Single-pixel image restoration method and system, identification method and storage medium
By training and optimizing the latent feature encoder and reconstruction network, and combining the U-Net convolutional neural network and adversarial discriminator, the problem of poor imaging quality of single-pixel cameras at low sampling rates is solved, and efficient and accurate single-pixel image restoration and recognition are achieved.
Patent Information
- Application Number
- CN202510966829.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-14
- Publication Date
- 2025-10-28
AI Technical Summary
Traditional RGB visible light cameras have insufficient imaging quality for high-altitude target monitoring, especially in complex lighting conditions where it is difficult to identify birds or drones. Single-pixel cameras have poor imaging quality at low sampling rates, resulting in severe noise, blurriness, and artifacts in the reconstructed images, making it impossible to distinguish the basic outline of the target.
We employ a latent feature encoder and a reconstruction network, optimize the latent feature encoder and reconstruction network through training, and further optimize the latent feature encoder and reconstruction network using the L1 norm consistency loss function and the comprehensive loss function. We combine the U-Net convolutional neural network for image reconstruction and use an adversarial discriminator and a recognition model to improve image quality.
It achieves efficient and accurate reconstruction of single-pixel images under low sampling rate conditions, improves image recognition accuracy, reduces the probability of illusions in reconstructed images, and ensures high fidelity and clarity of images.
Smart Images
Figure CN120852196A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of single-pixel image restoration technology, and particularly relates to a single-pixel image restoration method and system, recognition method and storage medium. Background Technology
[0002] In aviation environments such as airports, the activities of birds and drones pose a serious threat to flight safety. Traditional monitoring methods mainly rely on RGB visible light cameras, which have significant limitations in practical applications. For example, when targets (such as birds or drones) are at high altitudes, due to the limitations of the camera's focal length, the targets appear as tiny black dots in the image, lacking sufficient detail (such as shape and trajectory), making it difficult for monitoring systems to accurately identify and issue warnings. This limitation is particularly pronounced under complex lighting conditions (such as backlighting or fog), severely restricting the effectiveness of flight safety management.
[0003] To address this issue, single-pixel imaging technology has received widespread attention in recent years. Unlike traditional cameras, single-pixel cameras utilize sensors in special wavelength bands such as infrared and terahertz, along with modulated optical paths, to reconstruct images by trading time for space. This technology is particularly suitable for capturing targets at long distances or in low-light conditions.
[0004] However, the core drawback of single-pixel cameras is that their image quality is highly dependent on the sampling rate. When the sampling rate is low (such as due to hardware limitations or real-time requirements), the reconstructed image will have severe noise, blur and artifacts, and may even be unable to distinguish the basic outline of the target. Summary of the Invention
[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a single-pixel image restoration method that can restore single-pixel images efficiently and accurately.
[0006] To achieve the above objectives, the present invention adopts the following technical solution:
[0007] A method for restoring a single-pixel image includes the following steps:
[0008] S1, after acquiring the first single-pixel image, obtain the latent feature vector of the first single-pixel image;
[0009] S2, after reconstructing the first single-pixel image based on the latent feature vector, obtain the reconstructed image corresponding to the current first single-pixel image.
[0010] Preferably, S1 further includes the following: obtaining the latent feature vector of the first single-pixel image using a trained latent feature encoder.
[0011] Preferably, training the latent feature encoder also includes the following:
[0012] S11, after degrading the sharp image, obtain the corresponding blurred single-pixel image; record a sharp image and its corresponding blurred single-pixel image as a pair of training images;
[0013] S12, using a latent feature encoder, obtain the latent feature vectors of the clear image and the corresponding blurred single-pixel image respectively; record a pair of training images and the corresponding pair of latent feature vectors as a training data;
[0014] S13, Calculate the first consistency loss function L for a pair of training images corresponding to their latent feature vectors. imp Then, towards the first consistency loss function L imp Gradient reduction direction optimizes the latent feature encoder: L imp =||e Y -e GT ||1; where e Y e represents the latent feature vector of a blurred single-pixel image Y; GT Let ||·||1 represent the latent feature vector of the sharp image GT; ||·||1 represents the L1 norm.
