CN101795344A - Digital hologram compression method and system, decoding method and system, and transmission method and system - Google Patents

Digital hologram compression method and system, decoding method and system, and transmission method and system Download PDF

Info

Publication number
CN101795344A
CN101795344A CN201010116678A CN201010116678A CN101795344A CN 101795344 A CN101795344 A CN 101795344A CN 201010116678 A CN201010116678 A CN 201010116678A CN 201010116678 A CN201010116678 A CN 201010116678A CN 101795344 A CN101795344 A CN 101795344A
Authority
CN
China
Prior art keywords
hologram
layer
output
input
neural net
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.)
Granted
Application number
CN201010116678A
Other languages
Chinese (zh)
Other versions
CN101795344B (en
Inventor
杨光临
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.)
Peking University
Original Assignee
Peking 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 Peking University filed Critical Peking University
Priority to CN 201010116678 priority Critical patent/CN101795344B/en
Publication of CN101795344A publication Critical patent/CN101795344A/en
Application granted granted Critical
Publication of CN101795344B publication Critical patent/CN101795344B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Abstract

The invention discloses a digital hologram compression method, a digital hologram compression system, a digital hologram decoding method, a digital hologram decoding system, a digital hologram transmission method and a digital hologram transmission system. The transmission method comprises the following steps: synthesizing a hologram from an acquired digital image by using a computer; compressing the hologram by using a ratio of neuron points in an input layer to neuron points in a hidden layer in a neural network to acquire a compressed and coded synthetic hologram; decoding the compressed and coded synthetic hologram by using the ratio of the neuron points in the hidden layer to neuron points in an output layer in the neural network to acquire a decoded hologram; and calculating field distribution on an imaging surface by using Fresnel conversion and transmitting the decoded hologram. The digital hologram is compressed by using a BP neural network, so the nonlinearity of information distribution of the digital hologram, information distribution similar random noise and compression disadvantages caused by a big dynamic range are effectively avoided and the compression efficiency of the digital hologram is effectively enhanced.

