1

Image Compression Based on Compressive Sensing:

End-to-End Comparison with JPEG
Xin Yuan, Senior Member, IEEE and Raziel Haimi-Cohen, Senior Member, IEEE

Abstract—We present an end-to-end image compression system CS is not a complete signal compression system because its
based on compressive sensing. The presented system integrates “compressed signal”, the measurements vector, is an array of
the conventional scheme of compressive sampling and recon- real numbers rather than a sequence of bits or bytes. Thus, in
struction with quantization and entropy coding. The compression
performance, in terms of decoded image quality versus data rate, order to build a complete system we added a source coding
is shown to be comparable with JPEG and significantly better at stage, in which the measurements are quantized, and a channel
arXiv:1706.01000v1 [cs.CV] 3 Jun 2017

the low rate range. We study the parameters that influence the coding stage, which the quantized measurements into a byte
system performance, including (i) the choice of sensing matrix, sequence, using entropy coding.
(ii) the trade-off between quantization and compression ratio,
and (iii) the reconstruction algorithms. We propose an effective
method to jointly control the quantization step and compression A. Related Work
ratio in order to achieve near optimal quality at any given bit
rate. Furthermore, our proposed image compression system can Goyal et al. [8] applied an information-theoretic approach
be directly used in the compressive sensing camera, e.g. the to assess the effectiveness of a CS based compression system
single pixel camera, to construct a hardware compressive sampling for sparse signals x ∈ RN , with only K non-zero entries,
system.
a.k.a. K-sparse. Their benchmark was the “baseline” method,
Index Terms—Compressive sensing, image compression, quan- where the coded bit sequence consisted of the sparsity pattern
tization, entropy coding, sparse coding, reconstruction, JPEG, (i.e. a specification of the indices of the non-zero entries in x)
JPEG2000. and the quantized and coded non-zero entries. They showed
that the rate-distortion functions of a CS-based compression
I. I NTRODUCTION system are considerably worse than those of the baseline
Compressive Sensing (CS) [2], [3], [4] has been proposed method for two reasons. First, the number of measurement
more than a decade ago as a method for dimensionality M required by CS to recover x is several times larger than
reduction of signals which are known to be sparse or com- K, M/K ≥ log(N/K), and the number of bits needed to
pressible in a specific basis representation. By “sparse” we represent the additional M − K quantized variables exceeds
mean that only a relatively small number of the coefficients the number of bits needed to specify the sparsity pattern,
of the representation are non-zero, whereas “compressible” especially when the number of bits per measurement is high.
indicates that the magnitude of the coefficients decays quickly, Second, the quantization noise in the baseline method is
according to a power law, hence the signal can be well proportional to K, whereas in CS with a random sensing
approximated by a sparse signal. In the CS paradigm, the matrix, it is proportional to M . Goyal et al. suggest that the
signal is projected onto a low-dimension space, resulting in use of distributed lossless coding and entropy-coded dithered
a measurements vector. If the signal is sparse, it is possible quantization might potentially alleviate those problems, but
to exactly reconstruct it from the measurements vector. If the complexity added by those methods would probably make
the measurements are noisy or the signal is not sparse but them impractical.
compressible the reconstruction yields an approximation to the Despite this pessimistic outlook, the research of the effect
original signal. Natural images are inherently compressible in of quantization on CS measurements got significant attention
the frequency or wavelet domain and therefore suitable for in the last few years. Laska and Baraniuk studied the trade-off
CS. The past five years saw an impressive progress in this between the number of measurements and quantization accu-
field with new reconstruction algorithms [5], [6], [7] achieving racy [9] and showed that the sensitivity of the reconstruction
better reconstructed image quality at a lower compression ratio accuracy to the quantization noise varies dramatically with
(the ratio of the dimension of the measurements vector to the the compression ratio. Laska et al. showed that even a small
number of pixels in the original image). These algorithms number of saturated measurements, i.e. measurements whose
go beyond sparsity and leverage other properties of natural quantization error is not bounded, may cause a considerable
images, such as having low rank [6], or being being capable degradation on reconstruction accuracy [10]. Methods for
of denoising [7]. accounting for the quantization effect in the reconstruction
Encouraged by these achievements, we set out to create an algorithm were studied in [11]. The extreme case of 1-bit
end-to-end image compression system based on CS. In itself quantizer was investigated in [9], [12], [11]. The asymptotic
normality and approximate independence of measurements
The authors are with Nokia Bell Labs, 600 Mountain Avenue, Murray Hill, generated by various sensing matrices were shown in [13].
NJ, 07974, USA, [email protected]. The MATLAB code used to generate
the results in this paper can be downloaded at CS vs. JPEG Demo and more Dai et al. compared various quantizer designs and studied their
results are available at [1]. effect on reconstruction accuracy [14], [15]. The distribution
2
Figure 1
of measurements generated by video and image CS systems,
which included quantization, were also described in [16], [17].
However this significant body of research was of limited value
for our purposes. First, these works assume a random, or a CS (1879 Bytes)

structurally random sensing matrix, while the sensing matrices

suitable for our purposes could be different (see Sec. II-A).
Second, most of these works did not assume any channel
coder and therefore did not study the resulting bit rates. Those
that considered a channel coder did not study the interaction
between the quantizer and the channel coder and the trade-offs
in their design.
Ground truth

JPEG (1903 Bytes)

B. Image Compression via JPEG and JPEG2000
Since its introduction in 1992, JPEG [18] has been one
of the most popular image compression methods in use.
JPEG2000 [19], which was released ten years later, has
superior compression properties at a wider bit rate/quality
range. A brief overview of these standards is given in Sec. II-F,
where their architectures are compared with that of the pro- Fig. 1. Comparison of JPEG and CS-based compression of the “Monarch”
image. Curves of structural similarity (SSIM) vs. compressed file size, as
posed compressive sensing based image compression system well as decoded images of similar file size are shown. Please zoom-in on red
(CSbIC). circles for details.
At the time of its introduction, the higher complexity of
JPEG2000 could be an impediment for its adoption, but
the sensors need to be inexpensive and have low energy
as computing capabilities improved, this was not longer an
consumption, the network operates at low power and is usually
issue for most cases. Nevertheless, JPEG remains the image
unreliable, and much prior information may be gathered while
compression tool of choice in a wide range of applications. It
the system is running. Therefore, CSbIC may be useful in a
appears that the advantages of JPEG2000 were not necessary
class of applications which are quite different from those that
for many targeted applications, which raises the question if
JPEG and JPEG2000 were designed for.
there is any need for yet another image compression scheme.
However, a compression scheme based on CS should have
some unique properties which are radically different from D. Contributions of This Work
those of any conventional signal compression scheme, which
This paper makes the following contributions:
may justify its adoption.
i) We provide an end-to-end architecture for a CSbIC
C. Uniqueness of CS based Image Compression system, including compressive sensing, source coding,
channel coding and a mechanism to adjust the system
A compression scheme based on CS enjoys the following parameters to control data rate and reconstruction quality.
unique properties which are primarily different from those of ii) We address the theoretical issues raised by Goyal et al. [8]
any conventional signal compression scheme. in a practical way, by using domain-specific knowledge:
• The encoder is of much lower complexity than the de- (i) we employ reconstruction algorithms that do not rely
coder. In fact, a significant part of the encoder processing, on sparsity but on other properties of natural images,
namely the measurements generation, can be done in the and (ii) we use deterministic sensing matrices which are
analog domain, e.g. by a single-pixel camera or a lensless known to be effective for this such signals.
camera [20], [21], [22], [23], [24]. iii) Having an end-to-end system enables us to measure its
• The decoder is not fully defined by the encoder. Different performance in terms of quality versus data rate, and
reconstruction algorithms can be applied to the same thus benchmarks it against the leading image compression
measurements vector at different situations, and as more standards, namely JPEG and JPEG2000. We show that
prior information becomes available the reconstruction our CSbIC system is on-par with JPEG, with a clear
algorithm can be improved to take it into account [25], advantage in the low data rate range. Please refer to Fig. 1
[6], [26], [27], [5], [7], [24], [28]. as an example.
• The loss of a small fraction of the measurements gen- iv) We describe various design choices in each of the system
erally results in only a minor degradation in reconstruc- components and study the effect that these choices have
tion quality. This may be used to achieve robustness to on the overall system performance.
channel impairments without the usual overhead of error The system that we describe is far from being fully optimized,
protection. and throughout the paper we point out where further improve-
These unique properties suggest that CSbIC would be very ment may be made. Nevertheless, even at this preliminary
useful in applications such as media sensor networks, where stage the performance of our system makes it a viable al-
 3

