TRANSFORMADAS DISCRETAS DE COMPRIMENTOS 8, 16 E 32 COM BAIXO CUSTO ARITMÉTICO PARA APLICAÇÕES DE COMPRESSÃO DE IMAGEM E VÍDEO

Low-complexity discrete transforms of lengths 8, 16, and 32 address the high computational complexity of DCT in JPEG and HEVC by eliminating multiplications, achieving efficient image and video compression with reduced energy consumption and improved coding performance.

BR102018014244B1Active Publication Date: 2026-11-10UNIVE FEDERAL DE SANTA MARIA UFSM +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
BR102018014244
Authority / Receiving Office
BR · BR
Patent Type
Patents
Current Assignee / Owner
Filing Date
2018-07-12
Publication Date
2026-11-10
Estimated Expiration
2038-07-12

AI Technical Summary

Technical Problem

Existing image and video compression technologies, such as JPEG and HEVC, rely on the discrete cosine transform (DCT) which requires floating-point arithmetic and non-trivial multiplications, leading to high computational complexity and energy consumption, especially for highly correlated signals like images.

Method used

A series of low-complexity discrete transforms of lengths 8, 16, and 32 are proposed, based on approximations of the DCT, which eliminate non-trivial multiplications and reduce computational complexity through fast algorithms, using only additions and shifts, and are scalable to larger lengths using the Jridi, Alfalou, and Meher method.

Benefits of technology

The proposed transforms achieve high processing speed, low hardware area consumption, and low energy consumption, while maintaining decorrelation and energy compaction properties, outperforming existing transforms in coding gain and efficiency metrics.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 00000020_0000
    Figure 00000020_0000
  • Figure 00000021_0000
    Figure 00000021_0000
  • Figure 00000021_0001
    Figure 00000021_0001
Patent Text Reader

Abstract

Discrete transforms of lengths 8, 16, and 32 with low arithmetic cost for image and video compression applications. Due to the increasing demand for high data processing rates, the scientific and technological communities have been paying close attention to methodologies that require low computational cost. One way to reduce computational cost in signal processing is to use transforms that have zero multiplicative cost. In the present invention, three new transforms of lengths 8, 16, and 32 are proposed. These transforms are approximations of the discrete cosine transform (dct) and are very useful in image and video compression, such as in the JPEG, MPEG-1, MPEG-2, H.261, H.263, H.264, and more recently, HEVC standards. The proposed scheme is based on an orthogonal approximation for the 8-length dct based on the series of approximations proposed by Ahmed, Bouguezel, and Swamy.The new 8-length transform was scaled up by the Jridi, Alfalou, and Meher method for lengths 16 and 32. Fast algorithms for these new approximations are also introduced. The proposed methods were evaluated using similarity metrics with the DCT and encoding measures, as well as image compression experiments, which establishes their good performance in applications for this purpose.
Need to check novelty before this filing date? Find Prior Art

Description

1 / 17 Discrete Transforms of Lengths 8, 16, and 32 with Low Arithmetic Cost for Image and Video Compression Applications FIELD OF APPLICATION

[001] The present invention has as its applicability the compression of digital data based on block transform. The image and video encoding system of this invention is similar to that used in the JPEG and HEVC standards. To use this invention, it is sufficient to replace the commonly used block transform in these compression standards with the proposed invention. The invention provides great decorrelation power for signals modeled by a first-order Markovian process. An example of data that fits this type of model is stationary images. The present invention has low arithmetic cost, which enhances its application in systems that require the processing of massive amounts of data, severe resource limitations, and power limitations. BACKGROUND OF THE INVENTION

[002] In the context of signal processing, the discrete cosine transform (DCT)

[44] is a central tool. In particular, in the context of image and video coding, the discrete cosine transform is fundamental [45, 46, 47]. The DCT has been widely used, as in the JPEG

[48] , H.264

[49] and HEVC

[50] standards. The Karhunen-Loève transform (KLT) is a transform that has optimal energy compression properties

[51] , with the DCT being an asymptotic approximation of the KLT. In this context, there is great interest in using the DCT, because, unlike the KLT, the DCT does not depend on the input signal

[45] .

[003] Nowadays, the transmission of large amounts of data is remarkable, therefore, this type of transmission has been deserving great attention. Thus, when applying DCT, a Petition 870260049988, dated 05 / 26 / 2026, page 3 / 30 2 / 17 Changes in the processing stage can significantly affect the speed and energy consumption of the process [52, 53]. When the data follows a highly correlated first-order Markovian process, as in the case of images, DCT is applied to blocks of the input signal for greater computational efficiency. However, DCT computation can become even more efficient with the use of fast algorithms, which cannot be applied to KLT due to their dependence on the input signal

[45] . For the calculation of fast DCT algorithms, there are several options in the literature [54, 55, 56, 57].