Description

Digital hologram compression, coding/decoding method and system, transmission method and system
Technical field
The present invention relates to the Image Compression field, relate in particular to a kind of digital hologram compression, coding/decoding method and system, transmission method and system.
Background technology
1948, physicist Denis. lid primary (Dennis Gabor 1900-1979) invented holography, i.e. wavefront reconstruction technology when the resolution of research microscope.This is a kind of two step imaging techniques, the steps include: with object of coherent source irradiation, the diffracted wave that object produces and another bundle coherent reference wave interference stack produce the spatial fringe that distributes according to certain rules, with optical film these spatial fringe distribution records are got off, pass through the series of physical chemical treatment step then, just formed the hologram that includes object full detail (amplitude information and phase information).When using when shining this hologram with a branch of reference light as playback light, the amplitude of object and phase place will be come out in space reconstruct.
Along with the development of digital technology and computing technique, there is the scholar to propose directly to utilize the computer manufacture hologram.Calculation holographic is compared optical holographic, and it does not need real object to exist, as long as provide the concrete mathematical description of object wave in advance, just can utilize digital computer and plotter comprehensively to go out computed hologram, and can be by this object wave of hologram reconstruction.The computer manufacture hologram has reduced in the optical holographic various errors in the restriction of experiment condition and the processing procedure the influence of hologram quality on the other hand.The calculation holographic technology is widely used in the optical information processing now, and comprehensive and optics computing generates the detection that special reference corrugated is used for optical element, or realizes various optical transforms as special wavefront transformation element as spatial filter.
One width of cloth hologram record full detail of object (comprising amplitude information and phase information), the light amplitude that every bit write down on the hologram all is that the each point diffracted wave is with the result of reference light coherent superposition on the object, so the every bit on the hologram all includes the global information of object.If the sub-piece of any one on the intercepting hologram can reproduce the complete picture of original objects, just the definition of reproduced image can reduce along with the area of the sub-piece of hologram and descend.This has proved absolutely that the information that hologram comprised has bigger redundancy.
The object of physical record contains abundant high-frequency information, in order to guarantee that reproduced image does not produce serious aliasing, requirement wants enough high to the sample rate of object wave, simultaneously in order to guarantee the definition of reproduced image, need hologram to have with great visual angle, the data volume that such width of cloth hologram comprises is very big, has brought a lot of inconvenience for the storage and the transmission of hologram.Compression processing method and technology that therefore a lot of scholar's research holographic informations are arranged.
But, because the nonlinear transformations characteristic distributions of computer-generated hologram and high dynamic range make traditional image coding technique be difficult to handle such problem.The computer-generated hologram of labyrinth object is that a kind of information of similar random noise distributes, bigger dynamic range is arranged, the probability that each pixel intensity occurs approaches normal distribution, because the structure of interference fringe is very meticulous, local pixel excursion is bigger, comprise a lot of high-frequency informations, information is distributed on the whole spatial frequency domain relatively average, and this makes that traditional lossless coding technique efficient on the compression hologram information is not high.
Summary of the invention
The object of the present invention is to provide a kind of digital hologram compression, coding/decoding method and system, transmission method and system.Based on the present invention, can well overcome non-linear, the big drawback of bringing of dynamic range of computer digit hologram image information distribution, well improve the compression efficiency of digital hologram.
The invention provides a kind of digital hologram compression method, comprise the steps: that the digital picture that will obtain adopts computer-generated hologram; Utilize input layer in the neural net and hide the ratio that neuron is counted in the layer, described hologram is compressed, obtain the resultant hologram of compressed encoding.
Above-mentioned digital hologram compression method, preferred described computer hologram making step comprises: the mathematic(al) representation of selecting the object wave of needs description; Calculate the fresnel diffraction field distribution of object wave on holographic facet; The hologram synthesis step is encoded into the transmittance function of hologram with optical field distribution, finishes the synthetic of hologram.
Above-mentioned digital hologram compression method, between preferred described computer hologram making step and the described neural network image compression step, also comprise the neural network learning training step, described neural metwork training step comprises: the neural network parameter initialization step, parameter to described neural net is carried out initialization, and each neuronic connection weights and bias are set to random number and are absorbed in local minimum to avoid the BP neural net; Input step is with a training sample (X among the training sample set S i, Y i) be input to described BP neural net, X iSend into input layer as input vector, Y iBe sent to output layer as teacher's vector; Hide layer output calculation procedure, utilize formula h j = f ( Σ i = 1 N w ji x i + b j ) Calculate the output h of each the neuron j that hides layer j, realize that described training sample is through described BP neural net propagated forward; Wherein, x iBe the input of input layer, h jFor hiding the output of layer, w JiBe the connection weights of input layer to hiding layer, bj is biasing; Output layer output calculation procedure is utilized formula y i = f ( Σ i = 1 K w ij ′ h j + b i ′ ) Calculate the output y of each neuron i of output layer jWherein, y iBe the output of output layer, w Ij' for hiding the connection weights of layer to output layer, b i' for setovering; Output layer error amount calculation procedure, the error value E of calculating output layer, described error E obtains E=∑ (Y by the difference of calculating actual output vector and teacher's vector i-y i) 2The error back propagation computing unit, what be used to calculate output layer and hide layer is connected the weights adjustment, and the computing formula of weights adjustment is: Δ w Ij=α δ jo i, wherein, α is the learning rate of network, o j=f (s j) be according to j neuronic input value s jThe output valve that calculates, δ jBe the partial derivative of the error E of output layer to j neuron input, the δ value of one deck can be calculated according to the δ value of back one deck before the network, and formula is δ j m - 1 = f ′ ( s j m - 1 ) Σ ω ij m δ i m , δ wherein j M-1Be m-1 layer j neuronic δ value, s j M-1Be m-1 layer j neuronic input value, f ' be the first derivative of neuronal activation function, and the weights adjustment ginseng value of described hiding layer is calculated from the anti-pass error of output layer, till the connection weights adjustment of each node is calculated in described hiding layer; Revise step, that utilizes output layer and hide layer is connected the weights modified computing formulae, revises and hides the connection weight w that layer arrives output layer Ij', input layer is to the connection weight w of hiding layer Ji
Above-mentioned digital hologram compression method, in the preferred described input step, described training sample set obtains as follows: segmentation procedure with the hologram of N * N pixel, is divided into size earlier and is the square data block of m * m; Training sample obtains subelement, converts the square data block of each described m * m to a m 2The vector of * 1 dimension, m 2Corresponding to input node number in the described BP neural net, the image of then described N * N pixel is divided into (N/m) 2Individual training vector is as the described training sample of described BP neural net, and N is the pixel size of image to be compressed, and m requires N to be divided exactly by m for pixel being carried out the pixel size of piecemeal;
Above-mentioned digital hologram compression method, in the preferred described cutting unit, the value of m is 8.
The invention also discloses a kind of digital hologram coding/decoding method, comprise the steps: the neural network image decoding step, utilize and hide the ratio that neuron is counted in layer and the output layer in the described neural net, the resultant hologram of described compressed encoding is decoded, obtain decoded hologram; Fresnel reproduces step, adopts the field distribution on the Fresnel conversion Calculation imaging surface, described more decoded hologram.
The invention also discloses a kind of digital hologram transmission method, comprise outside the digital hologram compression method based on artificial synthetic neural net, also comprise: the neural network image decoding step, utilize and hide the ratio that neuron is counted in layer and the output layer in the described neural net, resultant hologram to described compressed encoding is decoded, and obtains decoded hologram; Fresnel reproduces step, adopts the field distribution on the Fresnel conversion Calculation imaging surface, described more decoded hologram.
On the other hand, the invention also discloses a kind of digital hologram compressibility, comprising: computer hologram is made module, and the digital picture that is used for obtaining adopts computer-generated hologram; The neural network image compression module is used for utilizing the neural net input layer and hides the ratio that neuron is counted in the layer, and described hologram is compressed, and obtains the resultant hologram of compressed encoding.