CS Image Encoder CS Image Decoder

Joint bit rate/quality Control CS Reconstruction
coded non-linear iterative optimization
Original Recon-
Quantize
measurements
Projection of whole structed
Image Common stream image
image onto low Entropy coding Entropy
quantization unquantize
dimension space step M (<<N) decoding
measurements

Lossy Lossless
Possible implementation by CS camera

JPEG Encoder JPEG Decoder Decoded

Bit rate/quality Control inverse 2D-DCT image
JPEG On 8x8 Blocks
Original compressed
Image 2D-DCT on Quantize stream
8x8 blocks coefficient specific Entropy coding entropy
quantization step N unquantize
decoding
coefficients
Lossless Lossy Lossless

2 © Nokia 2016 Confidential

Fig. 2. Image compression architecture comparison between proposed CSbIC (top) and JPEG (bottom).

ternative to the conventional methods (refer to Figs 4-6 for which leverage this property [22], [29]. These matrices are not
comparison with JPEG). incoherent with the common sparsity bases of natural images,
The rest of this paper is organized as follows. Sec. II as classical CS theory would require for guaranteeing robust
describes the system architecture and discusses the design reconstruction [2], [4], but generally they perform better than
choices in each component. Sec. III provides the general the classical sensing matrices in our application, i.e. image
framework of reconstruction algorithms. Sec. IV presents the compression. For example, (1) can be implemented by per-
results of our performance testing and Sec. V discusses the forming 2D-DCT on the image pixels and then reordering the
implications of our results. resulting 2D-DCT coefficients in a “zig-zag” order (similar
to the one used in JPEG encoding) and selecting the first M
II. S YSTEM A RCHITECTURE low-frequency coefficients [22].
In some applications, such as the single-pixel camera and
A diagram of the system architecture is given in Fig. 2. lensless camera [20], [21], [22], a binary-valued matrix, e.g. a
Each of the encoding steps is matched by a corresponding matrix whose entries are only ±1 (or {0, 1}) is more suitable
decoding step (in reverse order), with the exception of the bit for hardware implementation. In this case we approximate the
rate/quality control block, which appears only in the encoder. 2D frequency decomposition by using a 2D Walsh-Hadamard
In the following we present a detailed description of each of transform (2D-WHT) [30]. Let
those processing steps.
W = Wh ⊗ Wv , (2)
A. Sensing Matrix and Measurements Generation
where ⊗ denotes the Kronecker product [31] and Wh , Wv
We consider monochromatic images of Nv ×Nh pixels. The are Walsh-Hadamard matrices in sequency order [32], that is,
pixels of the input image are organized column by column as the k th row of Wh , Wv , has k − 1 zero crossings (if Nv or
a pixel vector x ∈ RN (N = Nv Nh ). The pixel vector x is Nh is not a power of two, the image is padded by zero pixels).
multiplied by a sensing matrix Φ ∈ RM ×N , M  N , yielding Similar to the 2D-DCT case, the selected measurements are
the measurements vector the first M coefficients of the zig-zag ordered entries of
Wx [29]. Note that Wx can also be computed numerically in
y = Φx. (1)
an efficient way, using the Fast Walsh-Hadamard Transform.
Φ is quite a large matrix, even for small images. Therefore, We can get the CS theoretical guarantee for successful
for practical reasons, Φ is never stored, and the operation reconstruction, w.h.p. (with high probability), by replacing the
(1) is implemented as a fast transform. It is well known that deterministic matrices described above with random matrices,
a 2-dimensional discrete cosine transform (2D-DCT) is very whose entries are independent, identically distributed (IID)
effective in decorrelating an image, and most of the energy random variables (RVs), with Gaussian or Rademacher dis-
of the image is concentrated in the low frequency transform tributions [33], [34]. These matrices are universal, i.e. w.h.p.
coefficients. Recently, new sensing matrices were introduced they are incoherent with any given sparsity basis. Further-
4

