CN113194312B - Planetary science exploration image adaptive quantization coding system combined with visual saliency - Google Patents

Planetary science exploration image adaptive quantization coding system combined with visual saliency Download PDF

Info

Publication number
CN113194312B
CN113194312B CN202110460101.7A CN202110460101A CN113194312B CN 113194312 B CN113194312 B CN 113194312B CN 202110460101 A CN202110460101 A CN 202110460101A CN 113194312 B CN113194312 B CN 113194312B
Authority
CN
China
Prior art keywords
coding
saliency
quantization parameter
coding unit
module
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.)
Active
Application number
CN202110460101.7A
Other languages
Chinese (zh)
Other versions
CN113194312A (en
Inventor
戴育岐
薛长斌
周莉
李晓斌
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.)
National Space Science Center of CAS
Original Assignee
National Space Science Center of CAS
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 National Space Science Center of CAS filed Critical National Space Science Center of CAS
Priority to CN202110460101.7A priority Critical patent/CN113194312B/en
Publication of CN113194312A publication Critical patent/CN113194312A/en
Application granted granted Critical
Publication of CN113194312B publication Critical patent/CN113194312B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

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
    • H04N19/10—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
    • H04N19/124—Quantisation
    • 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
    • H04N19/12—Selection from among a plurality of transforms or standards, e.g. selection between discrete cosine transform [DCT] and sub-band transform or selection between H.263 and H.264
    • H04N19/122—Selection of transform size, e.g. 8x8 or 2x4x8 DCT; Selection of sub-band transforms of varying structure or type
    • 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
    • H04N19/134—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding characterised by the element, parameter or criterion affecting or controlling the adaptive coding
    • H04N19/146—Data rate or code amount at the encoder output
    • H04N19/147—Data rate or code amount at the encoder output according to rate distortion criteria
    • 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
    • H04N19/134—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding characterised by the element, parameter or criterion affecting or controlling the adaptive coding
    • H04N19/146—Data rate or code amount at the encoder output
    • H04N19/149—Data rate or code amount at the encoder output by estimating the code amount by means of a model, e.g. mathematical model or statistical model

Landscapes

  • Engineering & Computer Science (AREA)
  • Multimedia (AREA)
  • Signal Processing (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Algebra (AREA)
  • Mathematical Analysis (AREA)
  • Mathematical Optimization (AREA)
  • Pure & Applied Mathematics (AREA)
  • Discrete Mathematics (AREA)
  • Compression Or Coding Systems Of Tv Signals (AREA)

Abstract

The invention discloses a planet science exploration image self-adaptive quantization coding system combined with visual saliency, which is deployed on an exploration satellite, and comprises: the system comprises a significance detection module, an adaptive quantization parameter adjusting module and an HEVC coding module; the saliency detection module is used for carrying out saliency detection on the acquired planet scientific detection original image to obtain a global saliency map; the adaptive quantization parameter adjusting module is used for dividing the global saliency map according to the size of the coding unit to obtain a saliency sub-map corresponding to the coding unit, executing a perceptual quantization coding strategy combining visual saliency on the saliency sub-map, calculating to obtain a weighted quantization component and a perceptual quantization component of each coding unit, and further obtaining a quantization parameter offset of each coding unit; and the HEVC coding module is used for carrying out HEVC intra-frame coding on the original image according to the quantization parameter offset of each coding unit to obtain a compressed image and sending the compressed image to the ground.

