CN108737298B - SCMA blind detection method based on image processing - Google Patents

SCMA blind detection method based on image processing Download PDF

Info

Publication number
CN108737298B
CN108737298B CN201810293367.5A CN201810293367A CN108737298B CN 108737298 B CN108737298 B CN 108737298B CN 201810293367 A CN201810293367 A CN 201810293367A CN 108737298 B CN108737298 B CN 108737298B
Authority
CN
China
Prior art keywords
image
scma
dimensional
dimensional image
detection method
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201810293367.5A
Other languages
Chinese (zh)
Other versions
CN108737298A (en
Inventor
张川
杨超
徐炜鸿
尤肖虎
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.)
Southeast University
Original Assignee
Southeast 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 Southeast University filed Critical Southeast University
Priority to CN201810293367.5A priority Critical patent/CN108737298B/en
Publication of CN108737298A publication Critical patent/CN108737298A/en
Application granted granted Critical
Publication of CN108737298B publication Critical patent/CN108737298B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L25/00—Baseband systems
    • H04L25/02—Details ; arrangements for supplying electrical power along data transmission lines
    • H04L25/0202—Channel estimation
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L25/00—Baseband systems
    • H04L25/02—Details ; arrangements for supplying electrical power along data transmission lines
    • H04L25/0202—Channel estimation
    • H04L25/0238—Channel estimation using blind estimation

Landscapes

  • Engineering & Computer Science (AREA)
  • Power Engineering (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses an SCMA blind detection method based on image processing, which comprises the following steps: (1) setting a normalized SCMA system with (N, K) as a parameter, wherein the code dimension is K, the constellation dimension is N, and taking a node with the highest occurrence frequency as an adjacent pixel point in a two-dimensional image, constructing a corresponding mapping template to obtain a mapping template from a one-dimensional signal to the two-dimensional image; (2) after a two-dimensional image mapping template with system characteristics is formulated, mapping each code word in an original SCMA codebook into a KxK image by taking a corresponding two-dimensional image mapping template as a support; (3) and (3) pre-filtering and denoising the two-dimensional image obtained in the step (2), and carrying out SCMA (sparse code multiple access) code word detection on the processed image. The invention can greatly reduce the degree of dependence of the traditional detection method on channel estimation, and improve the error rate of detection in the presence of channel errors, and has higher throughput rate.

