Sunday, April 17, 2016

Compressive Sensing Camera with Pixel-wise Coded Exposure

OSA Optics Express publishes Johns Hopkins University paper "Compact all-CMOS spatiotemporal compressive sensing video camera with pixel-wise coded exposure" by Jie Zhang, Tao Xiong, Trac Tran, Sang Chin, and Ralph Etienne-Cummings. The paper presents a prototype 127 × 90 pixels image sensor that can reconstruct 100 fps videos from coded images sampled at 5 fps. With 20× reduction in readout speed, the novel CMOS sensor only consumes 14μW to provide 100 fps videos.

"Figure 1 shows the difference between a conventional camera with global exposure and a PCE camera. In a conventional global exposure camera, all the pixels are exposed for a fixed amount of time (T v) to readout one image at readout frame rate of 1/T v. This is compared to a PCE camera, in which pixels are exposed through a random short “single-on” exposure of fixed duration (Te) within T v. The readout circuit only samples the pixel value at the end of T v with readout speed of 1/T v. PCE essentially compresses a spatiotemporal video into a single coded image. Upon on receiving the coded image, PCE reconstructs the entire video from the single coded image using sparse spatiotemporal reconstruction with an over-complete dictionary.

Since the reconstructed framerate is 1/(unit time o f Te), PCE provides a high frame rate using the same readout speed as a conventional image sensor. PCE is also different from traditional spatial CS approach, which recovers one frame using multiple random spatial samples. Thus, PCE is more optimal for video applications because the sparse samples include both spatial and temporal information.


  1. This is very interesting, but it seems to me they are actually built up a high-speed camera and selected some wanted sub-frames to readout. For SLM based systems, readout only happened at the end of each frame. (e.g. for a frame period of 100ms, SLM system using LCoS chops it into many sub-frame periods for optical encoding, and readout when all sub-frames are finished at the end of 100ms. But in this paper, they actually readout for every needed sub-frames)

    1. Jie (Jack) ZhangMay 4, 2016 at 6:58 AM

      Hi Roy, in this design, the readout still uses conventional row scan method to update pixel memory. But the ADC only samples a pixel at the end of its exposure. When sampling is not needed, the ADC can be turned off to save power. So at the end, the amount of sampling for this sensor should be the same as SLM based system.

      You bring up a good point, I think by re-designing the SRAM addressing method, we might be able to slow down the row scan part of the sensor and save more power.


All comments are moderated to avoid spam.