专利汇可以提供Method and apparatus for picture data reduction for digital video signals专利检索,专利查询,专利分析的服务。并且Method and apparatus for the reduction of picture data for digital video signals comprising a processing of the signals by means of block by block transformation method so that a transformed and quantized signal which was generated at a time t-1 and placed in an image storage is subtracted from a transformed signal that occurs at a time t and whereby the difference signal obtained is subject to quantization and the quantized difference signal is subjected to an analysis and to a time delay which corresponds to the time requirement for the analysis for updating the content of the image storage. The signal delayed is added to the signal read out from the image storage which is also delayed and is added dependent on the addition condition signal obtained from the analysis and is subjected to an entropy coding dependent on the analysis results with the addition condition signals containing information as to whether a block which has been analyzed has been concluded is a moved or unmoved block and when the block is a moved block containing information regarding a coefficient group to be transmitted. The signal coded in such fashion is subjected to a buffering and depending on the degree of buffer filling a quantization stage and an analysis stage is influenced so that a signal from a buffer control is supplied to the quantization stage for the purpose of selecting one of a plurality of predetermined quantization characteristics and a second signal is supplied from the buffer control means to the analysis stage to select the maximum number of coefficient groups and a third signal is supplied to the analysis stage from the buffer control for the purpose of deciding whether a block is to be transmitted or not and the coefficients represent the digitized video signal transformed by block which is subdivided into coefficient groups.,下面是Method and apparatus for picture data reduction for digital video signals专利的具体信息内容。
We claim as our invention:1. A method for picture data reduction for digital video signals, comprising a pre-processing of the signals by means of block-by-block transformation method, whereby a transformed and quantized signal that was generated at a time of t-1 and deposited in an image store is subtracted from a transformed signal that occurs at a time t to obtain a difference signal and whereby the difference signal is subjected to a quantization, comprising the steps of, analyzing the quantitized difference signal and delaying for a time (VZ) which corresponds to the time requirement for the analysis (AS) said difference signal for updating the content of the image store, adding delayed signal to the signal read out from the image store (M) which has been correspondingly delayed, said adding being dependent on an addition condition signal acquired from said analysis, and applying an entropy coding (HC) to said signal which depends on the results of said analysis, said addition condition signals containing information as to whether a block whose analysis has been concluded is a "moved"or "unmoved" block and, in case said block is a "moved" block, containing information regarding a coefficient group to be transmitted; buffering the signal coded in such fashion with a buffer (B) for supplying an output signal channel a uniform data flow for transmission, obtaining said uniform data flow from the non-uniform data flow of the entropy coding; influencing depending on the degree of the fill of the buffer a quantization stage (Q) and an analysis stage (AS) whereby a signal from a buffer control means (BC) is supplied to the quantization stage (Q), and selecting one of a plurality of pre-determined quantization characteristics, supplying a second signal from the buffer control means (BC) to the analysis stage (AS) for selecting the maximum number of coefficient groups, and supplying a third signal to the analysis stage (AS) from the buffer control means (BC) for deciding whether a block is to be transmitted or is not to be transmitted; subdividing the coefficients representing the digitized video signal transformed block-by-block into coefficient groups according to prescribed rules; and identifying a value for each of the coefficient groups in a calculation stage (E), and determining from said value whether a super-group formed in a decision means (S) from neighboring coefficient groups is to be transmitted is selected, whereby coefficient groups which are not to be transmitted based on said identified value can also be arranged in such a supergroup, and secondly, classifying in a mediated fashion, following a step-by-step summationE.sub.I (i)=E.sub.I (i-1)+E(i)of all values respectively belonging to a block in an integrator (I), whereby i=2 . . . 31 preferably applies, whereby E(i) is the value for the coefficient group i and whereby E.sub.I (1)=E(1) applies, said classification serving the purpose to decide whether a block is to be transmitted and the manner in which a block to be transmitted is to be coded.2. A method according to claim 1, wherein the coefficient groups are formed such that coefficients (y(u, v)) whose matrix indices (u,v) meet the conditionu+v=i-1are respectively combined to form a coefficient group (i), whereby u, v=0 . . . 15 and whereby u, v are the horizontal or, respectively, vertical discrete frequencies.3. A method according to claims 1 or 2, characterized in that said value for the selection of a coefficient group and for classification is ##EQU6## whereby i is the number of a coefficient group, whereby k can have the values 1, 2, 3 . . . , and whereby .DELTA.y.sub.Q (u, i-1-u) is the quantized difference signal of the coefficients (u, i-1-u).4. A method according to claim 3, characterized in that the value for the selection of a coefficient group and for the classification is the energy ##EQU7## of a coefficient group (i), whereby .DELTA.y.sub.Q (u, i-1-u) is the quantized difference signal of the coefficients (u,i-1-u).5. A method according to claim 3, characterized in that the value for the selection of a coefficient group and for the classification is the sum of the absolute values ##EQU8## of a coefficient group (i), whereby .DELTA.y.sub.Q (u, i-1-u) is the quantized difference signal of the coefficients (u, i-1-u).6. A method according to claim 1, characterized in that the quantization is executed such that respectively like quantization intervals are employed for all amplitude ranges of the signal to be quantized, so that a linear quantization results.7. A method according to claim 6, characterized in that the quantization is executed dependent on the picture activity; and a selection of one of four quantization intervals which respectively differ by the factor "2" is accomplished.8. A method according to claim 1, characterized in that a block-wise classification in four classes is carried out, said classes serving the purpose of respectively selecting one of four coding allocation tables, whereby one of these four classes is an "unmoved" class.9. A method according to claim 8, characterized in that the classification is executed such that, in a first step, a decision is made from a data field respectively supplied by the integrator (I) as to whether the appertaining block is to be classified as "moved" or "unmoved", whereby, when the block is classified as "unmoved", a 2-bit code word is generated for said class 4 which is the "unmoved" class or, respectively, a decision is made in a second and, under given conditions, a third step concerning which of three "moved" classes, namely class 1 . . . class 3, said block is to be allocated to, whereby the energy belonging to the supergroup to be transmitted is identified from the data field (E.sub.I) and is successively compared to two thresholds read out and edited from first and second tables, a 2-bit code word being respectively generated under given conditions for said classes 1 . . . 