Above-mentioned digital hologram compressibility, preferred described computer hologram are made module and comprised: the mathematic(al) representation selected cell of object wave is used to select the mathematic(al) representation of the object wave of needs description; Fresnel diffraction field distribution computing unit is used to calculate the fresnel diffraction field distribution of object wave on holographic facet; The hologram synthesis unit is used for optical field distribution is encoded into the transmittance function of hologram, finishes the synthetic of hologram.
Above-mentioned digital hologram compressibility, preferred described computer hologram is made between module and the described neural network image compression module, also be connected with the neural network learning training module, specifically comprise: the neural network parameter initialization unit, be used for the parameter of described neural net is carried out initialization, each neuronic connection weights and bias are set to random number and are absorbed in local minimum to avoid the BP neural net; Input unit is used for a training sample (X with training sample set S i, Y i) be input to described BP neural net, X iSend into input layer as input vector, Y iBe sent to output layer as teacher's vector; Hide layer output computing unit, be used to utilize formula h j = f ( Σ i = 1 N w ji x i + b j ) Calculate the output h of each the neuron j that hides layer j, realize that described training sample is through described BP neural net propagated forward; Wherein, x iBe the input of input layer, h jFor hiding the output of layer, w JiBe the connection weights of input layer to hiding layer, b jBe biasing; Output layer output computing unit is used to utilize formula y i = f ( Σ i = 1 K w ij ′ h j + b i ′ ) Calculate the output y of each neuron i of output layer jWherein, y iBe the output of output layer, w Ij' for hiding the connection weights of layer to output layer, b i' for setovering; Output layer error amount computing unit is used to calculate the error value E of output layer, and described error E obtains E=∑ (Y by the difference of calculating actual output vector and teacher's vector i-y i) 2The error back propagation computing unit, what be used to calculate output layer and hide layer is connected the weights adjustment, and the computing formula of weights adjustment is: Δ w Ij=α δ jo i, wherein, α is the learning rate of network, o j=f (s j) be according to j neuronic input value s jThe output valve that calculates, δ jBe the partial derivative of the error E of output layer to j neuron input, the δ value of one deck can be calculated according to the δ value of back one deck before the network, and formula is δ j m - 1 = f ′ ( s j m - 1 ) Σ ω ij m δ i m , δ wherein j M-1Be m-1 layer j neuronic δ value, s j M-1Be m-1 layer j neuronic input value, f ' be the first derivative of neuronal activation function, and the weights adjustment ginseng value of described hiding layer is calculated from the anti-pass error of output layer, till the connection weights adjustment of each node is calculated in described hiding layer; Amending unit, what be used to utilize output layer and hide layer is connected the weights modified computing formulae, revises and hides the connection weight w of layer to output layer Ij', input layer is to the connection weight w of hiding layer Ji
Above-mentioned digital hologram compressibility, preferred described input unit comprises: cut apart subelement, with the image of N * N pixel, be divided into size earlier and be the square data block of m * m; Training sample obtains subelement, converts the square data block of each described m * m to a m 2The vector of * 1 dimension, m 2Corresponding to input node number in the described BP neural net, the image of then described N * N pixel is divided into (N/m) 2Individual training vector is as the described training sample of described BP neural net, and N is the pixel size of image to be compressed, and m requires N to be divided exactly by m for pixel being carried out the pixel size of piecemeal.
Above-mentioned digital hologram compressibility, in the preferred described cutting unit, the value of m is 8.
On the other hand, the present invention also provides a kind of digital hologram decode system, comprising:
The neural network image decoder module utilizes and hides the ratio that neuron is counted in layer and the output layer in the described neural net, and the resultant hologram of described compressed encoding is decoded, and obtains decoded hologram; Fresnel reproduces module, adopts the field distribution on the Fresnel conversion Calculation imaging surface, described more decoded hologram.
On the other hand, the present invention also provides a kind of digital hologram image transmission system, comprise digital hologram compressibility as described, also comprise: the neural network image decoder module, be used for utilizing neural net to hide the ratio that neuron is counted in layer and the output layer, resultant hologram to described compressed encoding is decoded, and obtains decoded hologram; Fresnel reproduces module, is used to adopt the field distribution on the Fresnel conversion Calculation imaging surface, described more decoded hologram.
The present invention adopts the BP neural net that digital hologram is compressed, and effectively overcome problems such as nonlinear degree height, dynamic range that digital hologram information distributes are big, effectively raises the compression efficiency of digital hologram.And, on the nonlinear holographic information distribution problem of processing, have intelligent and adaptivity, when compression ratio changed, the reproduced image quality that compression obtains had robustness preferably.
Description of drawings
Fig. 1 is the flow chart of steps according to the digital hologram processing method embodiment that the present invention is based on artificial synthetic neural net;
Fig. 2 is the flow chart of steps of computer-generated hologram;
Fig. 3 is the recording beam path schematic diagram of Fresnel hologram;
Fig. 4 a is three layers of BP neural network structure schematic diagram;
Fig. 4 b is neuronic mathematical description schematic diagram;
Fig. 5 is BP neural network learning and training step flow chart;
Fig. 6 is the schematic diagram of cutting apart of train samples collection;
Fig. 7 a is a Lena source images schematic diagram;
The Fresnel computer-generated hologram of Fig. 7 b for obtaining according to the calculation holographic principle;
Fig. 7 c-Fig. 7 j is respectively at 64:32 (R=50%), 64:16 (R=25%), and 64:4 (6.25%), under the situation of 64:1 (R=1.5625%), the reproduction figure of the hologram of the decompression after the process neural net compression that obtains and the hologram of decompression;
Fig. 8 a is the PSNR curve synoptic diagram of BP neural net compression algorithm;
Fig. 8 b is the MSE curve synoptic diagram of BP neural net compression algorithm;
Fig. 9 is DCT, DWT, the PSNR curve contrast schematic diagram of three kinds of compression algorithms of BP neural net;
Figure 10 is DCT, DWT, the MSE curve contrast schematic diagram of three kinds of compression algorithms of BP neural net;
Figure 11 a is the structural representation based on the digital hologram treatment system embodiment of artificial synthetic neural net;
Figure 11 b is the structural representation based on the digital hologram treatment system embodiment of artificial synthetic neural net from the angle extraction of transmitting terminal and receiving terminal.
Embodiment
For above-mentioned purpose of the present invention, feature and advantage can be become apparent more, the present invention is further detailed explanation below in conjunction with the drawings and specific embodiments.
Nonlinear characteristic at the distribution of digital hologram information, and the information distributional class is like random noise and the characteristics that have than great dynamic range, aspect the processing nonlinear problem adaptivity and intelligent advantage are being arranged in conjunction with the BP neural net, that has proposed a kind of Information Compression processing method of the computer-generated hologram that conversion is reproduced based on BP neural net and Fresnel the present invention: with computer one width of cloth digital picture is made into computer-generated hologram, according to the structure of neural net hologram is decomposed then and convert training sample to, after reaching convergence by the sample training neural net, neural net with this convergence is carried out compressed encoding and decoding to hologram, adopts Fresnel conversion to reproduce decoded hologram at last.
With reference to Fig. 1, Fig. 1 is the flow chart of steps according to the digital hologram processing method embodiment that the present invention is based on artificial synthetic neural net, comprising: computer hologram making step 110, adopt computer-generated hologram with the digital picture of obtaining; Neural network image compression step 120 utilizes input layer in the neural net and hides the ratio that neuron is counted in the layer, and described hologram is compressed, and obtains the resultant hologram of compressed encoding; Neural network image decoding step 130 is utilized and is hidden the ratio that neuron is counted in layer and the output layer in the neural net, and the resultant hologram of described compressed encoding is decoded, and obtains decoded hologram; Fresnel reproduces step 140, adopts the field distribution on the Fresnel conversion Calculation imaging surface, described more decoded hologram.
Wherein, with reference to Fig. 2, computer hologram making step 110 further comprises following three steps: the mathematic(al) representation of object wave is selected step 210, and selection needs the mathematic(al) representation of the object wave of description; Fresnel diffraction field distribution calculation procedure 220 is calculated the fresnel diffraction field distribution of object wave on holographic facet; Hologram synthesis step 230 is encoded into the transmittance function of hologram with optical field distribution, finishes the synthetic of hologram.Above-mentioned steps is promptly made the computer-generated hologram of digital picture in conjunction with the calculation holographic principle.According to principle of holography, object wave can be expressed as Parallel reference light can be expressed as Rexp[j2 π (α x+ β y)], so on the holographic recording face resulting distribution of light intensity distribute be hologram transmittance function h (x y) can be expressed as:
h ( x , y ) = | O ( x , y ) + R ( x , y ) | 2
Figure GSA00000046408600103
Figure GSA00000046408600104
