CN112218094A - JPEG image decompression effect removing method based on DCT coefficient prediction - Google Patents

JPEG image decompression effect removing method based on DCT coefficient prediction Download PDF

Info

Publication number
CN112218094A
CN112218094A CN201910622339.8A CN201910622339A CN112218094A CN 112218094 A CN112218094 A CN 112218094A CN 201910622339 A CN201910622339 A CN 201910622339A CN 112218094 A CN112218094 A CN 112218094A
Authority
CN
China
Prior art keywords
image
network
dct coefficient
dct
decompression effect
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN201910622339.8A
Other languages
Chinese (zh)
Inventor
熊淑华
孙梦笛
何小海
李兴龙
任超
普拉迪普卡恩
滕奇志
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Sichuan University
Original Assignee
Sichuan University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Sichuan University filed Critical Sichuan University
Priority to CN201910622339.8A priority Critical patent/CN112218094A/en
Publication of CN112218094A publication Critical patent/CN112218094A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
    • H04N19/60—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using transform coding
    • H04N19/625—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using transform coding using discrete cosine transform [DCT]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
    • H04N19/10—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding
    • H04N19/102—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding characterised by the element, parameter or selection affected or controlled by the adaptive coding
    • H04N19/117—Filters, e.g. for pre-processing or post-processing
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
    • H04N19/85—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using pre-processing or post-processing specially adapted for video compression
    • H04N19/86—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using pre-processing or post-processing specially adapted for video compression involving reduction of coding artifacts, e.g. of blockiness

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Signal Processing (AREA)
  • Multimedia (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Computational Linguistics (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • Discrete Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Compression Or Coding Systems Of Tv Signals (AREA)

Abstract

The invention discloses a JPEG image decompression effect removing method based on DCT coefficient prediction. The method comprises the following steps: giving a compressed image, and sequentially extracting overlapped image blocks from the upper left corner of the image through the convolution operation of a network; performing DCT (discrete cosine transformation) on each image block through a convolutional neural network to obtain a corresponding DCT coefficient; learning a nonlinear mapping relation of DCT coefficients by using a wide-activation residual error network; correcting DCT coefficients learned through a network a priori through quantization constraint; carrying out IDCT transformation on the learned DCT coefficient by using a network, and transforming the DCT coefficient back to a pixel domain; the output of the network is recombined into an image by the inverse of the image block extraction operation. The JPEG image decompression effect removing method can effectively remove the compression effect of the JPEG image and can recover part of image detail information lost due to compression. Therefore, the invention is an effective JPEG image decompression effect method.

