You are on page 1of 4

CMOS image sensor with adaptive readout scheme

for low power applications


2021 IEEE International Midwest Symposium on Circuits and Systems (MWSCAS) | 978-1-6654-2461-5/21/$31.00 ©2021 IEEE | DOI: 10.1109/MWSCAS47672.2021.9531910

Varun Kumar, Bibhudatta Satapathy, Wilfred Kisku, Amandeep Kaur, and Deepak Mishra
Indian Institute of Technology Jodhpur, Rajasthan, India.
Email: {kumar.129, satapathy.1, kisku.1, amandeepkaur, dmishra}@iitj.ac.in

Abstract—An adaptive readout scheme for low-power CMOS using numerical optimization algorithms. The work reported
image sensors is proposed in this paper. In the first step, the entire in [6] reads the difference between consecutive pixels. The
pixel array is partitioned into square blocks. Subsequently, the current pixel is read-only if the difference is greater than the
intensity range of each block is divided into a fix number of bins
which are adapted to the intensity range of the block. For this quantization step of the ADC. In [7], power saving is achieved
purpose, the maximum and minimum intensity values in each at the column level by reducing the number of conversion
block are obtained using winner take all circuit.The proposed cycles. The current pixel is operated only for δ difference
scheme is implemented and simulated on standard test images. between the two pixels. The most significant bits (MSBs)
The observed peak signal-to-noise ratio for the Cameraman and are copied from the previous reference pixel and the ADC
Thermal image is 40.37 dB and 41.16 dB, respectively for 8×8
block size and 16 bins, with 54% saving in the total power is operated only for δ difference, which reduces the total
compared to the conventional readout. An additional power power consumption of the ADC. Other techniques to reduce
saving can be observed by reducing the total number of threshold power are reported in [8] and [9], where the concepts of re-
levels. configurable ADC and on-array compressive acquisition are
employed to reduce the power.
I. I NTRODUCTION
Motivated from the above studies that use pixel correlation,
The applications of CMOS image sensors (CIS) have in- this work proposes an adaptive readout scheme to reduce the
creased manifold ranging from battery-operated devices like ADC power. The pixel array is divided into a given block
mobile phones, medical devices to IoT applications like size which is sensed for the minimum and maximum intensity
surveillance, agriculture, remote health monitoring, and others. values. The values are divided into the desired number of bin
A wider range of wireless applications imposes requirements thresholds which are only resolved by the ADC to reduce the
on power consumption, bandwidth, and data processing [1]. A power. This is analogous to rule-based artificial intelligence
CIS consists of a pixel array, readout circuit, memories, and (AI) systems where the comparators can be observed as
output buffers. Among these components, the most power- implementing different rules to infer the value of the pixel
hungry element is the readout electronics which consists under consideration. The paper is organized as follows. Section
of double sampling circuits followed by analog to digital II explains the proposed algorithm, section III reports the
converters (ADCs). A wide range of literature is available to simulation results, and the paper is concluded in section IV.
reduce readout power at the circuit level. For instance, in cyclic
ADC, op-amps consume a large amount of power. The power II. P ROPOSED ALGORITHM
reduction can be achieved by using an inverter as an amplifier A CIS mainly consists of a pixel array, readout circuit,
[2]. Similarly, the power of digital to analog converter (DAC) memories, bias circuit, row, and column decoders as shown
in SAR ADC can be reduced by introducing a pilot DAC in Fig. 1 (a). The conventional approach of reading pixels is
switching scheme [3]. The power reduction in slope ADC is to read them sequentially through a common readout circuit
achieved by introducing multiple steps in slope ADC instead placed at the chip level. This approach has limitations in
of a single step [4]. However, in all the above techniques, the terms of high pixel resolution and high frame rate. In the
power reduction is achieved without taking advantage of the proposed readout scheme, the total number of pixels to be
characteristics of the input signal. read is reduced based on the amount of power saving required
Thinking beyond the traditional ways of reducing power at in the readout. In addition, the threshold values are block
the circuit level, a significant power reduction can be achieved adaptive which is decided by the difference in maximum (max)
at the architecture level using the properties of the input and the minimum (min) intensity for a particular block under
signal. In natural images, the neighboring pixels carry similar consideration. These thresholds are the only sampled values
information. Various techniques to achieve low power and to be read by the ADC to reduce power. The blocks with
reduced bandwidth requirements are reported in [5]–[9]. The a lesser intensity difference would need a lesser number of
reduction in power is achieved by reducing the total number thresholds and vice versa. In addition, thresholds or the bin
of pixels to be read or by reducing the number of bits to size are decided according to targeted values of compression.
be resolved. In [5], the number of pixels to be resolved by To implement this algorithm, the proposed CIS requires an
ADC are selected randomly and later reconstructed off-chip additional Winner Take All circuit and an interpolation circuit