In computer-generated hologram is made, earlier object wave function and reference wave function are carried out discretization sampling and quantification, calculate the computer-generated hologram that just can obtain object according to (1) formula.Because such hologram function is the nonnegative function of a reality, can carry out encoding process with computer easily.With reference to Fig. 3 a, wherein, 3a1 is a point-source of light, and 3a2 is an object, and 3a3 is a hologram plane.
Behind the hologram that step among Fig. 2 is obtained, after the compression of process neural net, the decompression, carry out the reproduction of hologram image according to the Fresnel conversion, when being execution in step 140, realize according to following principle: during the Fresnel hologram reconstruction, with the playback light irradiation hologram identical with reference light, the transmitted field amplitude on the holographic facet is so:
T(x，y)＝R(x，y)·h(x，y)＝R(x，y)·[|R(x，y)+O(x，y)| 2]
＝R·(R·R*+O·O*+R*·O+R·O*) (2)
＝R|R| 2+R|O| 2+RR*O+RRO*
(2) in the formula, first and second is noise item, field component that is proportional to original object wave of the 3rd expression, to form a virtual image of object in reproduction light source one side of hologram, the 4th is a field component that is proportional to original object wave conjugation, to form a real image of object in observer's one side of hologram, shown in Fig. 3 b.Wherein, light source is reproduced in the 3b1 representative, and 3b2 represents the virtual image, and 3b3 represents hologram, and 3b4 represents real image.
Below neural network image compression step 120, neural network image decoding step 130 are described in detail.
At first, neural net is introduced.Neural net is a kind of large-scale parallel distributed processors that is made of the simple process unit.Neural net has learning ability, can produce rational output to different training sample data collection.This information processing capability makes neural net can solve the challenge that some current techniques can't be handled.Computational process in the neural net is exactly the information transmission processing process in the neural net.From another point of view, the hologram that comprises object 3D information can be regarded as the set of a series of optics pixels, and these pixels are comprising amplitude information and phase information simultaneously.These optics pixels are followed identical optical propagation law when spatial transmission, so very high self-similarity is arranged between the pixel of hologram.This specific character is very similar with the neuronic contact in the neural net, so neural net can be used for handling the nonlinear transformations distribution problem of hologram.
Neural net is a kind of intellectuality and the higher algorithm of fitness, and one self is non-linear by the interconnected neural net that forms of non-linear neuron, and this specific character is more effective for handling nonlinear image information.The expressed relation of neural net is a kind of mapping relations that are input to output simultaneously, sample of picked at random is given network from a training set, network is just adjusted its connection weights, makes by input signal and approaches Expected Response to reach network convergence by the real response that network calculations produces with least-mean-square-error criterion.When handling hologram information with neural net method, because the self-similarity between each pixel of hologram, just in time corresponding to the annexation between each neuron in the neural net, so neural net method is used to handle nonlinear holographic information and distributes more suitable.
In actual applications, the neuron number that neural net comprised and the training sample of neural net and training time are subjected to certain restriction, therefore neural net is a kind of algorithm that diminishes when image is handled in compression, and the object 3D information that comprises of computer-generated hologram can be lost in the neural net compression process in generation like this.Along with reducing of compression ratio, losing of information can be along with increase, and this reproduced image quality to hologram can produce certain influence.
Because the information redundancy of hologram itself uses different compression methods, the degree of losing of the amount of information of hologram is different when compression ratio descends.Neural net itself has the characteristics of adaptability and fault-tolerance, and it can adjust the variation of the weights of self with the adaptation training sample set, is unlikely to badly influence the convergence capabilities of whole network when some or several neuronic abnormal behaviours.So neural net has certain robustness when handling hologram information.
With reference to Fig. 4 a, Fig. 4 b.Wherein, in Fig. 4 a, X represents input layer, and layer, Y output layer are hidden in the H representative.Typical three layers of BP (behind the Back Propagation to transmitting) neural network structure is with reference to Fig. 4 a, Fig. 4 b.In Fig. 4 a, X represents input layer, and layer, Y output layer are hidden in the H representative.Comprising input layer as Data Input Interface, the output layer of hiding layer and exporting as data.Basic processing unit in the neural net is a neuron, and the relation of neuronic input and output can be expressed as:
y ( t ) = f ( Σ i = 1 n w i x i ( t ) + θ ) - - - ( 3 )
Neuron is accepted a series of input x iAnd weight w iWeighted sum, θ is biasing, f is neuronic activation primitive, y is neuronic response output.Therefore in neural net, each neuronic output of hiding layer and output layer can be expressed as:
h j = f ( Σ i = 1 N w ji x i + b j ) 1 ≤ j ≤ K
y i = f ( Σ i = 1 K w ij ′ h j + b i ′ ) 1 ≤ i ≤ N (4)
In formula (4), f (x) is the activation primitive of each layer, w JiBe to tie up connection weight value matrix, w to the K that hides layer * N from input layer Ij' be from hiding the N * K dimension connection weight value matrix of layer to output layer, b jAnd b i' be respectively each neuronic biasing.
In the reality, in described neural network image compression step 120 and the described neural network image decoding 130, described neural net is the BP neural net of convergence, obtains by following training and learning procedure:
Step 1, the parameter of BP neural net is carried out initialization, each neuronic connection weights and bias are set to little random number.Can avoid the BP neural net to be absorbed in local minimum like this, help the convergence of network.
Step 2, with a sample (X among the training sample set S i, Y i) be input to BP neural net, X iSend into input layer as input vector, Y iBe sent to output layer as teacher's vector.
Step 3, training sample promptly utilize formula through BP neural net propagated forward h j = f ( Σ i = 1 N w ji x i + b j ) Calculate the output h of each the neuron j that hides layer j
Step 4, utilize formula y i = f ( Σ i = 1 K w ij ′ h j + b i ′ ) Calculate the output y of each neuron i of output layer j
The error amount of step 5, calculating output layer, this error obtains E=‖ Y-y ‖=∑ (Y by the difference of calculating actual output vector and teacher's vector i-y i) 2
The error amount of layer is hidden in step 6, calculating δ j m - 1 = f ′ ( s j m - 1 ) Σ ω ij m δ i m . This error is calculated from the anti-pass error of output layer, till the error of each node is calculated in will hiding layer.
Step 7, utilize output layer and hide the connection weights that are connected each layer of weights modified computing formulae correction of layer.
Step 8 is returned step 2, is next one input learning sample repeating step 2 to 7.
Said process can be understood in conjunction with Fig. 5.
Image is transferred to the BP neural net to carry out needing image is done preliminary treatment before the encoding and decoding.According to the input neuron number of BP neural net, segment the image into the training sample set of little sub-piece as neural net.For size is the image of N * N pixel, is divided into size earlier and is the square data block of m * m.The square data block of each m * m converts a m to then 2The vector of * 1 dimension, m 2Corresponding to the input node number of BP network, this sampled images just is divided into (N/m) 2Individual training vector is as the learning sample set of network, as shown in Figure 6.
Three layers of BP neural net that typically are used for image compression, the structure of general input layer and output layer is symmetrical, identical neuron node number is arranged, the input layer that the vector that the training sample that obtains after the image segmentation preliminary treatment is concentrated is fed to the BP network also is fed to output layer as teacher signal simultaneously as input signal.After reaching convergence, preserve all weights and the codec of other network parameters of this moment as the compression hologram information by BP algorithm network training.
The BP network realizes that the principle of data compression is to realize by adjusting the neuron node number of hiding layer and input layer.If the nodal point number of hiding layer in the network is less than the nodal point number of input layer, image data transmission will reduce when hiding layer so, adds that the data total amount of neural network weight and other network parameter is less than from the image data amount of input layer input and has just realized data compression if hide view data that layer preserves.Can be at receiving terminal according to neural network weight that receives and the original BP neural net of network parameter reconstruct, view data is transferred to the recovery that output layer is realized view data from the hiding layer of the neural net of this reconstruct.
In BP neural net resultant hologram compression processing system, at first preliminary treatment obtains a training sample set to resultant hologram through image, when cutting apart hologram, the value of sub-block size parameter m should be chosen different values according to practical application, and what the m value was got is excessive or too small all improper, if the m value is excessive, then sample size is less, and the input node number of neural net increases, the complicated network structure, if the m value is too small, then sample size is more, makes the net training time lengthening, the constringency performance variation, by experiment, find that the m value is 8 proper, whole system has lower complexity and short training convergence time.