Description

JPEG image decompression effect removing method based on DCT coefficient prediction
Technical Field
The invention designs a JPEG image decompression effect removing method based on DCT coefficient prediction, and the convolutional neural network based on the DCT domain can better learn the priori knowledge of JPEG compression in the DCT domain, can better inhibit compression noise while keeping image edge detail information, and belongs to the field of digital image processing.
Background
In the current explosion of information, people are exposed to a great deal of information every day, which includes a lot of image information. In order to save transmission bandwidth and increase the transmission speed of information, compression techniques are often used to compress images, wherein JPEG is one of the most common image compression methods at present. With the increasing requirements of people on the definition of images, how to improve the definition of compressed images becomes an urgent problem to be solved. The decompression effect method for compressing an image has a good image restoration effect and does not increase a high hardware cost, so that the method has been widely paid attention by domestic and foreign scholars, and has become a research hotspot in the fields of image processing, computer vision and the like in recent years. The compression coding technique of JPEG can be briefly described as: first, an image to be compressed is divided into non-overlapping image blocks of equal size, and then each image block is subjected to independent discrete cosine transform (BDCT) transform, quantization and encoding. Although this method is simpler and has a higher compression rate, the decoded image often has a significant compression effect due to the irreversible quantization process, such as: blocking effects and ringing effects, which severely affect the quality of the image.
The decompression effect algorithm of the compressed image is a post-processing algorithm independent of a codec, and is used for improving the quality of the compressed image. The decompression effect algorithms for a visually compressed image can be classified into reconstruction-based algorithms and learning-based algorithms. Reconstruction-based methods typically utilize the compressed image's own prior information and the codec's compressed information for post-processing, such as conventional convex set projection algorithms and methods that transform the image to the wavelet domain using wavelet transforms or the like, and then process the wavelet domain coefficients in its transform domain to remove the compression effect. But reconstruction-based methods tend to be more complex and tend to obscure some details. In the learning-based method, a sparse representation algorithm and a convolution neural network-based decompression effect algorithm are representative. While convolutional neural network based methods tend to be more efficient and may better restore the texture of the image.
Disclosure of Invention
Aiming at the characteristics of the decompression effect algorithm, the invention provides a convolution neural network based on DCT coefficient prediction to learn the mapping relation between the compressed image and the original image. The convolutional neural network can fully learn the prior knowledge of JPEG compression in a DCT domain, the receptive field of the network is increased in a block overlapping mode, the larger the receptive field is, the larger the range of an input image which can be contacted by the convolutional neural network is, and the convolutional neural network possibly contains more global features with higher semantic level. The overlapping block-taking mode can also eliminate discontinuity among image blocks, thereby further removing the compression effect in the compressed image.
The invention provides a JPEG image decompression effect removing method based on DCT coefficient prediction, which mainly comprises the following operation steps of:
(1) giving a compressed image, sequentially extracting overlapped image blocks with the size of 8 multiplied by 8 from the upper left corner of the image through the convolution operation of a network, wherein the sliding step length of the extraction operation is 6 pixels;
(2) performing DCT (discrete cosine transformation) on each image block through a convolutional neural network to obtain a corresponding DCT coefficient;
(3) learning a nonlinear mapping relation of DCT coefficients by using a wide-activation residual error network to reduce quantization errors;
(4) the DCT coefficient learned by the network is corrected in a priori mode through quantization constraint, and the range of the DCT coefficient is constrained to improve the robustness of the network;
(5) performing IDCT transformation on the learned DCT coefficient, and transforming the DCT coefficient back to a pixel domain;
(6) the output of the network is recombined into an image by the inverse of the image block extraction operation.
Drawings
FIG. 1 is a block diagram of the JPEG image decompression effect removing method based on DCT coefficient prediction according to the present invention
FIG. 2 is a schematic diagram of a filter bank in a block extraction operation of the present invention
FIG. 3 is a schematic diagram of a wide active residual block structure utilized in the present invention
FIG. 4 is a graph comparing the results of the decompression effect of the present invention and eight methods on an image "Barbara
FIG. 5 is a graph comparing the decompression effect of the eight methods on the image "Lighthouse 3" according to the present invention
FIG. 6 is a frame comparison diagram of the decompression effect result of the eight methods on the image "Sailing 3" according to the present invention
Detailed Description
The JPEG image decompression effect removing method based on DCT coefficient prediction mainly comprises the following steps:
(1) giving a compressed image, sequentially extracting overlapped image blocks with the size of 8 multiplied by 8 from the upper left corner of the image through the convolution operation of a network, wherein the sliding step length of the extraction operation is 6 pixels;
(2) performing DCT (discrete cosine transformation) on each image block through a convolutional neural network to obtain a corresponding DCT coefficient;
(3) learning a nonlinear mapping relation of DCT coefficients by using a wide-activation residual error network to reduce quantization errors;
(4) the DCT coefficient learned by the network is corrected in a priori mode through quantization constraint, and the range of the DCT coefficient is constrained to improve the robustness of the network;
(5) performing IDCT transformation on the learned DCT coefficient, and transforming the DCT coefficient back to a pixel domain;
