All articles published by MDPI are made immediately available worldwide under an open access license. No special
permission is required to reuse all or part of the article published by MDPI, including figures and tables. For
articles published under an open access Creative Common CC BY license, any part of the article may be reused without
permission provided that the original article is clearly cited. For more information, please refer to
https://www.mdpi.com/openaccess.
Feature papers represent the most advanced research with significant potential for high impact in the field. A Feature
Paper should be a substantial original Article that involves several techniques or approaches, provides an outlook for
future research directions and describes possible research applications.
Feature papers are submitted upon individual invitation or recommendation by the scientific editors and must receive
positive feedback from the reviewers.
Editor’s Choice articles are based on recommendations by the scientific editors of MDPI journals from around the world.
Editors select a small number of articles recently published in the journal that they believe will be particularly
interesting to readers, or important in the respective research area. The aim is to provide a snapshot of some of the
most exciting work published in the various research areas of the journal.
Discrete wavelet transforms are widely used in signal processing, data compression and spectral analysis. For discrete data with finite sizes, one always pads the data with zeros or extends the data into periodic data before performing the discrete periodic wavelet transform. Due to discontinuity on the boundaries of the original data, the obtained wavelet coefficients always decay slowly, leading to data compression ratios that are significantly lower. In order to solve this issue, in this study, we coupled polynomial fitting into classic discrete periodic wavelet transforms to mitigate these boundary effects.
The application of wavelets in data compression, denoising and time-frequency representation [1,2,3,4,5,6] requires that the wavelets are real-valued, (anti)symmetric and compactly supported. Classic orthogonal wavelets cannot achieve these properties, so the orthogonality in wavelets is relaxed to biorthogonality [7,8,9,10,11,12]. Let and be a pair of real-valued compactly supported biorthonomal wavelets in that are generated by real-valued compactly supported scaling functions and . The frequently used wavelets in higher dimensions just consist of tensor products of one-dimensional wavelets, e.g., we can generate the tensor product of and as follows:
Similarly, taking the tensor products of and , we can obtain , and . Then, and are a pair of two-dimensional biorthonomal wavelets of .
Denote
Any function can be expanded into a biorthonomal wavelet series:
where
If , then wavelet coefficients of any smooth function satisfy [1,2]
The associated wavelet filter banks and are
Then, the one-dimensional wavelet coefficients satisfy the following:
and the two-dimensional wavelet coefficients satisfy
see [1]. Formulas (2) and (3) are the one-dimensional and two-dimensional discrete wavelet transforms (DWTs), respectively [1,2].
Biorthonomal wavelets can be extended to deal with periodic data. The periodization of any function is denoted by
The families generated by the periodization of , , and , namely
are a pair of biorthonomal periodic wavelet bases for [1,2]. Any periodic function can be expanded into a biorthonomal periodic wavelet series:
where
If , then periodic wavelet coefficients of any smooth periodic function satisfy
Noticing that and are compactly supported, there exists a positive integer N such that
We take and then define periodic sequences and such that
For , the one-dimensional periodic wavelet coefficients satisfy the following:
and the two-dimensional periodic wavelet coefficients satisfy
see [1]. Formulas (5) and (6) are one-dimensional and two-dimensional discrete periodic wavelet transforms (DPWTs), respectively [1,2]. Since the convolution of finite length data are always computed approximately by circular convolution in signal processing and data analysis, there is no big difference between (2) and (3) and (5) and (6); hence, the DPWT is considered to be the same as the DWT.
The DWT/DPWT has become a routine fast algorithm in various signal processing fields such as data compression, spectral analysis, denoising, mechanical fault diagnosis, etc. Zeng [13] used two-dimensional DWTs to extract weak information from the gamma spectrum; Alaifari et al. [14] developed a recovery algorithm of square-integrable signals from the absolute values of their wavelet transforms; Lin [15] applied the DWT to the simulation of corrosion fields on buried pipelines; Torrence and Compo [15] developed an empirical formula for the statistical significance testing of DWTs of signals against red noise. Later on, Zhang [16] extended this empirical formula to the generalized case and gave the proof in the rigid statistical framework. Since all data used in this research are of a finite size, we always needed to either pad the data with zeros before performing the DWT or extend the data into periodic data before performing the DPWT. Due to discontinuity on the data boundary, the associated wavelet coefficients always decayed slowly, leading to data compression ratios that were significantly lower and wavelet spectra that were distorted. Although the DWT/DPWT has been widely used in different fields (e.g., [13,14,15,16,17,18]), until now there has been no method to delete or mitigate boundary effects in the DWT/DPWT.