3.10. A method according to claim 1, characterized in that seven code tables of variable word length are employed for the purpose of coding, whereby on code table (1) is provided for extremely small signal variances and one code table (7) is provided for extremely large signal variances, whereby the code table (i) respectively differs from the code table (i+1) in that it is generated for a signal variance which is greater by the factor "4".11. A method according to claims 9 or 10, characterized in that the allocation of the code tables to the individual coefficients of a block is done with the assistance of three allocation tables of which one is selected by means of one of the identified classes for a "moved" block.12. A method according to claim 1, characterized in that an eighth code table is provided for coding the supergroup to be transmitted; and an auxiliary table is provided which is employed to select therefrom the code word number of the eighth code table allocated to this supergroup for every combination of the values (N.sub.O and N.sub.D) for limiting the super-group to be transmitted.13. A method according to claim 10, characterized in that the seven code tables are Huffman code tables.14. A method according to claim 12 wherein the eighth code table is a Huffman code table.15. A method according to claim 1 characterized in that a buffer control means (BC) defines the maximum allowable number of coefficient groups as well as the threshold for the decision "moved"/"unmoved" and modifies the quantization resolution "memory-free" dependent on the degree that the buffer memory is filled.16. A method according to claim 15, characterized in that the maximum number of coefficient groups (N.sub.NMAX) is defined by the relationship ##EQU9## wherein trun (truncated) is the "cut-off" function of the illustrated function and whereby B.sub.N preferably lies in the range of 0.3 . . . 0.6; and in that the threshold (T) for the decision "moved"/"unmoved" is varied by the relationship ##EQU10## wherein B.sub.T preferably lies in the range of 0.6 . . . 0.8; and in the quantizer resolution .DELTA. is varied with an intermediate value .DELTA. * which varies with the degree of the filling of the buffer memory and is related to the buffer memory length, which varies by the relationship ##EQU11## whereby the quantizer resolution is established from this intermediate value .DELTA. * under the following conditions:when 0.75.DELTA..sub.o .ltoreq..DELTA.* applies, then .DELTA.=.DELTA..sub.o applies,when 0.375.DELTA..sub.o .ltoreq..DELTA.*<0.75.DELTA..sub.o, then .DELTA.=.DELTA..sub.o /2 applies,when 0.1875.DELTA..sub.o .ltoreq..DELTA.*<0.375.DELTA..sub.o applies, then .DELTA.=.DELTA..sub.o /4 applies,when .DELTA.*<0.1875.DELTA..sub.o applies, then.DELTA.=.DELTA..sub.o /8 applies,whereby B.sub..DELTA. preferably has values between 0.4 and 0.6.17. A method according to claim 16, characterized in the tabular values are determined for the relationships for N.sub.DMAX, T and .DELTA., and these tabular values are addressed on the basis of the six most significant bits of the degree of buffer filling.18. Apparatus for picture data reduction for digital video signals comprising, a receiver with a receiver buffer memory (B.sub.E), a decoder (DC) connected to said receiver buffer memory B.sub.E, a reconstruction means (R) connected to said decoder, a receiver buffer control means (BC.sub.E) connected between said receiver buffer memory B.sub.E and said reconstruction means (R), a receiver summing element (+.sub.E) receiving inputs from said decoder and said reconstruction means, and a receiver image store (M.sub.E) connected to said receiver summing element, the data-reduced digital video signal is supplied to said receiver buffer memory (B.sub.E) after a channel decoding; said decoder (DC) reconstructs a signal with constant word length from the preferably Huffman-coded signal intermediately stored in the receiver buffer memory (B.sub.E); the reconstruction means (R) reproduce representative values from the numbers for representative values coded with constant word length and from a signal supplied by the receiver buffer control means (BC.sub.E) for the selection of one of a plurality of quantizer tables, said receiver buffer control means (BC.sub.E) produces the output signal as the sourceside quantization signal, and said receiver buffer control means receives a signal indicative of the degree of filling (B) of the buffer memory from the receiver buffer memory (B.sub.E); the representative values from the reconstruction means (R) are supplied to the receiver summing element (+.sub.E), and addition condition signals (N.sub.O,N.sub.D and a signal for "moved"/"unmoved") from the decoder (DC) as well as the picture signal from the receiver image store (M.sub.E) reconstructed at time t-1 are also supplied, and the reconstructed difference signal at a time t from the reconstruction means (R) is added to the picture signal reconstructed at a time t-1, with the addition accomplished depending on the addition condition signals; and the reconstructed transformed picture signal obtained by said addition supplied to a inverse transformation stage (IT) and, also is supplied to said receiver image store (M.sub.E).19. Apparatus according to claim 18, characterized in that a counter is provided which measures the degree of filling of said buffer memory.20. Apparatus according to claim 18, characterized in that a shift register is provided for delaying by n blocks the video signals.21. Apparatus according to claim 18, characterized in that a gate chain is provided for delaying by n blocks said video data.22. Apparatus according to claim 18, characterized in that at least one ROM memory is used for storing tables.
BACKGROUND OF THE INVENTION
1. Field of the Invention
The present invention relates to a method for picture data reduction for digital video signals comprising a preprocessing of the signals by means of block-by-block transformation whereby a transformed and quantatized signal which was generated at a time t-1 and deposited in an image storage is subtracted from a transformed signal which occurs at a time t and whereby the difference signal obtained in such manner is subjected to a quantization.
2. Description of the Prior Art
Prior art methods for picture data reduction can be subdivided into:
1. DPCM (Differential Pulse Code Modulation) methods-transformation methods; and
2. Hybrid methods.
In DPCM methods, the difference between an estimate determined from samples that have already been transmitted and the actual sample is respectively identified. In pure DPCM coders, this prediction occurs three-dimensionally, in other words, both within a frame or picture as well as from frame to frame.
In transformation methods, an imaging of the frame into the transformation region occurs. Due to the high cost, only two dimensional transformations have previously been realized in practice.
The present invention relates to a hybrid method. The principles of a hybrid method is illustrated in FIG. 1. In FIG. 1, a digitized signal x (k, e, t) is supplied to a transformation stage and produces a transformation coefficient signal y(u, v, t) which is supplied to a quantitizer Q which produces a signal Ya(u, v, t) which is supplied through an adder to a coder C which produces a signal Yc(u,v,t) which is supplied as the channel signal. The output of the quantitizer Q is also supplied to a predictor and memory P+M which supplies a signal yp (u, v,t-1) to an adder to add the signal to the output of the transformation stage before supplying it to the quantitizer Q.
Hybrid coding represents a mixture of transformation and DPCM. The transformation within a frame occurs two-dimensionally, block size 16×16 or 8×8 picture points, whereas DPCM operates from frame to frame. The signal decorrelated by transformation and/or DPCM is quantitized and transmitted.