[004] However, even with the use of fast algorithms and the block system, we still have floating-point arithmetic for the DCT calculation, due to non-trivial multiplications that are required in its computation

[45] . Thus, alternatively, transforms that are approximations of the DCT can be used. If the input values ​​of such transforms are dyadic rationals, approximations free of non-trivial multiplications will be obtained, as well as in approximations based on lifting schemes

[55] . Alternatively, input values ​​in {ü, ± 1, ±1, ±2} can be used, thus being free of non-trivial multiplications [59, 60, 61, 62, 63, 64]. With this, the computation of these approximate transforms will involve only bit additions and shifts. In the literature, there are several works that address these approximations [60, 61, 63, 64, 65, 66, 67, 68, 69, 10, 71, 72, 73, 74].According to figures of merit relevant to the problem, such transforms are considered close to the DCT. Instead of providing the exact computation of the DCT, they offer an approximation at a very low computational cost.

[005] The most traditional zero-multiplication complexity transform in the literature for image processing is the Walsh-Hadamard transform (WHT)

[75] , however it has not been proposed as an approximation of DCT. The first approximation proposed for DCT was the signaled DCT (SDCT)

[59] . SDCT is obtained from the Petition 870260049988, dated 05 / 26 / 2026, page 4 / 30 3 / 17 application of the sign function to each element of the exact DCT matrix. Computing the SDCT of length 8 with a fast algorithm requires 24 additions. The SDCT is a non-orthogonal transform, but it has a simple inverse and its computation with the fast algorithm proposed in

[76] requires 20 additions and 4 bit shifts. Thus, the SDCT, even though non-orthogonal, has an inverse with low computational complexity.

[006] After SDCT was proposed, Lengwehasatit-Ortega

[77] proposed a series of low-complexity DCT approximations. Among these approximations, the “level 1 approximation” stands out as a transform of length 8 with zero multiplicative complexity. This transform is orthogonal, that is, its inverse is equal to its transpose, requiring 24 additions and 2 bit shifts for its computation. This transform has good performance in terms of peak signal-to-noise ratio (PSNR) in image compression.

[007] A group of computationally low-complexity transforms of length 8 are the transforms proposed by Bouguezel-Ahmed-Swamy (BAS) [63, 64, 71, 72, 73, 74]. These transforms are characterized by not being very close to DCT in terms of mean squared error

[45] and total energy error

[61] , but they have high unified coding gain

[78] and efficiency

[79] , making them suitable for image compression applications. In addition to the transforms of length 8, the transform proposed in

[74] was scaled to lengths 16 and 32. In

[64] a general expression was given for the proposed transform in order to obtain its scaling for any length N. This group of transforms, in general, consists of variations of other low-complexity transforms from the literature, mainly SDCT.

[008] The rounded DCT (RDCT)

[61] is a length 8 transform that is obtained by applying a rounding function to the exact DCT. A Petition 870260049988, dated 05 / 26 / 2026, page 5 / 30 4 / 17 The computational complexity of its fast algorithm is 22 additions. The modified RDCT (MRDCT)

[60] is a transform obtained from the convenient modification of the RDCT and requires only 14 additions for its implementation. The MRDT is the transform with the lowest computational complexity found in the literature.

[009] Another approximation of the DCT is a transform created for application in radio-frequency images

[65] , obtained from the minimum argument of the total energy error, being the result of an optimization problem. In

[80] , a transform is proposed that is a modification of the matrix proposed by Brahimi and Bouguezel

[81] , this transform requires only 16 additions for its computation and has good measures of coding gain and efficiency in relation to its low arithmetic cost.

[010] Applying integer functions to the input values ​​of the exact DCT of length 8 resulted in a class of discrete transforms of low computational complexity

[70] . This class considers a search parameter that varies within a given range and this generates a large number of new transforms. By introducing the orthogonality and quasi-orthogonality constraints, and also considering the absence of one of these two constraints, the inverse transforms are simple and have low complexity. In this way, the large number of transforms generated by the method is reduced to a few new transforms with good properties.

