DNA image transmission method and device based on deep joint source channel coding

Abstract

The invention discloses a DNA image transmission method and device based on deep joint source channel coding, and relates to the field of image coding, and the method comprises the steps: S1, constructing an image coding model, a channel simulation and sequence preprocessing module and an image decoding model; s2, acquiring a training data set; s3, training and optimizing the image coding model and the image decoding model; s4, acquiring an image to be transmitted; s5, encoding the image encoding model; s6, a channel simulation and sequence preprocessing module performs channel simulation and sequence preprocessing to obtain a numerical matrix; s7, the image decoding model decodes and reconstructs the image; according to the method, through an end-to-end integrated design, image compression, base mapping, DNA channel simulation and biological constraint optimization are integrated into the same neural network framework, deep collaboration of source coding and channel coding is realized, a high-fidelity image can still be recovered in a noise environment, and the method is suitable for large-scale popularization and application. Therefore, the overall efficiency and reliability of the DNA image storage process are greatly improved.

Description

Technical Field

[0001] The present invention relates to the field of image coding, and in particular to a DNA image transmission method and device based on deep joint source-channel coding. Background Art

[0002] As the volume of digital information continues to grow exponentially, traditional storage media such as disks, tapes, and optical disks are facing serious bottlenecks in storage density, shelf life, and energy consumption, making it difficult to meet the needs of long-term, large-scale data archiving. In contrast, deoxyribonucleic acid (DNA), a naturally high-density, highly stable biological macromolecule, has the theoretical potential to store over 10 billion GB of data per gram and can be preserved passively for thousands of years or even longer, making it a research hotspot for next-generation archival storage media.

[0003] The typical DNA storage process involves mapping digital data into a base sequence (source encoding) through compression and error-correction coding; chemically synthesizing short oligonucleotides and amplifying them via polymerase chain reaction (PCR) to increase molecular abundance; using second- or third-generation sequencing technologies to read the noisy base sequence; and finally, recovering the digital information through channel decoding and source decoding. During the synthesis, amplification, sequencing, and sample processing processes, errors such as base insertions, deletions, and substitutions can occur, often requiring additional error-correction codes and sequence constraint encoding (such as run-length-limited or GC content constraints) to ensure reliability.

[0004] In the fields of wireless communications and image transmission, Deep Joint Source Channel Coding (Deep JSCC) utilizes neural networks to learn signal representations end-to-end, achieving performance advantages over traditional separate coding methods under bandwidth-constrained and noisy conditions. However, to date, no research has yet fully integrated Deep JSCC with DNA storage processes to form a comprehensive end-to-end joint encoding and decoding solution, which remains a gap in the DNA storage field.

[0005] Furthermore, DNA storage must meet a series of biological constraints: the GC content of the base sequence must be controlled between 40% and 60% to avoid sequencing bias caused by high or low GC content; the length of homopolymers (repeated copies of the same base) should be kept to less than 5 to reduce insertion / deletion errors. Existing methods often use hard constraints after encoding or offline sequence screening, lacking the ability to learn and optimize these constraints.

[0006] In summary, existing DNA storage technologies still face performance tradeoffs and efficiency bottlenecks in terms of separate coding and biological constraint processing. There is an urgent need for a deep joint coding scheme that can organically combine source coding, channel coding, and biological constraint optimization to achieve high-quality reconstruction and sequence biocompatibility with limited redundancy. Summary of the Invention

[0007] The purpose of the present invention is to design a DNA image transmission method and device based on deep joint source-channel coding in order to solve the above problems.

[0008] The present invention achieves the above-mentioned purpose through the following technical solutions:

[0009] The DNA image transmission method based on deep joint source-channel coding includes:

[0010] S1. Construct an image coding model, a channel simulation and sequence preprocessing module, and an image decoding model. The image coding model includes a normalization layer, a feature extraction network, and a base mapping module from input to output. The image decoding model includes a splicing layer, a multi-chain fusion module, a restoration network, and an inverse normalization layer from input to output. The feature extraction network is used to extract deep image features and flatten the one-dimensional vector to obtain a potential vector. The base mapping module is used to map the continuous potential vector output by the feature extraction network into a discrete DNA base sequence. The channel simulation and sequence preprocessing module is used to simulate the synthesis, PCR amplification, and sequencing errors in the DNA storage process to simulate the DNA base sequence, and standardize multiple simulated base sequences into a numerical matrix that can be processed by the neural network. The multi-chain fusion module fuses the sequence output by the splicing layer into a sequence of length The vector of the multi-chain fusion module is restored by the restoration network;

[0011] S2, obtain training data set;

[0012] S3. Perform end-to-end joint training optimization on the image encoding model and image decoding model using the training dataset;

[0013] S4, obtaining the image to be transmitted;

[0014] S5. Encode the image to be transmitted using the optimized image coding model to obtain a coding result;

[0015] S6. The encoding result is input into the channel simulation and sequence preprocessing module, and the channel simulation and sequence preprocessing are performed on the encoding result to obtain a numerical matrix;

[0016] S7. The numerical matrix is ​​input into the optimized image decoding model for decoding to obtain a reconstructed image.

[0017] A DNA image transmission device based on deep joint source-channel coding, comprising:

[0018] Storage; the storage stores at least one instruction, at least one program, code set or instruction set;

[0019] Processor; the at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by the processor to implement the above-mentioned DNA image transmission method based on deep joint source-channel coding.

[0020] A computer-readable storage medium stores at least one instruction, at least one program, a code set, or an instruction set, wherein the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by a processor to implement the above-mentioned DNA image transmission method based on deep joint source-channel coding.

[0021] The beneficial effect of the present invention is that: through an end-to-end integrated design, this method integrates image compression, base mapping, DNA channel simulation and biological constraint optimization into the same neural network framework, realizes the deep synergy of source coding and channel coding, and can still restore high-fidelity images in noisy environments, thereby greatly improving the overall efficiency and reliability of the DNA image storage process. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 Schematic diagram of the image coding model, channel simulation and sequence preprocessing module, and image decoding model in the present invention;

[0023] Figure 2 The biological constraints of the coding sequence are met under different weight coefficients in the loss function;

[0024] Figure 3 The image restoration quality comparison between the two variants of the present invention and the traditional coding scheme VAEU-QC at different coding rates, including PSNR and SSIM indicators. DETAILED DESCRIPTION

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings of the embodiments of the present invention. It should be understood that the described embodiments are only a portion of the embodiments of the present invention, not all of them. Generally, the components of the embodiments of the present invention described and illustrated in the drawings herein may be arranged and designed in a variety of different configurations.

[0026] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention as claimed, but rather merely represents selected embodiments of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort shall fall within the scope of protection of the present invention.

[0027] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings.

[0028] In the description of the present invention, it should be understood that the terms "upper", "lower", "inside", "outside", "left", "right", etc. indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, or are the orientations or positional relationships in which the inventive product is conventionally placed when in use, or are the orientations or positional relationships conventionally understood by those skilled in the art. These are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or component referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present invention.

[0029] Furthermore, the terms “first”, “second”, etc. are merely used for distinguishing descriptions and should not be understood as indicating or implying relative importance.

[0030] In the description of the present invention, it should also be noted that, unless otherwise expressly specified or limited, terms such as "disposed" and "connected" should be understood in a broad sense. For example, "connected" can mean a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can also mean internal communication between two components. Those skilled in the art will be able to understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0031] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0032] like Figure 1 、 Figure 2 、 Figure 3 As shown, the DNA image transmission method based on deep joint source-channel coding includes:

[0033] S1, build image coding model, channel simulation and sequence preprocessing module and image decoding model,

[0034] The graph coding model includes a normalization layer, a feature extraction network, and a base mapping module from input to output. The graph coding model maps the normalized original image into a latent representation vector of fixed length. , the coding rate is , the process can be expressed as: ,in, It is a pixel-by-pixel linear normalization operation, expressed as: The graphics coding model mainly consists of two parts: feature extraction: extracting deep image features through multi-layer Conv2d+BatchNorm+PReLU; vectorization: flattening the three-dimensional features output by the convolution layer into a one-dimensional vector , of length k, ; This latent vector not only contains the compressed image information, but also provides continuous and differentiable input for subsequent DNA base mapping.

[0035] The feature extraction network is used to extract deep image features and flatten the one-dimensional vector to obtain the latent vector. The feature extraction network includes the first convolution layer, the first activation layer, and the extraction module from input to output. The extraction module is connected to multiple extraction blocks in sequence. Each extraction block includes the second convolution layer, the first normalization layer, and the second activation layer from input to output.

[0036] The base mapping module is used to map the continuous potential vector output by the feature extraction network into a discrete DNA base sequence; specifically:

[0037] The base mapping module is used to convert the continuous latent vector output by the feature extraction network into Mapped to discrete DNA base sequences , specifically:

[0038] ①, Sigmoid activation and scaling: for each potential vector component Apply the Sigmoid function and map it to the (0,1) interval, expressed as: ,right Do the proportional amplification, expressed as: ;

[0039] ② Discretization: After magnification Perform the rounding operation to get the integer value, which is expressed as: ; To maintain the gradient path in the back propagation, The gradient is approximately 1;

[0040] ③、Integer→base mapping: Map the integer symbols {0,1,2,3} one by one to the four DNA bases to obtain the initial base sequence , the mapping is expressed as: ;

[0041] The base mapping module converts the continuous domain output of the feature extraction network into discrete bases that meet biological significance, ensuring the input correctness of subsequent DNA channel simulation and sequence processing.

[0042] The channel simulation and sequence preprocessing module is used to simulate the synthesis, PCR amplification and sequencing errors in the DNA storage process to simulate the DNA base sequence, and standardize multiple simulated base sequences into a numerical matrix that can be processed by the neural network;

[0043] The simulation is as follows: 、 and The insertion errors, deletion errors and substitution errors that may occur in the actual DNA synthesis, PCR amplification and DNA sequencing process are simulated on the DNA base sequence to obtain the simulation output sequence, which is expressed as ,copy Simulation output sequence ;

[0044] Insert error: at any position in the sequence, with probability Insert a random base, expressed as: ;

[0045] Deletion error: for each base By probability Delete, represented by: ;

[0046] Substitution error: for each base By probability Randomly replace with one of the other three bases, expressed as: ;

[0047] The specific standardization process is: for each simulation output sequence Perform trimming to unify each replicate sequence to the same length Matrix input; and map the character set {A, C, G, T, Q} to the integer set {1, 2, 3, 4, 0}; The length of each copy is The numerical sequence of stacked into a numerical matrix .

[0048] Channel simulation can realistically simulate the real errors in the DNA storage and reading process, and normalization processing ensures the consistency of all simulation outputs in length and numerical space, providing stable and batch-processable input for deep decoding networks.

[0049] The image decoding model includes the splicing layer, multi-chain fusion module, restoration network and inverse normalization layer from input to output; the image decoding model is to convert the preprocessed base sequence matrix Mapping back to a continuous image tensor , the whole process can be expressed as: ,

[0050] in, This is the opposite denormalization operation to the encoding operation, which pulls the network output back from [0, 1] to the image pixel range.

[0051] The multi-chain fusion module fuses the sequences output by the splicing layer into a sequence with a length of The multi-chain fusion module includes multiple parallel fusion modules. The outputs of the splicing layer are used as inputs of multiple fusion modules respectively. The outputs of multiple fusion modules are used as inputs of the restoration network. Each fusion module includes a third convolution layer and a third activation layer from input to output.

[0052] The restoration network converts the vector of the multi-chain fusion module into a three-dimensional feature map, and gradually restores the image through multiple layers of restoration blocks. The fifth activation layer limits the output to [0,1]. The restoration network includes the first transposed two-dimensional convolution layer, the fourth activation layer, multiple restoration blocks, the second transposed two-dimensional convolution layer and the fifth activation layer from input to output. Each restoration block includes the third transposed two-dimensional convolution layer, the second batch normalization layer and the sixth activation layer from input to output. The fifth activation layer uses the Sigmoid activation function, and the first activation layer, the second activation layer, the third activation layer and the fourth activation layer use the PReLU activation function.

[0053] S2, obtain training data set;

[0054] S0. Construct a composite loss function; specifically including:

[0055] (1) The encoder parameters and decoder parameters Considered as the variable to be optimized, the overall optimization objective is constructed and expressed as: ,in, is the image reconstruction quality loss, is the loss of DNA biological constraints, is the weight coefficient, which is used to balance the importance of the two. The value of produces variant models with different preferences; the constraints of this optimization problem are the mapping and simulation steps involved in the network forward process.

[0056] (2) Reconstruction quality loss: using pixel Level mean square error MSE measures the reconstructed image With the original image The difference between them is expressed as: ;

[0057] Biological constraint loss includes GC content deviation loss and homopolymer length penalty loss; GC content deviation loss is a penalty for the base sequence obtained by mapping ,make For GC content, it is expected to remain within the target , the GC content deviation loss is expressed as: ; The homopolymer length penalty loss is calculated as the set of all homopolymer lengths of the base sequence , for those exceeding the threshold The homopolymer length penalty loss is expressed as: , is the total number of sets, and the biological constraint loss is expressed as: ;

[0058] Then the composite loss function L is expressed as: .

[0059] By designing the composite loss function L, the network can be trained end-to-end under a single loss function, ensuring both high-quality image reconstruction and automatic generation of DNA base sequences that meet biological constraints.

[0060] S3. Perform end-to-end joint training and optimization of the image coding model and the image decoding model using the training dataset; specifically, uniformly train the parameters of the image coding model and the image decoding model to achieve collaborative optimization of source coding, channel simulation, and biological constraints. For each training batch, the normalized image is input into the image coding model, and the reconstructed image is obtained after base mapping, channel simulation, sequence preprocessing, and image decoding model. The composite loss of the current batch is calculated according to the composite loss function L. Based on this loss, the gradients of the image coding model and the image decoding model are simultaneously obtained through backpropagation, and all parameters are updated using the Adam optimizer at a preset learning rate. Repeat the above steps until the loss converges, completing the end-to-end joint training.

[0061] S4, obtaining the image to be transmitted;

[0062] S5. Encode the image to be transmitted using the optimized image coding model to obtain a coding result;

[0063] S6. The encoding result is input into the channel simulation and sequence preprocessing module, and the channel simulation and sequence preprocessing are performed on the encoding result to obtain a numerical matrix;

[0064] S7. The numerical matrix is ​​input into the optimized image decoding model for decoding to obtain a reconstructed image.

[0065] A DNA image transmission device based on deep joint source-channel coding, comprising:

[0066] Storage; the storage stores at least one instruction, at least one program, code set or instruction set;

[0067] Processor; the at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by the processor to implement the above-mentioned DNA image transmission method based on deep joint source-channel coding.

[0068] A computer-readable storage medium stores at least one instruction, at least one program, a code set, or an instruction set, wherein the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by a processor to implement the above-mentioned DNA image transmission method based on deep joint source-channel coding.

[0069] The technical solution of the present invention is not limited to the above-mentioned specific embodiments. Any technical variations made according to the technical solution of the present invention fall within the protection scope of the present invention.

Claims

1. A DNA image transmission method based on deep joint source-channel coding, characterized in that: include: S1. Construct an image coding model, a channel simulation and sequence preprocessing module, and an image decoding model. The image coding model includes a normalization layer, a feature extraction network, and a base mapping module from input to output. The image decoding model includes a splicing layer, a multi-chain fusion module, a restoration network, and an inverse normalization layer from input to output. The feature extraction network is used to extract deep image features and flatten the one-dimensional vector to obtain a potential vector. The base mapping module is used to map the continuous potential vector output by the feature extraction network into a discrete DNA base sequence. The channel simulation and sequence preprocessing module is used to simulate the synthesis, PCR amplification, and sequencing errors during the DNA storage process to simulate the DNA base sequence, and standardize multiple simulated base sequences into a numerical matrix that can be processed by the neural network. The multi-chain fusion module fuses the sequence output by the splicing layer into a vector of length k. The restoration network restores the vector of the multi-chain fusion module. S2, obtain training data set; S3. Perform end-to-end joint training optimization on the image encoding model and image decoding model using the training dataset; S4, obtaining the image to be transmitted; S5. Encode the image to be transmitted using the optimized image coding model to obtain a coding result; S6. The encoding result is input into the channel simulation and sequence preprocessing module, and the channel simulation and sequence preprocessing are performed on the encoding result to obtain a numerical matrix; S7. The numerical matrix is ​​input into the optimized image decoding model for decoding to obtain a reconstructed image.

2. The DNA image transmission method based on deep joint source-channel coding according to claim 1 is characterized in that: The feature extraction network includes the first convolution layer, the first activation layer and the extraction module from input to output. The extraction module is connected to multiple extraction blocks in sequence. Each extraction block includes the second convolution layer, the first batch of normalization layers and the second activation layer from input to output.

3. The DNA image transmission method based on deep joint source-channel coding according to claim 1 is characterized in that: The base mapping module is used to map the continuous potential vector output by the feature extraction network into a discrete DNA base sequence. Specifically: ①, Sigmoid activation and scaling: for each potential vector component Apply the Sigmoid function and map it to the (0,1) interval, expressed as: ;right Do the proportional amplification, expressed as: ; ② Discretization: After magnification Perform the rounding operation to get the integer value, which is expressed as: ; To maintain the gradient path in back propagation, the gradient of round is approximated to 1 in the reverse direction; ③、Integer→base mapping: Map the integer symbols {0,1,2,3} one by one to the four DNA bases to obtain the initial base sequence , the mapping is expressed as: .

4. The DNA image transmission method based on deep joint source channel coding according to claim 3 is characterized in that: The simulation is as follows: 、 and The insertion errors, deletion errors and substitution errors that may occur in the actual DNA synthesis, PCR amplification and DNA sequencing process are simulated on the DNA base sequence to obtain the simulation output sequence, which is expressed as , copy v copies of simulation output sequence ; Insert error: at any position in the sequence, with probability Insert a random base, expressed as: ; Deletion error: for each base By probability Delete, represented by: ; Substitution error: for each base By probability Randomly replace with one of the other three bases, expressed as: ; The specific standardization process is: for each simulation output sequence Perform pruning to unify each copy sequence into a matrix input of length K; and map the character set {A, C, G, T, Q} to the integer set {1, 2, 3, 4, 0}; The numerical sequences of length K are stacked into a numerical matrix .

5. The DNA image transmission method based on deep joint source-channel coding according to claim 3 is characterized in that: The multi-chain fusion module includes multiple parallel fusion modules. The output of the splicing layer is used as the input of multiple fusion modules respectively, and the output of multiple fusion modules is used as the input of the restoration network. Each fusion module includes a third convolution layer and a third activation layer from input to output.

6. The DNA image transmission method based on deep joint source-channel coding according to claim 3 is characterized in that: The restoration network includes the first transposed two-dimensional convolution layer, the fourth activation layer, multiple restoration blocks, the second transposed two-dimensional convolution layer and the fifth activation layer from input to output. Each restoration block includes the third transposed two-dimensional convolution layer, the second batch normalization layer and the sixth activation layer from input to output. The fifth activation layer uses the Sigmoid activation function.

7. The DNA image transmission method based on deep joint source-channel coding according to claim 1, characterized in that: Before S3, it includes: S0, constructing a composite loss function; in S3, using the composite loss function to perform end-to-end joint training optimization on the image encoding model and the image decoding model.

8. The DNA image transmission method based on deep joint source-channel coding according to claim 7 is characterized in that: Included in S0: (1) The encoder parameters and decoder parameters Considered as the variable to be optimized, the overall optimization objective is constructed and expressed as: ,in, is the image reconstruction quality loss, is the loss of DNA biological constraints, is the weight coefficient; (2) Reconstruction quality loss: using pixel Level mean square error MSE measures the reconstructed image The difference between it and the original image x is expressed as: ; Biological constraint loss includes GC content deviation loss and homopolymer length penalty loss; GC content deviation loss is a penalty for the base sequence obtained by mapping ,make For GC content, it is expected to remain within the target , the GC content deviation loss is expressed as: ; The homopolymer length penalty loss is calculated as the set of all homopolymer lengths of the base sequence , for those exceeding the threshold The homopolymer length penalty loss is expressed as: , k is the total number of sets, and the biological constraint loss is expressed as: ; Then the composite loss function L is expressed as: .

9. A DNA image transmission device based on deep joint source-channel coding, characterized in that: include: Storage; The memory stores at least one instruction, at least one program, a code set or an instruction set; Processor; the at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by the processor to implement the DNA image transmission method based on deep joint source-channel coding as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that The readable storage medium stores at least one instruction, at least one program, a code set, or an instruction set, and the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by a processor to implement the DNA image transmission method based on deep joint source-channel coding as described in any one of claims 1 to 8.

Images