Basically, all hybrid methods operate according to the diagram illustrated in FIG. 1. In developed methods, the functions Q, P and C are adaptively executed
European Patent Application No. 82.3070263 discloses a method which employs a coder having the following essential features:
Dynamic bit allocation--The bit rate is minimized and is selected from a plurality of Huffman code tables by means of a prediction algorithm for each coefficient to be coded.
Length of run coding--Zeros successively appearing along a defined scan direction are coded by lengths of run.
Constant Channel rate--Is achieved by coupling the quantitizer to the buffer filling. A PI controller with proportional integrating behavior is employed for this purpose.
The publication of F. May, "Codierung von Bildfolgen mit geringer Rate fur gestorte Uebertrangungskanale", NTG-Fachberichte, Vol. 74, pp. 379-388, describes a system for picture transmission using narrow-band radio channels with a transmission rate of 9.6K bit/s and a frame frequency of 0.5 frames. A plurality of bit allocation matrices are provided for this known method so that the optimum of the respective block is identified and transmitted in the form of a class affiliation. Optimum non-linear quantization characteristics are also employed with respect to the quadratic error. A constant channel rate is achieved by input buffer control, in other words, every frame is first analyzed, the number of coeficients to be transmitted is then modified until the channel rate is observed.
The publication of W. H. Chen, W. K. Pratt entitled "Scene Adaptive Coder", in the IEEE Trans. Comm., Vol. Com32, No. 3, of Mar. 1984, describes an adaptive band width compression technique which employs a discrete cosine transformation. This system is similar to that describes in European Patent Application No. 82.30 70 263 referenced above.
A publication of A. G. Tescher, entitled "Rate Adaptive Communication", appearing in the IEEE International Conference on Communication, of 1978, pages 1.1-19.1.6 describes a concept for a bit rate control in a source coding system.
The technical book publication of W. K. Pratt entitled Image Transformation Techniques, published by the Academic Press, New York, San Francisco, and London in 1979 provides overall discussion of the transformation techniques of the systems.
SUMMARY OF THE INVENTION
The object of the present invention is to provide a method of the species initially referenced which enables a picture quality which is improved significantly over known methods for the same or constant channel rate. In the invention, picture data reduction for digital video signals comprises preprocessing the signals using block-by-block transformation method whereby a transformed and quantitized signal that was generated at a time t-1 and placed in an image storer is subtracted from a transformed signal that occurs at a time t and whereby the difference acquired in this manner is subjected to a quantitization and the quantitized difference signal is subjected to an analysis and is subjected to a time delay VZ which corresponds to the time requirement for the analysis AS and on the one hand updating the content of the image storage and the signal which is delayed in this manner is added to the signal read out from the image storer M which is also correspondingly delayed and is added thereto dependent on the addition condition signal acquired from the analysis and on the other hand is subjected to an entropy coding HC depending on the analysis results. The addition condition signals containing information as to whether a block whose analysis has been concluded is a "moved" or a "unmoved" block and in case said block is a "moved" block containing information regarding a coefficient group to be transmitted, the coded signal is subjected to a buffering B which is intended to offer an output signal channel a uniform data flow for transmission and offering said uniform data flow from a nonuniform data flow of the entropy coding. Dependent on the degree of buffer filling, a quantization stage Q, an analysis stage AS is influenced so that a signal from a buffer control means BC is supplied to the quantization stage Q for selecting one of a plurality of predetermined quantization characteristics whereby a second signal is supplied from the buffer control means BC to the analysis stage AS for the purpose of selecting the maximum number of coefficient groups and where a third signal is supplied to the analysis stage AS from the buffer control means BC for deciding whether a block is to be transmitted or is not to be transmitted. The coefficients represent the digitized video signal transformed block-by-block which is subdivided into coefficient groups according to prescribed rules and a measurement scale for each of these coefficient groups is identified in a calculation stage E such that the scale first causes a supergroup to be formed in a decision means S from neighboring coefficient groups and to be transmitted and selected such that the coefficient groups which are not to be transmitted according to the identified scale can be embedded in a supergroup and by means of which a classification is executed by a following step-by-step summation of all the scales respectively belonging to a block in an integrator I where i=2 . . . 3is preferably applies and E(i) is the scale for the coefficient group i and whereby EI (1)=E (1) applies and the classification serves the purpose for deciding whether a block is to be transmitted and what way a block to be transmitted is to be coded.
Other objects, features and advantages of the invention will be readily apparent from the following description of certain preferred embodiments thereof taken in conjunction with the accompanying drawings although variations and modifications may be effected without departing from the spirit and scope of the novel concepts of the disclosure and in which:
BRIEF DESCRIPTION OF THE DRAWINGS
FIG. 1 is a schematic illustration illustrating the basic concept of prior art hybrid coding;
FIG. 2 is a block diagram of a complete transmission system according to a preferred exemplary embodiment according to the invention;
FIG. 3 is a block diagram of a transmitter of the exemplary embodiment of the transmission system shown in FIG. 2;
FIG. 4 is a block circuit diagram of a receiver according to the exemplary embodiment shown in FIG. 2;
FIG. 5a is a schematic illustration of a preferred exemplary embodiment of the manner in which a field comprising mxn coefficients is subdivided into coefficient groups in the form of imaginary diagonal strips;
FIG. 5b is a schematic illustration which shows how a buffer control in the method of the invention effects the coder output rate by limiting the number of coefficient groups to be transmitted;
FIGS. 5c and 5d show how neighboring coefficient groups are combined in a supergroup; and
FIGS. 6a, 6b and 6c illustrate the characteristics of a buffer control.
DESCRIPTION OF THE PREFERRED EMBODIMENTS
As shown in FIG. 2, a transmitter has a transformation stage T which transforms the signal with a discrete cosine transformation (DCT). The invention can be utilized with other transformations as well. The coding method occurs as shown in the block circuit diagrams of FIGS. 2 and 3 for the transmitter and FIGS. 2 and 4 for the receiver. As shown in FIG. 2, the transmitter has a transformation stage T which transforms the sign and supplies it to a subtractor which supplies an output to a quantizer Q. The quantizer supplies an input to an adder which supplies an output to a memory M which also supplies an input to the adder and the memory M also supplies an input to the subtractor. A coding device HC receives the output of the quantizer and also an output of an analysis stage AS which receives an input from the quantizer Q. The coding device supplies an output to the output buffer B which supplies an output to the channel encoding device. The output buffer also supplies an input to the buffer control means BC which supplies inputs to the quantizer Q and to the analysis stage AS.