[011] Based on the parameterization of the fast Feig-Winograd algorithm

[57] , in

[62] a new class of approximate DCT transforms of length 8 was obtained, being a parameter-dependent matrix subspace. This method consists of performing a computational search considering that the parameters of this matrix subspace have only values, {ü, ±2, ±1, ±2}. Thus, a large number of new transforms are found which are subjected to a multicriteria optimization problem [82, 83]. In this way, optimal transforms are found according to a set of distance and quality metrics, which includes the following measures: mean squared error

[45] , Petition 870260049988, dated 05 / 26 / 2026, page 6 / 30 5 / 17 total energy error

[61] , unified encoding gain

[78] and efficiency

[79] .

[012] Considering the minimization of angular measures, in

[84] a new orthogonal transform of length 8 is proposed. This transform has an arithmetic cost of 24 additions and 6 bit shifts. This transform was scaled to lengths 16 and 32 by the method of Jridi, Alfalou and Meher (JAM)

[85] , and their computations require 64 and 160 additions, respectively.

[013] JAM's method

[85] is a scaling method for transforms of smaller lengths to larger lengths. By this method it is possible to obtain transforms of length 2N, from transforms of length N. When proposing the method in

[85] , the authors also scale the RDCT transform to lengths 16 and 32, which require 60 and 152 additions, respectively.

[014] Other 16-length transforms are given in [68, 76, 87]. Replacing the DCT input values ​​with values ​​from the set {0, ±1} respecting orthogonality, symmetry with respect to DCT and good energy compaction properties,

[68] proposed an approximation for a 16-length DCT that requires 72 additions for its computation. The transform proposed in

[87] is obtained from integer approximations. This transform is orthogonal and requires 60 additions for its computation. In

[86] , a transform is proposed using a methodology similar to MRDCT and, to our knowledge, is the 16-length transform with the lowest arithmetic complexity in the literature, requiring only 44 additions for its computation. SUMMARY OF THE INVENTION

[015] The methodology for obtaining the proposed invention aims at a length 8 transform that approximates the DCT based on the transforms proposed by Bouguezel-Ahmad-Swamy [63, 64, 71, 72, 73, 74]. The approximate transform Petition 870260049988, dated 05 / 26 / 2026, page 7 / 30 The 6 / 17 found was scaled to lengths of 16 and 32 by the method of Jridi, Alfalou and Meher

[85] . The present invention requires a low arithmetic cost for image and video compression and is free of non-trivial multiplications. In this way, it is proposed to replace DCT in the JPEG and HEVC standards with the invention proposed here. Thus, the device implemented from the proposed invention will have high processing speed, low area consumption in hardware implementation and low energy consumption. BRIEF DESCRIPTION OF THE FIGURES

[016] Figure 1: Proposed low-complexity T8 matrix of length 8.

[017] Figure 2: Permutation matrix A1 of the proposed fast algorithm.

[018] Figure 3: Matrix A2 of the proposed fast algorithm.

[019] Figure 4: K-matrix of the proposed fast algorithm.

[020] Figure 5: Permutation matrix P of the fast algorithm.

[021] Figure 6: Proposed low-complexity T16 matrix of length 16. ΓηηΛΐ ΤΊ · Λ Γ , · ΓΤΊ 11- 1-111 · j Ο Π j

[022] Figure 7: Proposed low-complexity T32 matrix of length 32.

[023] Figure 8: Image compression experiment for different compression levels, with r coefficients retained, considering DCT and the proposed 8-length transform (C8).

[024] Figure 9: Image compression experiment for different compression levels, with r coefficients retained, considering DCT and the proposed length 16 (Cιβ)·

[025] Figure 10: Image compression experiment for different compression levels, with r coefficients retained, considering DCT and the proposed length 32 (C32) transform.

[026] Figure 11: Fast C8 transform algorithm in flowchart form Petition 870260049988, dated 05 / 26 / 2026, page 8 / 30 7 / 17 and sign decomposing the matrix T8, relating the input data xn, n = 0,1, 2,..., 7, to the output coefficients Xk, k = 0,1,2,..., 7. Dashed arrows represent multiplication by -1.

[027] Figure 12: Fast Cie transform algorithm in the form of a signal-flow diagram decomposing the Tie matrix, relating the input data xn, n = 0,1, 2,..., 15, to the output coefficients Xk, k = 0,1, 2,..., 15. Dashed arrows representing multiplication by -1.