more, the measurements generated by those random matri- If Q(y) = c and |c| < L, we define the dequantizer by
ces are mutually independent and asymptotically normally
Q−1 (c) = cs + µ, (5)
distributed, which is helpful in the quantization and coding
design. Such fully random matrices do not allow fast transform hence the quantization error is bounded by
implementation of (1), but similar desired properties and
performance guarantees were shown for structurally random |y − Q−1 (Q(y))| ≤ 0.5s. (6)
matrices (SRM) [13], [35], [36], where Φx is obtained by On the other hand, if |c| = L, the quantized measurement
applying a randomly selected permutation to x, computing a is saturated and the quantization error cannot be bounded.
fast transform, such as DCT or WHT, on the permuted vector, Even a small number of saturated measurements can cause
and randomly selecting M of the transform coefficients as severe quality degradation unless they are specially handled
measurements (the DC coefficient is always selected). We de- in the reconstruction [10]. The simplest way to do it is by
note these matrices SRM-DCT and SRM-WHT, respectively. not using the saturated measurements at all; attempts to mod-
ify the reconstruction algorithm to use those measurements
B. Quantization showed little gain over simply discarding them. Another option
(not considered in [10]) is to code the value of Q̃(y) for
The quantizer maps the measurements vector y into a each saturated measurement y in some ad hoc method and
finite sequence of codewords taken from a finite codebook transmit is as additional information. In both cases saturated
C and the dequantizer maps the codewords sequence into a measurements incur a penalty, either in the form of transmitted
measurements vector which is an approximation of the original codewords which are not used, or as ad hoc transmission of
measurements vector. If the sensing matrix is deterministic Q̃(y). Therefore, we select L large enough to make saturation
the measurements are highly uncorrelated; if it is a SRM the a rare event. In fact, L can be set sufficiently large to eliminate
measurements are nearly independent. Hence the advantage saturation completely, but a very large codebook may have an
of vector quantization [37] over scalar quantization is small adverse effect on channel coding (see Sec. II-D3). We found
and does not justify its added complexity [14]. Therefore, that a good trade-off is to select L so that quantizer’s range
we consider a scalar quantizer which maps each measurement s(L−0.5) is about 4 standard deviations of the measurements.
{yi }M
i=1 to a codeword qi = Qi (yi ) ∈ C, where Qi : R → C is However, the system is not very sensitive to this parameter
the quantizer of the ith measurement. In this work we use the — performance does not change much if the range is 3 or
same quantizer for all measurements, hence in the following 6 standard deviations. We also compared ignoring saturated
we omit the subscript i from Qi . measurements to sending Q̃(y) for each of them. We chose
The simplest scalar quantizer is the uniform quantizer. We the latter because it performed slightly better and had the
select the “mid-tread” type, defined by important advantage of not requiring any change in the re-
def construction algorithm.
Q(y) = max(−L, min(L, Q̃(y))), (3) With all the sensing matrices considered, the first measure-
def
Q̃(y) = b(y − µ)/s + 0.5c, (4) ment y0 is the DC coefficient, which is the sum of all pixels
in the image. Since the pixels are unsigned, y0 is much larger
1
PMbyc denotes the largest integer not exceeding y, µ =
where than the other measurements and is always saturated, therefore
M i=1 yi is the mean of the measurements, s is the quan- it requires special handling: y0 is excluded when calculating
tizer’s step, Q̃(y) is the unclipped quantized value, and L is the mean (µ), and the standard deviation of the measurements,
a positive integer which determines the range s(L − 0.5) of and Q(y0 ) = L is not included in the quantized measurement.
the actual quantizer Q(y). Consequently there are 2L + 1 Instead, Q̃(y0 ) is coded in an ad hoc fashion and transmitted
codewords, C = {−L, · · · , L}. Since the distribution of the separately.
measurements is highly non-uniform, the codewords distribu- Unless the quantization is very coarse, its effect on the
tion is also not uniform, hence in order to represent codewords measurements can be modeled as adding white noise, uni-
effectively by a bit sequence we need to use variable length formly distributed in [−s/2, s/2], which is uncorrelated with
coding (VLC) in the channel coder. On the other hand, an the measurements [37]. Hence the variance of the quantization
optimal quantizer (in the mean square sense) or an entropy noise in each measurement is
constrained quantizer [37] usually results in nearly equally
2
populated quantization regions, which makes it possible to σQ = s2 /12. (7)
use fixed length coding (FLC) with little data rate penalty.
The integral pixel values are generally obtained by sampling
Thus the design choice is between a simple quantizer with
an analog signal and rounding the samples values to the nearest
a sophisticated VLC, versus a sophisticated quantizer with a
integer. Hence the pixel values contain digitization noise with
simple FLC. We opted for the first option because designing
variance of 1/12. This noise appears in the measurement yj
an optimal quantizer requires knowledge of the measurements
with variance
distribution, which is difficult to estimate for a deterministic 2
σD = kφj k22 /12, (8)
sensing matrices. Another reason to use a uniform quantizer
is that the reconstruction may be sensitive to the presence of where φj is the j th row of Φ. In the sensing matrices that
even few measurements with large errors [10], which is often we consider kφj k2 is constant, kφj k2 = kΦk2 . Clearly,
the case with non-uniform quantizers. there is no point in spending bits to accurately represent the
 5

digitization noise, hence we need to have σQ ≥ σD and labels into a single label L, thus we have 2L codewords:
consequently s ≥ kΦk2 . Typical quantizer step sizes are −L + 1, . . . , L.
between kΦk2 to 50kΦk2 . Let pc , c ∈ C be the probability of a measurement to be
quantized to c. If the codewords {Q(yi )}M i=1 are IID random
C. Image Quality Control variables, then a tight lower bound on the expected number of
bits required to transmit these codewords is the entropy rate:
The compression ratio R and the quantizer step size s con-
trol the coded image size and the quality of the reconstructed def P
H = −M c∈C pc log2 pc . (10)
image. One can get to the same reconstruction quality with
various combinations of these parameters but the coded image Arithmetic coding (AC) [38] represents the codeword sequence
size varies greatly. Our experiments (Section IV-B) showed by a bit sequence, the length of which can get arbitrarily close
that at any given quality the lowest bit rate is achieved when to the entropy rate for a large M . We use AC to encode the
codewords sequence. The AC bit sequence is coded as a bit
Rs = CkΦk2 , (9) array.
where C is a constant. The optimal value of C depends on Since the probabilities pc , c ∈ C are not known a priori,
the type of sensing matrix and varies from image to image. they need to be estimated and sent to the receiver, in addition
However, we found that using C = 2.0 is a good general to the AC bit sequence. These probability estimates can be
value for all pictures. Thus, in our tests s is determined by R obtained in two ways: For SRMs, the measurements are
using (9), with C = 2.0. This quantization step is sufficiently approximately normally distributed [13], hence for |c| < L,
fine to allow modeling the quantization noise as uncorrelated, pc is the normal probabilities of the quantization Pintervals
uniformly distributed white noise. [cs + µ − 0.5s, cs + µ + 0.5s), and pL = 1 − |c|<L pc .
Thus, all that needs to be sent to the receiver is the estimated
D. Lossless Coding standard deviation of the measurements, which is coded as a
real number. For deterministic sensing matrices, it is neces-
The lossless encoder encodes the codeword sequence gen- sary to compute a histogram of the quantized measurements
erated by the quantizer, as well as some miscellaneous infor- sequence, use it to determine the probabilities and then code
mation (e.g. µ, s, and the ad hoc representation of saturated the histogram and send it to the receiver along with the AC
measurements), as a bit sequence. The lossless decoder ex- bit sequence. Sending the histogram is an overhead, but it
actly decodes the codeword sequence and the miscellaneous is small in comparison to the gain achieved with arithmetic
information from the bit sequence. coding. In fact, even with measurements generated by SRM,
1) Coded Numbers Format : Various types of numbers are in many cases the total bit rate achieved when using AC with
coded by the lossless encoder (and decoded by the lossless a histogram is better than the total bit rate when using the
decoder). Each type is encoded in a different way: normal distribution assumption, because the gain obtained by
Unbounded signed or unsigned integers are integers accurately describing the actual codeword frequencies is more
whose maximal possible magnitude is not known in advance. than the overhead required of transmitting the histogram.
They are represented by byte sequences bytes, where the most In natural images the magnitudes of the coefficients of 2D-
significant bit (MSB) In each byte is a continuation bit — it DCT or 2D-WHT decay quickly, hence measurements gener-
is clear in the last byte of the sequence and set in all other ated using a deterministic sensing matrix are not identically
bytes. The rest of the bits are payload bits which represent distributed, which violates the assumptions under which AC
the integer. The number of bytes is the minimal number that is asymptotically optimal. In order to handle this problem
has enough payload bits to fully represent the unsigned or we partition the codeword sequence into sections, and for
unsigned integer. each section we compute a histogram and an AC sequence
Real numbers, which are natively stored in single or double separately. This, of course, makes the overhead of coding the
precision floating point format [38] are coded as pairs of histograms significant. In the following we describe how the
unbounded signed integers representing the mantissas and the histograms are coded efficiently and how to select a locally
exponents in the floating point format. optimal partition of the codewords sequence.
Bit arrays are zero padded to a length which is a multiple 3) Coding of Histograms : In order to be efficient, the code
of 8 and coded as a sequence of bytes, 8 bits per bytes. of histograms of short codeword sequences should be short
Bounded unsigned integer arrays are arrays of unsigned as well. Fortunately, such histograms often have many zero
integers, each of which may be represented by a fixed number counts, which can be used for efficient coding. A histogram
of bits. An array of n integers, each of which can be repre- is coded in one of three ways:
sented by b bits, is encoded as a bit array of bn bits. Full histogram: A sequence of 2L unbounded unsigned in-
Each of these number formats can be easily decoded. Note tegers, containing the counts for each codeword. This method
that these formats are byte aligned for simpler implementation. is effective when most counts are non-zero.
2) Entropy Coding : The codewords ±L represent saturated Flagged histogram: A bit array of 2L bits indicates for
measurements. Whether those measurements are ignored or each codeword whether the corresponding count is non-zero,
transmitted separately, there is no distinction between sat- and a sequence of unbounded unsigned integers contains the
uration from above or below, hence we merge these two non-zero counts. This method is effective when a significant
6