In this study, we coupled simple polynomial fitting with a classic DWT/DPWT to mitigate boundary effects on wavelet coefficients. In Section 2, since our proposed algorithm is complex, we first briefly introduce the simplest version of our improved DWT/DPWT algorithm and show how boundary discontinuities are mitigated in our improvement. In Section 3, we discuss the continuous version of our improvement and show the advantages of our improvement. In Section 4, we establish, step by step, a full version of our improvement of the DWT/DPWT. In Section 5, we give details of some numerical experiments that we used to test our improvement algorithm. In Section 6, we summarize the advantages of our improvement algorithm and discuss some potential applications.
2. Our Improvement of the DWT: The Simplest Version
In order to mitigate boundary effects in DWT/DPWT, we proposed the following algorithm.
First, the discrete data, , are decomposed into three parts:
where
The computation of depends on the values , , and . The computation of depends on the boundary data , , and , which can be further decomposed as follows:
Next, an odd extension, and then a periodic extension, for , , , and are performed. By (2.1) and (2.2), it is clear that for
so we can easily deduce that the above odd and periodic extensions guarantee continuity and differentiability on the data boundary. Finally, one-dimensional and two-dimensional DWTs/DPWTs are performed for these extension data.
Due to the smooth extension in our improvement, by (1) and (4), the decay rate of the obtained wavelet coefficients in our improvement algorithm is faster than that of traditional (periodic) wavelet coefficients. This means that our improvement can compress the data better than the traditional DWT/DPWT. Our proposed algorithm can be further modified to achieve higher compression ratios. The full version of our improved algorithm is summarized at the end of Section 4.
3. Continuous Version of Our Improved Wavelet Algorithm
In order to explain the advantages of our improved algorithm over traditional wavelet algorithms, we established a step by step continuous version of our improved wavelet algorithm.
Denote . If is continuous on the region for all , we say . Let the fundamental polynomial be a univariate polynomial of degree satisfying
where and . Then, can be represented as follows:
where the coefficients, , satisfy
For any function , define the interpolation polynomial of f at the nodes 0 and :
where and represent the integral part. Then, satisfies
Let . Then,
For , an odd extension is performed, which is denoted by . Again, a 1-periodic extension of is performed to obtain . By (10), it follows that , i.e., the univariate function can be decomposed as follows:
where is the interpolation polynomial of f at the nodes 0 and , is a 1-periodic odd function and .
For a bivariate smooth function , denote and . Define as the interpolation polynomial of f at the vertices of :
such that for and ,
Define
Theorem1.
Let . Then, the following decomposition formula holds:
and can be extended into a 1-periodic odd function and .
Proof.
By (9), we know that and . Again, by (11), for ,
By (12), we deduce that for and ,
Therefore, for ,
By (13), for ,
Similarly, we can deduce that for ,
Finally
For r, an odd extension is performed, denoted by ,
From this and (14), it follows that . Again, a periodic extension is performed, denoted by ,
so is a 1-periodic odd function and , and it can be reconstructed well using its two-dimensional periodic wavelet coefficients. □
By (4), the periodic wavelet coefficients of decay as fast as . Compared with this, if we directly compute traditional (periodic) wavelet coefficients of , due to discontinuity on the boundary, the obtained (periodic) wavelet coefficients decay as fast as .
By (12), is determined by four derivative functions of f on the boundary of :
Similarly to the above process, these four functions can be decomposed as
From this, we have
After odd extensions and then 1-periodic extensions for and , we obtain four 1-periodic smooth odd functions . Similarly to the argument in (4), the one-dimensional periodic wavelet coefficients of these four periodic functions decay as fast as . So, can be reconstructed well using the value of the derivative of f on the four vertices of and the one-dimensional periodic wavelet coefficients. By (11), can be reconstructed using the value of the derivative of f on the four vertices of . Therefore, f can be reconstructed well using the value of the derivative of f on the four vertices of and the one-dimensional and two-dimensional fast-decaying periodic wavelet coefficients.
4. Our Improvement of the DWT: Full Version
In this section, the traditional DWT/DPWT is improved so that the impacts of the data boundary on the wavelet coefficients can be mitigated well. The full version of our improved algorithm is stated as follows.
For , assume that the sampling of f on is given by the following:
The derivatives on the boundary of the square are denoted by
Based on (11) and (12), we define and as
The discrete data, , can be decomposed as follows:
By (17)–(19), we have
An odd extension for the data is performed as follows:
where . Following this, the periodic extension for is performed to obtain .
Proposition1.
For a large J, the periodic wavelet coefficients of satisfy
where and is stated in Theorem 1.
Proof.