[028] Figure 13: Fast C32 transform algorithm in the form of a signal-flow diagram decomposing the T32 matrix, relating the input data xn, n = 0,1, 2,..., 31, to the output coefficients Xk, k = 0,1, 2,..., 31. Dashed arrows representing multiplication by -1. DETAILED DESCRIPTION

[029] Starting from a low-complexity matrix of length 8 (T8), an approximate transform (C8), also of length 8, is obtained. The matrix T8 has only elements from the set {0, ±·, ±1, ±2}. To obtain C8, the polar decomposition method

[88] is applied to T8, and then a diagonal adjustment matrix (S) is obtained. Considering image compression applications, the matrix S can be suppressed in the quantization step, so the arithmetic cost of applying the approximate transform C8 is given only by the arithmetic cost of applying the matrix T8. The approximate transform (C8) can be written as the product of the diagonal adjustment matrix (S) and the low-complexity matrix (T8).

[030] Thus, the approximate transform (C8) is explicitly given by C8 = S · T8, where T8 can be visualized in Figure 1 and S = diag (^-,1, -^-, -^-, , -^-, -^-, 2-

[031] Applying T8 to a one-dimensional vector requires 42 additions and 10 ι 1 7 · í ΤΊ C i 1 i ΓTΊ 1 · 1 1 j bit shifts. However, factoring the T8 matrix, similarly to Petition 870260049988, dated 05 / 26 / 2026, page 9 / 30 8 / 17 decimation at the base-2 Cooley-Tukey frequency

[89] , the additive complexity can be reduced by approximately 52.38%, requiring only 20 additions for its computation. The bit shift complexity can be reduced by 70%, requiring only 3 bit shifts. This factorization is called the fast algorithm and is given as follows, T8 = P · K · A2 · A1, the matrices A1, A2, K and P can be visualized in Figures 2, 3, 4, and 5, respectively, with P being a permutation matrix. The fast algorithm of the C8 transform in the form of a signal-flow diagram decomposing the T8 matrix is ​​presented in Figure 11.

[032] Applying the method of Jridi, Alfalou and Meher

[85] to T8, we have the following low-complexity matrix of length 16, called T16, which can be seen in Figure 6. Computing this transform requires 186 additions of 32 bit shifts, but using the scheme proposed by Jridi, Alfalou and Meher and factoring the T16 matrix, only 56 additions and 6 bit shifts are needed. In this way, the additive complexity is reduced by approximately 69.89% and the bit shift complexity by 81.25%. The fast algorithm for the C16 transform in the form of a signal-flow diagram decomposing the T16 matrix is ​​presented in Figure 12.

[033] Scaling the T16 matrix using the method of Jridi, Alfalou and Meher, we find the following low-complexity matrix of length 32, called T32, given in Figure 7. T32 requires 768 additions and 160 bit shifts. However, by recursively applying the scheme proposed by Jridi, Alfalou and Meher, the T32 matrix requires only 144 additions and 12 bit shifts for its computation. Thus, the additive complexity is reduced by 81.25% and the bit shift complexity is reduced by 92.5%. The fast C32 transform algorithm in the form of a signal-flow diagram decomposing the T32 matrix is ​​presented in Figure 13.

[034] To find the diagonal fitting matrices relating to matrices T16 and T32, Petition 870260049988, dated 05 / 26 / 2026, page 10 / 30 9 / 17 just do it, Sí = / [Tj · TJ, with j = 16, 32. APPLICATION IN IMAGE COMPRESSION

[035] In order to evaluate the performance of the proposed methods, we adopted JPEG image compression based on the Lena image from the public image bank “The USC-SIPI Image Database”

[90] with a size of 512 x 512 pixels in grayscale with 8 bits. The image compression method used follows the following step-by-step: 1. Each image A was divided into blocks of length 8, 16, and 32, according to the length of the transformation considered; 2. Subsequently, each block was subjected to the proposed transformations of j JJ At in the following way, B = C · A · C1, where B is the transformed block; 3. The standard zig-zag sequence

[91] is used in B, and the initial coefficients of each block are retained and the image

[61] is reconstructed; 4. The remaining coefficients are assigned a value of zero, thus obtaining B'; 5. The inverse transform is applied to each block as follows, -rA' = Ct· B' · C; 6. The A' blocks are reassembled in the reverse order to that described in item 1.

[036] Transforms with length 8 provide 64 coefficients in the transform domain for each block, while transforms with length 16 provide 256 coefficients and transforms with length 32 provide 1024 coefficients for each block. The image compression percentage (PCI) is given by, PCI =1 ——. N2

[037] An example of this process can be seen in Figures 8, 9 and 10. The images are compressed using the exact DCT and the proposed transformations for different values ​​of r. COMPARISON WITH OTHER INVENTIONS Petition 870260049988, dated 05 / 26 / 2026, page 11 / 30 10 / 17

[038] For comparison purposes, orthogonal transformations of length 8, 16 and 32 were considered, which can be seen in Tables 1, 2 and 3.