share of the counts is zero. Algorithm 1 partitioning the codeword sequence into entropy
Indexed histogram: A bounded integer indicates the num- coded sections.
ber of non-zero counts, an array of bounded integers contains 1. Initialization:
the indices of the non-zero counts, and a sequence of un- a. Partition the codewords sequence {Q(yj )}M j=1 into sec-
K
bounded unsigned integers contains the non-zero counts. This tions {Sk }k=1 such that each section contains only one
method is effective when most of the counts are zero. In the codeword (RLE).
extreme case of a single non-zero count the AC bit sequence b. Compute the histograms’ counts
def
is of zero length, hence this histogram coding is effectively a hk (c) = hSk (c), c ∈ C, k = 1 . . . K.
run length encoding (RLE). c. Compute the coded sections’ lengths using (14):
def
The histogram is coded in these three ways and the shortest lk = L̂(Sk ), ∀k = 1, . . . , K.
code is transmitted. A 2-bit histogram format selector (HFS) 2. For a fixed m > 0 Let
indicates which representation was chosen. The HFSs of all
sections of the codeword sequence are coded as a bit array. P = {(k, j)|1 ≤ k < j ≤ min(K, k + m − 1)} (11)
Thus, each section of the codeword sequence is represented and for each pair (k, j) ∈ P, let Sk,j be the section obtained
by the HFS, the selected histogram representation and the AC by merging sections Sk , . . . , Sj .
bit sequence.
a. Compute the histograms counts of Sk,j :
Pj
hk,j (c) = r=k hr (c) , c ∈ C.
b. Using the histogram and (14), compute the gain of
Raw image parameters
Header

merging Sk , . . . , Sj :
Pj
Algorithmic parameters gk,j = r=k lr − L̂(Sk,j ).
3. If ∀(k, j) ∈ P : gk,j < 0, exit.
Global image parameters 4. Update the partition:
No. of sections HCS of each section a. Let (k ∗ , j ∗ ) = arg max(k,j)∈P gk,j .
b. Merge sections Sk∗ , . . . , Sj ∗ into one section, with
… Histogram section K
body

Histogram section 1 histogram and code length of {hk∗ ,j ∗ (c), c ∈ C} and

Pj ∗
AC sequence section 1 … AC sequence section K
r=k∗ lr − gk ,j , respectively.
∗ ∗

c. Go to step 2.
Saturated measurements
our experiments. When the algorithm starts all the bits are
Fig. 3. Coded image structure. spent on HFSs and histogram representation, and none on
AC bit sequences. As the algorithm progresses and sections
4) Partitioning into AC sections: Partitioning the codeword are merged, more bits are spent on AC, and the histograms
sequence into AC sections requires estimating the number of become fewer in number, but aving more non-zero counts. It
bits in each coded section. Let {hs (c), c ∈ C} be the histogram is plausible that we could get even better compression if we
of section S, where hs (c) is the count for codeword c. In order used more efficient ways for histogram representation, e.g. by
to avoid repeated AC encoding computation, the number of using more parametric models of approximated histograms Bell Labs (in
bits5 in the AC sequence of S is estimated, based on (10), by © Nokia addition
Internal Use to the normal distribution).

def P  5) Coded Image Structure: The data structure of the coded

L̂AC (s) = 8 c∈C hs (c) log2 (Ms /hs (c))/8 , (12) image is shown in Fig. 3. The coded image begins with a
where dae is the least integer not smaller than a, and header specifying the image parameters (size, bits per pixel,
P color scheme, etc.) and the encoder’s algorithmic choices
Ms = c∈C hs (c) (13) (e.g. sensing matrix type, compression ratio, quantizer param-
eters). The body of the code consists of some global image
is the number of the codewords in the section. The total section
parameters (µ, s, L, Q̃(y0 ), etc.) followed by the entropy
code length is estimated by
coded quantized measurements, which are the bulk of the
L̂(s) = L̂AC (s) + LH (s) + 2, (14) transmitted data: the HFSs, the histograms and the AC se-
quences for each section. If the values of Q̃(y) are transmitted
where LH (s) is the number of bits used for the histogram for saturated measurements (corresponding to terms equaling
coding and the last term on the right hand side is the bits L in the codeword sequence), these are coded as an array of
used for the HFS. unbounded signed integers.
The partition of the codeword sequence into sections is
preformed using a greedy algorithm (Algorithm 1). We begin E. Decoding
by partitioning the sequence into RLE sections, and in each
iteration we merge up to m consecutive sections, selected so Decoding is done in reverse order of encoding, as follows:
that the merging yields the greatest possible reduction in total • The numerical parameters and bit arrays in the coded
coded sections length. m is a constant which was set to 4 in image are parsed.
 7