The output of the channel encoding means of the transmitter is supplied to the receiver wherein a channel decoding means receives the incoming signal and supplies it to a receiver buffer BE which supplies an output to a decoder DC which supplies an output to a reconstruction means R. A receiver buffer control means BCE receives an output from the receiver buffer BE and supplies an input to the reconstructions means R. A receiver summing means +E receives the output of the reconstruction means and also an input from the decoder DC. The receiver summing means supplies an output to the innertransformation stage IT which produces the reconstructed signal. The receiver summing means +E also supplies an input to a receiver image storer ME which supplies an input to the receiver summing element.
FIG. 3 illustrates in greater detail portions of the transmitter where the output of the transformation stage T is supplied to the subtractor which supplies an output to the quantizer Q which supplies an output to the first time delay VZ which supplies an output to the entropy coding device HC. An adder also receives output from the first time delay VZ as well as an output of a decision means S and an output of a classification device K. The adder supplies an output to a memory M which supplies an output to the subtractor. A second time delay VZ receives the output of the memory and supplies an input to the adder. The analysis stage AS comprises a calculation stage E which receives the output of the quantizer Q and supplies an input to a first network L1 and a second network L2. A decision means S receives the output of the first network L1 as well as an input NDMAX from the buffer control BC. An integrator I receives the output of the second network L2 and supplies an input to the classification stage K which receives an input T from the buffer control BC as well as an input NDMAX from the buffer control BC. The buffer control also supplies an input ≠ to the quantizer Q as illustrated. The entropy coding device HC supplies an output to the output buffer B which produces the output channel signal which is to be transmitted and also supplies an input to the buffer control BC.
The receiver buffer BE receives the incoming channel signal and supplies it to a decoder DC which supplies an output to the reconstruction means R. A receiver buffer control means BCE receives an output from the receiver buffer BE and supplies an input to the reconstruction means R. The decoder DC supplies an input to a receiver summing means +E which also receives the output of reconstruction means R. The transformed signal appears at the output of the receiver summing means +E and the output of the receiver summing means +E is supplied to a receiver image storer ME which also supplies an input to the receiver summing means +E.
The incoming frames are two-dimensionally cosine transformed in blocks (block size 16×16 picture points). The block size 8×8 can be simply realized by modification of Huffman code tables 1B and of the bit allocation matrices, Table 2 attached. The difference between the spectral coefficients thus obtained and the corresponding coefficients in the DPCM memory M is then quantized in block Q according to the quantization interval Δ determined by the buffer control.
The energy calculation stage E is then defined for each coefficient group as illustrated in FIG. 5A from the quanzation prediction error signal ΔyQ (u, v, t). ##EQU1##
It is assured by the limit function fA (x) that the result t of ΔYQ 2 is not represented with more bits than needed for further processing. The accumulator employed for the summation likewise has only twelve bits whereby a thirteenth bit is set to "1" and remains as soon as overflow has once occured.
The energies E(i) obtained in this manner are forwarded to the decision means S through a network L1. L1 limits the amplitude range to E*(i)·(0≦E*(i)≦16) so that E*(i) can be represented with 5 bits.
Whether a coefficient group is to be transmitted is determined in the stage S for every coefficient group on the basis of its energy by comparison to thresholds deposited in table form. The number of the first coefficient group to be transmitted supplies NO whereas the number of the last coefficient group to be transmitted supplies ND. When NO <4 then it is equated with "1". In case no coefficient group to be transmitted has been found, NO and ND are equated with "1". It is therefore assured that the block is classified as unmoved given the classification K as well. The buffer control can influence the rate by assigning the maximum plurality of coefficient groups.
In case that ND is greater than a value NDMAX prescribed by the buffer control, then ND=NDMAX is to be set.
The output of the decision means S is forwarded for classification K to the classification means K and to a coding means HC and to a conditioned adder (+).
The output of the calculation stage E supplies the energies E(i) to an integrator I through a network L2 which cuts off or truncates the least four significant bits. The integrator forms the signal EI(i) from E(i) according to the following equation.
EI (i)=EI (i-1)+E(i) i=2 2, . . . , 31 and EI (1)=E(1) (2 )
Only the bits having the significance of 0 . . . 7 are thereby taken into consideration in the addition, whereas bit 8, OR-operated with the overflow bit of the adder, yields the bit 8 of the accumulator, so that 9-bit code words are again present at the output of the integrator I.
The classification stage K executes the following operations:
EI (ND)>T (barrier T=0,1,2,3 prescribed by the buffer control) → block moved
EI (ND)≦T→block unmoved (3)
By cutting off or truncating the four least-significant bits in the energy calculation, the four values of the barrier or threshold T specified in the relationship of equation (3) result from FIG. 6a as shown by curve K1.
When the block is unmoved, it is assigned to the "unmoved" class 4. When the block is moved and thus, is to be transmitted, then the energy of the supergroup to be transmitted is defined as: ##EQU2## and with the assistance of EG, the block is assigned to one of three "moved" classes.
The two necessary class boundaries G(NO, ND, 1) and G(NO, ND, 2) are identified in the following fashion: ##EQU3## where EH is the mean energy variance presumed in the generation of the Huffman code tables B(i, j, k) is the allocation matrix f of the Huffman code tables for class K (table 2),
EHg is the energy up to the diagonal ND averaged over class k and k+1. ##EQU4##
The case discrimination and the calculation of G(NO, ND,k) and EG results that the like component in all classes is coded with the same Huffman code table for maximum variance. Its energy therefore remains unconsidered in the classification. The supergroup to be transmitted is then coded in the entropy coding means HC and written into the output buffer B. The code tables 1-7 of table 1 are employed therefore and these being selected for every coefficient via the allocation matrices in table 2. So-called "modified" Huffman codes are employed in the coding. Values /y/≦yesc are thereby Huffman-coded. Given /y/>yesc, an escape word is transmitted followed by the value of y in the natural code. The quantization interval can assume the values Δo, Δo /2,Δo /4,Δo /8. Amplitude levels of 255, 511, 1023 and 2047 correspond to these values. These natural code words therefore have different lengths (8, 9, 10, 11 bits).
The class affiliation and the supergroup (NO, ND) must be additionally transmitted for every block. The following bit rates are required for this overhead:
First case: 2 bits when k=4 ("unmoved" class)
Second case: 2 bits + the average word length indicated in table 1B ("Huffman" code tables for supergroup and class when k=1 through 3.