(6) the output of the network is recombined into an image by the inverse of the image block extraction operation.
Specifically, the step (1) is to perform an extraction operation of an image block (m is usually 8 in the JPEG standard) with a size of m × m in the image by convolution:
Figure RE-GDA0002203842020000031
where e (y) denotes performing an image block extraction operation on the image y,
Figure RE-GDA0002203842020000032
denoted as convolution operation, WeIs a weight of m2A filter bank of size m x m. After such convolution, each image block of m × m size in the image y will become m in size2The image block tensor of × 1 × 1 (the expression format of the multi-dimensional data in this specification is: number × channel × column × row). Specifically, in this network, the image block extraction layer contains 64 8 × 8 filters, i.e., the parameter m is set to 8
The specific operation process is described as follows:
to extract the first pixel in the top left corner of the image block, the first element in the top left corner of the m × m filter is set to 1, and the remaining elements are set to 0, while its learning rate is set to 0 at the time of network training. And so on, when the elements of the first row and the second column of the m × m filter are set to 1, and the rest elements are kept to 0, the filter extracts the pixels of the first row and the second column in the current image block.
In order to save more coding bits, the JPEG standard carries out non-overlapping block taking on a coded image, but the non-overlapping block taking causes obvious blocking effect in a compressed image. In the network provided by the invention, the image blocks are subjected to overlapping block fetching, and the blocking effect can be effectively removed by the overlapping block fetching mode, so that the performance of the decompression effect removing network is improved. Meanwhile, through the operation of overlapping and extracting the image blocks, the receptive field of the network is indirectly increased, so that the result output by the network can utilize more information of the input image, the pixel to be reconstructed is reconstructed for multiple times to a certain extent, and the robustness of the decompression effect network is increased.
In the step (2), performing DCT transform on each image block to obtain a corresponding DCT coefficient, and implementing DCT transform in the network using convolution operation in order to increase the training speed of the network and reduce the training complexity of the network. The DCT transformation in the network is expressed as operation D:
Figure RE-GDA0002203842020000033
wherein the content of the first and second substances,
Figure RE-GDA0002203842020000034
representing convolution operations, weight WDIs one comprising m2Each size is m2 X 1 filter bank, ypIs the result obtained in step (1). Initializing weights W by DCT transform matrixDAnd setting the learning rate to be zero in the training process of the network. Specifically, 64 filters of 64 × 1 × 1 are included in the DCT transform operation.
In the step (3), a wide-activation residual error network is used for learning the nonlinear mapping relation of the DCT coefficients. The DCT coefficients are input into a wide active residual block, the output through the residual block can be represented as:
yout＝F{xin,(Wi)}+xin (3)
where F denotes the residual mapping to be learned, xinAnd youtRepresenting the input and output of the residual network layer, respectively. It can be seen that when the learned mapping relationship F is 0, the mapping of the residual block is the equivalent mapping.
In a residual block of a wide-activation residual network, channel expansion is carried out on features before ReLU activation, and channel restoration is carried out on the activated features, so that the network prediction performance can be effectively improved, and more network parameters and calculated amount are not introduced. In the present invention, the convolutional layer before ReLU has 128 filters, and the convolutional layer after ReLU has 64 filters. The network contains 20 wide active residual blocks.
In the step (4), the DCT coefficient learned through the network is corrected in a priori manner through quantization constraint, and the specific method of constraint is as follows:
first, the quantization process of the DCT coefficients of an 8 × 8 image block can be expressed as follows:
Figure RE-GDA0002203842020000041
wherein x isdAnd ydAnd representing DCT coefficients obtained by DCT transformation of the original image block and the compressed image block, wherein subscripts k and l respectively represent the row and the column of the pixel in the image block. M denotes a quantization matrix.
Equation (4) is further transformed from the rounded properties:
Figure RE-GDA0002203842020000042
the range of values of the DCT coefficients from which the original image block can be obtained is:
Figure RE-GDA0002203842020000043
the above expression represents the value range of the DCT coefficient of the original uncompressed image block, i.e. the upper and lower boundaries of the DCT coefficient of the image block after the decompression effect is restrained.
Therefore, the DCT coefficient residual of the network prediction, i.e. the difference between the original image DCT coefficient and the compressed image DCT coefficient, should satisfy:
Figure RE-GDA0002203842020000044
wherein, Δ xdDenotes the residual of the DCT coefficients, M denotes a quantization matrix, and k and l denote rows and columns, respectively, in the quantization matrix.
Step (5) performing IDCT conversion on the learned DCT coefficient to express D-1：
Figure RE-GDA0002203842020000045
Wherein the content of the first and second substances,
Figure RE-GDA0002203842020000046
representing convolution operations, weights
Figure RE-GDA0002203842020000047
Is m2An IDCT transform filter bank, ypIs the result obtained in step (1).
D-1Also realized by the convolutional layer of the network, the weights of the filter bank are initialized by the IDCT transformation matrix, and gradient updating is not carried out in the training process. Specifically, 64 filters of 64 × 1 × 1 are included in the IDCT transform operation.
And (6) recombining the output of the network into an image through the inverse operation of the image block extraction operation. And averaging the network output results of the overlapping areas and outputting the network output results.
