In machine vision and other imaging applications, a scene may contain useful information beyond what a conventional visible-light camera records. A multi-storied photodiode CMOS image sensor is designed to capture visible red, green and blue (RGB) information alongside near-infrared (NIR) signals in a single sensor. Its two aligned photodiode arrays sit in separate stacked substrates, allowing each layer to collect a different portion of the incoming light.
This approach offers an alternative to combining separate cameras or reserving individual pixels for infrared. The sensor architecture also introduces optical design considerations: light reaching the lower array passes through the upper substrate and intervening layers, so the structure between the arrays affects the signal that can be detected.
How a multi-storied photodiode CMOS image sensor works
The sensor uses two photodiode (PD) arrays, one above the other. The top array primarily detects visible light. The bottom array receives light that passes through the top semiconductor layer and its color filters, including longer-wavelength light such as NIR.
Because the arrays are aligned, corresponding pixels in the upper and lower layers observe the same position in the scene. The top layer does not simply block all light from the lower one: it absorbs some wavelengths while allowing a portion of the incident light to pass through. The lower array then converts the transmitted light into electrical signals.
In the demonstrated design, the upper array has red, green and blue color-filter pixels. The lower array is aligned beneath those pixels, so its responses differ according to the filter above each pixel as well as the optical properties of the layers. The result is several spectral signals from the two arrays, rather than only a conventional RGB image.
Why stack the arrays instead of combining cameras?
Capturing RGB and NIR with two separate cameras can require a beam-splitting optical component and careful alignment to bring both images into the same frame. A single sensor with RGB and NIR-designated pixels avoids that two-camera arrangement, but pixels assigned to NIR are unavailable for direct RGB sampling. Reconstructing a full RGB image then requires interpolation, and the shared pixel design must accommodate competing optical and electrical requirements.
A stacked design takes a different route: it places the visible and transmitted-light detection arrays in separate substrates while aligning their pixels. This avoids dedicating positions in the same array solely to NIR. The demonstrated sensor captured RGB and NIR images simultaneously without additional optical components, using signals from the two layers to produce each image.
The approach can also support different uses of the signals. In the reported demonstration, an IR-reduction algorithm combined information from the top and bottom arrays to reduce the effect of NIR on the reconstructed RGB image. The bottom-array signals could also be used to produce an NIR image. These results illustrate how a sensor can provide more than one spectral view from the same imaging device.
Optical filtering and signal processing trade-offs
The lower array’s response is shaped by more than the color filter above it. Light must pass through the top semiconductor, wiring layers and bonding layer before reaching the lower photodiodes. The material, thickness and layout of these layers influence how much light is transmitted and the angles at which it reaches the bottom pixels. In this architecture, the wiring and bonding structure can therefore act as part of the optical filter.
Measurements showed different spectral responses for lower pixels beneath the three color-filter types. Those responses followed the expected tendency based on the optical properties of the filters and substrate materials. This indicates that the intervening structure can be considered as part of the sensor’s spectral design, rather than as electrically functional packaging alone.
There are also sensitivity costs. The reported device’s wiring geometry limited the angle of light reaching the lower photodiodes, and measurements showed reduced sensitivity at larger incident angles. The authors identified wiring layout and layer structure as ways to adjust this angle dependence. However, the reported sensitivity figures and possible improvements apply to that specific design; they should not be treated as general performance guarantees for stacked sensors.
Signal processing brings another consideration. The demonstrated RGB reconstruction subtracts a scaled contribution from the lower-array signal to reduce NIR contamination. Since the lower signal has its own noise, using it in the subtraction also affects the noise of the resulting RGB signal. The coefficient depends on the arrays’ quantum efficiencies and exposure times, so the processing strategy involves balancing NIR reduction against the resulting signal-to-noise behavior.
What the architecture may mean for machine vision
For industrial imaging designers, the central idea is that spectral separation can be built into a stacked sensor structure. A single aligned sensor can provide visible and NIR information without the separate-camera alignment challenge described for conventional arrangements or dedicated NIR pixels within the same array. This can be relevant when an application needs both types of image data from the same scene and imaging position.
Evaluation should consider the complete optical and electrical system, not just the presence of two photodiode layers. Relevant design questions include:
- Which wavelengths should each layer detect, and how do the filters and semiconductor layers shape those responses?
- How do wiring and bonding materials, thickness and layout affect lower-layer transmission and angle sensitivity?
- How will exposure settings and signal processing combine the two arrays’ outputs?
- What signal-to-noise trade-offs arise when lower-array data is used to correct the visible image?
The architecture is also presented as a potential platform for other functions, including distance measurement or phase-difference detection for lens focusing, using appropriate pixels and circuits. These are possible extensions of the concept, distinct from the RGB and NIR imaging demonstrated in the reported device.
Conclusion
A multi-storied photodiode CMOS image sensor captures visible and near-infrared information with aligned photodiode arrays in separate stacked substrates. Its value lies in combining multiple spectral views within one sensor, while its practical performance depends on optical transmission, layer geometry, incident angle and signal processing. For automation and machine-vision teams assessing multispectral imaging, those factors are essential to matching the sensor architecture to the required image bands and system constraints.