• The quantization codewords are recovered by arithmetic from a quantization table and is different for each co-
decoding. efficient, resulting in quantization noise shaping, which
• The unsaturated quantized measurements are computed may give JPEG an advantage at higher quality/data rate
using (5). For the saturated measurements (those having (Fig. 4). JPEG2000 also performs quantization noise
a codeword of L), if values of Q̃(y) are transmitted, they shaping through varying step size, and in addition, its
are used. Otherwise, the values of the quantized saturated quantizer is not exactly uniform—the quantization inter-
measurements are set to zero. val around zero (the “dead-zone”) is larger than the other
• The sensing matrix is determined according to the al- quantization intervals, effectively forcing small wavelet
gorithmic choices in the coded image. If there was no coefficients to zero and reducing the amount of bits spent
ad hoc transmission of Q̃(y) for saturated measurements, on coding them. It is plausible that using noise-shaping
the rows corresponding to these measurements are set to and slight non-uniformities in the quantization would
zero. In practice this is done by replacing the original improve the performance of CSbIC as well.
sensing matrix Φ by DΦ, where D is a M ×M diagonal • The bit rate and quality trade-off in JPEG and JPEG2000
matrix whose diagonal elements are zero for saturated is controlled by tuning the operation of a single module
measurements and one for unsaturated measurements. — the quantizer. In contrast, the CSbIC this trade-off is
• The image is reconstructed using the sensing matrix and controlled by jointly tuning two different modules: The
the quantized measurements, as described in detail in projection, or measurement capturing module is tuned by
Sec. III. changing the compression ratio, and the quantizer is tuned
by changing the quantization step.
• In JPEG, entropy coding is based on Huffman coding
F. Module Comparison of CSbIC with JPEG and JPEG2000 and RLE. JPEG2000 uses arithmetic coding, with a so-
We now compare the architectures of CSbIC, JPEG and phisticated adaptive algorithm to determine the associated
JPEG2000 and consider the aspects which may lead to perfor- probabilities. CSbIC uses arithmetic coding, which is
mance differences. Fig. 2 compares the architecture of CSbIC known to be better than Huffman coding, but instead
with that of JPEG side by side. The main common points and of using adaptive estimation of the probabilities, the
differences are: codewords are partitioned into sections and for each
section a histograms of codewords is computed and sent
• JPEG, JPEG2000 and CSbIC begin with a linear projec-
as side information. The overhead of the transmitted
tion of the image onto a different space. However:
histograms may be a disadvantage of CSbIC relative to
– JPEG2000 may partition the image into tiles of JPEG2000.
varying sizes, which are processed separately. This • In JPEG and JPEG2000, the decoder generates the im-
is equivalent to using a block-diagonal projection age from the dequantized transform coefficients by an
matrix, where each block corresponds to a tile. In inverse 2D-DCT or wavelet transform, respectively — a
JPEG the tiles (referred to as blocks) are of fixed simple linear operation which does not rely on any prior
size of 8 × 8 pixels. In CSbIC the projection is done information not included in the received data. In contrast,
on the whole image, which was adequate for the the CS reconstruction in CSbIC is an iterative, non-linear
image sizes which we experimented with. For larger optimization, which relies on prior assumptions about the
images, adding tiling should be straight forward. structure of the image (e.g. sparsity).
– In both JPEG and JPEG2000, each tile/block is
projected on a space of the same dimension, N ,
hence there is no data loss in this operation, whereas III. I MAGE R ECONSTRUCTION
in CSbIC the projection is lossy since it is on a M - Since the seminal works of Candès et al. [2] and
dimensional space, M  N . Donoho [3], various reconstruction algorithms have been de-
– In JPEG the projection is a 2D-DCT with the output veloped [26], [39], [40], [5], [41], [42]. The early recon-
organized in zig-zag order, while in JPEG2000 it is struction algorithms leveraged the property of natural images
a 2D-wavelet transform. In CSbIC, a 2D-DCT based of being compressible when projected by a suitable sparsity
projection is one of several options. operator D:
– JPEG uses block to block prediction of the DC
f = Dx, (15)
coefficient in order to deal with the issue of the
DC coefficient being much larger than the other where f denotes the projected vector and it is usually forced
coefficients. In JPEG2000 this is done by subtracting to be sparse. D can be a pre-defined basis (DCT or wavelet),
a fixed value from all pixels before the projection. In or learned on the fly [43]. Another popular sparsity operator is
contrast, CSbIC takes care of this issue in the quan- the Total Variation (TV), where D is a projection on a higher
tization stage. The effect of these different methods dimension space [5], [25].
is similar and has little impact on performance. Recently, better results were obtained by algorithms such
• CSbIC uses a simple uniform scalar quantizer with an as D-AMP [7] and NLR-CS [6], which exploit established
identical step size for all measurements. JPEG uses the image denoising methods or the natural images property of
same type of quantizer, but the step size is selected having a low rank on small patches. While using different
Figure 4
8

Barbara Boats Cameraman Foreman

House Lenna Monarch Parrots

Fig. 4. Performance diagrams, SSIM vs. compressed file size (in bytes), comparing JPEG (black solid curves), JPEG2000 (black dash curves) with CSbIC
compression using different sensing matrices – 2D-DCT (solid) and 2D-WHT (dash), and different reconstruction algorithms — GAP-TV (blue), NLR-CS
(red) and D-AMP (green).

projection operators D, most of these algorithms compute x̂, and the reconstructed image quality, measured by structural
the estimated signal, by solving the minimization problem similarity (SSIM) [46]. We preferred SSIM over Peak-Signal-
to-Noise-Ratio (PSNR) as a measure of quality because SSIM
x̂ = argminx kDxkp , s.t. y = Φx, (16) better matches the quality perception of human visual system
where p can be 0 (kf k0 denotes the number of non-zero (refer to Fig. 5 and [1] for PSNR results, which provide similar
components in f ) or 1, using kf k1 as a computationally- observations to SSIM). This observation is verified by the
tractable approximation to kf k0 [44]. Alternatively, k kp can exemplar reconstructed images compares with JPEG images
stand for the nuclear norm k k∗ to impose a low rank in Fig. 6. Results are presented as points in SSIM vs. coded
assumption [6]. image size diagram. By connecting the points corresponding to
test cases with different rate conntrol parameters (but identical
Problem (16) is usually solved iteratively, where each itera-
in all other parametrs) we get performance curves which
tion consists of two steps [24]. Beginning with an initial guess,
can be compared with each other in order to determine the
the first step in each iteration projects the current guess onto
better operating parameters. Performance curves for JPEG or
the subspace of images which are consistent with the measure-
JPEG2000 were obtained using the MATLAB implementation
ments, and the second step denoises the results obtained in the
of those standards (the “imwrite” function). All the results
first step. Various projection and denoising algorithms [43],
presented in this section and more results can be found at [1],
[45] can be employed in this general framework in order to
and these results can be reproduced by the code downloadable
achieve excellent results.
from [1].
In our experiments we have used an improved variant of TV
known as Generalized Alternating Projection Total Variation
(GAP-TV) [28], as well as D-AMP and NLR-CS. A. Effect of The Choice of Sensing Matrix
The performance with the deterministic sensing matrices,
2D-DCT and 2D-WHT, was always significantly better than
IV. P ERFORMANCE T ESTING
with the SRMs with the same fast transforms, SRM-DCT
Our test material consisted of 8 monochrome images of and SRM-WHT, respectively (Fig. 7), regardless of the re-
256x256 8-bit pixels (Barbara, Boats, Cameraman, Foreman, construction algorithm which was used. Within each of those
House, Lenna, Monarch, Parrots — see Fig. 4). These images groups (deterministic matrices and SRMs), the sensing ma-
were processed in a variety of test conditions, specified by trices based on DCT generally yielded better performance
encoder and reconstruction parameters. The outcome of pro- than the those based on WHT (Fig. 7 right). However, (i) the
cessing an image in a particular test condition is the data rate, difference, in terms of SSIM for the same file size, is generally
as expressed by the size of the coded image file (in bytes), smaller than the performance difference between deterministic
 9

Fig. 5. PSNR vs. compressed file size (in bytes), comparing JPEG (black
solid curves), JPEG2000 (black dash curves) with CSbIC compression using Fig. 7. Performance plots (SSIM vs. compressed file size in bytes) for the
different sensing matrices – 2D-DCT (solid) and 2D-WHT (dash), and two images of CS with different sensing matrices. D-AMP (left) and NLR-CS
different reconstruction algorithms — GAP-TV (blue), NLR-CS (red) and (right) are used for reconstruction. Results with 8 images and 3 algorithms
D-AMP (green). Results with 8 images are available at [1]. are available at [1].