[0015] Preferably, S2 further includes the following: using the trained reconstruction network, reconstructing the first single-pixel image through the latent feature vector, and then obtaining the reconstructed image corresponding to the current first single-pixel image.
[0016] Preferably, training the reconstruction network also includes the following:
[0017] S21, The reconstruction network obtains the reconstructed image X based on the training data: X = f rec (Y,e Y ); where f rec (Y,e Y ) indicates the use of reconstructed network f rec Based on the latent feature vector e Y Reconstruct the blurred single-pixel image Y;
[0018] S22, After calculating the overall loss function Loss of the reconstructed network, optimize the reconstructed network in the direction that reduces the gradient of the overall loss function Loss:
[0019] Loss=α·L1+β·L ADV +γ·L img L img =||f class (X)-f class (GT)||1;
[0020] L ADV =E[logD(GT)]+E[logD(1-D(X))];
[0021] Where α, β, and γ represent the first, second, and third parameters, respectively; L1 represents the absolute value loss function; L ADV Let L represent the adversarial loss function; D represents the adversarial discriminator; D(·) represents the output value after processing with the adversarial discriminator; the bases of logD(GT) and logD(1-D(X)) are 10; E(·) represents the expected value; L img f represents the second consistency loss function; class (X) represents the recognition result vector of the reconstructed image X using a pre-trained recognition model; f class (GT) represents the recognition result vector of the sharp image GT corresponding to the reconstructed image X using a pre-trained recognition model; the adversarial discriminator D uses a VGG network.
[0022] Preferably, the reconstructed network f rec For the U-Net convolutional neural network, reconstruct the network f rec The first sampling block consists of six cascaded downsampling blocks and six cascaded upsampling blocks; each sampling block contains four cascaded fully connected layers FC1, FC2, FC3, and FC4; the input to the first downsampling block is the blurred single-pixel image Y and the latent feature vector e. Y The inputs of the other sampling blocks, excluding the latent feature vector e Y In addition, it includes the modulated image output from the previous sampling block; the modulated image output from the last upsampling block is the reconstructed image X.
[0023] Preferably, the sampling rate of the blurred single-pixel image is less than 5%.
[0024] The present invention also provides a single-pixel image recognition method, comprising: after obtaining a reconstructed image using a single-pixel image restoration method as described above, recognizing the content in the reconstructed image.
[0025] This invention also provides a single-pixel image restoration system, comprising: a latent feature encoder module and a reconstruction network module; after the acquired first single-pixel image is fed into the latent feature encoder to obtain the corresponding latent feature vector, the latent feature encoder module feeds the first single-pixel image and the corresponding latent feature vector into the reconstruction network module; the reconstruction network module reconstructs the first single-pixel image according to the latent feature vector and outputs the corresponding reconstructed image; each module is programmed or configured to execute the steps of the above-described single-pixel image restoration method.
[0026] The present invention also provides a computer-readable storage medium storing a computer program programmed or configured to perform a single-pixel image restoration method as described above.
[0027] The beneficial effects of this invention are as follows:
[0028] (1) The single-pixel image restoration method of the present invention can efficiently restore an accurate and high-fidelity reconstructed image regardless of the sampling rate of the first single-pixel image.
[0029] (2) In the single-pixel image restoration method of the present invention, the latent feature encoder is optimized in the direction of decreasing gradient of the first consistency loss function. The first consistency loss function is the L1 norm of the difference between the latent feature vector of the blurred single-pixel image Y and the latent feature vector of the corresponding clear image GT. This makes the latent feature vectors extracted by the latent feature encoder in the latent space before and after the degradation of the clear image have a lot of consistent discriminative information. It also makes the latent feature vectors corresponding to the clear image before and after degradation more and more consistent with the optimization of the latent feature encoder, which is more conducive to improving the accuracy of the subsequent reconstruction network to reconstruct the blurred single-pixel image Y based on the latent feature vector.