Last, the DPCM memory is brought to the current reading. The supergroup and the class affiliation are therefore to be considered as: ##EQU5## Y'(u,v,t)=Y'(u,v,t-1)+ΔyQ (u,v t) otherwise.
(Buffer Control)
As set forth above, a constant channel rate is achieved by modification of the barrier "moved"/"unmoved" T, of the quantization interval Δ and of the diagonal NDMAX to be maximally transmitted. The values T, Δ, NDMAX illustrated in FIG. 6b are identified by the non-linear characteristics K1 through K3 depending on the filling of the buffer.
In the region Bn <B(k)≦1 (characteristic K3), the rate is controlled via NDMAX.
No control results for BΔ≦B(k)≦1. When 0≦B(k)<BΔ occurs, then the buffer control results via the quantization interval Δ shown in FIG. 6c. Δ can thereby only assume values that meet the following inequality.
0≦int (1dΔo/Δ))≦3 (8)
As previously set forth, the quantization interval must also be considered in the coding.
Assuming a very full buffer B(k)>BT illustrated by the characteristic K1 in FIG. 6a, the barrier T is also raised by a quadratic characteristic K1. The raising of the barrier "moved"/"unmoved" then occurs in two ways:
(a) increase of T by the characteristic K1
(b) reduction of NDMAX and, thus, of the total energy (EG) by characteristic K3.
A very efficient noise suppression with full buffer is achieved by means of (b).
Significant innovations over known methods of the present invention are:
(1) Buffer Control
Way of limiting the bit rate by omitting coefficient groups when the buffer runs full. The number of coefficient groups is therefore controlled with a proportional controller.
Control of the rate by way of the quantization interval given a buffer running empty with a proportional controller. The quantization interval can therefore only assume the values of Δo, Δo /2,Δo /4,Δo /8.
Way of recognizing altered blocks by calculation of the energy of the quantitized signal. This is identical to a coupling of the "moved"/"unmoved" T to the quantization interval. A good value for T is T=N2 /12. This T is constant over a wide range of the buffer filling.
Raising the barrier T given buffer running full by way of a quadratic characteristic.
The controller for the number of diagonals to be maximally transmitted and the controller for the quantization interval never work in common but only respectively one operates dependent on the fill of the buffer.
The fact that only the coefficient groups that are transmitted are taken into consideration for the modification recognition for blocks is an advantage.
2. Coding
Adaptive Huffman coding by fixed allocation of the Huffman code tables for three "moved" categories (previously there has been dynamic allocation of the Huffman code tables (HCT) /1/ and fixed allocation of non-linear optimum n-bit maxquantizers /2/).
Classification on the basis of the quantized signal.
Way of identifying the class boundaries from the allocation of the Huffman code tables and the variance for which the HCT are generated.
Way of recognizing and coding modified supergroups within a block (in /1/, by length of run coding and end of block code word).
The fact that only the coefficient groups that are transmitted are considered for the modified recognition of blocks.
The tables which are utilized in this invention follow.
TABLE 1______________________________________Huffman Code Table______________________________________(A) For CoefficientsGiven code word numbers unequal to zero and Huffman codeof the Operation sign is appended to the tables:(1) Less Than Zero VZ = 1(2) Greater Than Zero VZ = 0(The code word length in the corresponding code words istherefore greater than the length of the code in the table.)Code Table Number: 1Number of words: 511Scanner: 0.75Residual Probability 0.00100Actual Residual Probility 0.00021Mean Word Length: 1.8520Entropy: 1.6386______________________________________Code Word Number Huffman Code Word Length Probability______________________________________0 0 1 0.6097491 10 3 0.1654092 110 4 0.0251913 1110 5 0.0038374 11110 6 0.0005845 Escape Word 11111 14 0.0000896 ESC 14 0.0000147 ESC 14 0.0000028 ESC 14 0.0000009 ESC 14 0.00000010 ESC 14 0.000000and so forth until NW/2 = 255______________________________________Code Table Number 2Number of words 511Scanner 1.50Residual Probability 0.00100Actual Residual Probability 0.00086Mean Word Length 2.6491Entropy 2.5680______________________________________Code Word Number Huffman Code Word length Probability______________________________________0 00 2 0.3752991 1 2 0.1904552 010 4 0.0743253 0110 5 0.0290064 01110 6 0.0113195 011110 7 0.0044176 0111110 8 0.0017247 01111110 9 0.0006738 Escape Word 01111111 17 0.0002639 ESC 17 0.00010210 ESC 17 0.00004011 ESC 17 0.00001612 ESC 17 0.00000613 ESC 17 0.000002and so forth until NW/2 = 255______________________________________Code Table Number: 3Number of Words: 511Scanner: 3.00Residual Probability 0.00100Actual Residual Probability 0.00068Entropy 