978-1-6654-2461-5/21/$31.00
Authorized licensed use limited to: Indian ©2021
Institute ofIEEE 766 on January 06,2024 at 17:17:33 UTC from IEEE Xplore. Restrictions apply.
Technology - Jodhpur. Downloaded
The obtained VCmax and VCmin values are fed to the
interpolation circuit, shown in Fig. 3 which consists of a
resistor ladder followed by a set of comparators. The two ends
of the interpolation circuit are connected to the predetermined
maximum and minimum pixel voltages and the remaining
threshold voltages are generated internally through the ladder.
At a time, the number of thresholds for all the blocks is
fixed and is determined for a given power saving factor. For
instance, if the power saving is 75% and block size = 8×8, the
number of threshold levels obtained will be 16. The generated
threshold values are sequentially processed using 10-bit chip-
level ADC and the corresponding digital bits are stored in the
memory. At the same time, the pixel values of a given block
are compared with the thresholds to generate a decision bit.
For a given pixel value, the output logic level of comparator
Vcom,k is observed and the corresponding 10-bit digital code
for the threshold value where Vcom,k is low is filled.
Accordingly, all the pixels of a given block are filled with
the digitized values stored in the memory, and, therefore, saves
Fig. 1. Block diagram considering (8 × 8) block size and 16 threshold level
the readout power. The same winner take all circuit and the
interpolation circuit will be then reused by another 8 × 8
block for resolving other set of pixel values. In this example,
compared to the traditional read-out technique where the ADC
was used 64 times considering 8 × 8 block, the proposed
scheme operates the ADC only 16 times, which significantly
reduces the power. The proposed algorithm is summarized as
follows:

Algorithm 1: Proposed Algorithm


Result: newImage
blockRow ← 8;
blockCol ← 8;
levels ← ((100 - compression)/100)*(blockRow *
blockCol);
img ← imageAquisition();
for i ← 0 to imgRow do
for j ← 0 to imgCol do
imgBlock ← extractImageBlock(img,i,j);
Fig. 2. Winner take all circuit for (8 × 8) block size max, min ← Of(imgBlock);
diff ← {max, min};
bins ← getBins(diff, levels);
as shown in Fig. 1 (b). The fixed bin size is achieved using for ii ← 0 to blockRow do
the interpolation circuit shown in Fig. 3. The absolute value for kk ← 0 to blockCol do
of thresholds is decided based on the minimum and maximum newBin ← compare{pixels,bins};
intensity value present in the block using winner take all circuit newPixelValue ← newBin;
[10] as shown in Fig. 2. The circuit used to measure the end
maximum intensity consists of a number of N-input source end
followers (SFs), shown in Fig. 2(A). The number of SFs is end
decided by the block size, for example, 8 × 8. All the sources end
are shorted and the bias current is generated through the simple
current mirror circuit. Similarly, a number of P-input source
followers (SFs), shown in Fig. 2(B) are used to determine the
minimum intensity present in the block. All the pixel voltages III. R ESULTS AND A NALYSIS
in a given block are sensed by the winner take all circuit This section presents simulation results of the proposed
and the resultant maximum (VCmax ) and minimum voltages adaptive readout scheme. The algorithm is implemented and
(VCmin ) are stored in the load capacitor. evaluated in MATLAB as a proof of concept for the set

767 on January 06,2024 at 17:17:33 UTC from IEEE Xplore. Restrictions apply.
Authorized licensed use limited to: Indian Institute of Technology - Jodhpur. Downloaded
Fig. 3. Interpolation circuit for 16 threshold levels.

Fig. 6. PSNR in dB with respect to threshold levels for varying block size,
plotted for a cameraman image.