scaling factor, in all four matrices. Therefore, when speaking

about signal energy and measurements we refer only to the
AC component and exclude the first measurement which is the
DC component. Since SRMs randomize and whiten the signal
prior to applying the transform, their measurements capture
about M/N of the signal energy. The 2D-DCT measurements
capture a much larger fraction of the signal energy because
most of the signal energy is concentrated in the low frequen-
cies. Since 2D-WHT is a crude approximation of spectral
decomposition, its low “frequency” components capture less
energy than those of 2D-DCT, but much more than the M/N
of the SRM matrices. This argument is certainly true, but it
is only a part of the explanation. Consider the sensing matrix
given by
Φ̃ = ΘΦ, (17)
where Φ is a M × N 2D-DCT or 2D-WHT deterministic
sensing matrix and Θ is given by
 
1 01×(M −1)
Θ= , (18)
0(M −1)×1 Θ̃

where Θ̃ is a random orthonormal (M − 1) × (M − 1) matrix,

Fig. 6. Left: SSIM vs. compressed files size (in bytes) D-AMP+2D-WHT is defined by
used in CSbIC. Right: exemplar reconstructed images from CSbIC and JPEG Θ̃ = SHR, (19)
with similar file size. More results are available at [1].
where R is a diagonal matrix whose diagonal entries are
IID random variables which get the values of ±1 with equal
and structurally random matrices, (ii) the difference between probability; H is the DCT transform matrix of order M − 1,
SRM-DCT and SRM-WHT is generally much smaller than with rows scaled to have unit norm; and S is a random
the difference between 2D-DCT and 2D-WHT, and (iii) the selection matrix, i.e. a matrix in which each row and each
magnitude of the difference varies with the image and the column contains one entry of 1 and the rest are zero.
reconstruction algorithm which is used. In particular, with The measurements computed by Φ̃ are obtained by a
GAP-TV the performance difference between DCT and WHT random orthonormal transformation (ROT) of the measure-
based matrices was most pronounced, whereas with D-AMP ments computed by Φ, hence we denote the sensing matrices
WHT performance was very close to that of DCT, and in a Φ̃ ROT-2D-DCT and ROT-2D-WHT, respectively. The ROT
few cases the WHT based matrices slightly outperformed the matrices capture the same components of the signal as the
DCT based ones. corresponding deterministic matrices and the addition of the
An obvious reason for these differences is the fraction of orthonormal transformation Θ should not matter to any of
the signal energy which is captured by the measurements with the reconstruction algorithms. Furthermore, the quantization
each sensing matrix. The DC measurement is the same, up to a step, given by eq. (9), is the same and so are the quantization
10

ized by randomly permuting its entries, but both variants have

a similar effect on the distribution of the measurements: when
a vector is multiplied by either one of these matrices, the distri-
bution of the entries of the result is approximately zero mean
gaussian or a mixture of zero mean gaussians with similar
covariances [13]. Figure 9 shows the effect of multiplication by
Θ̃ on the AC measurements. The measurements generated by
the deterministic matrices have a narrowly peaked distribution
with a long tail, while the measurements generated by the
ROT matrices have a much wider, gaussian like distribution,
with a shorter tail. As a result the entropy coding needs
significantly more bits per measurement for coding the ROT
measurements. The distribution of the measurements generated
by SRM-DCT and SRM-WHT has also a Gaussian like shape,
similar to that of the ROT generated measurements. Therefore,
Fig. 8. Reconstructed image quality vs. file size of “Foreman” image we may conclude that the difference between the SRM to the
with deterministic, SRM and ROT sensing matrices and with GAP-TV
reconstruction. corresponding deterministic matrices has two components: one
component is the different fraction of signal energy captured
by each method, and is represented by the improvement from
the SRM matrices performance to that of the ROT matrices. A
second component relates to the different distributions of the
measurements in the deterministic case and it is represented
by the improvement from the ROT matrices performance to
that of the corresponding deterministic matrices.
One can notice in Figure 9 that the histogram of 2D-WHT
is a little wider than that of 2D-DCT. Thus, the advantage of
2D-DCT over 2D-WHT seems also to be not only because of
the amount of signal energy captured in each case but also
because of differences in the measurements distribution.
The arguments above fail to explain why, with some recon-
struction algorithms, SRM-DCT performs better than SRM-
WHT, since the captured fraction of signal energy, and the
measurements distribution are similar in both cases. Further-
more, they do not explain why this difference is large with
some reconstruction algorithms and hardly noticeable with
others. Thus, there appear to be some subtle interactions
between the prior assumptions used by the reconstruction
algorithm and the type of the sensing matrix.
Fig. 9. Codeword histograms of of “Foreman” for deterministic and ROT
sensing matrices and with compression ratio of 0.1. The distribution of the
codewords in the deterministic case is much tighter than for the ROT matrices, B. Analysis of Joint Quality Control
and it is tighter for 2D-DCT.
The effect of quantization on performance can be seen in
Fig. 10. In general, as the quantization step is decreased, the
noise and digitization noise, given by eq. (8). Nevertheless, the compressed file size and the reconstructed image quality are
performance of the ROT matrices, while significantly better increased. However, the rate of increase is not constant. For
than the SRM matrices, was not as good as the deterministic each compression ratio there seems to be three distinct regions
ones. The difference in SSIM between the performance curves for the quantization step. In the lower quality/data rate region,
of a deterministic matrix and the corresponding ROT matrix decreasing the quantization step results in a significant gain in
was about a third or a quarter of the difference between the quality with little increase in the compressed file size; in the
performance curves of the same deterministic matrix and the higher quality/data rate region the opposite is true: decreasing
corresponding SRM matrix (see example in Fig. 8). Thus, the quantization step increases the file size significantly with
multiplying the measurements vector Φx by Θ, causes a hardly any quality gain; and in the narrow intermediate region
significant performance degradation. By its definition, Θ̃ is between those two we get moderate increase in quality and
a DCT based SRM with a compression ratio of 1, which uses file size as the quantization step is decreased. The intermediate
local randomization, that is, the input signal is randomized by region occurs at higher quantization step for lower compres-
randomly toggling the signs of its entries. This randomization sion ratios. As is clear from Fig. 10, for any required quality
method is different from the global randomization used in the or file size, the best combination of compression ratio and
SRM-DCT sensing matrix, where the input signal is random- quantization step is the one at which the constraint is met when
 11

entropy coding), there is no meaningful “compressed image

file”. Instead, we draw curves of quality, with and without
quantization, vs. the compression ratio (CSr). In Fig. 11 we
plot the SSIM of reconstructed images processed by CSbIC
(including quantization), with that of images reconstructed
directly from the unquantized measurements vectors. As ex-
pected, the results without quantization are consistently better
than the results with quantization. In SSIM terms, the total
decoding error is the difference between 1, which represents
perfect reconstruction, and the SSIM value of the image de-
coded by the CSbIC. The error component due to quantization
is the gap between curves with and without quantization. This
gap varies with the reconstruction algorithms and compression
ratio, both in absolute terms and relative to the total decoding
error. GAP-TV seems to be not that sensitive to quantization
error, with a SSIM decrease of 0.02–0.07 due to quantization.
NLR-CS seems to be the most sensitive, with SSIM decrease
of 0.04–0.10. The behavior of D-AMP is not uniform: at
Fig. 10. Reconstructed image quality vs. file size for various combinations
compression ratios below 0.10 the SSIM decrease is similar to
of compression ratio and quantization step (Barbara, 2D-WHT, GAP-TV). NLR-CS, but at higher compression ratios the SSIM difference
Each curve corresponds to a particular compression ratio {0.02, 0.03, 0.04, decreases, down to 0.01–0.04 at a compression ratio of 0.25.
0.05, 0.07, 0.10, 0.15, 0.20} and the points (markers) on each curve represent
different quantization steps.