[039] Similarity and efficiency measures of the transform, arithmetic complexity (A) and bit shift complexity (S) were considered for all transforms. Mean squared error (MSE)

[45] and total energy error (e)

[61] are similarity measures to the DCT and empirically measure the distance between the approximate matrices and the exact DCT. Coding gain (C*)

[78] and transform efficiency (η)

[79] are coding measures and measure the power of decorrelation and energy compaction. Table 1 considers all transforms of length 8 used for comparison with the proposed transform. Table 2 considers all transforms of length 16. Table 3 considers those of length 32. Table 1: Evaluation measures for the transforms of the BAS approximation series and other transforms considered for comparison between the transforms of length 8. Transform and MSE C* η AS BASs(ü)

[63] 26.86 0.07 7.91 85.64 16 0 FWe

[62] 3.32 0.02 6.05 83.08 18 0 BAS3

[73] 6.85 0.03 7.91 85.38 18 0 BAS5(1)

[63] 26.86 0.07 7.91 85.38 18 0 BAS5 (2)

[63] 26.40 0.07 8.12 86.86 18 2 BASi

[82] 5.93 0.02 8.12 86.86 18 2 FW5

[62] 7.41 0.05 7.58 83.08 20 10 RDCT

[61] 1.79 0.01 8.18 87.43 22 0 BASe

[64] 35.06 0.10 7.95 85.31 24 0 Level 1 (LO) approximation

[77] 0.87 0.01 8.39 88.70 24 2 BAS4

[74] 4.09 0.02 8.33 88.22 24 4 ABM

[84] 1.22 0.01 8.63 90.46 24 6 C8 = S · T8 (proposed) 4.12 0.02 8.12 86.73 20 3

[040] According to Table 1, it is noted that the proposed length transformation Petition 870260049988, dated 05 / 26 / 2026, page 12 / 30 The 11 / 17 transform has an intermediate computational cost compared to the other transforms considered. In general, it performs better than transforms with lower arithmetic complexity and worse than transforms with higher arithmetic complexity. However, compared to the FW5 transform, which has the same additive complexity, the proposed transform performs better in all comparison metrics. Table 2: Evaluation measures for the transforms of the BAS approximation series and other transforms considered for comparison between transforms of length 16. Transform and MSE Cg η AS SOBCM

[86] 41.00 0.09 7.86 67.61 44 0 SBCKMK

[87] 30.32 0.06 8.29 70.83 60 0 JAM

[85] 14.74 0.05 8.43 72.23 60 0 BAS-2013

[64] 54.62 0.13 8.19 70.64 64 0 BAS-2010

[74] 16.41 0.056 8.52 73.63 64 8 ABM (JAM)

[84] 13.70 0.05 8.88 76.81 64 12 BCEM

[68] 8.08 0.05 7.84 65.28 72 0 C16 = S · T16 (proposal) 18.77 0.06 8.37 72.34 56 6

[041] From Table 2, it can be observed that the proposed transform of length 16 has greater arithmetic complexity only than the SOBCM transform, but it has superior performance in all metrics considered. Furthermore, the proposed transform has superior performance to the SBCKMK and BAS-2013 transforms in relation to all metrics considered; it has superior performance to the BCEM transform in coding gain and efficiency; it has superior performance to the JAM transform in relation to efficiency, despite having lower arithmetic complexity.

[042] Viewing Table 3, it can be seen that the proposed length 32 transform has the lowest arithmetic complexity among the length 32 transforms. Petition 870260049988, dated 05 / 26 / 2026, page 13 / 30 12 / 17 cataloged in the literature. Although the proposed transform has lower arithmetic complexity, it has superior performance to the BAS-2010 transform in relation to all metrics considered and also has superior performance to the JAM transform of length 32 in relation to efficiency. Table 3: Evaluation measures for the transforms of the BAS approximation series and other transforms considered for length comparison 32. Transform and MSE Cg η AS JAM

[85] 48.10 0.11 8.50 56.97 152 0 BAS-2013

[64] 192.18 0.76 8.27 55.91 160 0 BAS-2010

[74] 117.07 0.24 8.50 58.50 160 16 ABM (JAM)