To illustrate the effectiveness of the present invention, the present invention will take 3 common test images for experiments, namely "Barbara", "Lighthouse 3" and "Sailing 3". The image was used to generate a compressed image of QF 10 with the JPEG encoder of MATLAB. The comparison method selects eight decompression effect algorithms of CONCOLOR, D2SD, SSRQC, ARCNN, TNRD, DnCNN-3, MemNet and DPW-SDNet. The experimental results are shown in FIG. 4, FIG. 5 and FIG. 6, wherein (b) to (k) represent JPEG, CONCOLOR, D2SD, SSRQC, ARCNN, TNRD, DnCNN-3, MemNet, PW-SDNet and the image results after the inventive process, respectively.
The eight video super-resolution reconstruction algorithms are respectively as follows:
the method comprises the following steps: the method proposed by Zhang et al. Reference is made to "Zhang J, Xiong R, Chen Z, et al" CONCOLOR: constrained non-constrained low-rank model for Image deblocking "-IEEE Transactions on Image Processing, vol.25, No.3, pp.1246-1259,2016".
The second method comprises the following steps: the method proposed by Liu et al. Reference is made to "Liu X, Wu X, Zhou J, et." Data-driven soft decoding of compressed images in dual transform-pixel domain. "IEEE Transactions on Image Processing, vol.25, No.4, pp.1649-1659,2016".
The third method comprises the following steps: the method proposed by Turgay et al. The references "E.Turgay, G.B.Akar", "Reducing image compression aspects by structural prediction and qualification constraint application" IEEE Transactions on Circuits and Systems for Video Technology, vol.27, No.10, pp.2057-2071,2017 ".
The method four comprises the following steps: the method proposed by Dong et al. Reference is made to "Dong C, Deng Y, Change Loy C, et al," Compression aspects reduction by a depth comparative network, "IEEE International Conference on Computer Vision, pp.576-584,2015".
The method five comprises the following steps: the method proposed by Chen et al. Reference is made to "Chen Y, pack T." convertible nonlinear interaction differentiation A flexible frame for fast and effective image representation "IEEE transactions on pattern analysis and machine interaction, vol.39, No.6, pp.1256-1272,2017".
The method six: the method proposed by Zhang et al. "Zhang K, Zuo W, Chen Y, et al," Beyond a Gaussian noise: Residual learning of deep CNN for Image noise. "IEEE Transactions on Image Processing, vol.26, No.7, pp.3142-3155,2017".
The method comprises the following steps: the method proposed by Tai et al. "Tai Y, Yang J, Liu X, et al," Mernet: A permanent memory network for image retrieval, "IEEE International Conference on Computer Vision, pp.4539-4547, 2017".
The method eight: the method proposed by Chen et al. "Chen H, He X, Qiang L, et al," DPW-SDNet: Dual Pixel-Wavelet Domain Deep CNNs for Soft Decoding of JPEG-Compressed images, "IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2018".
The contents of the comparative experiment are as follows:
experiment 1 was performed by JPEG compression of images "Barbara", "Lighthouse 3" and "Sailing 3" by QF ═ 10 using MATLAB, and the experimental results are shown in fig. 4(b), fig. 5(b) and fig. 6 (b).
Experiment 2 was conducted by the method 1 to perform the decompression effect on the compressed images "barbarbara", "Lighthouse 3" and "Sailing 3", and the experimental results are shown in fig. 4(c), fig. 5(c) and fig. 6 (c).
Experiment 3 was conducted by the method 2 to perform the decompression effect on the compressed images "barbarbara", "Lighthouse 3" and "Sailing 3", and the experimental results are shown in fig. 4(d), fig. 5(d) and fig. 6 (d).
Experiment 4 was performed by decompressing the compressed images "Barbara", "Lighthouse 3" and "Sailing 3" by method 3, and the experimental results are shown in fig. 4(e), fig. 5(e) and fig. 6 (e).
Experiment 5 was conducted by the method 4 to perform the decompression effect on the compressed images "barbarbara", "Lighthouse 3" and "Sailing 3", and the experimental results are shown in fig. 4(f), fig. 5(f) and fig. 6 (f).
Experiment 6 was conducted by the method 5 to perform the decompression effect on the compressed images "barbarbara", "Lighthouse 3" and "Sailing 3", and the experimental results are shown in fig. 4(g), fig. 5(g) and fig. 6 (g).
Experiment 7 was conducted by the method 6 to perform the decompression effect on the compressed images "barbarbara", "Lighthouse 3" and "Sailing 3", and the experimental results are shown in fig. 4(h), fig. 5(h) and fig. 6 (h).
Experiment 8 was conducted by the method 7 to perform the decompression effect on the compressed images "barbarbara", "Lighthouse 3" and "Sailing 3", and the experimental results are shown in fig. 4(i), fig. 5(i) and fig. 6 (i).
Experiment 9 was conducted by the method 8 to perform the decompression effect on the compressed images "barbarbara", "Lighthouse 3" and "Sailing 3", and the experimental results are shown in fig. 4(j), 5(j) and 6 (j).
Experiment 10 was conducted by the present invention to perform the decompression effect on the compressed images "barbarbara", "Lighthouse 3" and "Sailing 3", and the experimental results are shown in fig. 4(k), fig. 5(k) and fig. 6 (k).
Several groups of comparison experiments show that the details of the JPEG-compressed image are very fuzzy, and the image has serious blocking effect and artifacts; as shown in the enlarged region of fig. 4 "barbarbara", the deblocking effect of method 1, method 2, method 3, method 4, method 5 and method 6 is removed to some extent, but there is still some apparent detail ambiguity; although the stripe information of the region can be simply distinguished, the stripe is distorted; in the method 8, more obvious artifacts still exist; the image obtained by the method for removing the compression effect obtains good results in terms of removing the blocking effect and maintaining detailed information. Similar comparison results can be obtained in fig. 5 and 6.
In order to objectively evaluate the effectiveness of the present invention, common Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) are used as objective evaluation criteria, and an objective parameter evaluation criterion PSNR-B commonly used in the field of decompression effect is added, wherein the first table shows the average PSNR, the average SSIM and the average PSNR-B of various algorithms, and the unit of the PSNR and the PSNR-B is dB.
Table one shows that for the three objective criteria, the present invention has the highest values, and for the images "barbarara", "Lighthouse 3" and "Sailing 3", the average PSNR values of the present invention are respectively higher than method 8 by 0.20dB, 0.16dB, 0.12dB, the average SSIM values are respectively higher than method 8 by 0.0078, 0.0021, 0.0036, and the average PSNR-B values are respectively higher than method 8 by 0.20dB, 0.32dB, 0.12 dB.
In conclusion, the method has a certain degree of improvement in subjective visual effect and has obvious advantages in objective evaluation criteria. Therefore, the invention is an effective JPEG image decompression effect method.
Watch 1
Figure RE-GDA0002203842020000071