(a) (b)
Fig. 4. (a) Original cameraman image, (b) Image obtained using the proposed
scheme with 16 threshold levels and 8 × 8 block size. The observed PSNR
is 40.37 dB

Fig. 7. PSNR in dB with respect to threshold levels for varying block size,
plotted for a thermal image.

of thresholds at different block sizes and the observations are


plotted in Fig. 6 and Fig. 7. It can be seen from the plots that
(a) (b) for the given block size, the PSNR increases with an increase
Fig. 5. (a) Considered a thermal image from FLIR Dataset (b) Corresponding in the number of thresholds as the pixel values are filled with
image obtained using the proposed scheme with 16 threshold levels and 8 × 8
block size. The observed PSNR is 41.16 dB. better estimates. Similarly, for a given number of thresholds,
the PSNR is higher for smaller block sizes in most cases. Small
block sizes tend to have a small range of intensity resulting in
of standard images and circuit simulations are performed in small bins. However, a small block size also leads to a large
CADENCE. Two sample images (cameraman and a thermal number of blocks resulting in higher power consumption due
image) are shown in Fig. 4 (a) and Fig. 5 (a), respectively. to a linear increment in the total number of ADC runs. As a
The images obtained using the proposed scheme considering trade-off 8 × 8 block size is used for the other test images.
8×8 block size and 16 threshold levels are shown in Fig. 4 (b) The observed PSNR values are listed in Table I and com-
and Fig. 5 (b), respectively. The observed peak signal-to-noise pared with [11], [12], [13], and [14] the different amounts of
ratio for the cameraman and thermal images are 40.37 dB compression or saving in ADC runs. A number of thresholds
and 41.16 dB, respectively, reflecting a good quality of the are calculated using the targeted amount of saving. The
produced image. proposed technique outperforms the existing techniques in all
Further, the variation in PSNR is studied for varying number experiments except for the House image where it shows the