Description

SCMA blind detection method based on image processing
Technical Field
The invention relates to the technical field of wireless communication, in particular to an SCMA blind detection method based on image processing.
Background
In recent years, with the wide application of wireless communication technology in various fields, social communication requirements are rapidly increased rapidly, and the traditional communication technology cannot meet the social development requirements increasingly. According to the prediction of main operators and authoritative consultants, the mobile broadband service flow will increase by 1000 times in the next 10 years. To cope with the enormous communication pressure in the future, 5G has come to be produced as a completely new mobile communication technology. Among them, "Gbps user experience rate" will be the most critical technical indicator of 5G. To realize the ultra-high transmission rate, 5G is applied to important technologies such as large-scale antenna arrays, novel multiple access, ultra-dense networking, novel network architecture, full-spectrum access and the like. And a new multiple access technology, as one of the key technologies for 5G implementation, will play a crucial role in the whole system. The novel multiple access technology can carry out high-efficiency superposition transmission on the sending signals so as to further improve the access capability of the system, thereby ensuring the large-scale equipment connection requirement of the 5G network.
Retrospective multiple access techniques, which have undergone considerable evolution and change since their creation, have an irreplaceable significance in modern wireless communications. Conventional multiple access techniques include Frequency Division Multiple Access (FDMA), Time Division Multiple Access (TDMA), Code Division Multiple Access (CDMA), Space Division Multiple Access (SDMA), etc., while Orthogonal Frequency Division Multiple Access (OFDMA) is used in 4G techniques. Many of these multiple access techniques are multiple access techniques at the orthogonal level, and are greatly limited by the number of resources. In contrast, in the non-orthogonal multiple access technology, the number of access users can be multiplied by the number of resources, thereby effectively solving the bottleneck.
Sparse code division multiple access (SCMA) is a new type of multiple access technology expected to be used in 5G communications, and has important non-orthogonal characteristics. Theoretical analysis shows that the SCMA has extremely excellent overload bearing capacity and resource reuse capacity, compared with the traditional multiple access technology, the access amount of the SCMA can be improved by at least 50%, and the access amount can be further improved along with the increase of the number of the interconnection in the system. In the related document regarding SCMA, a decoding strategy based on maximum a posteriori probability is proposed for decoding, and an SCMA system using the decoding strategy can maintain a superior error rate level while ensuring a high level of access, but has an extremely high requirement on the accuracy of channel estimation, so that a blind detection technique that does not rely on channel estimation will exert great advantages in practical applications.
Disclosure of Invention
The technical problem to be solved by the invention is to provide an SCMA blind detection method based on image processing, which can greatly reduce the dependence degree of the traditional detection method on channel estimation, improve the error rate of detection when channel errors exist, and has high throughput rate.
In order to solve the technical problem, the invention provides an SCMA blind detection method based on image processing, which comprises the following steps:
(1) setting a normalized SCMA system with (N, K) as a parameter, wherein the code dimension is K, the constellation dimension is N, and taking a node with the highest occurrence frequency as an adjacent pixel point in a two-dimensional image, constructing a corresponding mapping template to obtain a mapping template from a one-dimensional signal to the two-dimensional image;
(2) after a two-dimensional image mapping template with system characteristics is formulated, mapping each code word in an original SCMA codebook into a KxK image by taking a corresponding two-dimensional image mapping template as a support;
(3) and (3) pre-filtering and denoising the two-dimensional image obtained in the step (2), and carrying out SCMA (sparse code multiple access) code word detection on the processed image.
Preferably, in step (3), the pre-filtering and denoising of the two-dimensional image includes two methods: full-variation two-dimensional filtering and image training of a convolutional neural network.
Preferably, the fully-variational two-dimensional filtering specifically comprises: d represents an area to be modified in an image, E represents a boundary area of the area to be modified, and assuming that the whole image is Ω ═ douc E, the following energy function can be defined:
R[u]＝∫∫Ωr[|Δu(x,y)|]dxdy