E. Comparison with JPEG and JPEG 2000

the quantization step is at the transition region. We determined CSbIC was tested with various sensing matrices and various
empirically that this is satisfied when the compression ratio reconstruction algorithms. As might be expected, the best
and quantization step satisfy (9) While the optimal value for combination depends on the particular image as well as on the
the constant in the right hand side varies from picture to quality or data rate operating point. However, in general the
picture, we found that a constant of C = 2.0 gave good results best performance was achieved with 2D-DCT+NLR-CS or 2D-
for all images and we used it in our experiments. WHT+D-AMP. In almost all test images CSbIC is decidedly
better than JPEG at low bit rates (Fig. 4). A visual inspection
C. Effect of the Reconstruction Algorithm shows (Fig. 6) that at low bit rates JPEG images have blocking
artifacts, while in CSbIC the artifacts are more subtle.
We found that the more modern reconstruction algorithms,
In half of the images CS compression is better than JPEG
NLR-CS and D-AMP, give a better performance than GAP-
for all quality levels. In the rest of the images JPEG is better
TV (Fig. 4). In most cases NLR-CS is a little better than
for the higher quality cases, with the crossover happening at
D-AMP, especially at low-quality/low bit rate settings. We
SSIM of 0.7–0.9 (0.8–0.9 for two out of the four images).
note however, that both NLR-CS and D-AMP require careful
manual parameter tuning and data normalization in order to At this stage, the performance CSbIC is inferior to that of
perform well, so it is possible that the results would be JPEG2000, but at the low quality regions, below about 0.7
somewhat different with better tuning. SSIM, their performance is quite similar (Fig. 4).
The performance gains of D-AMP, and NLR-CS come at the
cost of a much higher complexity. The reconstruction time of V. D ISCUSSION
GAP-TV, D-AMP and NLR-CS is 0.6, 35 and 147 seconds
receptively, with the same compression rate on the same Intel Both JPEG and CS leverage the inherent compressibility of
i7 CPU with 24G memory. The algorithms were implemented natural images, but in different ways: In JPEG this is done at
in Matlab, but it appears that even with optimized implemen- the encoder — the 2D DCT decomposition on 8x8 pixel blocks
tation the last two algorithms would run considerably slower essentially represents the signal according to a sparsity basis;
than GAP-TV, because they are inherently more complex. in classical CS the compressibility assumption comes into play
only at the reconstruction. JPEG also exploits another property
of natural images, namely that the high magnitude coefficients
D. Effect of Quantization are the low frequency ones, therefore JPEG allocates more bits
In order to get a better insight into the performance of our to their accurate representation. This domain-specific insight
system, we isolate the component of the error which is due to was built into CSbIC by using the deterministic sensing
quantization by running tests where the measurements vectors matrices described in Sec. II-A, which improved performance
are reconstructed without quantization. Of course, in this case significantly in comparison to the SRMs of classical CS.
we cannot draw the usual performance curves of SSIM vs. JPEG achieves lower bit rates solely through quantization,
compressed image file size, because without quantization (and while CSbIC does it by quantization, and in addition, by
12

Barbara Boats Cameraman Foreman

House Lenna Monarch Parrots

Fig. 11. SSIM vs. compression ratio (CSr), comparing the reconstructed image with quantization (dash lines) and without quantization (solid lines) for
different reconstruction algorithms.

discarding most of the transform coefficients during the mea- R EFERENCES

surements selection. This is possible because CS reconstruc-
tion recovers the missing coefficients by invoking various [1] “CS vs. JPEG,” https://sites.google.com/site/eiexyuan/csvsjpeg, ac-
cessed: 2017-02-24.
structural assumptions about the original image. This gives [2] E. J. Candès, J. Romberg, and T. Tao, “Robust uncertainty principles:
CSbIC a significant advantage at low bit rates, where excessive Exact signal reconstruction from highly incomplete frequency informa-
quantization results in poor performance of JPEG. Another tion,” IEEE Transactions on Information Theory, vol. 52, no. 2, pp.
489–509, February 2006.
possible explanation for that is that JPEG processes 8x8 pixel [3] D. L. Donoho, “Compressed sensing,” IEEE Transactions on Informa-
blocks nearly independently. At low bit rates, only a small tion Theory, vol. 52, no. 4, pp. 1289–1306, April 2006.
number of DCT coefficients are adequately represented in [4] M. Elad, Sparse and Redundant Representations: From Theory to Ap-
plications in Signal and Image Processing, 1st ed. Springer Publishing
each JPEG block, which may be too little for blocks with Company, Incorporated, 2010.
much detail and causing the blocking artifacts which were [5] C. Li, W. Yin, H. Jiang, and Y. Zhang, “An efficient augmented
observed. In contrast, all CS measurements contribute equally lagrangian method with applications to total variation minimization,”
Computational Optimization and Applications, vol. 56, no. 3, pp. 507–
to the description of the whole image. 530, 2013.
CSbIC yields better performance with certain types of [6] W. Dong, G. Shi, X. Li, Y. Ma, and F. Huang, “Compressive sensing
sensing matrices and reconstruction algorithms. However, the via nonlocal low-rank regularization,” IEEE Transactions on Image
Processing, vol. 23, no. 8, pp. 3618–3632, 2014.
less successful candidates also performed decently, and they [7] C. A. Metzler, A. Maleki, and R. G. Baraniuk, “From denoising to
may be the preferred choice in some situations. Binary-valued compressed sensing,” IEEE Transactions on Information Theory, vol. 62,
sensing matrices may be preferred for capturing measurements no. 9, pp. 5117–5144, Sept 2016.
[8] V. K. Goyal, A. K. Fletcher, and S. Rangan, “Compressive sampling and
in the analog domain, and GAP-TV reconstruction may be lossy compression,” IEEE Signal Processing Magazine, vol. 25, no. 2,
useful when speed is important, for example, for quick image pp. 48–96, 2008.
browsing. [9] J. N. Laska and R. G. Baraniuk, “Regime change: Bit-depth versus
measurement-rate in compressive sensing,” IEEE Transactions on Signal
JPEG is the culmination of many years of research into Processing, vol. 60, no. 7, pp. 3496–3505, July 2012.
the properties of natural images. Wherever we could we [10] J. N. Laska, P. T. Boufounos, M. A. Davenport, and R. G. Baraniuk,
leveraged this body of knowledge in CSbIC. However, much “Democracy in action: Quantization, saturation, and compressive sens-
ing,” Applied and Computational Harmonic Analysis, vol. 31, no. 3, pp.
improvement can still be made, both in the quantization and 429 – 443, 2011.
entropy coding aspects of the system and in its CS aspects. The [11] A. Zymnis, S. Boyd, and E. Candes, “Compressed sensing with quan-
latter includes the sensing matrix design, the reconstruction tized measurements,” IEEE Signal Processing Letters, vol. 17, no. 2, pp.
149–152, Feb 2010.
algorithm, and the derivation of theoretical guarantees for
[12] J. N. Laska, Z. Wen, W. Yin, and R. G. Baraniuk, “Trust, but verify: Fast
successful reconstruction. and accurate signal recovery from 1-bit compressive measurements,”
Signal Processing, IEEE Transactions on, vol. 59, no. 11, pp. 5289–
5301, Nov. 2011.
 13