Description

Planetary science exploration image adaptive quantization coding system combined with visual saliency
Technical Field
The invention relates to the technical field of computer vision and image compression, in particular to a planet science exploration image self-adaptive quantization coding system combining visual saliency.
Background
The optical load data acquired in the planetary detection process has great research value, and the acquisition and analysis of the original image data have very important significance for scientists to analyze and explore unknown landforms. The original high-resolution image data is large in volume and often contains various redundant information, large storage and transmission space needs to be occupied, and meanwhile, the return rate of scientific data is severely limited by a complex and special deep space communication environment, so that an efficient and reliable image coding scheme needs to be adopted to carry out on-orbit processing and compression on the original image data to the maximum extent, the data transmission volume is reduced, the pressure of a transmission system is reduced, and the satellite communication efficiency is improved.
From the perspective of information theory, the data describing the image information source consists of two parts, namely effective information quantity and redundancy quantity, and the essence of the image compression algorithm is to achieve the purpose of reducing the data quantity by eliminating various redundant information such as spatial domain, time domain, vision, statistics and the like in the data. In the past decades, image compression coding technology has been under rapid development, and new technology and new standards are continuously updated and iterated to adapt to application requirements of various bandwidth scenes. From the conventional JPEG and JPEG2000 still image compression standards to video compression standards such as MPEG series and h.26x series supporting efficient image compression, a hybrid coding framework formed by prediction, transformation, quantization and entropy coding has become the mainstream design direction, and a more optimal design of each module is the focus of research. The development of the existing coding performance is becoming more and more difficult, researchers begin to pay attention to content-based image compression coding and deeply mine visual redundancy so as to further increase the compression ratio and improve the coding efficiency, and the subjective quality of an image is still ensured while the code rate is low. One important direction Of research is Region-Of-Interest (ROI) coding, which combines Visual saliency detection with conventional coding. The ROI coding scheme for saliency driving generally performs saliency detection on an input image first according to the following procedure, and then improves the coding quality of the image by adjusting coding parameters to control the distortion degrees of different regions respectively.
Most of the traditional saliency detection methods are based on manually designed feature description operators, generally represent low-level features of images, and cannot identify and understand rich semantic object information in the images. In recent years, thanks to the rapid increase of hardware computing power, Artificial Intelligence (AI) technology represented by a deep neural network has been widely applied to visual tasks such as image classification, object detection, and target tracking. The significance detection method based on the neural network breaks through the limitation of the traditional manual characteristics and displays strong characterization learning capacity.
The method for optimizing the coding by combining the visual saliency can be applied to any video image compression coding standard because the syntax structure of a code stream does not need to be changed. The ROI coding method for designing the semantic understanding unit in the coding preprocessing stage considers the selectivity and the bias of a human visual system for image content perception, can effectively filter and extract complex visual input information, optimizes the coding process from the perspective of improving the visual perception efficiency, improves the quality of a coded image and increases the complexity of the whole processing flow, so that the design of an efficient image significance detection algorithm and a reasonable coding resource distribution method are the keys for improving the coding efficiency.
Disclosure of Invention
The invention provides an adaptive quantization coding system for planet science probe images, which aims to overcome the defects of the prior art and combines visual saliency.
In order to achieve the above object, the present invention provides an adaptive quantization coding system for planet science exploration images in combination with visual saliency, deployed on an exploration satellite, the system comprising: the system comprises a significance detection module, an adaptive quantization parameter adjusting module and an HEVC coding module; wherein the content of the first and second substances,
the saliency detection module is used for carrying out saliency detection on the acquired planet scientific detection original image to obtain a global saliency map with the same size as the original image and inputting the global saliency map into the self-adaptive quantitative parameter adjustment module;
the adaptive quantization parameter adjusting module is used for dividing the global saliency map according to the size of the coding unit to obtain saliency subgraphs corresponding to the coding unit, executing a perceptual quantization coding strategy combining visual saliency on the saliency subgraphs, calculating to obtain a weighted quantization component and a perceptual quantization component of each coding unit, further obtaining a quantization parameter offset of each coding unit and inputting the quantization parameter offset into the HEVC coding module;