Claims (4)

1. A JPEG image decompression effect removing method based on DCT coefficient prediction mainly comprises the following operation steps:
the method comprises the following steps: giving a compressed image, sequentially extracting overlapped image blocks with the size of 8 multiplied by 8 from the upper left corner of the image through the convolution operation of a network, wherein the sliding step length of the extraction operation is 6 pixels;
step two: performing DCT (discrete cosine transformation) on each image block through a convolutional neural network to obtain a corresponding DCT coefficient;
step three: learning a nonlinear mapping relation of DCT coefficients by using a wide-activation residual error network to reduce quantization errors;
step four: the DCT coefficient learned by the network is corrected in a priori mode through quantization constraint, and the range of the DCT coefficient is constrained to improve the robustness of the network;
step five: performing IDCT transformation on the learned DCT coefficient, and transforming the DCT coefficient back to a pixel domain;
step six: the output of the network is recombined into an image by the inverse of the image block extraction operation.
2. The JPEG image decompression effect method based on DCT coefficient prediction as claimed in claim 1, wherein the DCT coefficients are quantization-constrained in step four, and learning of the convolutional neural network is guided by calculating DCT coefficient loss by using an L1 loss function in the network training process.
3. The method according to claim 1, wherein the DCT transform and the IDCT transform are performed by convolution operation in the second and the fifth steps, the weight initialization is performed by the DCT transform matrix and the IDCT transform matrix respectively, and no gradient update is performed during the network training process, so that the whole convolutional neural network can be trained end-to-end.
4. The JPEG image decompression effect method based on DCT coefficient prediction according to claim 1, wherein the constructed convolutional neural network achieves the purpose of JPEG image decompression effect by predicting DCT coefficients based on the essence that distortion is generated by JPEG compression.
CN201910622339.8A 2019-07-11 2019-07-11 JPEG image decompression effect removing method based on DCT coefficient prediction Pending CN112218094A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910622339.8A CN112218094A (en) 2019-07-11 2019-07-11 JPEG image decompression effect removing method based on DCT coefficient prediction