[13] R. Haimi-Cohen and Y. M. Lai, “Compressive measurements generated of multipliers,” Found. Trends Mach. Learn., vol. 3, no. 1, pp. 1–122,
by structurally random matrices: Asymptotic normality and quantiza- January 2011.
tion.” Signal Processing, vol. 120, pp. 71–87, 2016. [40] M. A. Figueiredo, J. M. Bioucas-Dias, and R. D. Nowak, “Majoriza-
[14] W. Dai, H. V. Pham, and O. Milenkovic, “A comparative study of tionminimization algorithms for wavelet-based image restoration,” IEEE
quantized compressive sensing schemes,” in 2009 IEEE International Transactions on Image Processing, vol. 16, no. 12, pp. 2980–2991, 2007.
Symposium on Information Theory, June 2009, pp. 11–15. [41] Y. Wang, J. Yang, W. Yin, and Y. Zhang, “A new alternating minimiza-
[15] ——, “Distortion-rate functions for quantized compressive sensing,” in tion algorithm for total variation image reconstruction,” SIAM Journal
Networking and Information Theory, 2009. ITW 2009. IEEE Information on Imaging Sciences, vol. 1, no. 3, pp. 248–272, 2008.
Theory Workshop on, June 2009, pp. 171–175. [42] X. Liao, H. Li, and L. Carin, “Generalized alternating projection for
[16] Y. Baig, E. M. K. Lai, and J. P. Lewis, “Quantization effects on weighted-`2,1 minimization with applications to model-based compres-
compressed sensing video,” in Telecommunications (ICT), 2010 IEEE sive sensing,” SIAM Journal on Imaging Sciences, vol. 7, no. 2, pp.
17th International Conference on, April 2010, pp. 935–940. 797–823, 2014.
[17] D. Venkatraman and A. Makur, “A compressive sensing approach to [43] M. Aharon, M. Elad, and A. Bruckstein, “K-SVD: An algorithm for
object-based surveillance video coding,” in 2009 IEEE International designing overcomplete dictionaries for sparse representation,” IEEE
Conference on Acoustics, Speech and Signal Processing, April 2009, Transactions on Signal Processing, vol. 54, no. 11, pp. 4311–4322, 2006.
pp. 3513–3516. [44] E. Candes, M. Wakin, and S. Boyd, “Enhancing sparsity by reweighted
[18] Information Technology – Digital Compression And Coding Of Con- `1 minimization,” Journal of Fourier Analysis and Applications, vol. 14,
tinuous Tone Still Images – Requirements And Guidelines, “CCITT no. 5, pp. 877–905, 2008.
Recommendation T.81,” 1992. [45] K. Dabov, A. Foi, V. Katkovnik, and K. Egiazarian, “Image denoising by
[19] Information technology – JPEG 2000 image coding system: Core coding sparse 3d transform-domain collaborative filtering,” IEEE Transactions
system, “ITU-T Recommendation T.800,” 2002. on Image Processing, vol. 16, no. 8, pp. 2080–2095, August 2007.
[20] M. F. Duarte, M. A. Davenport, D. Takhar, J. N. Laska, T. Sun, [46] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image
K. F. Kelly, and R. G. Baraniuk, “Single-pixel imaging via compressive quality assessment: From error visibility to structural similarity,” IEEE
sampling,” IEEE Signal Processing Magazine, vol. 25, no. 2, pp. 83–91, Transactions on Image Processing, vol. 13, no. 4, pp. 600–612, 2004.
2008.
[21] G. Huang, H. Jiang, K. Matthews, and P. Wilford, “Lensless imaging
by compressive sensing,” pp. 2101–2105, Sept 2013.
[22] J. Romberg, “Imaging via compressive sampling,” IEEE Signal Process-
ing Magazine, vol. 25, no. 2, pp. 14–20, 2008.
[23] X. Yuan, H. Jiang, G. Huang, and P. Wilford, “Lensless compressive
imaging,” arXiv:1508.03498, 2015.
[24] ——, “SLOPE: Shrinkage of local overlapping patches estimator for
lensless compressive imaging,” IEEE Sensors Journal, vol. 16, no. 22,
pp. 8091–8102, November 2016.
[25] J. Bioucas-Dias and M. Figueiredo, “A new TwIST: Two-step iterative
shrinkage/thresholding algorithms for image restoration,” IEEE Trans-
actions on Image Processing, vol. 16, no. 12, pp. 2992–3004, December
2007.
[26] M. A. T. Figueiredo, R. D. Nowak, and S. J. Wright, “Gradient projection
for sparse reconstruction: Application to compressed sensing and other
inverse problems,” IEEE Journal of Selected Topics in Signal Processing,
pp. 586–597, Dec. 2007.
[27] S. Ji, Y. Xue, and L. Carin, “Bayesian compressive sensing,” IEEE
Transactions on Signal Processing, vol. 56, no. 6, pp. 2346–2356, June
2008.
[28] X. Yuan, “Generalized alternating projection based total variation mini-
mization for compressive sensing,” in 2016 IEEE International Confer-
ence on Image Processing (ICIP), Sept 2016, pp. 2539–2543.
[29] J.-H. Ahn, “Compressive sensing and recovery for binary images,” IEEE
Transactions on Image Processing, vol. 25, no. 10, pp. 4796–4802, 2016.
[30] W. K. Pratt, J. Kane, and H. C. Andrews, “Hadamard transform image
coding,” Proceedings of the IEEE, vol. 57, no. 1, pp. 58–68, Jan 1969.
[31] M. F. Duarte and R. G. Baraniuk, “Kronecker compressive sensing,”
IEEE Transactions on Image Processing, vol. 21, no. 2, pp. 494–504,
Feb 2012.
[32] B. J. Fino and V. R. Algazi, “Unified matrix treatment of the fast
walsh-hadamard transform,” IEEE Transactions on Computers, vol. C-
25, no. 11, pp. 1142–1146, Nov 1976.
[33] E. J. Candes and T. Tao, “Near-optimal signal recovery from random
projections: Universal encoding strategies?” IEEE Trans. Inf. Theor.,
vol. 52, no. 12, pp. 5406–5425, Dec. 2006.
[34] ——, “Decoding by linear programming,” IEEE Transactions on Infor-
mation Theory, vol. 51, no. 12, pp. 4203–4215, Dec 2005.
[35] T. T. Do, L. Gan, N. H. Nguyen, and T. D. Tran, “Fast and efficient com-
pressive sensing using structurally random matrices,” IEEE Transactions
on Signal Processing, vol. 60, no. 1, pp. 139–154, Jan 2012.
[36] T. T. Do, T. D. Tran, and L. Gan, “Fast compressive sampling with
structurally random matrices,” in 2008 IEEE International Conference
on Acoustics, Speech and Signal Processing, March 2008, pp. 3369–
3372.
[37] A. Gersho and R. M. Gray, Vector Quantization and Signal Compression.
Norwell, MA, USA: Kluwer Academic Publishers, 1991.
[38] G. G. Langdon, “Arithmetic coding,” IBM J. Res. Develop, vol. 23, pp.
149–162, 1979.
[39] S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein, “Distributed
optimization and statistical learning via the alternating direction method