where Δ u (x, y) represents the gradient of the image function and r represents a non-negative real function; considering the influence of noise on the original image in practice, the expression is further converted into a constraint form, and the constraint problem can be processed by extreme value solution under an unconstrained condition, and finally the Euler-Lagrange equation of total variation de-noising is obtained:
Figure BDA0001618148120000021
the euler-lagrange equation is a nonlinear partial differential equation, so that the processing form of the euler-lagrange equation needs to be simplified through a digitalized differential equation, a half-lattice point approximation mode is adopted, and the step length h is required to be 1, so that the following iterative solution formula can be obtained:
Figure BDA0001618148120000022
Figure BDA0001618148120000031
in the practical application process, the iteration number is set to be N equal to 100, the integration step length is set to be dt equal to 0.01, the constant K is 50 and is used for adjusting the calculation of the gradient value, after the global noise reduction processing of total variation is carried out, the image is restored to be a one-dimensional signal by a receiving end, and the one-dimensional signal is detected by using the traditional DMPA detection method, wherein the DMPA detects the channel environment N0Set to a constant for processing.
Preferably, the image training of the convolutional neural network specifically comprises: the convolutional neural network is a multi-layer structure, where each layer performs the following operations on its input:
Figure BDA0001618148120000032
wherein Wi,jAn ith row and a jth column of the trained filter coefficient matrix are shown, X represents the input image matrix of the layer, and after convolution operation and bias addition, the nonlinear operation is carried out on the matrix, wherein sigma (i) is max (0, i);
the filtering effect on the SCMA received image is determined by the filter coefficients and the bias, in order to obtain better filter coefficients, the noise reduction convolutional neural network is usually trained by using reverse error propagation, and the filter effect is measured by using the following loss function:
Figure BDA0001618148120000033
where F (x; θ) represents the noise estimate for input x obtained at filter coefficient θ ═ W, bias ], and y represents the original noise-free signal transmitted; the training aim is to minimize the loss function, and the neural network is trained by utilizing back propagation in deep learning and a Mini-batch stochastic gradient descent algorithm to obtain the optimal combination of filter coefficients and bias theta [ W, bias ].
Preferably, the SCMA codeword detection on the processed image specifically includes: the SCMA code word detection hardware structure comprises an image processing module and a DMPA detection algorithm module, wherein the image processing module comprises an image training module and an image conversion module; the received two-dimensional image is subjected to convolutional neural network to obtain a denoised 4 multiplied by 4 preprocessed image, which relates to multiplication and addition operations; the preprocessed image reaches an image conversion module, the repeated 4 pixel values are averaged, and the average value is compressed and converted into a one-dimensional 1 multiplied by 4 complex signal; the complex signal enters a DMPA detection algorithm module to perform four-step detection processes and judge and output the original sending code word.
Preferably, the four-step detection algorithm comprises the following steps:
(31) initializing; if an (K, N) SCMA system is set, the SCMA code dimension is K, and the constellation dimension is N, the maximum number of user layers of the system is
Figure BDA0001618148120000041
Wherein, the (K, N) SCMA system is (4,2), and the maximum number of users is 6;
Figure BDA0001618148120000042
where y iskRepresenting the k-th bit, x, of the received signalk,1,xk,2,xk,3Respectively representing 3 bits, N, of the overlapping of the 3 user layers connected to the kth resource node0,kThe expressed power density is the power density of Gaussian noise in the environment corresponding to the kth resource node;
(32) updating the resource nodes;
Figure BDA0001618148120000043
Figure BDA0001618148120000044
Figure BDA0001618148120000045
wherein R iskRepresented by the kth resource node, m1,m2,m 31, M represents different transmit symbols for each of three user layers connected to the resource node;
Figure BDA0001618148120000046
representing confidence values passed to the resource node from those user layers connected to the resource node,
Figure BDA0001618148120000047
the confidence values that represent the communication from those resource nodes connected to the user layer are in the opposite communication direction;
(33) updating layer nodes;
Figure BDA0001618148120000048
Figure BDA0001618148120000049
wherein, M is 1.. M represents different symbols in the transmission symbol set; after the step (3) is finished, returning to the step (2), and forming one iteration; when the iteration number meets the convergence requirement, the step (3) enters the step (4);
(4) probability calculation and symbol judgment;
Figure BDA0001618148120000051
