LiSens— A Scalable Architecture for Video Compressive Sensing
Proc. ICCP 2015
The measurement rate of cameras that take spatially multiplexed measurements by using spatial light modulators (SLM) is often limited by the switching speed of the SLMs.
The measurement rate of cameras that take spatially multiplexed measurements by using spatial light modulators (SLM) is often limited by the switching speed of the SLMs. This is especially true for single-pixel cameras where the photodetector operates at a rate that is many orders-of-magnitude greater than the SLM. We study the factors that determine the measurement rate for such spatial multiplexing cameras (SMC) and show that increasing the number of pixels in the device improves the measurement rate, but there is an optimum number of pixels (typically, few thousands) beyond which the measurement rate does not increase. This motivates the design of LiSens, a novel imaging architecture, that replaces the photodetector in the single-pixel camera with a 1D linear array or a line-sensor. We illustrate the optical architecture underlying LiSens, build a prototype, and demonstrate results of a range of indoor and outdoor scenes. LiSens delivers on the promise of SMCs: imaging at a megapixel resolution, at video rate, using an inexpensive low-resolution sensor.
Proc. ICCP 2015
A classical work in compressive sensing imaging is the single pixel camera (SPC). The right figure shows one of the best-known results on real hardware to date. This result is only 128*128 images at video rate. We should stop and think for a second that why cannot we get megapixel-images at video rate using the SPC.
The answer for this is two folds.
(1) Natural scenes are increasingly compressible when resolved at higher resolutions. As a consequence, while high-resolution images are highly compressible, the same is not true for lower-resolution ones.
(2) SPC has very poor measurement rate and hence, is not conducive to sensing at high spatial and temporal resolutions. Here the measurement rate is defined as the number of measurements that a device can obtain in unit time.
We collect a bunch of 90 megapixel images of natural scenes. We do wavelet transform to images first, then keep the largest K wavelet coefficients unchanged and set the others to zero, and use the modified coefficients to reconstruct the image. Suppose the image has N pixels, so we have N wavelet coefficients. We define K/N as the non-zero ratio.
When setting nonzero ratio to 1%, we found that for higher-resolutions, the reconstruction SNR is much larger. When setting the nonzero ratio to 10%, we observe the same thing.
Next, we require the reconstruction SNR to be greater than 25dB, and find the “range of nonzero ratio” for different resolutions. We found that for higher-resolutions, the required nonzero ratio is lower.
We found that “Natural images are increasingly compressible when resolved at higher (spatial) resolutions” which implies that we can expect compressive sensing techniques to perform better when we sense signals at higher resolutions. The challenge of compressive sensing at high-resolutions is that resolving scenes at high spatial-temporal resolutions, even with a compressive camera, requires very high measurement rate.
A natural extension to the SPC design is to replace the single pixel by multiple pixels, so that we can achieve measurement rates that are orders of magnitude larger, and therefore realize imaging systems capable of sensing at high spatial and temporal resolutions.
There is a minimum # of pixels at which we can achieve the max possible measurement rate. When pixels are costly, as in the case of SWIR imaging, we want to operate at this particular point because it gives us a sensor that’s of very low resolution and inexpensive but yet gives us a measurement rate that's as high as what's possible given the architecture constraints.
The optimal # of pixels is ADC rate over DMD rate, which is typically 1000s of pixels.
The implication to this is that we can obtain measurement rate of a full-frame sensor but with a fraction of the number of pixels! (less than 0.1% pixels)
We build Line-sensor-based compressive camera, in short LiSens, to demonstrates this.
We make two modifications to the SPC architecture.
First, we use a linear array of pixels or a line-sensor to replace the photodetector.
Second, we add a cylindrical lens to tightly focus the DMD onto the line-sensor.
We can use a line sensor or a 2D sensor to replace the photo-detector. We choose to use line sensor for two reasons.
First, alignment is significantly easier. Because we use the same pattern for each row, slight crosstalk doesn’t matter.
Second, because a 1D sensor only uses a single line, there are a lot of space in top and bottom sides. Without losing the 100% fill factor of each pixel, we can have additional technology like frame transfer, or per-pixel ADC which can boost the measurement rate from MHZ to GHZ. All of these are only possible for 1D sensor because we have lots of space for circuitry in two directions which is not true for a 2D sensor.