and the HEVC coding module is used for carrying out HEVC intra-frame self-adaptive coding on the original image according to the quantization parameter offset of each coding unit to obtain a compressed image and sending the compressed image to the ground.
As an improvement of the above system, the significance detection module comprises a pre-established and trained ResNet50 model, a sorting unit, a weight vector calculation unit, a feature map weighted summation unit and an upsampling unit which are connected in sequence; wherein the content of the first and second substances,
the ResNet50 model is used for extracting multilayer depth semantic features from an original image; the ResNet50 model comprises a convolutional layer and a full-connection layer which are connected in sequence;
the sorting unit is used for numerically sorting the output tensor y of the full connection layer of the ResNet50 model to obtain a category vector c of the top five ranks:
Figure BDA0003041941290000021
the weight vector calculation unit is used for calculating the weight vector according to
Figure BDA0003041941290000031
Calculating gradient average of each channel characteristic graph output by the convolution layer in sequence to obtain corresponding weight vector
Figure BDA0003041941290000032
Figure BDA0003041941290000033
Wherein the content of the first and second substances,
Figure BDA0003041941290000034
the weight corresponding to the kth channel feature map is:
Figure BDA0003041941290000035
wherein z represents the sum of pixel values of the kth channel feature map,
Figure BDA0003041941290000036
represents the k channelCharacteristic diagram AkA pixel value at the middle (m, n) coordinate position;
the feature map weighted summation unit is configured to perform linear weighted fusion on the channel feature maps according to the weight corresponding to each channel feature map, so as to obtain a saliency map Sal:
Figure BDA0003041941290000037
where ReLU () represents a linear correction function:
Figure BDA0003041941290000038
wherein v represents a variable of the linear correction function;
the up-sampling unit is used for carrying out up-sampling interpolation processing on the saliency map Sal to obtain an output global saliency map Sal with the same size as the original imageout：
Salout＝Upsample(Sal)
Wherein upsamplale is an upsampling interpolation function.
As an improvement of the above system, the specific processing procedure of the adaptive quantization parameter adjustment module includes:
according to the global saliency map SaloutCalculating to obtain the global average salienceavg；
Dividing the global saliency map according to the size of the coding unit to obtain M × N saliency subgraphs corresponding to the M × N coding units;
traversing each significant subgraph in combination with a preset quantization parameter QPinitAnd a significance control factor beta, obtaining quantization parameter offsets corresponding to each coding unit by calculating a global significance comparison result and a local significance perception result, and inputting the quantization parameter offsets into the HEVC coding module.
As an improvement of the above system, the traversal of each significant subgraph is combined with a preset quantization parameter QPinitAnd a saliency control factor beta by calculating a global saliencyObtaining quantization parameter offsets corresponding to each coding unit according to the saliency comparison result and the local saliency sensing result, and inputting the quantization parameter offsets into an HEVC coding module; the method specifically comprises the following steps:
calculating average salience of the significant subgraph of the s-th row and the t-th columncu_avgAnd sum of significance Salcu_sumAnd calculating to obtain the significance weight omega of the coding unit of the ith row and the tth column:
Figure BDA0003041941290000041
wherein s is more than or equal to 1 and less than or equal to M, and t is more than or equal to 1 and less than or equal to N;
based on the significance weight omega corresponding to the coding unit and a preset quantization parameter QPinitCalculating the weighted quantization component QP of the coding unitweightedComprises the following steps:
Figure BDA0003041941290000042
based on preset significance control factor beta and corresponding significance sum Sal of the coding unitcu_sumCalculating the perceptual quantization component QP of the coding unitsal_offsetComprises the following steps:
QPsal_offset＝β*logSalcu_sum
the quantization parameter offset delta DQP of the coding unit of the s-th row and the t-th column is obtained by calculation according to the following formulastComprises the following steps:
ΔDQPst＝QPweighted+QPsal_offset-QPinit；
will delta DQPstInput to an HEVC coding module.
Compared with the prior art, the invention has the advantages that:
1. the system of the invention carries out image significance detection based on a multilayer neural network to obtain an image significance map, and the Mars surface has a large number of targets such as rocks, meteorite craters and the like with abundant textures, various sizes and single forms;
2. the system of the invention adaptively adjusts the corresponding quantization parameters for different coding units according to the corresponding significant subgraphs, the setting of the quantization parameters directly influences the reconstruction quality of the image content in the coding process, and the adaptive quantization parameter adjusting module of the invention comprehensively considers the global significance comparison result and the local significance sensing result, finely quantizes the region with higher significance, coarsely quantizes the non-significant region and reasonably distributes the coding bit resources.
Drawings
FIG. 1 is a block diagram of the overall structure of the planetary scientific exploration image adaptive quantization coding system combined with visual saliency according to the present invention;