wherein L isjRepresents the jth user layer; and selecting the symbol with the maximum probability value of each user layer, namely the finally estimated original transmission symbol.
The invention has the beneficial effects that: the filtering and noise reduction effects are enhanced by utilizing rich relevance among pixel points in a two-dimensional image, the high dependence of the traditional DMPA detection of an SCMA system on channel information is greatly reduced, the detection performance of the DMPA is greatly improved under the condition that channel estimation has deviation, series optimization on time sequence and resource reuse is carried out aiming at a hardware framework of a blind detection algorithm, and a detector framework with high universality is obtained.
Drawings
FIG. 1 is a factor diagram of a 6-user normalized SCMA system of the present invention.
FIG. 2 is a schematic diagram of a 4 × 4 sudoku form template after mapping into a two-dimensional image according to the present invention.
FIG. 3(a) is a schematic diagram illustrating the principle of pre-filtering by using a full-variational image processing technique according to the present invention.
FIG. 3(b) is a schematic diagram illustrating the principle of pre-filtering by using a full-variational image processing technique according to the present invention.
FIG. 4 is a schematic diagram illustrating the principle of image training denoising by using a convolutional neural network according to the present invention.
Fig. 5 is a schematic diagram illustrating comparison between the blind detection method based on the image processing technology and the conventional detection method.
FIG. 6 is a diagram of a general hardware architecture of the SCMA blind detection technique based on image processing according to the present invention.
Detailed Description
An SCMA blind detection method based on image processing comprises the following steps:
(1) setting a normalized SCMA system with (N, K) as a parameter, wherein the code dimension is K, the constellation dimension is N, and taking a node with the highest occurrence frequency as an adjacent pixel point in a two-dimensional image, constructing a corresponding mapping template to obtain a mapping template from a one-dimensional signal to the two-dimensional image;
(2) after a two-dimensional image mapping template with system characteristics is formulated, mapping each code word in an original SCMA codebook into a KxK image by taking a corresponding two-dimensional image mapping template as a support;
(3) and (3) pre-filtering and denoising the two-dimensional image obtained in the step (2), and carrying out SCMA (sparse code multiple access) code word detection on the processed image.
Let the normalized SCMA system with (N, K) as parameters, the code dimension is K and the constellation dimension is N. For convenience of explanation, N-4, K-2, i.e., 6-user SCMA standardized system is taken as an example. Figure 1 shows a 6-user SCMA system factor graph showing that each node has the same degree of ingress and egress due to high degree of standardization, and each node is indirectly connected to each other node. In order to observe that the connected nodes appear in the form of adjacent pixel points on the two-dimensional image, the other three nodes should be fairly distributed around each node, so that the 4 × 4 independent construction form template shown in fig. 2 can be obtained, the form of the template is not unique, and the SCMA system with other different parameters can construct the corresponding mapping template by using the node with the highest frequency of occurrence as the adjacent pixel point in the two-dimensional image through a similar method. During construction of the template, the following rules may simply be followed:
1) the size of the template needs to be adapted to the actual requirement of a corresponding SCMA system, and huge operation complexity caused by overlarge size is avoided, or the undersize size cannot cover enough node interconnection relations of one-dimensional signals;
2) the node interconnection relation of the one-dimensional signals needs to be fully embodied on the two-dimensional image after the mapping is completed, the characteristic information needs to be kept as complete as possible, and the loss of information of the constructed image is avoided;
3) the two-dimensional image after mapping needs to have a proper processable structure, the content of the two-dimensional image includes removal of redundancy and reasonable repetition of key information, and the two-dimensional image is embodied in practical application, namely node information is repeated to a certain degree, and occurrence of 0 pixel in a constructed image is avoided as far as possible.
After a two-dimensional image template with system characteristics is formulated, each code word in the original SCMA codebook is mapped into a 4 x 4 image by taking a corresponding template as a support, wherein only the two values, namely R and G, in the three values corresponding to RGB are used for respectively storing a real number part and an imaginary number part of the corresponding original code word. The specific implementation means of the method can be divided into full-variation image modification and convolutional neural network image training.