768 on January 06,2024 at 17:17:33 UTC from IEEE Xplore. Restrictions apply.
Authorized licensed use limited to: Indian Institute of Technology - Jodhpur. Downloaded
TABLE I R EFERENCES
PSNR FOR RECONSTRUCTION OF 512 × 512 TEST IMAGES WITH
COMPRESSION OF BLOCKS SIZED 8 × 8 AND COMPARISONS [1] X. Zhong, B. Zhang, A. Bermak, C.-Y. Tsui, and M.-K. Law, “A low-
power compression-based cmos image sensor with microshift-guided sar
Image Boat House Lena Peppers adc,” IEEE Transactions on Circuits and Systems II: Express Briefs,
(Compression) (85%) (95%) (89%) (89%) vol. 65, no. 10, pp. 1350–1354, 2018.
[11] 22.93 17.79 18.17 20.67 [2] F. Tang, B. Wang, A. Bermak, X. Zhou, S. Hu, and X. He, “A column-
[12] 22.81 20.10 20.67 20.25 parallel inverter-based cyclic adc for cmos image sensor with capacitance
and clock scaling,” IEEE Transactions on Electron Devices, vol. 63,
[13] 24.34 24.33 24.44 23.84
no. 1, pp. 162–167, 2015.
[14] 30.49 33.36 32.33 31.41
[3] D. G. Chen, F. Tang, and A. Bermak, “A low-power pilot-dac based
Ours 35.48 28.59 34.33 33.81 column parallel 8b sar adc with forward error correction for cmos image
sensors,” IEEE Transactions on Circuits and Systems I: Regular Papers,
vol. 60, no. 10, pp. 2572–2583, 2013.
[4] S. Lim, J. Lee, D. Kim, and G. Han, “A high-speed cmos image sensor
with column-parallel two-step single-slope adcs,” IEEE Transactions on
Electron Devices, vol. 56, no. 3, pp. 393–398, 2009.
[5] Y. Oike and A. El Gamal, “Cmos image sensor with per-column σ δ
adc and programmable compressed sensing,” IEEE Journal of Solid-
State Circuits, vol. 48, no. 1, pp. 318–328, 2012.
[6] M. Yue, D. Wu, and Z. Wang, “Data compression for image sensor
arrays using a 15-bit two-step sigma–delta adc,” IEEE Sensors Journal,
vol. 14, no. 9, pp. 2989–2998, 2014.
Fig. 8. Power consumption/block (mWatt) with respect to threshold levels [7] H.-J. Kim, S.-I. Hwang, J.-W. Kwon, D.-H. Jin, B.-S. Choi, S.-G. Lee,
for a fixed block size of (8 × 8) for both conventional and proposed scheme. J.-H. Park, J.-K. Shin, and S.-T. Ryu, “A delta-readout scheme for low-
power cmos image sensors with multi-column-parallel sar adcs,” IEEE
Journal of Solid-State Circuits, vol. 51, no. 10, pp. 2262–2273, 2016.
[8] A. Kaur, D. Mishra, and M. Sarkar, “A power efficient image sensor
second-best performance. readout with on-chip delta-interpolation using reconfigurable adc,”
IEEE Sensors Journal, vol. 20, no. 13, pp. 6833–6840, 2019.
The power-saving of the proposed scheme is evaluated by [9] A. Kaur, D. Mishra, K. Amogh, and M. Sarkar, “On-array compressive
using the ADC and comparator data reported [15]. The pro- acquisition in cmos image sensors using accumulated spatial gradients,”
posed scheme requires an additional winner take all circuit and IEEE Transactions on Circuits and Systems for Video Technology,
vol. 31, no. 2, pp. 523 – 532, 2020.
the interpolation circuit which consumes power of 3.52 mW [10] J. A. Starzyk and Y.-W. Jan, “A voltage based winner takes all circuit for
and 2.30 mW, respectively for 64 pixels. Compared to the analog neural networks,” Proceedings of the 39th Midwest Symposium
conventional readout scheme which consumes 27.84 mW of on Circuits and Systems, vol. 1, pp. 501–504, 1996.
[11] D. Mishra, A. Kaur, and M. Sarkar, “A super-pixel based on-chip image
power for 64 pixels, the proposed scheme consumes only compression for high speed cmos image sensors,” in 2017 International
12.78 mW of power for 16 thresholds, resulting in approx- Conference on Electron Devices and Solid-State Circuits (EDSSC), 2017,
imately 54% power saving. pp. 1–2.
[12] A. Kaur, D. Mishra, S. Jain, and M. Sarkar, “Content driven on-
Furthermore, at the system level, the ADC power saving is chip compression and time efficient reconstruction for image sensor
observed for a given 8 × 8 block size with respect to threshold applications,” IEEE Sensors Journal, vol. 18, no. 22, pp. 9169–9179,
levels. The conventional CIS consumes a constant power as 2018.
[13] N. Ansari and A. Gupta, “Image reconstruction using matched wavelet
all the pixels are read sequentially, which is represented by estimated from data sensed compressively using partial canonical iden-
a constant line in the graph shown in Fig. 8. However, using tity matrix,” IEEE Transactions on Image Processing, vol. 26, no. 8, pp.
the adaptive readout scheme, a significant reduction in power 3680–3695, 2017.
[14] A. Kaur, D. Mishra, K. M. Amogh, and M. Sarkar, “On-array com-
is observed, plotted in the same figure. The power reduces pressive acquisition in cmos image sensors using accumulated spatial
from 27.84 mW (conventional image sensor) to 14.52 mW gradients,” IEEE Transactions on Circuits and Systems for Video Tech-
using the proposed technique. Further power saving is also nology, vol. 31, no. 2, pp. 523–532, 2021.
[15] A. Kaur, D. Mishra, and M. Sarkar, “A 12-bit, 2.5-bit/phase column-
observed with respect to reduction in threshold levels. The parallel cyclic adc,” IEEE Transactions on Very Large Scale Integration
power is reduced to 7.56 mW when the number of threshold (VLSI) Systems, vol. 27, no. 1, pp. 248–252, 2018.
levels is reduced to four. The above graph is plotted using a
reference ADC power reported in [15].

IV. C ONCLUSION
In this paper, an adaptive readout scheme based CMOS
image sensor is proposed wherein the bin thresholds are
adapted to the intensity range of the given block. The ADC is
operated only for these bin threshold values which results in
reduction in readout power. The algorithm is implemented and
simulated in MATLAB and Cadence as a proof of concept. For
8 × 8 block, the proposed scheme reduces the conventional
CIS power of 27.84 mW to 12.78 mW for 16 thresholds
resulting in 54% power saving. The adaptive readout scheme
is, therefore, suitable for low power applications.

769 on January 06,2024 at 17:17:33 UTC from IEEE Xplore. Restrictions apply.
Authorized licensed use limited to: Indian Institute of Technology - Jodhpur. Downloaded

You might also like