FIG. 2 is a schematic diagram of the significance detection module of the present invention;
FIG. 3 is a flow diagram of an adaptive quantization parameter adjustment module process of the present invention;
fig. 4(a) is a comparison of the rate-distortion performance of the present invention with the standard HEVC method at a fixed QP (based on the PSNR objective evaluation method);
fig. 4(b) is a comparison of the rate-distortion performance of the present invention with the standard HEVC method at a fixed QP (based on PSNR-HVS objective evaluation method);
FIG. 4(c) is a comparison of the rate-distortion performance of the present invention with the standard HEVC method at a fixed QP (based on the PSNR-HVS-M objective evaluation method);
fig. 4(d) is a comparison of the rate-distortion performance of the present invention with the standard HEVC method at a fixed QP (based on the SSIM objective evaluation method);
fig. 4(e) is a comparison of the rate-distortion performance of the present invention with the standard HEVC method at a fixed QP (based on the MS-SSIM objective evaluation method);
fig. 4(f) is a comparison of the rate-distortion performance of the present invention with the standard HEVC method at a fixed QP (based on the VIFP objective evaluation method);
fig. 5(a) is a comparison of the rate-distortion performance of the present invention with the standard HEVC method at a fixed code rate (based on the PSNR objective evaluation method);
fig. 5(b) is a comparison of the rate-distortion performance of the present invention with the standard HEVC method at a fixed code rate (based on PSNR-HVS objective evaluation method);
FIG. 5(c) is a comparison of the rate-distortion performance of the present invention with the standard HEVC method at a fixed code rate (based on the PSNR-HVS-M objective evaluation method);
fig. 5(d) is a comparison of the rate-distortion performance of the present invention with the standard HEVC method at a fixed code rate (based on the SSIM objective evaluation method);
fig. 5(e) is a comparison of the rate-distortion performance of the present invention with the standard HEVC method at a fixed code rate (based on the MS-SSIM objective evaluation method);
fig. 5(f) is a comparison of the rate-distortion performance of the present invention with the standard HEVC method at a fixed code rate (based on the VIFP objective evaluation method).
Detailed Description
The invention provides a set of self-adaptive quantization coding system combined with visual saliency for compressing planet scientific exploration images. Firstly, a multilayer neural network-based saliency detection algorithm is designed, a global saliency map is extracted from an original image, then a quantization parameter adjustment algorithm combining global saliency contrast and local saliency perception is provided, different quantization parameters are distributed to each coding unit according to the saliency, the same quantization parameter scheme is used for the whole image instead, and more coding bits and calculation complexity are distributed to a saliency region more conforming to the visual characteristics of human eyes. Meanwhile, the perception characteristics of human vision are considered, and a plurality of objective distortion measurement methods combined with visual perception mechanisms are introduced to carry out combined evaluation on the image quality after compression reconstruction.
The verification experiment result carried out on the published Mars image data shows that compared with the standard HEVC implementation scheme, the algorithm provided by the invention saves the coding bit rate by 4.10% on average under the condition of certain initial quantization parameters and realizes better visual quality; the image quality after the compression reconstruction is obviously improved under the condition of the same code rate.
The whole self-adaptive quantization coding system comprises an image acquisition module, a significance detection module, a self-adaptive quantization parameter adjustment module and an HEVC coding module.
The whole algorithm design comprises the following steps:
(1) firstly, inputting acquired image data into a saliency detection module, and performing saliency detection on an image to obtain a global saliency map with the same size as an original image;
(2) inputting the saliency map result into a self-adaptive quantization parameter adjustment module, dividing a global saliency map according to the size of a coding unit to obtain a plurality of saliency subgraphs, sequentially traversing the saliency subgraphs corresponding to each coding unit, calculating saliency weight and local saliency perception to obtain quantization parameter offset corresponding to each coding unit, and finally obtaining a quantization parameter offset matrix with the same division size as the image coding unit;
(3) and inputting the original image and the quantization parameter offset matrix into an HEVC (high efficiency video coding) coding module for image compression to obtain compressed image data and sending the compressed image data to the ground.
The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and examples.
Example 1
The overall structure block diagram of the invention is shown in fig. 1, and the input data comes from the camera responsible for image data acquisition or the stored raw image data.
The significance detection module is calculated as shown in fig. 2, and mainly comprises the following implementation steps:
(1) calculating an output tensor A of the last convolutional layer of the neural network model and an output tensor y of the full connection layer according to the input image by using a ResNet50 model trained on a large classified data set ImageNet;
(2) and (3) carrying out numerical sorting on the output tensor y to obtain a category vector c of the top five ranks:
Figure BDA0003041941290000061
(3) according to
Figure BDA0003041941290000071
Calculating the gradient of each channel characteristic diagram output by the convolution layer in sequenceAveraging to obtain corresponding weight vector