Fig. 3(a) shows a schematic diagram of pre-filtering by using a fully-variational image processing technique, where D denotes an area to be modified in an image, E denotes a boundary area of the area to be modified, and assuming that the whole image is Ω ═ due, the following energy function can be defined:
R[u]＝∫∫Ωr[|Δu(x,y)|]dxdy
where Δ u (x, y) represents the gradient of the image function and r represents a non-negative real function. Considering the influence of noise on the original image in practice, the expression is further converted into a constraint form, and the constraint problem can be processed by extreme value solution under an unconstrained condition, and finally the Euler-Lagrange equation of total variation de-noising is obtained:
Figure BDA0001618148120000071
the euler-lagrange equation is a nonlinear partial differential equation, and thus it is necessary to simplify its processing form by a digitized differential equation. The principle of this simplified form is shown in fig. 3(b), and an approximation of half-grid point is adopted, and the step length h is 1, so that the following iterative solution formula can be obtained:
Figure BDA0001618148120000072
Figure BDA0001618148120000073
in the practical application process, the iteration number is set to be N equal to 100, the integration step length is set to be dt equal to 0.01, the constant K is 50 to adjust the calculation of the gradient value, after the global noise reduction processing of total variation is performed, the image is restored to be a one-dimensional signal by the receiving end, and the detection is performed by using the traditional DMPA detection method, but the DMPA is used for detecting the channel environment N0The requirement of (2) is not harsh, the channel estimation can be not needed, and the setting is broadly a constant for processing. A specific performance comparison is shown in fig. 5.
The convolutional neural network is a multi-layer structure, where each layer performs the following operations on its input:
Figure BDA0001618148120000074
wherein Wi,jThe ith row and the jth column of the trained filter coefficient matrix are represented, X represents the image matrix of the layer input, and the image matrix is subjected to convolution operation and added with biasAfter bias, we perform a nonlinear operation σ (·), where σ (i) is max (0, i).
The filtering effect on the SCMA received image is determined by the filter coefficients and the bias, and in order to obtain better filter coefficients, the noise reduction convolutional neural network is usually trained by using inverse error propagation. The following loss function is used to measure the filter effect:
Figure BDA0001618148120000075
where F (x; θ) represents the noise estimate for input x obtained at filter coefficient θ ═ W, bias ], and y represents the original noise-free signal transmitted. The goal of the training is to minimize this loss function. And (3) training the neural network by utilizing Back propagation (Back propagation) and Mini-batch stochastic gradient descent (Mini-batch stochastic gradient device) algorithms in deep learning, and obtaining the optimal combination theta of the filter coefficient and the bias as [ W, bias ].
Fig. 6 shows a general hardware architecture of the proposed SCMA blind detection technique, which can be roughly divided into an image processing module and a DMPA detection algorithm module, i.e. an upper part and a lower part in the figure, where the image processing module specifically includes an image training module and an image transformation module. The general process flow of the architecture can be described as follows: the received two-dimensional image is subjected to convolutional neural network to obtain a denoised 4 multiplied by 4 preprocessed image, which mainly relates to multiplication and addition operations; the preprocessed image reaches an image conversion unit, the repeated 4 pixel values are averaged, and the average value is compressed and converted into a one-dimensional 1 multiplied by 4 complex signal; the complex signal enters a DMPA detection algorithm module to perform four-step detection processes and judge and output the original sending code word. The main occupation module of the hardware resource is a DMPA detection emission module, the detection mode directly determines the design form of the hardware architecture, and the detection algorithm of the four steps can be described as follows:
(1) initializing; if an (K, N) SCMA system is set, the SCMA code dimension is K, and the constellation dimension is N, the maximum number of user layers of the system is
Figure BDA0001618148120000081
Wherein, the (K, N) SCMA system is (4,2), and the maximum number of users is 6;
Figure BDA0001618148120000082
where y iskRepresenting the k-th bit, x, of the received signalk,1,xk,2,xk,3Respectively representing 3 bits, N, of the overlapping of the 3 user layers connected to the kth resource node0,kThe expressed power density is the power density of Gaussian noise in the environment corresponding to the kth resource node;
(2) updating the resource nodes;
Figure BDA0001618148120000083
Figure BDA0001618148120000084
Figure BDA0001618148120000085
