Abstract: The neural network quantization is highly desired
procedure to perform before running neural networks on mobile
devices. Quantization without fine-tuning leads to accuracy drop of
the model, whereas commonly used training with quantization is done
on the full set of the labeled data and therefore is both time- and
resource-consuming. Real life applications require simplification and
acceleration of quantization procedure that will maintain accuracy of
full-precision neural network, especially for modern mobile neural
network architectures like Mobilenet-v1, MobileNet-v2 and MNAS. Here we present a method to significantly optimize training with
quantization procedure by introducing the trained scale factors for
discretization thresholds that are separate for each filter. Using the
proposed technique, we quantize the modern mobile architectures of
neural networks with the set of train data of only ∼ 10% of the
total ImageNet 2012 sample. Such reduction of train dataset size and
small number of trainable parameters allow to fine-tune the network
for several hours while maintaining the high accuracy of quantized
model (accuracy drop was less than 0.5%). Ready-for-use models and
code are available in the GitHub repository.
Abstract: Hammerstein–Wiener model is a block-oriented model
where a linear dynamic system is surrounded by two static
nonlinearities at its input and output and could be used to model
various processes. This paper contains an optimization approach
method for analysing the problem of Hammerstein–Wiener systems
identification. The method relies on reformulate the identification
problem; solve it as constraint quadratic problem and analysing its
solutions. During the formulation of the problem, effects of adding
noise to both input and output signals of nonlinear blocks and
disturbance to linear block, in the emerged equations are discussed.
Additionally, the possible parametric form of matrix operations
to reduce the equation size is presented. To analyse the possible
solutions to the mentioned system of equations, a method to reduce
the difference between the number of equations and number of
unknown variables by formulate and importing existing knowledge
about nonlinear functions is presented. Obtained equations are applied
to an instance H–W system to validate the results and illustrate the
proposed method.
Abstract: The random dither quantization method enables us
to achieve much better performance than the simple uniform
quantization method for the design of quantized control systems.
Motivated by this fact, the stochastic model predictive control
method in which a performance index is minimized subject to
probabilistic constraints imposed on the state variables of systems
has been proposed for linear feedback control systems with random
dither quantization. In other words, a method for solving optimal
control problems subject to probabilistic state constraints for linear
discrete-time control systems with random dither quantization has
been already established. To our best knowledge, however, the
feasibility of such a kind of optimal control problems has not
yet been studied. Our objective in this paper is to investigate the
feasibility of stochastic model predictive control problems for linear
discrete-time control systems with random dither quantization. To
this end, we provide the results of numerical simulations that verify
the feasibility of stochastic model predictive control problems for
linear discrete-time control systems with random dither quantization.
Abstract: Recently, the effectiveness of random dither
quantization method for linear feedback control systems has
been shown in several papers. However, the random dither
quantization method has not yet been applied to nonlinear feedback
control systems. The objective of this paper is to verify the
effectiveness of random dither quantization method for nonlinear
feedback control systems. For this purpose, we consider the attitude
stabilization problem of satellites using discrete-level actuators.
Namely, this paper provides a control method based on the random
dither quantization method for stabilizing the attitude of satellites
using discrete-level actuators.
Abstract: Recently, feedback control systems using random dither
quantizers have been proposed for linear discrete-time systems.
However, the constraints imposed on state and control variables
have not yet been taken into account for the design of feedback
control systems with random dither quantization. Model predictive
control is a kind of optimal feedback control in which control
performance over a finite future is optimized with a performance
index that has a moving initial and terminal time. An important
advantage of model predictive control is its ability to handle
constraints imposed on state and control variables. Based on the
model predictive control approach, the objective of this paper is to
present a control method that satisfies probabilistic state constraints
for linear discrete-time feedback control systems with random dither
quantization. In other words, this paper provides a method for
solving the optimal control problems subject to probabilistic state
constraints for linear discrete-time feedback control systems with
random dither quantization.
Abstract: Detecting changes in multiple images of the same
scene has recently seen increased interest due to the many
contemporary applications including smart security systems, smart
homes, remote sensing, surveillance, medical diagnosis, weather
forecasting, speed and distance measurement, post-disaster forensics
and much more. These applications differ in the scale, nature, and
speed of change. This paper presents an application of image
processing techniques to implement a real-time change detection
system. Change is identified by comparing the RGB representation of
two consecutive frames captured in real-time. The detection threshold
can be controlled to account for various luminance levels. The
comparison result is passed through a filter before decision making to
reduce false positives, especially at lower luminance conditions. The
system is implemented with a MATLAB Graphical User interface
with several controls to manage its operation and performance.