3.5416______________________________________Code Word Number Huffman Code Word Length Probability______________________________________0 00 2 0.2096201 10 3 0.1483142 110 4 0.0926523 010 4 0.0578804 1110 5 0.0361585 0110 5 0.0225886 11110 6 0.0141117 01110 6 0.0088158 111110 7 0.0055079 011110 7 0.00344010 1111110 8 0.00214911 0111110 8 0.00134212 11111110 9 0.00083913 01111110 9 0.00052414 111111110 10 0.00032715 111111111 10 0.00020416 Escape Word 01111111 17 0.00012817 ESC 17 0.00008018 ESC 17 0.00005019 ESC 17 0.00003120 ESC 17 0.00001921 ESC 17 0.000012and so forth until NW/2 = 255______________________________________Code Table Number: 4Number of Words: 511Scanner: 6.00Residual Probability 0.00800Actual Residual Probability 0.00636Mean Word Length 4.5988Entropy: 4,5335______________________________________0 000 3 0.1109671 01 3 0.0931792 100 4 0.0736473 110 4 0.0582094 1010 5 0.0460075 1110 5 0.0363636 0010 5 0.0287417 10110 6 0.0227168 11110 6 0.0179549 00110 6 0.01419110 101110 7 0.01121611 111110 7 0.00886512 001110 7 0.00700713 1011110 8 0.00553814 1111110 8 0.00437715 0011110 8 0.00346016 10111110 9 0.00273417 11111110 9 0.00216118 101111110 10 0.00170819 101111111 10 0.00135020 111111110 10 0.00106721 111111111 10 0.00084322 Escape Word 0011111 16 0.00066723 ESC 16 0.00052724 ESC 16 0.00041625 ESC 16 0.00032926 ESC 16 0.00026027 ESC 16 0.000206and so forth until NW/2 = 255______________________________________Code Table Number: 5Number of Words: 511Scanner: 12.00Residual Probability 0.00800Actual Residual Probability 0.00759Mean Word Length 5.5889Entropy 5.5313______________________________________0 0000 4 0.0571141 001 4 0.0523142 010 4 0.0465093 1000 5 0.0413484 1010 5 0.0367605 1100 5 0.0326816 1110 5 0.0290547 0110 5 0.0258308 10010 6 0.0229649 10110 6 0.02041610 11010 6 0.01815011 11110 6 0.01613612 00010 6 0.01434613 01110 6 0.01275414 100110 7 0.01133915 101110 7 0.01008016 110110 7 0.00896217 111110 7 0.00796718 000110 7 0.00708319 011110 7 0.00629720 1001110 8 0.00559821 1011110 8 0.00497722 1101110 8 0.00442523 1111110 8 0.00393424 0001110 8 0.00349725 0111110 8 0.00310926 10011110 9 0.00276427 10111110 9 0.00245728 11011110 9 0.00218529 00011110 9 0.00194230 00011111 9 0.00172731 01111110 9 0.00153532 100111110 10 0.00136533 101111110 10 0.00121334 101111111 10 0.00107935 110111110 10 0.00095936 011111110 10 0.00085337 011111111 10 0.00075838 1001111110 11 0.00067439 1001111111 11 0.00059940 1101111110 11 0.00053341 1101111111 11 0.00047442 Escape Word 1111111 16 0.00042143 ESC 16 0.00037444 ESC 16 0.00033345 ESC 16 0.00029646 ESC 16 0.00026347 ESC 16 0.000234and so forth until NW/2 = 255______________________________________Code Table Number: 6Number of Words: 511Scanner: 24.00Residual Probability 0.01000Actual Residual Probability 0.00989Mean Word Length: 6.5860Entropy: 6.5307______________________________________Code Word Number Huffman Code Word Length Probability______________________________________0 00000 5 0.0289761 0001 5 0.0277292 0010 5 0.0261453 0100 5 0.0246524 0110 5 0.0232445 10000 6 0.0219176 10010 6 0.0206657 10100 6 0.0194858 10110 6 0.0183729 11000 6 0.01732310 11010 6 0.01633311 11100 6 0.01540012 11110 6 0.01452113 00110 6 0.01369214 01010 6 0.01291015 01110 6 0.01217216 100010 7 0.01147717 100110 7 0.01082218 101010 7 0.01020319 101110 7 0.00962120 110010 7 0.00907121 110110 7 0.00855322 111010 7 0.00806523 111110 7 0.00760424 000010 7 0.00717025 001110 7 0.00676026 010110 7 0.00637427 011110 7 0.00601028 1000110 8 0.00566729 1001110 8 0.00534330 1010110 8 0.00503831 1011110 8 0.00475032 1100110 8 0.00447933 1101110 8 0.00422334 1110110 8 0.00398235 1111110 8 0.00375536 0000110 8 0.00354037 0011110 8 0.00333838 0101110 8 0.00314739 0111110 8 0.00296840 10001110 9 0.00279841 10001111 9 0.00263842 10011110 9 0.00248843 10111110 9 0.00234644 11001110 9 0.00221245 11011110 9 0.00208546 11101110 9 0.00196647 11111110 9 0.00185448 00001110 9 0.00174849 00111110 9 0.00164850 00111111 9 0.00155451 01011110 9 0.00146552 01111110 9 0.00138253 100111110 10 0.00130354 101111110 10 0.00122855 110011110 10 0.00115856 110111110 10 0.00109257 110111111 10 0.00103058 111011110 10 0.00097159 111111110 10 0.00091560 000011110 10 0.00086361 000011111 10 0.00081462 010111110 10 0.00076763 011111110 10 0.00072365 1001111111 11 0.00064366 1011111110 11 0.00060667 1100111110 11 0.00057268 1100111111 11 0.00053969 1110111110 11 0.00050870 1110111111 11 0.00047971 1111111110 11 0.00045272 1111111111 11 0.00042673 0101111110 11 0.00040274 0101111111 11 0.00037975 0111111110 11 0.00035776 0111111111 11 0.00033777 10111111110 12 0.00031878 10111111111 12 0.00029979 Escape Word 1010111 16 0.00028280 ESC 16 0.00026681 ESC 16 0.00025182 ESC 16 0.00023783 ESC 16 0.00022384 ESC 16 0.000210and so forth until NW/2 = 255______________________________________Code Table Number: 7Number of Words: 511Scanner: 48.00Residual Probability 0.03000Actual Residual Probability 0.02943Mean Word Length 7.6155Entropy: 