After the convergence of BP neural net reached and stablizes, hologram data to be compressed was finished the compressed encoding of hologram information then earlier by the image preliminary treatment to the transmission of hiding layer from input layer.By adjusting the ratio of the neuron nodal point number of hiding layer and input layer, can obtain different data compression rates.It is few more to hide layer nodal point number, and compression ratio is low more, otherwise it is many more to hide layer nodal point number, and compression ratio is high more.Finish the decoding of hologram information from hiding layer to the transmission of output layer, decoded image is through post processing of image, and promptly the pretreated inverse process of image reduces and obtains complete hologram.
Because hologram comprises the 3D information of object, the information that this and general pattern comprised has than big difference.In the quality evaluation for the hologram image processing, do not have unified standard.We can by human eye vision observe original hologram relatively and handle after hologram, but such subjective assessment can not characterize the hologram picture quality after processed objectively.Therefore, this paper adopts traditional quality of image processing evaluation criterion, promptly uses compression ratio (R), mean square error (MSE), and Y-PSNR (PSNR) is estimated compression efficiency and distortion measurement.R, MSE, PSNR is defined as:
R = S c S o = Q P - - - ( 5 )
MSE = 1 M × N Σ x = 1 M Σ y = 1 N [ f ( x , y ) - f ~ ( x , y ) ] 2 - - - ( 6 )
PSNR = 10 log 10 [ x p 2 MSE ] ( dB ) - - - ( 7 )
In formula (5), S cBe the computer-generated hologram size of data after the compression, S oIt is original computer-generated hologram size of data.Q is a neuron number of hiding layer, and P is the neuron number of input layer.In formula (6), M * N is the pixel size (as M * N=256 * 256) of image, f (x y) is the reproduced image function of original hologram,
Figure GSA00000046408600161
It is the reproduced image function of the hologram after overcompression is handled.In formula (7), x pIt is the peak-to-peak value of view data.
In the various compression processing method experiments of computer-generated hologram, the present invention adopts the hologram reconstruction picture quality after above three evaluation functions are estimated processing.
Fig. 7 (a) is that (65kb), Fig. 7 (b) is the Fresnel computer-generated hologram that obtains according to the calculation holographic principle to the Lena source images for 256 * 256 pixel sizes, 8bpp/256.Fig. 7 (c)-Fig. 7 (j) is respectively at 64:32 (R=50%), 64:16 (R=25%), 64:4 (6.25%), under the situation of 64:1 (R=1.5625%), the reproduction figure of the hologram of the decompression after the process neural net compression that obtains and the hologram of decompression.
The MSE of experimental result and the reproduced image under various compression ratios and PSNR experimental data are respectively shown in table 1 and Fig. 8 a, Fig. 8 b.
R, PSNR and the MSE experimental data of table 1BP neural net compression algorithm
Figure GSA00000046408600162
Analyze from table 1, can find that the PSNR of reproduced image and compression ratio are proportional.Because the BP neural net is not a model that accuracy is very high, in the compressed encoding of neural net, the neuron number of hiding layer is few more, and the compression ratio that obtains is low more, and the ability to express of network is just poor more, so information from input layer to the transmittance process of hiding layer, lose just many more.The degree direct influence of information loss reproduced image quality and PSNR value.Fig. 8 a, 8b are the PSNR curve and the MSE curve of BP neural net compression algorithm.Abscissa is a compression ratio.
Image function is very complicated, and is difficult to accurately describe with mathematical expression, and neural net can be approached any mathematical function approx, and the approximation capability of neural net is directly related to its network configuration and the neuronic number that is comprised.In general, the training sample set of larger amt and multilayer neural network structure can be than the training sample set of limited quantity and simple three-layer neural network structure token image functions better.In the experiment of this paper, the BP neural net that is used for doing compression is a simple three-layer neural network structure, comprises 64 input neurons and 64 output neurons, and training sample set is a width of cloth computer-generated hologram picture.These have all limited the ability of neural net, so neural net is accurate inadequately to the expression that nonlinear holographic information distributes under higher compression ratios.But from Fig. 7 (i), Fig. 7 (j) under reach 1: 64 in the compression ratio situation of (R=1.5625%), although the hologram blocking artifact after decompressing clearly, still can access identifiable reproduced image.
Below, advantage of the present invention is described in further detail.
For relatively discrete cosine (DCT) compression better, discrete wavelet (DWT) compression and the compression of BP neural net can be drawn in the PSNR-R curve and the MSE-R curve of three kinds of compression methods in the same coordinate system, and be as Fig. 9, shown in Figure 10.
From Fig. 9, Figure 10 analysis can find that in the PSNR of three kinds of compression methods curve and MSE curve, other two kinds of methods of PSNR curve ratio of neural net compression are more steady.Along with the decline of compression ratio R, the PSNR value of DCT compression and DWT compression descends very fast, and the PSNR of BP neural net compression then descends slowly relatively.This explanation DCT compression and DWT compression are higher than the compression of BP neural net to the influence degree of information loss and reproduced image quality.This that is to say that with respect to DCT compression and DWT compression, neural net has intelligent and adaptivity, and different compression ratios is had robustness on the nonlinear holographic information distribution problem of processing.In addition, under very low compression ratio, neural net compression ratio DCT compression and DWT are compressed with better reproduced image quality and littler reconstruction error.Reach 4: 64 at compression ratio when (R=6.25%), it is approaching that the PSNR value of the reproduced image that the neural net compression obtains and DCT compress the PSNR value that obtains, and when compression ratio further descends, neural net compression ratio DCT compresses more effective.Reach 1: 64 in compression ratio when (R=1.5625%), the neural net compression also than the DWT compression effectively.As seen, requiring under the situation of little compressible very, removing to handle hologram with neural net is a kind of more efficient methods.
On the other hand, the present invention also provides a kind of digital hologram treatment system based on artificial synthetic neural net, and with reference to Figure 11 a, this system comprises:
Computer hologram is made module 1101, and the digital picture that is used for obtaining adopts computer-generated hologram; Neural network image compression module 1102 is used for utilizing the neural net input layer and hides the ratio that neuron is counted in the layer, and described hologram is compressed, and obtains the resultant hologram of compressed encoding; Neural network image decoder module 1103 is used for utilizing neural net to hide the ratio that neuron is counted in layer and the output layer, and the resultant hologram of described compressed encoding is decoded, and obtains decoded hologram; Fresnel reproduces module 1104, is used to adopt the field distribution on the Fresnel conversion Calculation imaging surface, described more decoded hologram.
Identical among the operation principle of above-mentioned each module and the method embodiment, do not repeat them here, relevant part is mutually with reference to getting final product.
With reference to Figure 11 b, Figure 11 b is the structural representation based on the digital hologram treatment system embodiment of artificial synthetic neural net from the angle extraction of transmitting terminal (image compression system) and receiving terminal (picture decoding system).Wherein, 11a represents transmitting terminal, and 11b represents receiving terminal.Be that transmitting terminal 11a comprises computer hologram making module 1101, neural network image compression module 1102; Receiving terminal 11b comprises that neural network image decoder module 1103, Fresnel reproduce module 1104.
Neural network image compression module 1102 among the transmitting terminal 11a utilizes input layer in the neural net and hides the ratio that neuron is counted in the layer, described hologram is compressed, after obtaining the resultant hologram of compressed encoding, synthetic hologram is sent to receiving terminal 11b; Simultaneously, the relevant parameter that also needs will be in the neural net to hide layer, output layer also be sent to receiving terminal 11b, and like this, neural network image decoder module 1103 could be decoded by the parameter information that the receives hologram after to compressed encoding.
Identical among the operation principle of above-mentioned each module and the method embodiment, do not repeat them here, relevant part is mutually with reference to getting final product.
In sum, the present invention is in order to construct a kind of general holographic information compressibility, has intelligent and adaptivity on the nonlinear holographic information distribution problem handling, and when promptly compression ratio changed, the reproduced image quality that compression obtains had robustness preferably.Main advantage is:
The first, utilize the learning ability of neural net,, make neural net have the ability of expressing the hologram nonlinear Distribution by the sample training collection of hologram.
The second, utilize the intelligent of neural net and self adaptation characteristics, go to handle complicated hologram information distribution problem, big as dynamic range, the nonlinear degree height.
Three, utilize that relevance between the hologram pixel and the neuronic relevance in the neural network structure are very similar to be handled nonlinear holographic information and distribute.
More than a kind of digital hologram compression provided by the present invention, coding/decoding method and system, transmission method and system are introduced in detail, used specific embodiment herein principle of the present invention and execution mode are set forth, the explanation of above embodiment just is used for helping to understand method of the present invention and core concept thereof; Simultaneously, for one of ordinary skill in the art, according to thought of the present invention, part in specific embodiments and applications all can change.In sum, this description should not be construed as limitation of the present invention.