Figure BDA0003041941290000072
Wherein the content of the first and second substances,
Figure BDA0003041941290000073
the weight corresponding to the kth channel feature map is as follows, and the gradient average obtained by reversely deriving the convolutional layer feature map is used as a weight parameter:
Figure BDA0003041941290000074
wherein z represents the sum of pixel values of the kth channel feature map,
Figure BDA0003041941290000075
shows the kth channel profile AkA pixel value at the middle (m, n) coordinate position;
(4) according to the weight corresponding to each channel feature map, performing linear weighted fusion on each channel feature map according to the following formula to obtain a saliency map Sal:
Figure BDA0003041941290000076
where ReLU () represents a linear correction function, only outputs greater than 0 are retained,
Figure BDA0003041941290000077
wherein v represents a variable of the linear correction function;
(5) carrying out up-sampling interpolation processing on the fusion feature map to obtain an output global saliency map Sal with the same size as the input imageout：
Salout＝Upsample(Sal)
Wherein upsamplale is an upsampling interpolation function.
The specific algorithm flow of the adaptive quantization parameter adjustment module is shown in fig. 3, and mainly includes the following calculation steps:
the adjustment of quantization parameters in the coding process directly influences the reconstruction quality of image content, and the semantic perception quantization coding strategy combining visual saliency is providedweightedAnd perceptual quantization component QPsal_offsetThe two parts further obtain the quantization parameter offset value corresponding to each coding unit, and represent the quantization parameter of the current coding unit and the quantization parameter QP of the initial imageinitThe difference of (a) is calculated by the following formula:
ΔDQPst＝QPweighted+QPsal_offset-QPinit(1≤s≤M，1≤t≤N)
and finally obtaining a two-dimensional quantization parameter offset matrix delta DQP with the size being consistent with the division size of the image coding unit, wherein the size is M multiplied by N.
Weighted quantization component QP corresponding to each coding unitweightedAnd perceptual quantization component QPsal_offsetThe calculation steps are as follows: (1) QPweightedAccording to average salience of each coding unitcu_avgAnd global average sali of significanceavgThe ratio is calculated by calculating a global significance contrast weight ω as follows:
Figure BDA0003041941290000081
Figure BDA0003041941290000082
(2)QPsal_offsetaccording to salience sum Sal of each coding unitcu_sumIs subjected to logarithm operation to obtain
QPsal_offset＝β*logSalcu_sum
To verify the effectiveness of the present invention, the following comparative experiments were designed and implemented. The hardware configuration of the experimental verification is lntel (R) core (TM) i7-6700 CPU @3.40GHz 3.41GHz, and the memory is 36G; the software is configured as a Microsoft Visual Studio 2015 platform, the HEVC coding and decoding test model is HM16.20, the real mars surface image data provided by NASA (NASA. gov.com) is used in the test, the image resolution is 1344x1200, and the encoding setting adopts an All I frame (AI-Main) configuration mode. In contrast, we set up the result of the fixed quantization parameter of HM software as the standard HEVC coding standard according to the official profile encoder _ intra _ main.cfg, and perform compression coding using the set of {22, 25, 28, 32, 35, 38, 41, 44, 47, 51} quantization parameters to obtain ten sets of comparison results. In order to more comprehensively evaluate the coding performance of the algorithm, the coding results of the algorithm provided by the invention are compared, tested and verified on 6 groups of objective image quality evaluation indexes, and the comparison results are visually represented in a broken line graph form as shown in fig. 4(a) - (f). It can be obviously seen that the performance index curves of the algorithm provided by the invention are all above the curve of the original HEVC algorithm, and the experimental results show that the algorithm provided by the invention can effectively reduce the code rate on the premise of maintaining the same subjective quality.
When the output code rate is fixed, the HEVC encoder design is based on a rate distortion optimization theory, the optimal rate distortion performance of the encoder is obtained by traversing all rate distortion performance points for comparison, then an optimal encoding parameter set is selected, bit allocation of different units is completed in sequence, then corresponding quantization parameters are obtained by estimation according to a mathematical model established between the code rate and Lagrangian factors, and code rate control is achieved. In this document, based on HM16.20, the output bit rate (bbp) is fixed, and the comparison experiment result is shown in the form of histogram as shown in fig. 5(a) - (f), which shows that, compared with the standard HEVC encoder, each index of image quality after the algorithm is encoded is obviously improved.
In conclusion, when the quantization parameter is fixed, the HEVC intra-frame adaptive quantization coding algorithm in the present invention can save more code rates on the premise of keeping the image visual quality stable, and under the condition of a fixed code rate, the algorithm herein can effectively improve the image quality, and each index realizes significant gain respectively.
Finally, it should be noted that the above embodiments are only used for illustrating the technical solutions of the present invention and are not limited. Although the present invention has been described in detail with reference to the embodiments, it will be understood by those skilled in the art that various changes may be made and equivalents may be substituted without departing from the spirit and scope of the invention as defined in the appended claims.