7.5384______________________________________Code Word Number Huffman Code Word Length Probability______________________________________0 000000 6 0.0144531 00010 6 0.0141412 00100 6 0.0137353 00110 6 0.0133424 00111 6 0.0129595 01000 6 0.0125876 01010 6 0.0122267 01100 6 0.0118768 01110 6 0.0115359 100000 7 0.01120410 100010 7 0.01088311 100100 7 0.01057112 100110 7 0.01026813 101000 7 0.00997314 101010 7 0.00968715 101100 7 0.00940916 101110 7 0.00913917 110000 7 0.00887718 110010 7 0.00862319 110100 7 0.00837520 110110 7 0.00813521 111000 7 0.00790222 111010 7 0.00767523 111100 7 0.00745524 000001 7 0.00724125 000010 7 0.00703426 000110 7 0.00683227 001010 7 0.00663628 010010 7 0.00644629 010110 7 0.00626130 011010 7 0.00608131 011110 7 0.00590732 1000010 8 0.00573733 1000110 8 0.00557335 1001110 8 0.00525836 1010010 8 0.00510737 1010110 8 0.00496138 1011010 8 0.00481839 1011011 8 0.00468040 1011110 8 0.00454641 1100010 8 0.00441642 1100110 8 0.00428943 1101010 8 0.00416644 1101110 8 0.00404645 1110010 8 0.00393046 1110110 8 0.00381847 1111010 8 0.00370848 0000110 8 0.00360249 0001110 8 0.00349850 0010110 8 0.00339851 0100110 8 0.00330152 0100111 8 0.00320653 0101110 8 0.00311454 0110110 8 0.00302555 0111110 8 0.00293856 10000110 9 0.00285457 10001110 9 0.00277258 10010110 9 0.00269259 10011110 9 0.00261560 10011111 9 0.00254061 10100110 9 0.00246762 10101110 9 0.00239763 10111110 9 0.00232864 11000110 9 0.00226165 11001110 9 0.00219666 11010110 9 0.00213367 11011110 9 0.00207268 11011111 9 0.00201369 11100110 9 0.00195570 11101110 9 0.00189971 11110110 9 0.00184472 00001110 9 0.00179273 00011110 9 0.00174074 00101110 9 0.00169075 00101111 9 0.00164276 01011110 9 0.00159577 01101110 9 0.00154978 01111110 9 0.00150579 01111111 9 0.00146180 100001110 10 0.00141981 100011110 10 0.00137982 1001k01110 10 0.00133983 1001011111 10 0.00130184 101001110 10 0.00126385 101011110 10 0.00122786 101111110 10 0.00119287 101111111 10 0.00115888 110001110 10 0.00112589 110011110 10 0.00109290 110101110 10 0.00106191 110101111 10 0.00103192 111001110 10 0.00100193 111011110 10 0.00097294 111101110 10 0.00094495 111101111 10 0.00091796 000011110 10 0.00089197 000111110 10 0.00086698 000111111 10 0.00084199 010111110 10 0.000817100 010111111 10 0.000793101 011011110 10 0.000770102 011011111 10 0.000748103 1000011110 11 0.000727104 1000011111 11 0.000706105 1000011110 11 0.000686106 1000111111 11 0.000666107 1010011110 11 0.000647108 1010011111 11 0.000628109 1010111110 11 0.000610110 1010111111 11 0.000593111 1100011110 11 0.000576112 1100011111 11 0.000559113 1100111110 11 0.000543114 1100111111 11 0.000528115 1110011110 11 0.000513116 1110011111 11 0.000498117 1110111110 11 0.000484118 1110111111 11 0.000470119 0000111110 11 0.000456120 0000111111 11 0.000443121 Escape Word 11111 14 0.000431122 ESC 14 0.000418123 ESC 14 0.000406124 ESC 14 0.000395125 ESC 14 0.000383126 ESC 14 0.000372and so forth until NW/2 = 255______________________________________(B) Code tables for transmitted subregion and class affiliation:11 unmoved class00 Greatest Detail content01 Mean Detail content10 Smallest Detail contentIn the moved classes subregion is codes as follows:(1) NO = 1No. of diagonals equal code word number in"Huffman Code Table for subregion(2) Code Word number 32 escape word for:ND > 16 and simultaneous NO ≧ 4Escape word is transmitted first and ND is then transmittedwith 4 bits and NO transmitted with 5 bits total 16 bits.(3) Following Table valid for4 ≦ NO ≦ 16 and simultaneous 4 ≦ ND≦ 16The code word number in "Huffman code table for subregion"Possible combination for ND and NO then for subregions.______________________________________NO↓ND→ 4 5 6 7 8 9 10 11 12 13 14 15 16 4 33 34 35 36 37 38 39 40 41 42 43 44 45 5 46 47 48 49 50 51 52 53 54 55 56 57 6 58 59 60 61 62 63 64 65 66 67 68 7 69 70 71 72 73 74 75 76 77 78 8 79 80 81 82 83 84 85 86 87 9 88 89 90 91 92 93 94 9510 96 97 98 99 100 101 10211 103 104 105 106 107 10812 109 110 111 112 11313 114 115 116 11714 118 119 12015 121 12216 123______________________________________Huffman Code Table for subregionDivision Content: 6.94251Entropy: 5.29115Mid word length 5.34360______________________________________Code Word Number Huffman Code Word Length Probability______________________________________1 0000 4 0.0555312 0100 4 0.0555313 0101 4 0.0555314 0110 4 0.0555315 0111 4 0.0555316 1000 4 0.0555317 1001 4 0.0555318 1010 4 0.0555319 1011 4 0.05553110 1100 4 0.05553111 00010 5 0.02776612 11010 5 0.02776613 11011 5 0.02776614 11100 5 0.02776615 11101 5 0.02776616 11110 5 0.02776617 000110 6 0.01388318 111110 6 0.01388319 00011100 8 0.00347120 00011110 8 0.00347121 0001110100 10 0.00086822 0001110110 10 0.00086823 0001110111 10 0.00086824 0001111100 10 0.00086825 00011101010 11 0.00043426 00011111010 11 0.00043427 00011111011 11 0.00043428 00011111100 11 0.00043429 00011111101 11 0.00043430 00011111110 11 0.00043431 00011111111 11 0.00043432 1111110 7 0.00694133 001000 6 0.02776634 1111111 7 0.00694135 00100100 