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910622339.8A CN112218094A (en) 2019-07-11 2019-07-11 JPEG image decompression effect removing method based on DCT coefficient prediction

Publications (1)

Publication Number Publication Date
CN112218094A true CN112218094A (en) 2021-01-12

Family

ID=74047622

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910622339.8A Pending CN112218094A (en) 2019-07-11 2019-07-11 JPEG image decompression effect removing method based on DCT coefficient prediction

Country Status (1)

Country Link
CN (1) CN112218094A (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113052258A (en) * 2021-04-13 2021-06-29 南京大学 Convolution method, model and computer equipment based on middle layer characteristic diagram compression
CN114862687A (en) * 2021-02-04 2022-08-05 四川大学 Self-adaptive compressed image restoration method driven by depth deblocking operator
WO2022167659A1 (en) * 2021-02-05 2022-08-11 Deepmind Technologies Limited Generating images using sparse representations
CN114913072A (en) * 2022-05-16 2022-08-16 中国第一汽车股份有限公司 Image processing method and device, storage medium and processor

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR100298942B1 (en) * 1998-10-19 2001-10-27 문성룡 Fuzzy Crystal Neural Network Implementation of Quantization Fuzzy Trigonometric Functions
US20160328645A1 (en) * 2015-05-08 2016-11-10 Qualcomm Incorporated Reduced computational complexity for fixed point neural network
CN107463989A (en) * 2017-07-25 2017-12-12 福建帝视信息科技有限公司 A kind of image based on deep learning goes compression artefacts method
CN109495741A (en) * 2018-11-29 2019-03-19 四川大学 Method for compressing image based on adaptive down-sampling and deep learning
CN109922346A (en) * 2019-02-28 2019-06-21 兰州交通大学 A kind of convolutional neural networks for the reconstruct of compressed sensing picture signal