Abstract: Discrete Cosine Transform (DCT) based transform coding is very popular in image, video and speech compression due to its good energy compaction and decorrelating properties. However, at low bit rates, the reconstructed images generally suffer from visually annoying blocking artifacts as a result of coarse quantization. Lapped transform was proposed as an alternative to the DCT with reduced blocking artifacts and increased coding gain. Lapped transforms are popular for their good performance, robustness against oversmoothing and availability of fast implementation algorithms. However, there is no proper study reported in the literature regarding the statistical distributions of block Lapped Orthogonal Transform (LOT) and Lapped Biorthogonal Transform (LBT) coefficients. This study performs two goodness-of-fit tests, the Kolmogorov-Smirnov (KS) test and the 2- test, to determine the distribution that best fits the LOT and LBT coefficients. The experimental results show that the distribution of a majority of the significant AC coefficients can be modeled by the Generalized Gaussian distribution. The knowledge of the statistical distribution of transform coefficients greatly helps in the design of optimal quantizers that may lead to minimum distortion and hence achieve optimal coding efficiency.
Abstract: In this paper, a novel multipurpose audio watermarking
algorithm is proposed based on Vector Quantization (VQ) in Discrete
Cosine Transform (DCT) domain using the codeword labeling and
index-bit constrained method. By using this algorithm, it can fulfill the
requirements of both the copyright protection and content integrity
authentication at the same time for the multimedia artworks. The
robust watermark is embedded in the middle frequency coefficients of
the DCT transform during the labeled codeword vector quantization
procedure. The fragile watermark is embedded into the indices of the
high frequency coefficients of the DCT transform by using the
constrained index vector quantization method for the purpose of
integrity authentication of the original audio signals. Both the robust
and the fragile watermarks can be extracted without the original audio
signals, and the simulation results show that our algorithm is effective
with regard to the transparency, robustness and the authentication
requirements
Abstract: The existing image coding standards generally degrades at low bit-rates because of the underlying block based Discrete Cosine Transform scheme. Over the past decade, the success of wavelets in solving many different problems has contributed to its unprecedented popularity. Due to implementation constraints scalar wavelets do not posses all the properties such as orthogonality, short support, linear phase symmetry, and a high order of approximation through vanishing moments simultaneously, which are very much essential for signal processing. New class of wavelets called 'Multiwavelets' which posses more than one scaling function overcomes this problem. This paper presents a new image coding scheme based on non linear approximation of multiwavelet coefficients along with multistage vector quantization. The performance of the proposed scheme is compared with the results obtained from scalar wavelets.
Abstract: The binary phase-only filter digital watermarking
embeds the phase information of the discrete Fourier transform of the
image into the corresponding magnitudes for better image authentication.
The paper proposed an approach of how to implement watermark
embedding by quantizing the magnitude, with discussing how to
regulate the quantization steps based on the frequencies of the magnitude
coefficients of the embedded watermark, and how to embed the
watermark at low frequency quantization. The theoretical analysis and
simulation results show that algorithm flexibility, security, watermark
imperceptibility and detection performance of the binary phase-only
filter digital watermarking can be effectively improved with quantization
based watermark embedding, and the robustness against JPEG
compression will also be increased to some extent.
Abstract: Although the level crossing concept has been the subject of intensive investigation over the last few years, certain problems of great interest remain unsolved. One of these concern is distribution of threshold levels. This paper presents a new threshold level allocation schemes for level crossing based on nonuniform sampling. Intuitively, it is more reasonable if the information rich regions of the signal are sampled finer and those with sparse information are sampled coarser. To achieve this objective, we propose non-linear quantization functions which dynamically assign the number of quantization levels depending on the importance of the given amplitude range. Two new approaches to determine the importance of the given amplitude segment are presented. The proposed methods are based on exponential and logarithmic functions. Various aspects of proposed techniques are discussed and experimentally validated. Its efficacy is investigated by comparison with uniform sampling.
Abstract: In this paper, image compression using hybrid vector
quantization scheme such as Multistage Vector Quantization
(MSVQ) and Pyramid Vector Quantization (PVQ) are introduced. A
combined MSVQ and PVQ are utilized to take advantages provided
by both of them. In the wavelet decomposition of the image, most of
the information often resides in the lowest frequency subband.
MSVQ is applied to significant low frequency coefficients. PVQ is
utilized to quantize the coefficients of other high frequency
subbands. The wavelet coefficients are derived using lifting scheme.
The main aim of the proposed scheme is to achieve high compression
ratio without much compromise in the image quality. The results are
compared with the existing image compression scheme using MSVQ.
Abstract: This paper presents a new fingerprint coding technique
based on contourlet transform and multistage vector quantization.
Wavelets have shown their ability in representing natural images that
contain smooth areas separated with edges. However, wavelets
cannot efficiently take advantage of the fact that the edges usually
found in fingerprints are smooth curves. This issue is addressed by
directional transforms, known as contourlets, which have the
property of preserving edges. The contourlet transform is a new
extension to the wavelet transform in two dimensions using
nonseparable and directional filter banks. The computation and
storage requirements are the major difficulty in implementing a
vector quantizer. In the full-search algorithm, the computation and
storage complexity is an exponential function of the number of bits
used in quantizing each frame of spectral information. The storage
requirement in multistage vector quantization is less when compared
to full search vector quantization. The coefficients of contourlet
transform are quantized by multistage vector quantization. The
quantized coefficients are encoded by Huffman coding. The results
obtained are tabulated and compared with the existing wavelet based
ones.