-point pixels use a piecewise-linear response. Bright pixels are reset to a defined level, after which a flatter response phase adds further exposure capacity. This can capture an image without post-processing, but color reproduction is a stated limitation.
Logarithmic modes combine linear and logarithmic response regions; the knee point and slope may be adjustable. They damp the response to bright areas and can operate in one capture without post-processing. The trade-offs identified include color reproduction and fixed-pattern noise (FPN). Logarithmic pixel designs and solar-cell-style pixels offer other ways to extend response, but each has its own constraints. The cited designs include issues such as rolling readout, gain shading, image lag, temperature dependence, black-level drift and analog output requirements.
New pixel architectures: split photodiodes and dual gain
Newer HDR sensor designs combine pixel structures with multiple exposure or conversion-gain readouts. A split-pixel design uses a large and a small photodiode to handle different light levels. Dual conversion gain provides a high-gain (HCG) and low-gain (LCG) readout, supporting sensitivity to lower signals alongside greater capacity for brighter signals. Depending on the sensor, the exposures may be captured simultaneously or through separate readouts.
The whitepaper’s examples illustrate several implementation patterns rather than one universal design:
- STMicroelectronics VG6768: a dual-photodiode sensor listed with one exposure and three images, dual conversion gain, on-board HDR merging and anti-flicker support.
- Sony IMX390: a dual-photodiode design with dual gain for the small pixel, three images from one exposure, and on-board HDR processing. The document also describes the IMX490 as supporting dual gain on both photodiodes.
- OmniVision OX02A10: uses two simultaneous exposures with a dual-photodiode structure and dual gain on the large photodiode; its listed output options include combined linear data and RAW or YUV.
- ON Semiconductor AR0233: uses a single photodiode with dual exposure, dual conversion gain and separate readout of the photodiode and an electron “catch basin.”
These examples show why a headline dynamic-range figure alone is not enough to select a sensor. Engineers should also examine how many exposures or images are produced, whether exposures are simultaneous, how the HDR output is merged, available data formats, frame rate, anti-flicker provisions and the processing required downstream.
Motion, lighting and system integration
LED flicker is a particular concern in automotive imaging discussed in the source. A sequencer using multiple exposure times can produce flicker-related artifacts, while chopping the readout of shorter exposures may mitigate flicker but worsen motion artifacts. This highlights a broader integration challenge: exposure timing interacts with both scene motion and illumination behavior.
Newer pixel architectures can improve the capture options, but they do not remove the need to design the whole imaging chain. The source identifies an external ISP for HDR computation and sensor control as necessary for the modern sensors discussed. Camera selection should therefore account for sensor control, image merging and the format delivered to the vision system, not just the pixel architecture.
Choosing an HDR approach for machine vision
Start with the scene and the required output. For stable scenes, sequential exposures may be practical if reduced effective frame rate and motion artifacts are acceptable. Where movement matters, single-capture or simultaneous-exposure approaches may be more suitable, but interleaving can reduce resolution and newer pixel designs still require careful evaluation. Color fidelity, noise, readout behavior, lighting flicker and ISP requirements should all be considered alongside dynamic range.
HDR machine vision is best treated as a system-level design decision. Traditional techniques and newer sensor architectures each address the limits of linear pixels, but no approach is free of trade-offs. Matching exposure strategy and pixel design to scene dynamics—and planning for the required processing and optics—gives automation teams a more reliable basis for selecting an imaging solution.