[0030] (3) In the single-pixel image restoration method of the present invention, the comprehensive loss function Loss used to optimize the reconstruction network is composed of three loss functions. These three loss functions are used to measure the consistency of image information between the reconstructed image X and the corresponding clear image in three different discriminative dimensions, so as to ensure that no erroneous content is generated in the reconstructed image X due to the generation of illusions, and to ensure the high fidelity of the reconstructed image X relative to the corresponding clear image. As the reconstruction network is optimized, the corresponding pixels between the reconstructed image X and the corresponding clear image should also tend to be the same or highly similar, the features of the corresponding data between the reconstructed image X and the corresponding clear image should also tend to be the same or highly similar, and the recognition result vectors of the reconstructed image X and the corresponding clear image should also tend to be the same or highly similar.
[0031] (4) The present invention trains and optimizes the reconstruction network by using the comprehensive loss function Loss, which is composed of consistency constraints of image information in three different discriminative dimensions. The optimized reconstruction network can not only efficiently restore single-pixel images, but also significantly reduce the probability of hallucinations in the reconstructed images, and the details in the reconstructed images are also very clear.
[0032] (5) The single-pixel image recognition method of the present invention can significantly improve the recognition accuracy of single-pixel images. Attached Figure Description
[0033] Figure 1 This is a flowchart of a single-pixel image restoration method according to the present invention;
[0034] Figure 2 Two sets of clear images;
[0035] Figure 3 To and Figure 2 The corresponding two sets of blurred single-pixel images;
[0036] Figure 4 To and Figure 3 The two sets of reconstructed images correspond to each other. Detailed Implementation
[0037] To make the technical solution of the present invention clearer and more explicit, the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Solutions derived by those skilled in the art through equivalent substitution and conventional reasoning of the technical features of the present invention without creative effort all fall within the protection scope of the present invention.
[0038] Example 1
[0039] like Figure 1 The flowchart shown illustrates a single-pixel image restoration method, which includes the following steps:
[0040] S1, after acquiring the first single-pixel image, obtain the latent feature vector of the first single-pixel image;
[0041] S2, after reconstructing the first single-pixel image based on the latent feature vector, obtain the reconstructed image corresponding to the current first single-pixel image.
[0042] S1 also includes the following: using a trained latent feature encoder to obtain the latent feature vector of the first single-pixel image.
[0043] The sampling rate of the first single-pixel image may be a low-sampling-rate single-pixel image.
[0044] Training the latent feature encoder also includes the following sub-steps:
[0045] S11, after degrading the clear image, obtain the corresponding blurred single-pixel image; record a clear image and its corresponding blurred single-pixel image as a pair of training images.
[0046] In S11, methods such as compressed sensing single pixel and Radon single pixel transformation are used to degrade the clear image and obtain a blurred single pixel image with a low sampling rate.
[0047] In this embodiment, the sampling rate of the blurred single-pixel image is less than 5%. We hope that the single-pixel image restoration method of the present invention will also have a good restoration effect on single-pixel images with extremely low sampling rates (i.e., sampling rates less than 5%).
[0048] S12, using a latent feature encoder, obtain the latent feature vectors of the clear image and the corresponding blurred single-pixel image respectively; record a pair of training images and the corresponding pair of latent feature vectors as a training data.
[0049] In S12, the latent feature encoder is a common encoder network, specifically containing five encoder units. Each unit contains a convolutional layer with a stride of 2 and a kernel size of 3x3, a batch normalization layer, and a ReLU activation function.
[0050] S13, Calculate the first consistency loss function L for a pair of training images corresponding to their latent feature vectors. imp Then, towards the first consistency loss function L imp Gradient reduction direction optimizes the latent feature encoder: L imp =||e Y -e GT ||1; where e Y e represents the latent feature vector of a blurred single-pixel image Y; GT Let ||·||1 represent the latent feature vector of the sharp image GT; ||·||1 represents the L1 norm.