Claims (14)

1. a digital hologram compression method is characterized in that, comprises the steps: the computer hologram making step, and the digital picture of obtaining is adopted computer-generated hologram;
The neural network image compression step utilizes input layer in the neural net and hides the ratio that neuron is counted in the layer, and described hologram is compressed, and obtains the resultant hologram of compressed encoding.
2. digital hologram compression method according to claim 1 is characterized in that, described computer hologram making step comprises:
The mathematic(al) representation of object wave is selected step, and selection needs the mathematic(al) representation of the object wave of description;
Fresnel diffraction field distribution calculation procedure is calculated the fresnel diffraction field distribution of object wave on holographic facet;
The hologram synthesis step is encoded into the transmittance function of hologram with optical field distribution, finishes the synthetic of hologram.
3. digital hologram compression method according to claim 1, it is characterized in that, between described computer hologram making step and the described neural network image compression step, also comprise the neural network learning training step, described neural metwork training step comprises:
The neural network parameter initialization step carries out initialization to the parameter of described neural net, and each neuronic connection weights and bias are set to random number and are absorbed in local minimum to avoid the BP neural net;
Input step is with a training sample (X among the training sample set S i, Y i) be input to described BP neural net, X iSend into input layer as input vector, Y iBe sent to output layer as teacher's vector;
Hide layer output calculation procedure, utilize formula h j = f ( Σ i = 1 N w ji x i + b j ) Calculate the output h of each the neuron j that hides layer j, realize that described training sample is through described BP neural net propagated forward; Wherein, x iBe the input of input layer, h jFor hiding the output of layer, w JiBe the connection weights of input layer to hiding layer, b jBe biasing;
Output layer output calculation procedure is utilized formula y i = f ( Σ i = 1 K w ij ′ h j + b i ′ ) Calculate the output y of each neuron i of output layer jWherein, y iBe the output of output layer, w Ij' for hiding the connection weights of layer to output layer, b i' for setovering;
Output layer error amount calculation procedure, the error value E of calculating output layer, described error E obtains E=∑ (Y by the difference of calculating actual output vector and teacher's vector i-y i) 2
The error back propagation computing unit, what be used to calculate output layer and hide layer is connected the weights adjustment, and the computing formula of weights adjustment is: Δ w Ij=α δ jo i, wherein, α is the learning rate of network, o j=f (s j) be according to j neuronic input value s jThe output valve that calculates, δ jBe the partial derivative of the error E of output layer to j neuron input, the δ value of one deck can be calculated according to the δ value of back one deck before the network, and formula is δ j m - 1 = f ′ ( s j m - 1 ) Σ ω ij m δ i m , δ wherein j M-1Be m-1 layer j neuronic δ value, s j M-1Be m-1 layer j neuronic input value, f ' be the first derivative of neuronal activation function, and the weights adjustment ginseng value of described hiding layer is calculated from the anti-pass error of output layer, till the connection weights adjustment of each node is calculated in described hiding layer;
Revise step, that utilizes output layer and hide layer is connected the weights modified computing formulae, revises and hides the connection weight w that layer arrives output layer Ij', input layer is to the connection weight w of hiding layer Ji
4. digital hologram compression method according to claim 3 is characterized in that, in the described input step, described training sample set obtains as follows:
Segmentation procedure with the hologram of N * N pixel, is divided into size earlier and is the square data block of m * m;
Training sample obtains subelement, converts the square data block of each described m * m to a m 2The vector of * 1 dimension, m 2Corresponding to input node number in the described BP neural net, the image of then described N * N pixel is divided into (N/m) 2Individual training vector is as the described training sample of described BP neural net, and N is the pixel size of image to be compressed, and m requires N to be divided exactly by m for pixel being carried out the pixel size of piecemeal.
5. digital hologram compressibility according to claim 4 is characterized in that, in the described cutting unit, the value of m is 8.
6. a digital hologram coding/decoding method is characterized in that, comprises the steps:
The neural network image decoding step is utilized and is hidden the ratio that neuron is counted in layer and the output layer in the described neural net, and the resultant hologram of described compressed encoding is decoded, and obtains decoded hologram;
Fresnel reproduces step, adopts the field distribution on the Fresnel conversion Calculation imaging surface, described more decoded hologram.
7. a digital hologram transmission method is characterized in that, comprises as outside each described digital hologram compression method based on artificial synthetic neural net in the claim 1 to 5, also comprises:
The neural network image decoding step is utilized and is hidden the ratio that neuron is counted in layer and the output layer in the described neural net, and the resultant hologram of described compressed encoding is decoded, and obtains decoded hologram;
Fresnel reproduces step, adopts the field distribution on the Fresnel conversion Calculation imaging surface, described more decoded hologram.
8. a digital hologram compressibility is characterized in that, comprising:
Computer hologram is made module, and the digital picture that is used for obtaining adopts computer-generated hologram;
The neural network image compression module is used for utilizing the neural net input layer and hides the ratio that neuron is counted in the layer, and described hologram is compressed, and obtains the resultant hologram of compressed encoding.
9. digital hologram compressibility according to claim 8 is characterized in that, described computer hologram is made module and comprised:
The mathematic(al) representation selected cell of object wave is used to select the mathematic(al) representation of the object wave that needs describe;
Fresnel diffraction field distribution computing unit is used to calculate the fresnel diffraction field distribution of object wave on holographic facet;
The hologram synthesis unit is used for optical field distribution is encoded into the transmittance function of hologram, finishes the synthetic of hologram.
10. digital hologram compressibility according to claim 8 is characterized in that, described computer hologram is made between module and the described neural network image compression module, also is connected with the neural network learning training module, specifically comprises:
The neural network parameter initialization unit is used for the parameter of described neural net is carried out initialization, and each neuronic connection weights and bias are set to random number and are absorbed in local minimum to avoid the BP neural net;
Input unit is used for a training sample (X with training sample set S i, Y i) be input to described BP neural net, X iSend into input layer as input vector, Y iBe sent to output layer as teacher's vector;
Hide layer output computing unit, be used to utilize formula h j = f ( Σ i = 1 N w ji x i + b j ) Calculate the output h of each the neuron j that hides layer j, realize that described training sample is through described BP neural net propagated forward; Wherein, x iBe the input of input layer, h jFor hiding the output of layer, w JiBe the connection weights of input layer to hiding layer, b jBe biasing;
Output layer output computing unit is used to utilize formula y i = f ( Σ i = 1 K w ij ′ h j + b i ′ ) Calculate the output y of each neuron i of output layer jWherein, y iBe the output of output layer, w Ij' for hiding the connection weights of layer to output layer, b i' for setovering;
Output layer error amount computing unit is used to calculate the error value E of output layer, and described error E obtains E=∑ (Y by the difference of calculating actual output vector and teacher's vector i-y i) 2
The error back propagation computing unit, what be used to calculate output layer and hide layer is connected the weights adjustment, and the computing formula of weights adjustment is: Δ w Ij=α δ jo i, wherein, α is the learning rate of network, o j=f (s j) be according to j neuronic input value s jThe output valve that calculates, δ jBe the partial derivative of the error E of output layer to j neuron input, the δ value of one deck can be calculated according to the δ value of back one deck before the network, and formula is δ j m - 1 = f ′ ( s j m - 1 ) Σ ω ij m δ i m , δ wherein j M-1Be m-1 layer j neuronic δ value, s j M-1Be m-1 layer j neuronic input value, f ' be the first derivative of neuronal activation function, and the weights adjustment ginseng value of described hiding layer is calculated from the anti-pass error of output layer, till the connection weights adjustment of each node is calculated in described hiding layer;
Amending unit, what be used to utilize output layer and hide layer is connected the weights modified computing formulae, revises and hides the connection weight w of layer to output layer Ij', input layer is to the connection weight w of hiding layer Ij
11. digital hologram compressibility according to claim 10 is characterized in that, described input unit comprises:
Cut apart subelement,, be divided into size earlier and be the square data block of m * m the image of N * N pixel;
Training sample obtains subelement, converts the square data block of each described m * m to a m 2The vector of * 1 dimension, m 2Corresponding to input node number in the described BP neural net, the image of then described N * N pixel is divided into (N/m) 2Individual training vector is as the described training sample of described BP neural net, and N is the pixel size of image to be compressed, and m requires N to be divided exactly by m for pixel being carried out the pixel size of piecemeal.
12. digital hologram compressibility according to claim 11 is characterized in that, in the described cutting unit, the value of m is 8.
13. a digital hologram decode system is characterized in that, comprising:
The neural network image decoder module utilizes and hides the ratio that neuron is counted in layer and the output layer in the described neural net, and the resultant hologram of described compressed encoding is decoded, and obtains decoded hologram;
Fresnel reproduces module, adopts the field distribution on the Fresnel conversion Calculation imaging surface, described more decoded hologram.
14. a digital hologram image transmission system is characterized in that, comprises as each described digital hologram compressibility in the claim 8 to 11, also comprises:
The neural network image decoder module is used for utilizing neural net to hide the ratio that neuron is counted in layer and the output layer, and the resultant hologram of described compressed encoding is decoded, and obtains decoded hologram;
Fresnel reproduces module, is used to adopt the field distribution on the Fresnel conversion Calculation imaging surface, described more decoded hologram.
CN 201010116678 2010-03-02 2010-03-02 Digital hologram compression method and system, decoding method and system, and transmission method and system Active CN101795344B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN 201010116678 CN101795344B (en) 2010-03-02 2010-03-02 Digital hologram compression method and system, decoding method and system, and transmission method and system