Claims (2)

1. An adaptive quantization coding system for planet science exploration images, which is combined with visual saliency and is deployed on an exploration satellite, wherein the system comprises: the system comprises a significance detection module, an adaptive quantization parameter adjusting module and an HEVC coding module; wherein the content of the first and second substances,
the saliency detection module is used for carrying out saliency detection on the acquired planet scientific detection original image to obtain a global saliency map with the same size as the original image and inputting the global saliency map into the self-adaptive quantitative parameter adjustment module;
the adaptive quantization parameter adjusting module is used for dividing the global saliency map according to the size of the coding unit to obtain saliency subgraphs corresponding to the coding unit, executing a perceptual quantization coding strategy combining visual saliency on the saliency subgraphs, calculating to obtain a weighted quantization component and a perceptual quantization component of each coding unit, further obtaining a quantization parameter offset of each coding unit and inputting the quantization parameter offset into the HEVC coding module;
the HEVC coding module is used for carrying out HEVC intra-frame self-adaptive coding on the original image according to the quantization parameter offset of each coding unit to obtain a compressed image and sending the compressed image to the ground;
the specific processing procedure of the adaptive quantization parameter adjustment module comprises:
according to the global saliency map SaloutCalculating to obtain the global average salienceavg；
Dividing the global saliency map according to the size of the coding unit to obtain M × N saliency subgraphs corresponding to the M × N coding units;
traversing each significant subgraph, incorporating the pre-predictionSet quantization parameter QPinitAnd a significance control factor beta, obtaining quantization parameter offsets corresponding to each coding unit by calculating a global significance comparison result and a local significance perception result, and inputting the quantization parameter offsets into the HEVC coding module;
traversing each significant subgraph in combination with a preset quantization parameter QPinitAnd a significance control factor beta, obtaining quantization parameter offsets corresponding to each coding unit by calculating a global significance comparison result and a local significance perception result, and inputting the quantization parameter offsets into the HEVC coding module; the method specifically comprises the following steps:
calculating average salience of the significant subgraph of the s-th row and the t-th columncu_avgAnd sum of significance Salcu_sumAnd calculating to obtain the significance weight omega of the coding unit of the ith row and the tth column:
Figure FDA0003295381300000011
wherein s is more than or equal to 1 and less than or equal to M, and t is more than or equal to 1 and less than or equal to N;
based on the significance weight omega corresponding to the coding unit and a preset quantization parameter QPinitCalculating the weighted quantization component QP of the coding unitweightedComprises the following steps:
Figure FDA0003295381300000021
based on preset significance control factor beta and corresponding significance sum Sal of the coding unitcu_sumCalculating the perceptual quantization component QP of the coding unitsal_offsetComprises the following steps:
QPsal_offset＝β*logSalcu_sum
the quantization parameter offset delta DQP of the coding unit of the s-th row and the t-th column is obtained by calculation according to the following formulastComprises the following steps:
ΔDQPst＝QPweighted+QPsal_offset-QPinit；
will delta DQPstInput HEVC codingAnd a code module.
2. The planetary scientific exploration image adaptive quantization coding system combining visual saliency according to claim 1, characterized in that the saliency detection module comprises a ResNet50 model, a sorting unit, a weight vector calculation unit, a feature map weighted summation unit and an upsampling unit which are connected in sequence and are established in advance and trained; wherein the content of the first and second substances,
the ResNet50 model is used for extracting multilayer depth semantic features from an original image; the ResNet50 model comprises a convolutional layer and a full-connection layer which are connected in sequence;
the sorting unit is used for numerically sorting the output tensor y of the full connection layer of the ResNet50 model to obtain a category vector c of the top five ranks:
c＝{ci},i＝1,2,3,4,5
the weight vector calculation unit is used for calculating the weight vector according to
Figure FDA0003295381300000022
Calculating gradient average of each channel characteristic graph output by the convolution layer in sequence to obtain corresponding weight vector
Figure FDA0003295381300000023
Figure FDA0003295381300000024
Wherein the content of the first and second substances,
Figure FDA0003295381300000025
the weight corresponding to the kth channel feature map is:
Figure FDA0003295381300000026
wherein z represents the sum of pixel values of the kth channel feature map,
Figure FDA0003295381300000027
shows the kth channel profile AkA pixel value at the middle (m, n) coordinate position;
the feature map weighted summation unit is configured to perform linear weighted fusion on the channel feature maps according to the weight corresponding to each channel feature map, so as to obtain a saliency map Sal:
Figure FDA0003295381300000031
where ReLU () represents a linear correction function:
Figure FDA0003295381300000032
wherein v represents a variable of the linear correction function;
the up-sampling unit is used for carrying out up-sampling interpolation processing on the saliency map Sal to obtain an output global saliency map Sal with the same size as the original imageout：
Salout＝Upsample(Sal)
Wherein upsamplale is an upsampling interpolation function.
CN202110460101.7A 2021-04-27 2021-04-27 Planetary science exploration image adaptive quantization coding system combined with visual saliency Active CN113194312B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110460101.7A CN113194312B (en) 2021-04-27 2021-04-27 Planetary science exploration image adaptive quantization coding system combined with visual saliency