[84] 46.27 0.11 8.95 61.03 160 24 C32 = S · T32 (proposal) 55.27 0.12 8.44 57.33 144 12 CONCLUSIONS

[043] In this document, a new approximate orthogonal DCT transform of length 8 with low computation cost was proposed. The fast algorithm requires only 20 additions, 3 bit shifts, and no multiplication for its computation. Furthermore, this transform was scaled to lengths 16 and 32 using the method of Jridi, Alfalou, and Meher. Based on the scheme proposed by Jridi, Alfalou, and Meher, the scaled transform of length 16 requires 56 additions and 6 bit shifts for its computation, and the scaled transform of length 32 requires 144 additions and 12 bit shifts for its computation. The proposed methods were compared with other inventions regarding similarity measures and coding performance, highlighting the applicability of the proposed transform in digital media coding. Reinforcing this point, an application in the context of image compression was presented. REFERENCES Petição 870260049988, de 26 / 05 / 2026, pág. 14 / 30 13 / 17

[44] N. U. Ahmed and K. R. Rao. Orthogonal Transforms for Digital Signal Processing. Springer-Verlag New York, Inc., Secaucus, NJ, USA, 1975.

[45] V. Britanak, P. Yip, and K. R. Rao. Discrete Cosine and Sine Transforms. Academic Press, 2007.

[46] R. C. Gonzalez and R. E. Woods. Digital Image Processing (3rd Edition). Prentice-Hall, Inc., Upper Saddle River, NJ, USA, 2006.

[47] W. K. Pratt. Digital Image Processing: PIKS Inside. John Wiley & Sons, Inc., New York, NY, USA, 3rd edition, 2001.

[48] JPEG. Joint photographic experts group, 2012.

[49] T. Wiegand, G. J. Sullivan, G. Bjontegaard, and A. Luthra. Overview of the H.264 / AVC video coding standard. IEEE Transactions on Circuits and Systems for Video Technology, 13(7):560-576, jul 2003.

[50] M. T. Pourazad, C. Doutre, M. Azimi, and P. Nasiopoulos. HEVC: The new gold standard for video compression: How does HEVC compare with H.264 / AVC? IEEE Consumer Electronics Magazine, 1(3):36-46, July 2012.

[51] R. J. Clark. Relation between Karhunen-Loève and cosine transform. In IEE Proceedings-F, Communications, Radar and Signal Processing, volume 128, pages 359-360, 1981.

[52] U. S. Potluri, A. Madanayake, R. J. Cintra, F. M. Bayer, S. Kulasekera, and A. Edirisuriya. Improved 8-point approximate dct for image and video compression requiring only 14 additions. IEEE Transactions on Circuits and Systems I, 61(6):1727-1740, June 2014.

[53] V. A. Coutinho, R. J. Cintra, F. M. Bayer, S. Kulasekera, and A. Madanayake. A multiplierless pruned DCT-like transformation for image and video compression that requires ten additions only. Journal of Real-Time Image Processing, 12(2):247-255, aug 2016. Petição 870260049988, de 26 / 05 / 2026, pág. 15 / 30 14 / 17

[54] W. H. Chen, C. Smith, and S. Fralick. A fast computational algorithm for the discrete cosine transform. IEEE Transactions on Communications, 25(9):1004-1009, 1977.

[55] B. G. Lee. A new algorithm for computing the discrete cosine transform. IEEE Transactions on Acoustics, Speech and Signal Processing, ASSP-32:1243-1245, December 1984.

[56] C. Loeffler, A. Ligtenberg, and G. Moschytz. Practical fast 1D DCT algorithms with 11 multiplications. In Proceedings of the International Conference on Acoustics, Speech, and Signal Processing, pages 988-991, 1989.

[57] F. Feig and S. Winograd. Fast algorithms for the discrete cosine transform. IEEE Transactions on Signal Processing, 40(9), 1992.

[58] J. Liang and T. D. Tran. Fast multiplierless approximations of the DCT with the lifting scheme. IEEE Transactions on Signal Processing, 49(12):3032-3044, Dec 2001.

[59] T. I. Haweel. A new square wave transform based on the DCT. Signal Processing, 82:2309-2319, 2001.

[60] F. M. Bayer and R. J. Cintra. DCT-like transform for image compression requires 14 additions only. Electronics Letters, 48(15):919-921, 2012.

[61] R. J. Cintra and F. M. Bayer. A DCT approximation for image compression. IEEE Signal Processing Letters, 18(10):579-582, 2011.