8 0.00694136 001001010 9 0.00347137 001001011 9 0.00347138 001001100 9 0.00347139 001001101 9 0.00347140 001001110 9 0.00347141 0010011110 10 0.00173542 0010011111 10 0.00173543 00101000000 11 0.00086844 00101000001 11 0.00086846 0010101 7 0.01388347 00101001 8 0.00694148 001010001 9 0.00347149 001011000 9 0.00347150 001011001 9 0.00347151 0010110100 10 0.00173552 00101000011 11 0.00086853 00101101010 11 0.00086854 00101101011 11 0.00086855 00101101100 11 0.00086856 001010000101 12 0.00043457 001011011010 12 0.00043458 0010111 7 0.01388359 001100000 9 0.00347160 001100001 9 0.00347161 001100010 9 0.00347162 0010110111 10 0.00173563 00110001100 11 0.00086864 00110001101 11 0.00086865 00110001110 11 0.00086866 00110001111 11 0.00086867 001011011011 12 0.00043468 001100100000 12 0.00043469 0011010 7 0.01388370 001100101 9 0.00347171 001100110 9 0.00347172 0011001001 10 0.00173573 00110010001 11 0.00086874 00110011100 11 0.00086875 00110011101 11 0.00086876 00110011110 11 0.00086877 001100100001 12 0.00043478 001100111110 12 0.00043479 00110110 8 0.00694180 001101110 9 0.00347181 0011011110 10 0.00173582 00110111110 11 0.00086883 00110111111 11 0.00086884 00111000000 11 0.00086885 00111000001 11 0.00086886 001100111111 12 0.00043487 001110000100 12 0.00043488 00111001 8 0.00694189 0011100010 10 0.00173590 00111000011 11 0.00086891 00111000110 11 0.00086892 00111000111 11 0.00086893 00111010000 11 0.00086894 001110000101 12 0.00043495 001110100010 12 0.00043496 00111011 8 0.00694197 00111010010 11 0.00086898 00111010011 11 0.00086899 00111010100 11 0.000868100 00111010101 11 0.000868101 001110100011 12 0.000434102 001110101100 12 0.000434103 00111000 9 0.003471104 00111010111 11 0.000868105 00111100100 11 0.000868106 00111100101 11 0.000868107 001110101101 12 0.000434108 001111001100 12 0.000434109 001111010 9 0.003471110 00111100111 11 0.000868111 00111101100 11 0.000868113 001111011010 12 0.000434114 001111100 9 0.003471115 00111101110 11 0.000868116 001111011011 12 0.000434117 001111011110 12 0.000434118 001111101 9 0.003471119 001111011111 12 0.000434120 000111010110 12 0.000434121 001111110 9 0.003471122 000111010111 12 0.000434123 001111111 9 0.003471______________________________________
TABLE 2______________________________________Allocation Matrices Fixed Allocation Huffman CodeTable for the Three Moved Classes______________________________________CLASS 17 6 5 4 4 3 3 3 2 2 2 2 1 1 1 16 5 4 4 3 3 3 2 2 2 2 1 1 1 1 15 4 4 3 3 3 2 2 2 2 1 1 1 1 1 14 4 3 3 3 2 2 2 2 1 1 1 1 1 1 14 3 3 3 2 2 2 2 1 1 1 1 1 1 1 13 3 3 2 2 2 2 1 1 1 1 1 1 1 1 13 3 2 2 2 2 1 1 1 1 1 1 1 1 1 13 2 2 2 2 1 1 1 1 1 1 1 1 1 1 12 2 2 2 1 1 1 1 1 1 1 1 1 1 1 12 2 2 1 1 1 1 1 1 1 1 1 1 1 1 12 2 1 1 1 1 1 1 1 1 1 1 1 1 1 12 1 1 1 1 1 1 1 1 1 1 1 1 1 1 11 1 1 1 1 1 1 1 1 1 1 1 1 1 1 11 1 1 1 1 1 1 1 1 1 1 1 1 1 1 11 1 1 1 1 1 1 1 1 1 1 1 1 1 1 11 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1CLASS 27 5 4 3 3 2 2 2 1 1 1 1 1 1 1 15 4 3 3 2 2 2 1 1 1 1 1 1 1 1 14 3 3 2 2 2 1 1 1 1 1 1 1 1 1 13 3 2 2 2 1 1 1 1 1 1 1 1 1 1 13 2 2 2 1 1 1 1 1 1 1 1 1 1 1 12 2 2 1 1 1 1 1 1 1 1 1 1 1 1 12 2 1 1 1 1 1 1 1 1 1 1 1 1 1 12 1 1 1 1 1 1 1 1 1 1 1 1 1 1 11 1 1 1 1 1 1 1 1 1 1 1 1 1 1 11 1 1 1 1 1 1 1 1 1 1 1 1 1 1 11 1 1 1 1 1 1 1 1 1 1 1 1 1 1 11 1 1 1 1 1 1 1 1 1 1 1 1 1 1 11 1 1 1 1 1 1 1 1 1 1 1 1 1 1 11 1 1 1 1 1 1 1 1 1 1 1 1 1 1 11 1 1 1 1 1 1 1 1 1 1 1 1 1 1 11 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1CLASS 37 4 3 2 2 1 1 1 1 1 1 1 1 1 1 14 3 2 2 1 1 1 1 1 1 1 1 1 1 1 13 2 2 1 1 1 1 1 1 1 1 1 1 1 1 12 2 1 1 1 1 1 1 1 1 1 1 1 1 1 12 1 1 1 1 1 1 1 1 1 1 1 1 1 1 11 1 1 1 1 1 1 1 1 1 1 1 1 1 1 11 1 1 1 1 1 1 1 1 1 1 1 1 1 1 11 1 1 1 1 1 1 1 1 1 1 1 1 1 1 11 1 1 1 1 1 1 1 1 1 1 1 1 1 1 11 1 1 1 1 1 1 1 1 1 1 1 1 1 1 11 1 1 1 1 1 1 1 1 1 1 1 1 1 1 11 1 1 1 1 1 1 1 1 1 1 1 1 1 1 11 1 1 1 1 1 1 1 1 1 1 1 1 1 1 11 1 1 1 1 1 1 1 1 1 1 1 1 1 1 11 1 1 1 1 1 1 1 1 1 1 1 1 1 1 11 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1______________________________________
Although the invention has been described with respect to preferred embodiments, it is not to be so limited as changes and modifications can be made which are within the full intended scope of the invention as defined by the appended claims.
标题 | 发布/更新时间 | 阅读量 |
---|---|---|
图像编码装置 | 2020-05-08 | 212 |
基于网络实现注意力机制的图像压缩方法 | 2020-05-08 | 330 |
音频信号的频谱的频谱系数的编码 | 2020-05-11 | 449 |
压缩/解压缩的装置和系统、芯片、电子装置、方法 | 2020-05-12 | 463 |
一种高保真的H.264/AVC视频三系数可逆隐写方法 | 2020-05-12 | 561 |
一种基于视频流的对象识别装置 | 2020-05-11 | 916 |
一种光场焦点堆栈图像序列编、解码方法、装置及系统 | 2020-05-11 | 580 |
插值滤波器的训练方法、装置及视频图像编解码方法、编解码器 | 2020-05-08 | 392 |
深度图像压缩方法及其装置、设备和存储介质 | 2020-05-08 | 763 |
基于Attention机制的训练图片压缩网络的构建方法及系统 | 2020-05-08 | 623 |
高效检索全球专利专利汇是专利免费检索,专利查询,专利分析-国家发明专利查询检索分析平台,是提供专利分析,专利查询,专利检索等数据服务功能的知识产权数据服务商。
我们的产品包含105个国家的1.26亿组数据,免费查、免费专利分析。
专利汇分析报告产品可以对行业情报数据进行梳理分析,涉及维度包括行业专利基本状况分析、地域分析、技术分析、发明人分析、申请人分析、专利权人分析、失效分析、核心专利分析、法律分析、研发重点分析、企业专利处境分析、技术处境分析、专利寿命分析、企业定位分析、引证分析等超过60个分析角度,系统通过AI智能系统对图表进行解读,只需1分钟,一键生成行业专利分析报告。