[0051] In step S13, the present invention moves towards the first consistency loss function L imp The gradient reduction direction optimizes the latent feature encoder, while the first consistency loss function L... imp It is also the L1 norm of the difference between the latent feature vector of the blurred single-pixel image Y and the latent feature vector of the corresponding sharp image GT. This makes the latent feature vectors extracted by the latent feature encoder in the latent space before and after the sharp image degradation have a lot of consistent discriminative information. It also makes the latent feature vectors corresponding to the sharp image before and after degradation more and more consistent with the optimization of the latent feature encoder, which is more conducive to improving the accuracy of subsequent reconstruction networks in reconstructing the blurred single-pixel image Y based on the latent feature vector.
[0052] S2 also includes the following: using the trained reconstruction network, the first single-pixel image is reconstructed through the hidden feature vector, and the reconstructed image corresponding to the current first single-pixel image is obtained.
[0053] Training the reconstructed network also includes the following sub-steps:
[0054] S21, The reconstruction network obtains the reconstructed image X based on the training data: X = f rec (Y,e Y ); where f rec (Y,e Y ) indicates the use of reconstructed network f rec Based on the latent feature vector e Y Reconstruct the blurred single-pixel image Y.
[0055] S21 also includes the following:
[0056] Rebuild network f recThis is a U-Net convolutional neural network consisting of six cascaded downsampling blocks and six cascaded upsampling blocks; each sampling block contains four cascaded fully connected layers FC1, FC2, FC3, and FC4; the input to the first downsampling block is a blurred single-pixel image Y and a latent feature vector e. Y The inputs of the other sampling blocks, besides the latent feature vector e Y In addition, it includes the modulated image output from the previous sampling block; the modulated image output from the last upsampling block is the reconstructed image X.
[0057] Within the first downsampling block: the blurred single-pixel image Y is transformed into an intermediate image Y1 by convolution multiplied by a convolution kernel. Four cascaded fully connected layers are based on the latent feature vector e. Y After processing the intermediate image Y1, the modulated image Y1′ is obtained and fed into the second downsampling block. Except for the first downsampling block, each of the other downsampling blocks converts the input modulated image into its corresponding intermediate image through convolution. Then, four cascaded fully connected layers are applied based on the latent feature vector e. Y After processing the corresponding intermediate image, the corresponding modulated image is obtained. All upsampling blocks convert the input modulated image into the corresponding intermediate image through deconvolution. Four cascaded fully connected layers then convert the intermediate image based on the latent feature vector e. Y After processing the corresponding intermediate image, the corresponding modulated image is obtained.
[0058] Y i ′=Y i ×FC4(FC3(FC2(FC1(e Y ))))+FC4(FC3(FC2(FC1(e Y ))));wherein, FC1(·) represents
[0059] FC1 represents the modulation processing of the fully connected layer; FC2(·) represents the modulation processing of the fully connected layer; FC3(·) represents the modulation processing of the fully connected layer; FC4(·) represents the modulation processing of the fully connected layer; Y i Y represents the intermediate image in the i-th sampling block; i Y' represents the modulated image in the i-th sampling block; sampling blocks 1-6 are downsampling blocks, and sampling blocks 7-12 are upsampling blocks; 12 That is, to rebuild the network f. rec The output is the reconstructed image X.
[0060] S22, After calculating the overall loss function Loss of the reconstructed network, optimize the reconstructed network in the direction that reduces the gradient of the overall loss function Loss:
[0061] Loss=α·L1+β·L ADV+γ·L img L img =||f class (X)-f class (GT)||1;
[0062] L ADV =E[logD(GT)]+E[logD(1-D(X))];
[0063] Where α, β, and γ represent the first, second, and third parameters, respectively, which are preset by technicians; L1 represents the absolute value loss function, which calculates the L1 norm between the reconstructed image X and the corresponding clear image pixel by pixel; L ADV Let L represent the adversarial loss function; D represent the adversarial discriminator; D(·) represent the output value after processing with the adversarial discriminator, which is a scalar; the bases of logD(GT) and logD(1-D(X)) are 10; E(·) represent the expected value; L represents the adversarial loss function. ADV This is used to calculate the similarity of feature vectors between the reconstructed image X and the corresponding sharp image GT; the smaller the difference, the more similar they are. img f represents the second consistency loss function; class (X) represents the recognition result vector of the reconstructed image X using a pre-trained recognition model; f class (GT) represents the recognition result vector of the sharp image GT corresponding to the reconstructed image X using a pre-trained recognition model.