wherein R iskRepresented by the kth resource node, m1,m2,m 31, M represents different transmit symbols for each of three user layers connected to the resource node;
Figure BDA0001618148120000086
representing confidence values passed to the resource node from those user layers connected to the resource node,
Figure BDA0001618148120000091
the confidence values that represent the communication from those resource nodes connected to the user layer are in the opposite communication direction;
(3) updating layer nodes;
Figure BDA0001618148120000092
Figure BDA0001618148120000093
wherein, M is 1.. M represents different symbols in the transmission symbol set; after the step (3) is finished, returning to the step (2), and forming one iteration; when the iteration number meets the convergence requirement, the step (3) enters the step (4);
(4) probability calculation and symbol judgment;
Figure BDA0001618148120000094
wherein L isjRepresents the jth user layer; and selecting the symbol with the maximum probability value of each user layer, namely the finally estimated original transmission symbol.
The SCMA blind detection hardware architecture is designed to comprehensively use various hardware optimization design technologies including folding, retiming and pipelining, so that the general architecture enjoys high hardware efficiency and can meet the data throughput rate of practical application. The image processing module can be roughly divided into an image processing module and a DMPA detection algorithm module, and the image processing module specifically comprises an image training module and an image conversion module. The general process flow of the architecture can be described as follows: the received two-dimensional image is subjected to convolutional neural network to obtain a denoised 4 multiplied by 4 preprocessed image, which mainly relates to multiplication and addition operations; the preprocessed image reaches an image conversion unit, the repeated 4 pixel values are averaged, and the average value is compressed and converted into a one-dimensional 1 multiplied by 4 complex signal; the complex signal enters a DMPA detection algorithm module to perform four-step detection processes and judge and output the original sending code word. The main occupation module of the hardware resource is a DMPA detection and emission module, and the detection mode directly determines the design form of the hardware architecture.
According to the interconnection relationship of each node in the SCMA factor graph, the adjacent arrangement relationship between each pixel point after the node is mapped into the two-dimensional image is judged, and due to the node equivalence and the sufficient similarity of the normalized SCMA system, an individual node has the adjacent relationship with other nodes, so that the 4 x 4 pixel arrangement template of the independent structure with a non-unique form is obtained through deduction. Other SCMA systems can also use the highest occurrence rate as the adjacent pixel on the projection template by counting the entrance and exit of each node and marking the adjacent nodes on the factor graph, thereby obtaining the mapping template of the corresponding one-dimensional signal to the two-dimensional image.
Based on the obtained mapping template, all one-dimensional signals of a codebook in an original SCMA system are converted into corresponding two-dimensional image forms, and pre-filtering and de-noising processing are carried out before codeword detection by relying on an image processing technology. Under the processing of the two methods, the dependence degree of the original detection method on the accuracy of channel estimation is reduced.
The processing mode based on the total variation image retouching technology is as follows: the method comprises the steps of carrying out mathematical processing based on a first-order differential operator on the basis of a received noisy image, calculating gradient change between adjacent pixel points, calculating in an iteration mode by adopting a half-grid point approximation processing mode in specific operation, and continuously reducing pixel value difference of an adjacent area in iteration on the basis of the calculation, so that the effects of smoothing and noise reduction are achieved, and support is provided for follow-up blind detection.
The graph training based on the convolutional neural network has the processing mode as follows: firstly, a multilayer Convolutional Neural Network (CNN) is constructed, and the convolutional neural network is utilized to carry out two-dimensional convolution processing on a received noisy image for multiple times, so that self-adaptive filtering is realized, the output of the convolutional neural network is the image subjected to noise reduction, and the effects of smoothing and noise reduction are achieved.
The hardware architecture related to blind detection can be roughly divided into an image processing module and a traditional detection module, and a hardware design means including folding and retiming technologies is applied to obtain a universal blind detection architecture scheme. In the actual design, the flow direction of data is considered, and the framework is optimized in a pipeline operation mode, so that higher hardware efficiency and data throughput rate are ensured.