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110460101.7A CN113194312B (en) 2021-04-27 2021-04-27 Planetary science exploration image adaptive quantization coding system combined with visual saliency

Publications (2)

Publication Number Publication Date
CN113194312A CN113194312A (en) 2021-07-30
CN113194312B true CN113194312B (en) 2021-12-07

Family

ID=76979428

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110460101.7A Active CN113194312B (en) 2021-04-27 2021-04-27 Planetary science exploration image adaptive quantization coding system combined with visual saliency

Country Status (1)

Country Link
CN (1) CN113194312B (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116828184B (en) * 2023-08-28 2023-12-22 腾讯科技（深圳）有限公司 Video encoding method, video decoding method, video encoding device, video decoding device, computer equipment and storage medium

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106507116B (en) * 2016-10-12 2019-08-06 上海大学 A kind of 3D-HEVC coding method predicted based on 3D conspicuousness information and View Synthesis
CN110024387B (en) * 2016-11-28 2021-10-08 日本放送协会 Encoding device, decoding device, encoding method, and decoding method
CN111083477B (en) * 2019-12-11 2020-11-10 北京航空航天大学 HEVC (high efficiency video coding) optimization algorithm based on visual saliency
CN111432207B (en) * 2020-03-30 2020-11-24 北京航空航天大学 Perceptual high-definition video coding method based on salient target detection and salient guidance

Also Published As

Publication number Publication date
CN113194312A (en) 2021-07-30

Similar Documents

Publication Publication Date Title
Li et al. Learning convolutional networks for content-weighted image compression
Quach et al. Improved deep point cloud geometry compression
Cheng et al. Deep convolutional autoencoder-based lossy image compression
CN111726633B (en) Compressed video stream recoding method based on deep learning and significance perception
CN110717868B (en) Video high dynamic range inverse tone mapping model construction and mapping method and device
CN113822147B (en) Deep compression method for semantic tasks of collaborative machine
US20220277492A1 (en) Method and data processing system for lossy image or video encoding, transmission and decoding
CN114449276B (en) Super prior side information compensation image compression method based on learning
CN111681293A (en) SAR image compression method based on convolutional neural network
CN103501438A (en) Content self-adaptation image compression method based on PCA
US8249372B2 (en) Methods and devices for coding and decoding multidimensional digital signals
CN113194312B (en) Planetary science exploration image adaptive quantization coding system combined with visual saliency
CN110677644B (en) Video coding and decoding method and video coding intra-frame predictor
CN113947538A (en) Multi-scale efficient convolution self-attention single image rain removing method
CN116896638A (en) Data compression coding technology for transmission operation detection scene
CN114463449A (en) Hyperspectral image compression method based on edge guide
CN115147317A (en) Point cloud color quality enhancement method and system based on convolutional neural network
Cheng et al. Perceptual image compression using relativistic average least squares gans
CN111107360A (en) Spectrum-space dimension combined hyperspectral image lossless compression method and system
CN117459737B (en) Training method of image preprocessing network and image preprocessing method
CN114882133B (en) Image coding and decoding method, system, device and medium
Zhang et al. An Improved Quantization Algorithm for Electric Power Inspection
CN117853367A (en) Image deblocking effect method based on multi-task learning and edge-oriented heavy parameter network
Sadanandan et al. Image compression with modified skipline encoding and curve fitting
Lee et al. Implicit Neural Representation for Video Coding Through Progressive Feature Extraction

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
GR01 Patent grant
GR01 Patent grant