[62] C. J. Tablada, F. M. Bayer, and R. J. Cintra. A class of DCT approximations based on the feig-winograd algorithm. Signal Processing, 11:1-20, 2015.

[63] S. Bouguezel, M. O. Ahmad, and M. N. S. Swamy. A low-complexity parametric transform for image compression. In Proceedings of the 2011 IEEE International Symposium on Circuits and Systems, 2011.

[64] S. Bouguezel, M. O. Ahmad, and M. N. S. Swamy. Binary discrete cosine and Hartley transforms. IEEE Transactions on Circuits and Systems I, 60(4):989-1002, Petição 870260049988, de 26 / 05 / 2026, pág. 16 / 30 15 / 17 2013.

[65] U. S. Potluri, A. Madanayake, R. J. Cintra, F. M. Bayer, and N. Rajapaksha. Multiplier-free DCT approximations for RF multi-beam digital aperture-array space imaging and directional sensing. Measurement Science and Technology, 23(11):114003, 2012.

[66] U. S. Potluri, A. Madanayake, R. J. Cintra, F. M. Bayer, S. Kulasekera, and A. Edirisuriya. Improved 8-point approximate DCT for image and video compression requiring only 14 additions. IEEE Transactions on Circuits and Systems I, 2013.

[67] F. M. Bayer and R. J. Cintra. Image compression via a fast DCT approximation. IEEE Latin America Transactions, 8(6):708-713, December 2010.

[68] F. M. Bayer, R. J. Cintra, A. Edirisuriya, and A. Madanayake. A digital hardware fast algorithm and FPGA-based prototype for a novel 16-point approximate DCT for image compression applications. Measurement Science and Technology, 23(8):114010, 2012.

[69] F. M. Bayer, R. J. Cintra, A. Madanayake, and U. S. Potluri. Multiplierless approximate 4-point DCT VLSI architectures for transform block coding. Electronics Letters, 49(24):1532-1534, November 2013.

[70] R. J. Cintra, F. M. Bayer, and C. J. Tablada. Low-complexity 8-point DCT approximations based on integer functions. Signal Processing, 99:201-214, 2014.

[71] S. Bouguezel, M. O. Ahmad, and M. N. S. Swamy. Low-complexity 8x8 transform for image compression. Electronics Letters, 44(21):1249-1250, September 2008.

[72] S. Bouguezel, M. O. Ahmad, and M. N. S. Swamy. A multiplication-free transform for image compression. In 2nd International Conference on Signals, Circuits and Systems, pages 1-4, November 2008.

[73] S. Bouguezel, M. O. Ahmad, and M. N. S. Swamy. A fast 8x8 transform for image compression. In 2009 International Conference on Microelectronics (ICM), pages 74-77, Petição 870260049988, de 26 / 05 / 2026, pág. 17 / 30 16 / 17 December 2009.

[74] S. Bouguezel, M. O. Ahmad, and M. N. S. Swamy. A novel transform for image compression. In 53rd IEEE International Midwest Symposium on Circuits and Systems (MWSCAS), pages 509-512, August 2010.

[75] R. Tallia, P. Morello, and G. Castellano. The Walsh-Hadamard transform: an alternative means of obtaining phase and amplitude maps. Journal of Nuclear Medicine, 25(5):608-612, 1984.

[76] RS Oliveira and RJ Cintra. A factorization for reverse SDCT. Undergraduate research project at UFPE, 2013.

[77] K. Lengwehasatit and A. Ortega. Approximate forward DCT scalable variable complexity. IEEE Transactions on Circuits and Systems for Video Technology, 14(11):1236-1248, November 2004.

[78] N. S Jayant and P. Noll. Digital coding of waveforms: principles and applications to speech and video. Englewood Cliffs, NJ, pages 115-251, 1984.

[79] J. Takala and J. Nikara. Unified pipeline architecture for discrete sinis and cosine type IV transformations. In Proceedings of the 3rd International Conference on Information Communication and Signal Processing, 2001.

[80] RS Oliveira, RJ Cintra, FM Bayer and CJ Tablada. An orthogonal approximation for DCT. In: XXXI Brazilian Symposium on Telecommunications, 2013.

[81] N. Brahimi and S. Bouguezel. An efficient fast integer DCT transform for images compression with 16 additions only. Int'l Workshop on Systems, Signal Processing and their Applications, 2011. 71-74.

[82] K. Miettinen. Nonlinear Multiobjective Optimization, volume 12 of International Series in Operations Research and Management Science. Kluwer Academic Publishers, Dordrecht, 1999.