[0064] In this embodiment, the adversarial discriminator uses a VGG network; the pre-trained recognition model uses a VGG network trained on the ImageNet dataset. The recognition model classifies different image contents, and each dimension of the recognition result vector represents an image type.
[0065] This invention uses the penultimate layer features of a pre-trained recognition model to determine whether the image type of the reconstructed image X remains consistent with that of the clear image GT by using feature distance.
[0066] This invention does not employ solely a GAN (Generative Adversarial Network) model for end-to-end training to learn the mapping from low-quality to high-quality images. This is because while this approach can generate clear details from extremely low-sampled single-pixel images, there is a risk of generating phantom pixels during the restoration process. Specifically, the GAN model may over-complete non-existent features based on prior knowledge, for example, incorrectly restoring an extremely low-sampled single-pixel image that should be a drone as a bird. Such content distortion directly reduces the accuracy of subsequent single-pixel image recognition, posing a significant threat to the safety and reliability of aviation safety systems.
[0067] The comprehensive loss function Loss used in this invention to optimize the reconstruction network is a fusion of three loss functions. These three loss functions are used to measure the consistency of image information between the reconstructed image X and the corresponding sharp image in three different discriminative dimensions. Among them, the absolute value loss function L1 is a commonly used loss function in image reconstruction. In this invention, the L1 norm between the reconstructed image X and the corresponding sharp image is calculated pixel-by-pixel, representing the similarity of corresponding pixels between the reconstructed image X and the corresponding sharp image. As the reconstruction network is optimized, the corresponding pixels between the reconstructed image X and the corresponding sharp image should also tend to be the same or highly similar. The adversarial loss function L... ADV After extracting feature vectors from the corresponding data in the randomly selected reconstructed image X and the sharp image using an adversarial discriminator, the similarity between the two feature vectors is calculated to characterize the feature similarity between the reconstructed image X and the corresponding data in the sharp image; that is, as the reconstruction network is optimized, the features of the corresponding data between the reconstructed image X and the corresponding sharp image should also tend to be the same or highly similar. The second consistency loss function L... img The recognition result vector calculated in the process is used to subsequently determine the image type to which the image content belongs (such as different types such as birds, drones, etc.); that is, as the reconstruction network is optimized, the recognition result vectors of the reconstructed image X and the corresponding clear image should also tend to be the same or highly similar, so as to ensure that no erroneous content is generated in the reconstructed image X due to the generation of illusions, and to ensure the high fidelity of the reconstructed image X relative to the corresponding clear image.
[0068] This invention trains and optimizes the reconstruction network using a comprehensive loss function Loss, which is composed of consistency constraints of image information across three different discriminative dimensions. The optimized reconstruction network can not only efficiently restore single-pixel images, but also significantly reduce the probability of hallucinations in the reconstructed images, and the details in the reconstructed images are also very clear.
[0069] The present invention provides a single-pixel image restoration method that can efficiently restore an accurate and high-fidelity reconstructed image regardless of the sampling rate of the first single-pixel image.
[0070] like Figures 2-3 As shown, Figure 2 In the image, 2a and 2b represent two sets of clear images. Figure 3 In the image, 3a is the blurred single-pixel image corresponding to 2a. Figure 3 In this context, 3b represents the blurred single-pixel image corresponding to 2b. We use a single-pixel image restoration method from this invention, employing an optimized reconstruction network to respectively... Figure 3 The two sets of blurred single-pixel images in the image are used as the first single-pixel image for restoration and reconstruction. The reconstructed image after restoration and reconstruction is as follows: Figure 4 As shown, Figure 4 In the image, 4a is the reconstructed image corresponding to 3a. Figure 4 In the image, 4b corresponds to the reconstructed image of 3b. It can be seen that the reconstructed image obtained using this method is essentially the same as the clear image.