Comparing (13) and (19), it follows that is just the sampling of . Noticing that , by (2.6), it follows that
Denote the compact support of by . Noticing that and , we have
i.e.,
where
□
The wavelet coefficients can be decomposed as
Theorem2.
Suppose that φ and are symmetric at and ψ and are symmetric at and , respectively. Then periodic wavelet coefficients in (21) are symmetric:
Proof.
Since and are symmetric at and and are symmetric at and , respectively, by (6), it follows that
Based on the DPWT algorithm, by (20), we easily obtain the symmetry property of the wavelet coefficients. □
In (18), is determined by four sequences , , and , which can be decomposed further as follows:
After odd extensions and then 1-periodic extensions for , we obtain four -periodic smooth sequences . Again, by (5), we obtain one-dimensional periodic wavelet coefficients of :
Similar to Theorem 2, the above periodic wavelet coefficients are also symmetric.
Finally, we summarize our improvement of the discrete (periodic) wavelet transform for as follows:
Decomposition Algorithm
Step 1. By (17)–(19), we obtain . After that, an odd extension and a periodic extension are performed for to obtain . Finally, by Proposition 1 and (6), two-dimensional periodic wavelet coefficients are computed in (21). Again, by Theorem 2, the number of nonzero two-dimensional periodic wavelet coefficients that is necessary to store is .
Step 2. By (22), can be obtained. After odd extensions, and then periodic extensions, we obtain four -periodic smooth sequences . Again, by (5), one-dimensional periodic wavelet coefficients are obtained as those in (23). Similar to Theorem 2, due to the symmetric property of periodic wavelet coefficients, the number of nonzero two-dimensional periodic wavelet coefficients that is necessary to store is .
Step 3. The following values are stored:
Reconstruction Algorithm.
Step 1. By (17), using the values,
we obtain .
Step 2. Using the values in Step 1, (22), (23) and the one-dimensional inverse discrete periodic wavelet transform, we obtain
and, by (18), we obtain .
Step 3. By (21) and the two-dimensional inverse discrete periodic wavelet transform, we obtain . This means that can be computed.
Step 4. By (17), we can reconstruct f:
5. Numerical Experiments
We compared our improved DWT with the classic DWT through numerical experiments. The quality of data approximation was measured by the known peak signal-to-noise ratio (PSNR) [17]:
where is the original data, RMSE is the mean squared error between the original and the approximation divided by the square root of the total number of samples.
For the discrete data , we used the simplest version of our algorithm, which meant that no numerical derivative was computed to approximate . In the approximation and reconstruction processes, we retained a certain number of the largest coefficients in terms of energy from all one/two-dimensional wavelet coefficients and then reconstructed the original data from them. Figure 1 shows the quality of the approximations by using DWT and our improvement, where the depth of wavelet decomposition was set to the maximum level. The higher value of the PSNR indicated a better approximation performance. It was clear that the simplified version of our algorithm had a better approximation performance than that of the classic DWT. Noticing that the PSNR value is in terms of logarithmic scale, the smaller the number of retained coefficients, the clearer the performance difference became.
6. Discussion and Conclusions
The discrete wavelet transform and the discrete periodic wavelet transform is widely used in signal processing, data compression and wireless communication. Due to discontinuity on the boundary of the original data, the decay rate of the obtained wavelet coefficients was slow. In this study, we introduced a combination of polynomial interpolation and one-dimensional/two-dimensional discrete periodic wavelet transforms to mitigate these boundary effects. Due to the smooth extension in our algorithm, the one-dimensional and two-dimensional periodic wavelet coefficients in our algorithm decayed as fast as . Compared with this, if we applied the traditional discrete (periodic) wavelet transform, due to discontinuity on the boundary, the corresponding wavelet coefficients decayed as fast as . Even if we considered the simplest version of our algorithm (i.e., taking in (17) and (18), which meant that no numerical derivative was computed), the corresponding wavelet coefficients in our algorithm still decayed as fast as , which was faster than that of the traditional discrete (periodic) wavelet transform (i.e., ). Moreover, the numerical experiment also demonstrated that the simplest version of our algorithm could compress the data much better than the traditional discrete wavelet transform. The full version of our algorithm could be applied to compress smoother data, e.g., CMIP6 data [13]. The size of CMIP6 data increases sharply at the petabyte scale [19], but, even now, there is still no good algorithm to compress it. Since CMIP6 data are output data from numerical solutions of fluid equations and energy equations governing the Earth’s climate system [18], CMIP6 data is smooth and the derivative values on the data boundary are easily estimated, so our algorithm could compress this kind of data well.
Funding
This research was partially supported by European Commissions Horizon 2020 Framework Program No. 861584 and Taishan Distinguished Professor Fund.