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR100298942B1 (en) * 1998-10-19 2001-10-27 문성룡 Fuzzy Crystal Neural Network Implementation of Quantization Fuzzy Trigonometric Functions
US20160328645A1 (en) * 2015-05-08 2016-11-10 Qualcomm Incorporated Reduced computational complexity for fixed point neural network
CN107463989A (en) * 2017-07-25 2017-12-12 福建帝视信息科技有限公司 A kind of image based on deep learning goes compression artefacts method
CN109495741A (en) * 2018-11-29 2019-03-19 四川大学 Method for compressing image based on adaptive down-sampling and deep learning
CN109922346A (en) * 2019-02-28 2019-06-21 兰州交通大学 A kind of convolutional neural networks for the reconstruct of compressed sensing picture signal

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
索士尧，何小海，熊淑华，任超: "基于自使用高纬度非局部总变分的图像压缩去块效应算法", 《科学技术与工程》 *

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114862687A (en) * 2021-02-04 2022-08-05 四川大学 Self-adaptive compressed image restoration method driven by depth deblocking operator
CN114862687B (en) * 2021-02-04 2023-05-09 四川大学 Self-adaptive compressed image restoration method driven by depth deblocking operator
WO2022167659A1 (en) * 2021-02-05 2022-08-11 Deepmind Technologies Limited Generating images using sparse representations
CN113052258A (en) * 2021-04-13 2021-06-29 南京大学 Convolution method, model and computer equipment based on middle layer characteristic diagram compression
CN113052258B (en) * 2021-04-13 2024-05-31 南京大学 Convolution method, model and computer equipment based on middle layer feature map compression
CN114913072A (en) * 2022-05-16 2022-08-16 中国第一汽车股份有限公司 Image processing method and device, storage medium and processor

Similar Documents

Publication Publication Date Title
Zhang et al. DMCNN: Dual-domain multi-scale convolutional neural network for compression artifacts removal
Chen et al. DPW-SDNet: Dual pixel-wavelet domain deep CNNs for soft decoding of JPEG-compressed images
CN112218094A (en) JPEG image decompression effect removing method based on DCT coefficient prediction
CN108900848B (en) Video quality enhancement method based on self-adaptive separable convolution
Dong et al. Compression artifacts reduction by a deep convolutional network
Sun et al. Reduction of JPEG compression artifacts based on DCT coefficients prediction
CN112509094B (en) JPEG image compression artifact elimination algorithm based on cascade residual error coding and decoding network
CN112188217B (en) JPEG compressed image decompression effect removing method combining DCT domain and pixel domain learning
CN113055674A (en) Compressed video quality enhancement method based on two-stage multi-frame cooperation
Vo et al. Quality enhancement for motion JPEG using temporal redundancies
CN112150356A (en) Single compressed image super-resolution reconstruction method based on cascade framework
CN113962882A (en) JPEG image compression artifact eliminating method based on controllable pyramid wavelet network
KR100386639B1 (en) Method for decompression of images and video using regularized dequantizer
CN110730347A (en) Image compression method and device and electronic equipment
Hu et al. Combine traditional compression method with convolutional neural networks
Dumitrescu et al. Image compression and noise reduction through algorithms in wavelet domain
Albluwi et al. Artifacts reduction in jpeg-compressed images using cnns
CN114862687B (en) Self-adaptive compressed image restoration method driven by depth deblocking operator
Neethu et al. Improved quality of JPEG compressed image using approximate K-SVD algorithm
CN117459737B (en) Training method of image preprocessing network and image preprocessing method
Iriyama et al. Frequency domain-based bit-depth enhancement method and dealing with different bit-depth degradation
Boudechiche et al. Ensemble leaning-CNN for reducing JPEG artifacts
CN114071166B (en) HEVC compressed video quality improvement method combined with QP detection
CN113766250B (en) Compressed image quality improving method based on sampling reconstruction and feature enhancement
CN113542772B (en) Compressed image deblocking method based on enhanced wide-activation residual error network

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
WD01 Invention patent application deemed withdrawn after publication

Application publication date: 20210112

WD01 Invention patent application deemed withdrawn after publication