Claims (5)

1. An SCMA blind detection method based on image processing is characterized by comprising the following steps:
(1) setting a normalized SCMA system with (N, K) as a parameter, wherein the code dimension is K, the constellation dimension is N, and taking a node with the highest occurrence frequency as an adjacent pixel point in a two-dimensional image, constructing a corresponding mapping template to obtain a mapping template from a one-dimensional signal to the two-dimensional image;
(2) after a two-dimensional image mapping template with system characteristics is formulated, mapping each code word in an original SCMA codebook into a KxK image by taking a corresponding two-dimensional image mapping template as a support;
(3) pre-filtering and denoising the two-dimensional image obtained in the step (2), and carrying out SCMA (sparse code multiple access) code word detection on the processed image; the SCMA codeword detection of the processed image is specifically as follows: the SCMA code word detection hardware structure comprises an image processing module and a DMPA detection algorithm module, wherein the image processing module comprises an image training module and an image conversion module; the received two-dimensional image is subjected to convolutional neural network to obtain a denoised 4 multiplied by 4 preprocessed image, which relates to multiplication and addition operations; the preprocessed image reaches an image conversion module, the repeated 4 pixel values are averaged, and the average value is compressed and converted into a one-dimensional 1 multiplied by 4 complex signal; the complex signal enters a DMPA detection algorithm module to perform four-step detection processes and judge and output the original sending code word.
2. The image-processing-based SCMA blind detection method according to claim 1, wherein the pre-filtering and denoising of the two-dimensional image in step (3) comprises two methods: full-variation two-dimensional filtering and image training of a convolutional neural network.
3. The SCMA blind detection method based on image processing as claimed in claim 2, wherein the fully-variational two-dimensional filtering specifically is: d represents an area to be modified in an image, E represents a boundary area of the area to be modified, and assuming that the whole image is Ω ═ douc E, the following energy function can be defined:
R[u]＝∫∫Ωr[|Δu(x,y)|]dxdy
where Δ u (x, y) represents the gradient of the image function and r represents a non-negative real function; considering the influence of noise on the original image in practice, the expression is further converted into a constraint form, and the constraint problem can be processed by extreme value solution under an unconstrained condition, and finally the Euler-Lagrange equation of total variation de-noising is obtained:
Figure FDA0002820146350000011
the euler-lagrange equation is a nonlinear partial differential equation, so that the processing form of the euler-lagrange equation needs to be simplified through a digitalized differential equation, a half-lattice point approximation mode is adopted, and the step length h is required to be 1, so that the following iterative solution formula can be obtained:
Figure FDA0002820146350000021
Figure FDA0002820146350000022
in the practical application process, the iteration number is set to be N equal to 100, the integration step length is set to be dt equal to 0.01, the constant K is 50 and is used for adjusting the calculation of the gradient value, after the global noise reduction processing of total variation is carried out, the image is restored to be a one-dimensional signal by a receiving end, and the one-dimensional signal is detected by using the traditional DMPA detection method, wherein the DMPA detects the channel environment N0Set to a constant for processing.
4. The image-processing-based SCMA blind detection method according to claim 2, wherein the image training of the convolutional neural network is specifically: the convolutional neural network is a multi-layer structure, where each layer performs the following operations on its input:
Figure FDA0002820146350000023
wherein Wi,jAn ith row and a jth column of the trained filter coefficient matrix are shown, X represents the input image matrix of the layer, and after convolution operation and bias addition, the nonlinear operation is carried out on the matrix, wherein sigma (i) is max (0, i);
the filtering effect on the SCMA received image is determined by the filter coefficients and the bias, in order to obtain better filter coefficients, the noise reduction convolutional neural network is usually trained by using reverse error propagation, and the filter effect is measured by using the following loss function:
Figure FDA0002820146350000024
where F (x; θ) represents the noise estimate for input x obtained at filter coefficient θ ═ W, bias ], and y represents the original noise-free signal transmitted; the training aim is to minimize the loss function, and the neural network is trained by utilizing back propagation in deep learning and a Mini-batch stochastic gradient descent algorithm to obtain the optimal combination of filter coefficients and bias theta [ W, bias ].
5. The image processing based SCMA blind detection method according to claim 1, wherein its four-step detection algorithm comprises the following steps:
(31) initializing; if an (K, N) SCMA system is set, the SCMA code dimension is K, and the constellation dimension is N, the maximum number of user layers of the system is
Figure FDA0002820146350000031
Wherein, K is 4, N is 2, and the maximum number of users is 6;
Figure FDA0002820146350000032
where y iskRepresenting the k-th bit, x, of the received signalk,1,xk,2,xk,3Respectively representing 3 bits, N, of the overlapping of the 3 user layers connected to the kth resource node0,kThe expressed power density is the power density of Gaussian noise in the environment corresponding to the kth resource node;
(32) updating the resource nodes;
Figure FDA0002820146350000033
Figure FDA0002820146350000034
Figure FDA0002820146350000035
wherein R iskRepresented by the kth resource node, m1,m2,m31, M represents different transmit symbols for each of three user layers connected to the resource node;
Figure FDA0002820146350000036
representing confidence values passed to the resource node from those user layers connected to the resource node,
Figure FDA0002820146350000037
the confidence values that represent the communication from those resource nodes connected to the user layer are in the opposite communication direction;
(33) updating layer nodes;
Figure FDA0002820146350000038
Figure FDA0002820146350000039
wherein, M is 1.. M represents different symbols in the transmission symbol set; after the step (3) is finished, returning to the step (2), and forming one iteration; when the iteration number meets the convergence requirement, the step (3) enters the step (4); (4) probability calculation and symbol judgment;
Figure FDA0002820146350000041
wherein L isjRepresents the jth user layer; and selecting the symbol with the maximum probability value of each user layer, namely the finally estimated original transmission symbol.
CN201810293367.5A 2018-04-04 2018-04-04 SCMA blind detection method based on image processing Active CN108737298B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810293367.5A CN108737298B (en) 2018-04-04 2018-04-04 SCMA blind detection method based on image processing