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN 201010116678 CN101795344B (en) 2010-03-02 2010-03-02 Digital hologram compression method and system, decoding method and system, and transmission method and system

Publications (2)

Publication Number Publication Date
CN101795344A true CN101795344A (en) 2010-08-04
CN101795344B CN101795344B (en) 2013-03-27

Family

ID=42587759

Family Applications (1)

Application Number Title Priority Date Filing Date
CN 201010116678 Active CN101795344B (en) 2010-03-02 2010-03-02 Digital hologram compression method and system, decoding method and system, and transmission method and system

Country Status (1)

Country Link
CN (1) CN101795344B (en)

Cited By (22)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102122513A (en) * 2011-03-25 2011-07-13 上海大学 Coaxial recording device for digital hologram of transparent substance based on beam-splitting property of Fresnel biprism
CN103049083A (en) * 2012-12-17 2013-04-17 深圳市世纪天河科技有限公司 Method and system for controlling virtual holographic input and output
CN104765263A (en) * 2015-04-13 2015-07-08 北京大学 Compressing, coding and transmitting method for electronic holographic three-dimensional information
CN105976408A (en) * 2016-04-28 2016-09-28 北京大学 Digital holographic compression transmission method of quantum backward propagation nerve network
CN106254879A (en) * 2016-08-31 2016-12-21 广州精点计算机科技有限公司 A kind of Image Lossy Compression method applying own coding neutral net
CN106464936A (en) * 2014-05-01 2017-02-22 高通股份有限公司 Method and device for decoding multi-layer video data by determining the processing core of the decoder based on partition containing one or more layers
CN106485323A (en) * 2015-10-08 2017-03-08 上海兆芯集成电路有限公司 There is the neutral net unit that output buffer feedback is calculated with execution time recurrent neural network
CN106650918A (en) * 2016-11-25 2017-05-10 东软集团股份有限公司 System model building method and apparatus
CN107004157A (en) * 2015-01-22 2017-08-01 高通股份有限公司 Model compression and fine setting
CN107111782A (en) * 2014-11-26 2017-08-29 卡里尔斯公司 Neural network structure and its method
CN107463932A (en) * 2017-07-13 2017-12-12 央视国际网络无锡有限公司 A kind of method that picture feature is extracted using binary system bottleneck neutral net
CN107547773A (en) * 2017-07-26 2018-01-05 新华三技术有限公司 A kind of image processing method, device and equipment
WO2018107414A1 (en) * 2016-12-15 2018-06-21 上海寒武纪信息科技有限公司 Apparatus, equipment and method for compressing/decompressing neural network model
CN108701249A (en) * 2016-01-25 2018-10-23 渊慧科技有限公司 Image is generated using neural network
CN108881660A (en) * 2018-05-02 2018-11-23 北京大学 A method of computed hologram is compressed using the quantum nerve network of optimization initial weight
CN109074512A (en) * 2016-02-05 2018-12-21 渊慧科技有限公司 Image is compressed using neural network
CN109564636A (en) * 2016-05-31 2019-04-02 微软技术许可有限责任公司 Another neural network is trained using a neural network
CN109933676A (en) * 2019-02-14 2019-06-25 深兰科技（上海）有限公司 A kind of information base construction method and device
CN111353591A (en) * 2018-12-20 2020-06-30 中科寒武纪科技股份有限公司 Computing device and related product
WO2020232613A1 (en) * 2019-05-20 2020-11-26 深圳先进技术研究院 Video processing method and system, mobile terminal, server and storage medium
CN114782565A (en) * 2022-06-22 2022-07-22 武汉搜优数字科技有限公司 Digital archive image compression, storage and recovery method based on neural network
CN114998457A (en) * 2022-08-01 2022-09-02 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) Image compression method, image decompression method, related device and readable storage medium

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1097070A (en) * 1993-06-30 1995-01-04 惠州市新世纪发展总公司 Make the method for the synthetic rainbow holography master hologram of mold pressing with calculating holographic technique

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1097070A (en) * 1993-06-30 1995-01-04 惠州市新世纪发展总公司 Make the method for the synthetic rainbow holography master hologram of mold pressing with calculating holographic technique

Non-Patent Citations (5)

* Cited by examiner, † Cited by third party
Title
《JOURNAL OF DISPLAY TECHNOLOGY》 20061231 Alision E,etc. Compression of optically encrypted digital holograms using artificial neural networks 401-410 1-12 第2卷, 第4期 2 *
《信息技术》 20040430 胡金滨等 人工神经网络的BP算法及其应用 第1-4页 3-5,8-12,14 第28卷, 第4期 2 *
《兰州大学学报》 20050831 马义德等 一种基于分类的改进BP神经网络图像压缩方法 第70-72页 1-14 第41卷, 第4期 2 *
《工程图学学报》 20060228 张丽峰等 双曲正切非线性函数和加均值量化算法对BP神经网络图像压缩处理的影响 第110-115页 3-5,8-12,14 , 第1期 2 *
《江苏、山东、河南、江西、黑龙江五省光学（激光）联合学术05年会论文集》 20051231 苏蓉等 应用菲涅尔计算全息制作二维光子晶体的研究 摘要 2,6-14 , 2 *

Cited By (37)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102122513A (en) * 2011-03-25 2011-07-13 上海大学 Coaxial recording device for digital hologram of transparent substance based on beam-splitting property of Fresnel biprism
CN102122513B (en) * 2011-03-25 2012-06-13 上海大学 Coaxial recording device for digital hologram of transparent substance based on beam-splitting property of Fresnel biprism
CN103049083A (en) * 2012-12-17 2013-04-17 深圳市世纪天河科技有限公司 Method and system for controlling virtual holographic input and output
CN106464936A (en) * 2014-05-01 2017-02-22 高通股份有限公司 Method and device for decoding multi-layer video data by determining the processing core of the decoder based on partition containing one or more layers
US10390087B2 (en) 2014-05-01 2019-08-20 Qualcomm Incorporated Hypothetical reference decoder parameters for partitioning schemes in video coding
CN106464936B (en) * 2014-05-01 2019-06-14 高通股份有限公司 The method and apparatus that multi-layer video data are decoded by determining the processing core of decoder based on the subregion containing one or more layers
US11720795B2 (en) 2014-11-26 2023-08-08 Canary Capital Llc Neural network structure and a method thereto
CN107111782B (en) * 2014-11-26 2021-05-14 柯纳睿资本有限责任公司 Neural network structure and method thereof
CN107111782A (en) * 2014-11-26 2017-08-29 卡里尔斯公司 Neural network structure and its method
CN107004157A (en) * 2015-01-22 2017-08-01 高通股份有限公司 Model compression and fine setting
CN104765263A (en) * 2015-04-13 2015-07-08 北京大学 Compressing, coding and transmitting method for electronic holographic three-dimensional information
CN104765263B (en) * 2015-04-13 2017-09-19 北京大学 Electronic holographic three-dimensional information compressed encoding transmission method
CN106485323B (en) * 2015-10-08 2019-02-26 上海兆芯集成电路有限公司 It is fed back with output buffer to execute the neural network unit of time recurrent neural network calculating
CN106485323A (en) * 2015-10-08 2017-03-08 上海兆芯集成电路有限公司 There is the neutral net unit that output buffer feedback is calculated with execution time recurrent neural network
CN108701249A (en) * 2016-01-25 2018-10-23 渊慧科技有限公司 Image is generated using neural network
CN108701249B (en) * 2016-01-25 2023-04-14 渊慧科技有限公司 Generating images using neural networks
US11336908B2 (en) 2016-02-05 2022-05-17 Deepmind Technologies Limited Compressing images using neural networks
CN109074512A (en) * 2016-02-05 2018-12-21 渊慧科技有限公司 Image is compressed using neural network
CN105976408A (en) * 2016-04-28 2016-09-28 北京大学 Digital holographic compression transmission method of quantum backward propagation nerve network
CN109564636A (en) * 2016-05-31 2019-04-02 微软技术许可有限责任公司 Another neural network is trained using a neural network
CN109564636B (en) * 2016-05-31 2023-05-02 微软技术许可有限责任公司 Training one neural network using another neural network
CN106254879B (en) * 2016-08-31 2019-11-08 广东精点数据科技股份有限公司 A kind of application encodes the Image Lossy Compression method of neural network certainly
CN106254879A (en) * 2016-08-31 2016-12-21 广州精点计算机科技有限公司 A kind of Image Lossy Compression method applying own coding neutral net
CN106650918A (en) * 2016-11-25 2017-05-10 东软集团股份有限公司 System model building method and apparatus
CN106650918B (en) * 2016-11-25 2019-08-30 东软集团股份有限公司 The method and apparatus for constructing system model
WO2018107414A1 (en) * 2016-12-15 2018-06-21 上海寒武纪信息科技有限公司 Apparatus, equipment and method for compressing/decompressing neural network model
CN107463932A (en) * 2017-07-13 2017-12-12 央视国际网络无锡有限公司 A kind of method that picture feature is extracted using binary system bottleneck neutral net
CN107463932B (en) * 2017-07-13 2020-07-10 央视国际网络无锡有限公司 Method for extracting picture features by using binary bottleneck neural network
CN107547773B (en) * 2017-07-26 2020-01-03 新华三技术有限公司 Image processing method, device and equipment
CN107547773A (en) * 2017-07-26 2018-01-05 新华三技术有限公司 A kind of image processing method, device and equipment
CN108881660A (en) * 2018-05-02 2018-11-23 北京大学 A method of computed hologram is compressed using the quantum nerve network of optimization initial weight
CN111353591A (en) * 2018-12-20 2020-06-30 中科寒武纪科技股份有限公司 Computing device and related product
CN109933676A (en) * 2019-02-14 2019-06-25 深兰科技（上海）有限公司 A kind of information base construction method and device
WO2020232613A1 (en) * 2019-05-20 2020-11-26 深圳先进技术研究院 Video processing method and system, mobile terminal, server and storage medium
CN114782565A (en) * 2022-06-22 2022-07-22 武汉搜优数字科技有限公司 Digital archive image compression, storage and recovery method based on neural network
CN114998457A (en) * 2022-08-01 2022-09-02 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) Image compression method, image decompression method, related device and readable storage medium
CN114998457B (en) * 2022-08-01 2022-11-22 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) Image compression method, image decompression method, related device and readable storage medium

Also Published As

Publication number Publication date
CN101795344B (en) 2013-03-27

Similar Documents

Publication Publication Date Title
CN101795344B (en) Digital hologram compression method and system, decoding method and system, and transmission method and system
Namphol et al. Image compression with a hierarchical neural network
CN105976408A (en) Digital holographic compression transmission method of quantum backward propagation nerve network
CN104765263B (en) Electronic holographic three-dimensional information compressed encoding transmission method
CN108419083B (en) Image multilevel wavelet full subband compressed sensing coding method
CN108093264B (en) Core image compression, decompressing method and system based on splits' positions perception
CN113077377A (en) Color image steganography method based on generation countermeasure network
CN115063492B (en) Method for generating countermeasure sample for resisting JPEG compression
CN116228912A (en) Image compressed sensing reconstruction method based on U-Net multi-scale neural network
CN114894740A (en) Terahertz single-pixel imaging method and system
CN114581341A (en) Image style migration method and system based on deep learning
CN114138919A (en) Seismic data reconstruction method based on non-local attention convolution neural network
CN114022356A (en) River course flow water level remote sensing image super-resolution method and system based on wavelet domain
CN117037290A (en) Human face fake detection method based on global context structure difference
Zhuang et al. UCSNet: Priors Guided Adaptive Compressive Sensing Framework for Underwater Images
CN116797461A (en) Binocular image super-resolution reconstruction method based on multistage attention-strengthening mechanism
Ren et al. Information content compression and zero-order elimination of computer-generated hologram based on discrete cosine transform
CN113658330B (en) Holographic encoding method based on neural network
CN115797181A (en) Image super-resolution reconstruction method for mine fuzzy environment
CN104243986A (en) Compressed video capture and reconstruction system based on data drive tensor subspace
Kumari et al. Image quality estimation by entropy and redundancy calculation for various wavelet families
Entezarmahdi et al. Stationary image resolution enhancement on the basis of contourlet and wavelet transforms by means of the artificial neural network
Dandawate et al. Neuro-Wavelet based vector quantizer design for image compression
Ahmed et al. Image compression using Haar and modified Haar wavelet transform
CN112257509A (en) Stereo image single-stream visual saliency detection method based on joint information coding

Legal Events

Date Code Title Description
C06 Publication
PB01 Publication
C10 Entry into substantive examination
SE01 Entry into force of request for substantive examination
C14 Grant of patent or utility model
GR01 Patent grant