[83] M. Ehrgott. Multicriteria optimization. Lecture Notes in Economics and Petition 870260049988, dated 05 / 26 / 2026, page 18 / 30 17 / 17 Mathematical Systems. Springer-Verlag, 2000.

[84] RS Oliveira. Approximations for DCT based on angular measurement: Low complexity and image compression. Undergraduate thesis, UFPE, 2016.

[85] M. Jridi, A. Alfalou, and P. K. Meher. A generalized algorithm and reconfigurable architecture for efficient and scalable orthogonal approximation of DCT. IEEE Transactions on Circuits and Systems I, 62(2):449-457, 2015.

[86] T. L. T. Silveira, R. S. Oliveira, F. M. Bayer, R. J. Cintra, and A. Madanayake. Multiplierless 16-point DCT approximation for low-complexity image and video coding. Signal, Image and Video Processing, 11(2):227-233, 2017.

[87] T. L. T. Silveira, F. M. Bayer, R. J. Cintra, S. Kulasekera, A. Madanayake, and A. J. Kozakevicius. An orthogonal 16-point approximate DCT for image and video compression. Multidimensional Systems and Signal Processing, 27(1):87-104, 2016.

[88] N. J. Higham. Computing the polar decomposition—with applications. SIAM Journal on Scientific and Statistical Computing, 7(4):1160-1174, October 1986.

[89] R. E. Blahut. Fast Algorithms for Signal Processing. Cambridge University Press, 2010.

[90] The USC-SIPI image database, 2011.

[91] G. K. Wallace. The JPEG still picture compression standard. IEEE Transactions on Consumer Electronics, 30(1):xviii-xxxiv, 1992. Petição 870260049988, de 26 / 05 / 2026, pág. 19 / 30

Claims

1 / 2 CLAIMS 1. A computer-implemented method for encoding and decoding digital images and digital videos, in which blocks of pixels are transformed into blocks of coefficients and subsequently recomposed, characterized by comprising the following steps: (a) partitioning a digital image or a digital video frame into blocks A of dimension N x N, where NG {8, 16, 32}; (b) applying to each block A the direct transformation B = C · A · C 1 , where C = S · T, where T is an N x N matrix of elements belonging to the set {ü, ±2, ±1, ±2}, where S is an N x N diagonal matrix for scaling adjustment, and the application of the matrix T is performed exclusively by bit additions and shifts, without floating-point operations and without non-trivial multiplications; (c) retain, in each block B, the first r coefficients according to the zigzag sweep and assign a value of zero to the others, obtaining a block B'; (d) apply to each block B' the inverse transformation A = C 1 · B' · C;and (e) recompose the blocks A' into a reconstructed image or frame; wherein, for N = 8, the matrix T is the matrix T8 given by 111 ü ü -1 ü 1 -1 -1 1 1 -1 -1 ü1 1 2 Petition 870260049988, dated 05 / 26 / 2026, page 20 / 30 2 / 2; 2. Method, according to claim 1, characterized in that, for N = 8, the application of the T8 matrix is ​​performed by factorization T8 = P · K · A2 · A1, where A1, A2, K and P are 8 x 8 matrices that perform the fast factorization of the T8 matrix, with P being a permutation matrix, resulting in a computation of 20 additions and 3 bit shifts per input block of length 8.

3. Method, according to claim 1, characterized in that, for N = 16, the matrix T is the matrix T16 obtained by applying the Jridi-Alfalou-Meher scaling method to the matrix T8 as defined in claim 1, the computation of the matrix T16 being performed in 56 additions and 6 bit shifts per input block of length 16.

4. Method, according to claim 1, characterized in that, for N = 32, the matrix T is the matrix T32 obtained by recursively applying the Jridi-Alfalou-Meher scaling method to the matrix T16 as defined in claim 3, the computation of the matrix T32 being performed in 144 additions and 12 bit shifts per input block of length 32.

5. A method, according to any one of claims 1 to 4, characterized by further comprising a quantization step applied to block B, in which the scaling adjustment matrix S is incorporated into the quantization matrix, such that the arithmetic cost of the direct transformation of each block is reduced to the arithmetic cost of the isolated application of matrix T.

6. Digital processing system for encoding and decoding digital images and digital videos, characterized by comprising a processing unit configured to apply the T matrix by means of bit additions and shifts, without non-trivial multiplications, and to execute the method as defined in any one of claims 1 to 5. Petition 870260049988, dated 05 / 26 / 2026, p. 21 / 30