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810293367.5A CN108737298B (en) 2018-04-04 2018-04-04 SCMA blind detection method based on image processing

Publications (2)

Publication Number Publication Date
CN108737298A CN108737298A (en) 2018-11-02
CN108737298B true CN108737298B (en) 2021-01-26

Family

ID=63941201

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810293367.5A Active CN108737298B (en) 2018-04-04 2018-04-04 SCMA blind detection method based on image processing

Country Status (1)

Country Link
CN (1) CN108737298B (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113822825B (en) * 2021-11-25 2022-02-11 电子科技大学成都学院 Optical building target three-dimensional reconstruction method based on 3D-R2N2

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105721106A (en) * 2016-01-27 2016-06-29 电子科技大学 Multiuser detection method based on serial strategy for SCMA (Sparse Code Multiple Access) uplink communication system
CN106301683A (en) * 2016-08-04 2017-01-04 东南大学 A kind of DMPA interpretation method based on SCMA system and decoder architecture

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9461765B2 (en) * 2007-03-27 2016-10-04 Hughes Networks Systems, Llc Method and system for providing scrambled coded multiple access (SCMA)

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105721106A (en) * 2016-01-27 2016-06-29 电子科技大学 Multiuser detection method based on serial strategy for SCMA (Sparse Code Multiple Access) uplink communication system
CN106301683A (en) * 2016-08-04 2017-01-04 东南大学 A kind of DMPA interpretation method based on SCMA system and decoder architecture

Also Published As

Publication number Publication date
CN108737298A (en) 2018-11-02

Similar Documents

Publication Publication Date Title
CN109951214B (en) Signal detection method suitable for large-scale MIMO system
US20230299872A1 (en) Neural Network-Based Communication Method and Related Apparatus
CN104580039B (en) The receiver detection method of the radio MIMO system of lattice reduction algorithm auxiliary
CN106130688B (en) A kind of Sparse Code multiple access access detection method of low complex degree
CN108737022B (en) Low-complexity SCMA decoding method and device based on quantum computation
CN110659684A (en) Convolutional neural network-based STBC signal identification method
WO2021110054A1 (en) Multi-layer neural network assisted penalty dual decomposition channel decoding method
TWI422182B (en) Geometry based efficient decoder for underdetermined mimo systems and decoding method thereof
CN108737298B (en) SCMA blind detection method based on image processing
CN109831281B (en) Multi-user detection method and device for low-complexity sparse code multiple access system
Azari et al. Automated deep learning-based wide-band receiver
Huang et al. D $^ 2$-JSCC: Digital Deep Joint Source-channel Coding for Semantic Communications
Chen et al. Optimal receiver design for SCMA system
Cao et al. Learning to denoise and decode: A novel residual neural network decoder for polar codes
CN108322290B (en) Iterative detection method and system for wireless communication
CN116628524A (en) Community discovery method based on adaptive graph attention encoder
CN114611718A (en) Federal learning method and system for heterogeneous data
Lu et al. Attention-empowered residual autoencoder for end-to-end communication systems
CN106911431B (en) Improved partial edge information transmission method applied to demodulation process of sparse code multiple access system
Niazadeh et al. ISI sparse channel estimation based on SL0 and its application in ML sequence-by-sequence equalization
CN111313914A (en) SCL simplified decoding method based on neural network classifier
CN117354654B (en) Multi-parameter optimizing method for mixed forming constellation based on learning rate dynamic clipping
WO2018019109A1 (en) Decoding method, device and system based on overlapped multiplexing
Duan et al. A model‐driven robust deep learning wireless transceiver
CN116488970A (en) Balancing method for high-speed optical communication system

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant