Unsupervised auto-encoding network single-pixel imaging method and storage medium

Abstract

The application discloses a kind of unsupervised self-encoding network single-pixel imaging method and storage medium, its method includes pre-generation modulation speckle;Using modulation speckle to modulate the image of object, and the object information after modulation is discretized into digital signal;Reconstruct low-quality target object image jointly with the digital signal of discretization and modulation speckle;Low-quality image input self-encoding network, after coding and decoding, output reconstruction imaging result, set iteration number;According to network reconstruction imaging result and corresponding modulation speckle, calculate predicted intensity value;Calculate the error between predicted intensity value and the digital signal after discretization, adjust self-encoding network learnable parameter, carry out new round of iteration and output result;Loop iteration, until reach pre-set iteration number, output final reconstruction imaging result.The application is not based on specific training set, eliminates neural network output result generalization performance poor and so on Problem, while better inhibiting background noise, improve imaging quality.

Description

Technical Field

[0001] The present invention relates to the field of computer and optical imaging technology, and in particular to an unsupervised autoencoding network single-pixel imaging method. Background Art

[0002] Single-pixel imaging, a novel computational imaging technology, has important applications in biomedical imaging, remote imaging, optical remote sensing, and other fields. However, due to limitations in imaging mechanisms, single-pixel imaging with high spatial resolution is inefficient. Achieving high-quality single-pixel imaging at low sampling rates has been a hot topic of research.

[0003] In recent years, artificial intelligence methods, particularly deep learning, have been widely applied to single-pixel imaging. Typically, this technology uses a supervised approach to learn implicit priors from large labeled datasets, and the trained neural networks are then used for relevant imaging tasks. However, single-pixel imaging techniques based on supervised learning often suffer from challenges such as difficulty obtaining training sets and poor generalization performance of networks trained on specific datasets, significantly limiting or hindering their practical application. Summary of the Invention

[0004] The present invention proposes an unsupervised autoencoding network single-pixel imaging method, which can solve at least one of the above technical problems.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] An unsupervised self-encoding network single-pixel imaging method comprises the following steps:

[0007] Step S1: pre-generate modulated speckle;

[0008] Step S2: Modulating the imaged object using modulated speckle. The modulated object information is received by a single-pixel detector and discretized into a digital signal.

[0009] Step S3: combining the discretized digital signal with the modulated speckle to perform correlation operation to restore the low-quality target object image;

[0010] Step S4: Input the low-quality target object image described in S3 into the autoencoder network, set the number of iteration stops, initialize the network parameters, and output the reconstructed imaging result after the low-quality image is encoded and decoded by the network;

[0011] Step S5: Calculate the predicted intensity value of this iteration based on the imaging result in S4 and the corresponding modulated speckle;

[0012] Step S6: Calculate the predicted strength value in S5 and the discretized digital signal error in S2, and adjust the learnable parameter θ of the autoencoder network through the back propagation algorithm. *, and a new round of iteration is performed to output the reconstruction imaging result of this time;

[0013] Step S7: repeating steps S5 to S6, cyclic iteration, until the preset iteration number is reached, iteration stops, and the final reconstruction image is output.

[0014] Further, the modulated speckle in step S1 includes but is not limited to random speckle, orthogonal Hadamard basis speckle, Fourier basis speckle, wavelet basis speckle, etc.

[0015] Further, the related operation in step S3 is the conventional single-pixel image reconstruction algorithm for the current problem, including but not limited to the existing difference algorithm, iteration algorithm, singular value decomposition algorithm, and compressed sensing algorithm.

[0016] Further, the auto-encoding network comprises:

[0017] The auto-encoding network comprises three encoding layers, one double convolution layer, and three decoding layers;

[0018] That is, the encoding layer comprises convolution filter layer one, convolution filter layer two, and convolution filter layer three; and the decoding layer comprises decoding layer one, decoding layer two, and decoding layer three.

[0019] The convolution filter layer one, the convolution filter layer two, and the convolution filter layer three are used for feature extraction and compression in sequence, and the number of output channels of this layer is multiplied; the double convolution layer is used for further feature extraction; the decoding layer one is used for transposed convolution operation and amplification of the features transmitted by the double convolution layer, and receives the features transmitted by the previous layer and the features of the convolution filter layer three; the decoding layer two receives the features transmitted by the decoding layer one and the convolution filter layer two; the decoding layer three receives the features of the decoding layer two and the convolution filter layer one, adds an activation layer, and outputs the final reconstruction image.

[0020] The convolution filter layer one transmits a low-quality image with a resolution of n*n into the neural network, performs two convolution operations on the image by using a 5*5 two-dimensional convolution kernel in a same manner, and performs batch normalization processing, and the convolution extracts features with a size of n*n; then, the average pooling layer is used for pooling operation, and the extracted features are compressed to a size of n / 2*n / 2.

[0021] The convolution filter layer 2 repeats the above optimization operation on the features of the convolution filter layer 1, and the feature size is compressed to n / 4*n / 4 after pooling; the convolution filter layer 3 repeats the above optimization operation on the features of the convolution filter layer 2, and the feature size is compressed to n / 8*n / 8 after pooling; the double convolution layer performs two convolution operations and batch normalization, and the feature size remains unchanged; the decoding layer 1 performs a transposed convolution operation and batch normalization by a 5*5 convolution kernel to enlarge the image features to n / 4*n / 4, and at the same time, the concat layer links the features of the convolution filter layer 3, and then performs two convolution operations and batch normalization; the decoding layer 2 repeats the above optimization operation of the decoding layer 1, and the features are enlarged to n / 2*n / 2, and the concat layer links the features of the convolution filter layer 2; the decoding layer 3 repeats the operation of the decoding layer 2, and the features are enlarged to n*n. The concat layer links the features of the convolution filter layer 1, and at the same time is activated through the activation layer to map the reconstructed image.

[0022] Furthermore, step S4 includes: setting the corresponding optimization times, initializing the learnable parameters θ in the deep convolutional autoencoder network using the relevant initialization method * Initialization methods include but are not limited to random distribution initialization, normal distribution initialization, normal truncated distribution initialization, bilinear interpolation initialization, He initialization, and Xavier initialization. After the network inputs the low-quality correlation imaging result, it undergoes a series of feature operations such as convolution, downsampling, upsampling, and skip links to map it into a reconstructed image and output the corresponding image result. The result obtained by the autoencoder network processing is:

[0023] , x represents the preprocessing result of the traditional image reconstruction algorithm; is the autoencoder network model, x * Represents the imaging results reconstructed by the autoencoder network.

[0024] Furthermore, the process of calculating the predicted intensity value in step S5 is as follows: based on the single-pixel imaging physical model, the reconstructed image result x in step S4 is converted to * The predicted intensity value is obtained by doing the inner product with the corresponding modulated speckle H Among them H n is the nth modulated speckle, and a one-dimensional discrete sequence of predicted intensity values ​​is obtained by this method.

[0025] Furthermore, the error between the predicted strength value in step S6 and the discretized digital signal in step S2 is: loss = || I * -I|| 2 , where I * is the predicted intensity value, I is the digital signal after the actual measurement discretization in S2, and loss is the mean square error of the two intensity values. At the same time, according to the back propagation algorithm, the corresponding learnable parameter θ of the network is adjusted. *The value of is set and a new round of iteration is performed to output the imaging result of the new round of iterative reconstruction.

[0026] Furthermore, the optimization method in step S7 is as follows: based on the reconstructed image output in the previous iteration, the predicted intensity value of the image and the modulated speckle sequence is calculated, and the error between the calculated intensity value and the digital signal discretized in S2 is minimized. The process of steps S5 to S6 is repeated, and the reconstructed image is output. The iteration is continued until the number of iteration stops is reached and the final reconstructed image is output.

[0027] Furthermore, in step S7, the learning parameter θ * Through back propagation algorithm optimization, the optimized objective function is:

[0028]

[0029] Where x represents the correlation imaging result of the traditional single-pixel imaging reconstruction algorithm; It is an autoencoder network model; represents the physical model of single-pixel imaging; R(·) is the total variation associated with the function.

[0030] On the other hand, the present invention further discloses a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the steps of the above method.

[0031] It can be seen from the above technical solution that in order to achieve high-quality imaging under low sampling numbers and overcome the difficulties in obtaining data sets and the poor generalization ability of neural networks, the present invention proposes an unsupervised autoencoding network single-pixel imaging method. Compared with the data-driven reconstruction algorithm, the present invention embeds the single-pixel imaging physical model through an unsupervised autoencoding network. The resulting optimization algorithm can reconstruct the target image without training on any training set, which effectively overcomes the biggest drawback of the single-pixel imaging reconstruction method based on deep learning, that is, the difficulty in obtaining training sets. At the same time, compared with the existing methods, it has better imaging quality. It is conducive to solving the problem of high-quality image restoration in single-pixel imaging and has better reconstruction effect. The proposed method is not based on a specific training set, which eliminates problems such as poor generalization performance of the neural network output results, and better suppresses background noise. It provides an easy, effective and universal reconstruction scheme for single-pixel imaging. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 is a flowchart of a reconstruction method according to an embodiment of the present invention;

[0033] Figure 2 Schematic diagram of the autoencoding network structure according to an embodiment of the present invention;

[0034] Figure 3is a schematic diagram of a single-pixel imaging process according to an embodiment of the present invention;

[0035] Figure 4 It is an experimental effect diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0036] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.

[0037] like Figure 1 As shown, the unsupervised autoencoding network single-pixel imaging method described in this embodiment includes the following steps:

[0038] Step S1: pre-generate modulated speckle;

[0039] Step S2: modulating the image of the object using the modulated speckle pattern. The modulated object information is received by a single-pixel detector and discretized into a digital signal.

[0040] Step S3: combining the discretized digital signal with the modulated speckle to perform correlation operation to restore the low-quality target object image;

[0041] Step S4: inputting the low-quality target object image into the autoencoder network, setting the number of iteration stops, initializing the network parameters, encoding and decoding the low-quality image through the network, and outputting the reconstructed imaging result;

[0042] Step S5: Calculate the predicted intensity value of this iteration based on the reconstructed imaging result and the corresponding modulated speckle;

[0043] Step S6: Calculate the predicted strength value and the digital signal error after discretization in S2, and adjust the autoencoder network learnable parameter θ through the back propagation algorithm. * , perform a new round of iteration and output the reconstructed imaging result;

[0044] Step S7: Determine whether the number of network iteration stop times has been reached. If the number of network iteration stop times has not been reached, repeat steps S5 to S6.

[0045] Step S8: The iteration stops and the final reconstructed imaging result is output.

[0046] The following are detailed descriptions:

[0047] S1: Pre-generate modulated speckle: Generate a two-dimensional modulated speckle sequence with the same resolution as the imaging target by designing a coding pattern. The modulated speckle patterns include but are not limited to random speckle, orthogonal-based Hadamard speckle, Fourier speckle, and discrete wavelet-based speckle.

[0048] S2: Modulated speckle is used to modulate the imaging target. The modulated target information is received by a single-pixel detector and discretized into a digital signal:

[0049] The image of the imaging object is modulated using modulated speckle. The relevant intensity value is received by a single-pixel detector, and the optical signal is converted into an electrical signal. The electrical signal is then collected by a data acquisition card and discretized into digital information. The detection intensity value corresponds to the modulated speckle one-to-one. The detection intensity value I n =H n ·O (x,y) , where H n is the nth speckle in the modulated speckle sequence.

[0050] S3: Combine the discretized digital signal and the modulated speckle to perform correlation operation to restore the low-quality target object image;

[0051] The related operations in step S3 include one of the traditional iterative algorithm, differential iterative algorithm, normalized iterative algorithm, singular value decomposition algorithm, and compressed sensing algorithm in the single pixel imaging restoration algorithm.

[0052] In one embodiment of the present invention, the conventional single-pixel image reconstruction algorithm related to step S3 is specifically:

[0053]

[0054] Where x represents the low-quality image restored by the reconstruction method, <·> represents the average value, and At the same time, using Normalized speckle sequence.

[0055] S4: Input the low-quality target object image described in S3 into the autoencoder network, set the number of iteration stops, initialize the network parameters, design the network structure, hyperparameters, cost function, and optimizer, and output the reconstructed imaging result after network encoding and decoding;

[0056] Specifically, it includes: designing network structure, hyperparameters, cost function, and optimizer to improve the quality of the image reconstructed by the autoencoder network; after the network inputs a low-quality image, it performs a convolution operation with a two-dimensional 5×5 convolution kernel in the same way to extract features to ensure that the image resolution remains unchanged; then the extracted features are batch normalized; the normalized features are pooled through a pooling layer for down-sampling; at the same time, a part of the features are sent to the following upsampling layer using a jump link; after three convolutions, batch processing, and pooling, the obtained features are merged with the above-mentioned jump link features of the same resolution size, and an inverse convolution upsampling operation is performed; finally, it is mapped into a reconstructed image with the same resolution size as the original low-quality image.

[0057] like Figure 2 As shown, the autoencoding network structure is described as follows:

[0058] The autoencoding network includes 3 encoding layers, 1 double convolutional layer and 3 decoding layers;

[0059] That is, the encoding layer includes convolution filter layer 1, convolution filter layer 2, and convolution filter layer 3; the decoding layer includes decoding layer 1, decoding layer 2, and decoding layer 3;

[0060] The convolution filter layer 1, convolution filter layer 2, and convolution filter layer 3 are used in sequence for feature extraction and compression, and the number of output channels of the layer is multiplied; the double convolution layer is used to further extract features, the decoding layer 1 is used to perform a transposed convolution operation on the features input by the double convolution layer and amplify the features, receive the features input by the previous layer and receive the features of the convolution filter layer 3; the decoding layer 2 receives the features input by the decoding layer 1 and the convolution filter layer 2; the decoding layer 3 receives the features of the decoding layer 2 and the convolution filter layer 1, and adds an activation layer at the same time to output the final reconstructed image;

[0061] Among them, the convolution filter layer 1 passes a low-quality image with a resolution of n*n into the neural network, and performs two convolution operations on the image using a 5*5 two-dimensional convolution kernel in the same way and batch normalization. The convolution extracts features with a size of n*n; then the average pooling layer performs a pooling operation to compress the extracted features to a size of n / 2*n / 2.

[0062] The convolution filter layer 2 repeats the above optimization operation on the features of the convolution filter layer 1, and the feature size is compressed to n / 4*n / 4 after pooling; the convolution filter layer 3 repeats the above optimization operation on the features of the convolution filter layer 2, and the feature size is compressed to n / 8*n / 8 after pooling; the double convolution layer performs two convolution operations and batch normalization, and the feature size remains unchanged; the decoding layer 1 performs a transposed convolution operation and batch normalization by a 5*5 convolution kernel to enlarge the image features to n / 4*n / 4, and at the same time, the concat layer links the features of the convolution filter layer 3, and then performs two convolution operations and batch normalization; the decoding layer 2 repeats the above optimization operation of the decoding layer 1, and the features are enlarged to n / 2*n / 2, and the concat layer links the features of the convolution filter layer 2; the decoding layer 3 repeats the operation of the decoding layer 2, and the features are enlarged to n*n. The concat layer links the features of the convolution filter layer 1, and at the same time is activated through the activation layer to map the reconstructed image.

[0063] S5: Based on the imaging results described in S4 and the corresponding modulated speckle, calculate the predicted intensity value of this iteration:

[0064]

[0065] Where, I * represents the predicted intensity value, H represents the pre-modulated speckle, and x represents the result of the traditional single-pixel imaging algorithm. Represents the reconstructed imaging result output by the autoencoder network.

[0066] S6: Calculate the error between the digital signal in step S2 and the predicted strength value, adjust the autoencoder network through the back propagation algorithm and perform a new round of iteration. The specific error representation function is:

[0067] loss=||I * -I|| 2

[0068] In the formula, loss represents the loss function of the network, I * represents the predicted strength value, and I represents the digital signal in step S2. 2 Represents the error between the calculated predicted intensity value and the actual digital signal.

[0069] S7: Repeat steps S5 to S6, iterate in a loop until the preset number of iterations is reached, then stop the iteration and output the final reconstructed imaging result.

[0070] The process of optimizing the autoencoder network can be expressed as:

[0071]

[0072] Where, To optimize the resulting autoencoder network, the parameter θ can be optimized* The initial value of Hf is the hyperparameter obtained in step S4; θ (·) represents the physical model of single-pixel imaging; R(·) is the total variation associated with the function;

[0073] The autoencoder network reconstruction result is:

[0074]

[0075] Specifically, in one embodiment of the present invention, step S7 iteratively adjusts the network learnable parameter θ by back propagation algorithm. * , output the corresponding reconstructed image; calculate the intensity value of the reconstructed image and the modulated speckle sequence, and minimize the error between the calculated intensity value and the actual measured intensity value, and output is the final imaging result;

[0076] like Figure 3 As shown, the imaging system according to an embodiment of the present invention uses pre-generated modulated speckle as a mask for coding sampling, measuring a corresponding discretized digital signal. A traditional single-pixel imaging algorithm associates the signal with the modulated speckle to restore a low-quality image. The low-quality image is then fed into an autoencoder network, which, after encoding and decoding, outputs a reconstructed image. Based on a physical model, the reconstructed image and the modulated speckle are combined to calculate a predicted intensity value. By minimizing the error between the predicted intensity value and the actual measured digital signal, backpropagation is performed to adjust the network's learnable parameters and a new round of iteration is performed. As the iterations proceed, the autoencoder network outputs an image of higher quality.

[0077] All the above processes do not require any additional training data, and only drive the autoencoder network parameters θ through the physical model * The optimization iteration has good generalization ability and can effectively suppress background noise and eliminate artifacts.

[0078] Figure 4 It is the effect diagram of the present invention. Figure 4 The four figures on the left of Figures a, b, c, and d show the results of traditional single-pixel restoration using random speckle patterns (imaging resolution 64×64, number of samples 512), while the four figures on the right show the results of the method of the present invention. These four verification results demonstrate that the embodiments of the present invention achieve better imaging quality, facilitate high-quality image restoration in single-pixel imaging, and achieve superior reconstruction results.

[0079] In another aspect, the present invention further discloses a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of any of the above methods.

[0080] In still another aspect, the present application also discloses a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the computer program, when executed by the processor, causes the processor to perform the steps of any of the above methods.

[0081] In still another aspect, the present application also discloses a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the computer program, when executed by the processor, causes the processor to perform the steps of any of the above methods.

[0082] It is to be understood that the system provided by the embodiments of the present application corresponds to the method provided by the embodiments of the present application, and the explanation, examples and advantages of related contents can refer to the corresponding parts in the above method.

[0083] It is to be understood by those of ordinary skill in the art that all or part of the processes in the above-mentioned embodiments can be completed by a computer program instructing related hardware, and the program can be stored in a non-volatile computer readable storage medium. When the program is executed, it can include the processes of the above-mentioned embodiments. Any reference to memory, storage, database or other medium used in the embodiments provided by the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM) and memory bus dynamic RAM (RDRAM) and the like.

[0084] The technical features of the above embodiments can be combined in any manner. In order to make the description concise, not all possible combinations of the technical features in the above embodiments are described, however, as long as the combinations of the technical features do not exist contradictions, they should be considered as the scope of the present application.

[0085] The above examples are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that the technical solutions recorded in the foregoing examples can be modified, or some technical features can be replaced by equivalent features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An unsupervised autoencoder network single-pixel imaging method, characterized in that: The following steps are included: Step S1: pre-generate modulated speckle; Step S2: modulating the image of the object using the modulated speckle pattern. The modulated object information is received by a single-pixel detector and discretized into a digital signal. Step S3: combining the discretized digital signal with the modulated speckle to perform correlation operation to restore the low-quality target object image; Step S4: inputting the low-quality target object image into the autoencoder network, setting the number of iteration stops, initializing the network parameters, encoding and decoding the low-quality image through the network, and outputting the reconstructed imaging result; Step S5: Calculate the predicted intensity value of this iteration based on the reconstructed imaging result and the corresponding modulated speckle; Step S6: Calculate the predicted strength value and the digital signal error after discretization in S2, and adjust the learnable parameters of the autoencoder network through the back propagation algorithm. , perform a new round of iteration and output the reconstructed imaging result; The autoencoding network includes 3 encoding layers, 1 double convolutional layer and 3 decoding layers; That is, the encoding layer includes convolution filter layer 1, convolution filter layer 2, and convolution filter layer 3; the decoding layer includes decoding layer 1, decoding layer 2, and decoding layer 3; The convolution filter layer 1, convolution filter layer 2, and convolution filter layer 3 are used in sequence for feature extraction and compression, and the number of output channels of the layer is multiplied; the double convolution layer is used to further extract features, the decoding layer 1 is used to perform a transposed convolution operation on the features input by the double convolution layer and amplify the features, receiving the features input by the previous layer while receiving the features of the convolution filter layer 3; the decoding layer 2 receives the features input by the decoding layer 1 and the convolution filter layer 2; Decoding layer 3 receives the features of decoding layer 2 and convolution filter layer 1, adds an activation layer, and outputs the final reconstructed image; Among them, the convolution filter layer 1 passes a low-quality image with a resolution of n*n into the neural network, and performs two convolution operations on the image using a 5*5 two-dimensional convolution kernel in the same way and batch normalization. The convolution feature size is n*n; then the average pooling layer performs a pooling operation to compress the extracted features to a size of (n / 2)*(n / 2); The convolution filter layer 2 repeats the above optimization operation on the features of the convolution filter layer 1, and the feature size is compressed to (n / 4)*(n / 4) after pooling; the convolution filter layer 3 repeats the above optimization operation on the features of the convolution filter layer 2, and the feature size is compressed to (n / 8)*(n / 8) after pooling; the double convolution layer performs two convolution operations and batch normalization, and the feature size remains unchanged; the decoding layer 1 performs a transposed convolution operation and batch normalization by a 5*5 convolution kernel to enlarge the image features to (n / 4)*(n / 4), and at the same time, the concat layer links the features of the convolution filter layer 3, and then performs two convolution operations and batch normalization; the decoding layer 2 repeats the above optimization operation of the decoding layer 1, and the features are enlarged to (n / 2)*(n / 2), and the concat layer links the features of the convolution filter layer 2; the decoding layer 3 repeats the operation of the decoding layer 2, and the features are enlarged to n*n. The concat layer links the features of the convolution filter layer 1, and at the same time, it is activated through the activation layer to map the reconstructed image. Step S7: Determine whether the number of iterations to stop has been reached. If not, repeat steps S5 to S6. Step S8: The iteration stops and the final reconstructed imaging result is output.

2. The unsupervised autoencoder network single-pixel imaging method according to claim 1, characterized in that: The modulated speckle pattern in step S1 includes one of random speckle, orthogonal Hadamard-based speckle, Fourier-based speckle, and wavelet-based speckle.

3. The unsupervised autoencoder network single-pixel imaging method according to claim 1, characterized in that: The related operations in step S3 include one of the traditional iterative algorithm, differential iterative algorithm, normalized iterative algorithm, singular value decomposition algorithm, and compressed sensing algorithm in the single pixel imaging restoration algorithm.

4. The unsupervised autoencoder network single-pixel imaging method according to claim 1, characterized in that: The encoding and decoding process in step S4 specifically includes: First, a convolution operation is performed using the same two-dimensional convolution kernel to extract features from the input while ensuring that the image resolution remains unchanged. The extracted features are then batch normalized; the normalized features are pooled through the pooling layer, downsampled, and the relevant image features are compressed; At the same time, the corresponding image features are linked to the upsampling layer using a skip link method; After three convolutions, batch processing, and pooling, the obtained features are merged with the jump link features of the same resolution size, and three inverse convolution upsampling operations are performed; finally, it is mapped into a reconstructed image with the same resolution size as the original low-quality image.

5. The unsupervised autoencoder network single-pixel imaging method according to claim 1, characterized in that: The step S5 of calculating the iterative prediction strength value specifically includes: Where, represents the predicted strength value, represents the pre-generated modulated speckle, represents the reconstruction result of the traditional single-pixel imaging algorithm, Represents the reconstructed image result output by the autoencoder network.

6. The unsupervised autoencoder network single-pixel imaging method according to claim 1, characterized in that: In step S6, the error between the predicted strength value and the digital signal discretized in step S2 is calculated, wherein the error representation function is: Where, represents the loss function of the network, represents the predicted strength value, represents the digital signal in step S2, Represents the error between the calculated predicted intensity value and the actual digital signal.

7. The unsupervised autoencoder network single-pixel imaging method according to claim 6, characterized in that: Step S6 also includes using the back propagation algorithm to correct the weights of each layer of neurons in the autoencoder network according to the error signal, thereby updating the learnable parameters of the autoencoder network. And perform a new round of iterative calculation; The iterative calculation process is expressed as: Where, To optimize the autoencoder network, Optimizable parameters for neural networks; A physical model representing single-pixel imaging; is the total variation associated with the function; The autoencoder network reconstruction result is: 。 8. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 7.

Images