Adaptive resolution (AR) provides better imaging performance, greater field of view and can greatly reduce costs by eliminating the need for optics redesign due to the fact that you can reuse optics. Binning allows for improved data rate at the cost of resolution. Subsampling gives you improved data rate as well, so what is the difference between the three. In a recent post we discussed sampling approaches, focusing on adaptive resolution. In this installment we will focus on subsampling and how it is different from binning and adaptive resolution.
In binning you take a group of pixels and typically add them together as shown in the figure below.
Binning deals with whole pixels. This effectively reduces the amount of information that represents the image and reduces the image size, it reduces the number of pixels in the image. The white pixels shown in the figure above are not thrown away as in subsampling which we will get to shortly, but the pattern is simply repeated across the whole image sensor. Thus, in 2×2 binning each group of 4 pixels is summed and the reduced image transferred to the interface. This can be done for monochrome or Bayer color sensors. A Bayer sensor is shown (top right) below in the subsampling illustration.
In addition, binning the data may also reduce the impact of noise, but at the cost of resolution. Depending on how binning is implemented (FPGA or sensor) and what limits your data rate, the sensor or the interface, binning might be used to increase the frame rate. In CCD’s binning is done in the charge domain, so the effective number of electrons in the combined well is higher and the readout is defined by the readout section. This is known as “on chip” binning, so in the charge domain with a positive effect on dynamic range and noise. There is also “Off-chip” binning, meaning the addition is done after the charge to voltage conversion and even digital binning after the AD conversion. In this case there is just reduction of resolution, no noise gain effect or increase in full well.
Adaptive resolution (AR) is like binning, but you work with fractional or virtual pixels. It allows for image re-scaling to a desired format while utilizing a greater field of view. AR allows replacing an obsolete CCD sensor with CMOS and save money on system redesign because the optics and software can be reused while benefiting from the advantages of CMOS. AR allows you to create virtual pixels with better performance compared to the original pixels and allows the image to be down scaled to lower interface bandwidth. This was all discussed in the previous post.
Subsampling however takes one pixel or in the case of a Bayer sensor four pixels and uses it/them instead of a number of other pixels. 2×2 subsampling is throwing away 3 out of 4 pixels. Look at the top left image and consider an area of 8×8 pixels, so 64 pixels. In the bottom left image you only keep the 16 pixels, the rest are thrown away, not added, not averaged, not mixed, nothing, simply thrown away. Information is lost that cannot be regained, there is no full well or noise advantage. With subsampling the spatial distance between the pixels will increase so the MTF will be lower and you will see “jagged edges” or “staircase” diagonal lines in the image. In the Bayer phase it will be equal, so color is possible at a full area FOV, but at limited resolution and bandwidth. This is great for an overview picture at the double FOV and the option to switch towards half FOV with full resolution seeing all the details the individual pixel can address. The camera can do this fast but color balance can be slightly off due to the change in FOV (image content is different so several frames are needed to adjust to real color). The only benefit is that the data rate requirement on the interface is reduced. See the illustration below.
Thus, we see that subsampling is indeed different than binning and adaptive resolution and has its place for data rate reduction. Adaptive resolution has so many more benefits like improved image quality, larger field of view and cost savings due to the ability to reuse optics it is the hands down winner when the three are compared. However, all three sampling approaches have their advantages in specific applications.