Data Availability Statement
Not applicable.
Acknowledgments
The author thanks the reviewers and editors for their valuable comments.
Conflicts of Interest
The author declare no conflict of interest.
References
Mallat, S. A Wavelet Tour of Signal Processing; Academic Press: Cambridge, MA, USA, 2008. [Google Scholar]
Zhang, Z.; Jorgenson, P.E.T. Frame Theory in Data Science; Springer: Heidelberg, Germany, 2021. [Google Scholar]
Jha, C.K.; Kolekar, M.H. Tunable Q-wavelet based ECG data compression with validation using cardiac arrhythmia patterns. Biomed. Signal Process. Control2021, 66, 102464. [Google Scholar] [CrossRef]
Otsuka, A.; Muramatsu, K.; Nagata, F. Data-driven generation of random skin model shapes by using wavelet transformation. Comput. Ind. Eng.2020, 150, 106860. [Google Scholar] [CrossRef]
Chakraborty, J.; Nandy, A. Discrete wavelet transform based data representation in deep neural network for gait abnormality detection. Biomed. Signal Process. Control2020, 62, 102076. [Google Scholar] [CrossRef]
Kamgar, R.; Dadkhah, M.; Naderpour, H. Seismic response evaluation of structures using discrete wavelet transform through linear analysis. Structures2021, 29, 863–882. [Google Scholar] [CrossRef]
Deriaz, E.; Farge, M.; Schneider, K. Craya decomposition using compactly supported biorthogonal wavelets. Appl. Comput. Harmon. Anal.2010, 28, 267–284. [Google Scholar] [CrossRef] [Green Version]
Chen, Q.; Cao, H.; Shi, Z. Construction and decomposition of biorthogonal vector-valued wavelets with compact support Chaos. Solitons Fractals2009, 42, 2765–2778. [Google Scholar] [CrossRef]
Atreasa, N.; Karantzasb, N.; Papadakisb, M.; Stavropoulosc, T. On the design of multi-dimensional compactly supported Parseval framelets with directional characteristics. Linear Algebra Appl.2019, 582, 1C36. [Google Scholar] [CrossRef] [Green Version]
Dutkay, D.E.; Jorgensen, P.E.T.; Silvestrov, S. Decomposition of wavelet representations and Martin. J. Funct. Anal.2012, 262, 1043–1061. [Google Scholar] [CrossRef]
Averbuch, A.Z.; Neittaanmäki, P.; Zheludev, V.A. Spline and Spline Wavelet Methods with Applications to Signal and Image Processing; Springer: Berlin/Heidelberg, Germany, 2016. [Google Scholar]
Zeng, C.; Lai, W.; Lin, H.; Liu, G.; Qin, B.; Kang, Q.; Feng, X.; Yu, Y.; Gu, R.; Wu, J.; et al. Weak information extraction of gamma spectrum based on a two-dimensional wavelet transform. Radiat. Phys. Chem.2023, 208, 110914. [Google Scholar] [CrossRef]
Alaifari, R.; Bartolucci, F.; Wellershoff, M. Phase retrieval of bandlimited functions for the wavelet transform. Appl. Comput. Harmon. Anal.2023, 64, 102–117. [Google Scholar] [CrossRef]
Lin, J.; Zhou, W.; Cui, X.Z.; Hong, H.P. Application of wavelet transforms to the simulation of corrosion fields on buried pipelines. Comput. Struct.2023, 276, 106957. [Google Scholar] [CrossRef]
Torrence, C.; Compo, G.P. A practical guide to wavelet anal- ysis. Bull. Am. Meteorol. Soc.1998, 79, 61–78. [Google Scholar] [CrossRef]
Najmi, A.; Moon, T. Advanced Signal Processing; McGraw-Hill: New York, NY, USA, 2020. [Google Scholar]
Eyring, V.; Bony, S.; Meehl, G.A.; Senior, C.A.; Stevens, B.; Stouffer, R.J.; Taylor, K.E. Overview of the Coupled Model Intercomparison Project Phase 6 (CMIP6) experimental design and organization. Geosci. Model Dev.2016, 9, 1937–1958. [Google Scholar] [CrossRef] [Green Version]
Reichstein, M.; Camps-Valls, G.; Stevens, B.; Jung, M.; Denzler, J.; Carvalhais, N. Deep learning and process understanding for data-driven Earth system science. Nature2019, 566, 195C204. [Google Scholar] [CrossRef]
Figure 1.
Approximation quality of the discrete wavelet transform and our improved algorithm.
Figure 1.
Approximation quality of the discrete wavelet transform and our improved algorithm.
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.