[0071] Technicians then used the single-pixel image restoration method of this invention to restore 10,000 sets of blurred single-pixel images. The restored images were then compared and verified with their corresponding clear images. Technicians visually judged the accuracy of the restored images relative to the clear images, achieving an accuracy rate as high as 92.46%. The technicians also calculated the PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index) of the restored images. The PSNR was consistently above 19.9 dB, and the SSIM was consistently above 0.76. In contrast, the original blurred images with a 2% sampling rate had a PSNR of around 13 dB and an SSIM consistently below 0.28. It is evident that the reconstructed images obtained using the single-pixel image restoration method of this invention have high clarity and accuracy.
[0072] Example 2
[0073] The present invention also provides a single-pixel image recognition method, comprising: after obtaining a reconstructed image using a single-pixel image restoration method as described in Embodiment 1, recognizing the content in the reconstructed image.
[0074] The present invention provides a single-pixel image recognition method that can significantly improve the recognition accuracy of single-pixel images.
[0075] Example 3
[0076] The present invention also provides a single-pixel image restoration system, comprising:
[0077] The hidden feature encoder module and the reconstruction network module;
[0078] After the first single-pixel image is acquired, it is fed into the latent feature encoder to obtain the corresponding latent feature vector. The latent feature encoder module then sends the first single-pixel image and the corresponding latent feature vector into the reconstruction network module.
[0079] The reconstruction network module reconstructs the first single-pixel image based on the latent feature vector and outputs the corresponding reconstructed image.
[0080] Each module is programmed or configured to perform the steps of the single-pixel image restoration method described above.
[0081] The present invention also provides a computer-readable storage medium storing a computer program programmed or configured to perform a single-pixel image restoration method described in Embodiment 1.
[0082] The technologies, shapes, and structures not described in detail in this invention are all well-known technologies. It should also be noted that the above are merely preferred embodiments of this invention and are not intended to limit the scope of the invention. The components or steps in the embodiments of this invention can be decomposed and / or recombined, and these decompositions and / or recombinations should be considered equivalent solutions to this application and should all fall within the protection scope of this invention.
Claims
1. A method for restoring a single-pixel image, characterized in that, Includes the following steps: S1, after acquiring the first single-pixel image, obtain the latent feature vector of the first single-pixel image; S2, after reconstructing the first single-pixel image based on the latent feature vector, obtain the reconstructed image corresponding to the current first single-pixel image.
2. The single-pixel image restoration method according to claim 1, characterized in that, S1 also includes the following: using a trained latent feature encoder to obtain the latent feature vector of the first single-pixel image.
3. The single-pixel image restoration method according to claim 2, characterized in that, Training the latent feature encoder also includes the following: S11, after degrading the sharp image, obtain the corresponding blurred single-pixel image; record a sharp image and its corresponding blurred single-pixel image as a pair of training images; S12, using a latent feature encoder, obtain the latent feature vectors of the clear image and the corresponding blurred single-pixel image respectively; record a pair of training images and the corresponding pair of latent feature vectors as a training data; S13, Calculate the first consistency loss function L for a pair of training images corresponding to their latent feature vectors. imp Then, towards the first consistency loss function L imp Gradient reduction direction optimizes the latent feature encoder: L imp =||e Y -e GT ||1; Among them, e Y e represents the latent feature vector of a blurred single-pixel image Y; GT Let ||·||1 represent the latent feature vector of the sharp image GT; ||·||1 represents the L1 norm.
4. The single-pixel image restoration method according to claim 1, characterized in that, S2 also includes the following: using the trained reconstruction network, the first single-pixel image is reconstructed through the hidden feature vector, and the reconstructed image corresponding to the current first single-pixel image is obtained.
5. The single-pixel image restoration method according to claim 4, characterized in that, Training the reconstructed network also includes the following: S21, The reconstruction network obtains the reconstructed image X based on the training data: X = f rec (Y,e Y ); where f rec (Y,e Y ) indicates the use of reconstructed network f rec Based on the latent feature vector e Y Reconstruct the blurred single-pixel image Y; S22, After calculating the overall loss function Loss of the reconstructed network, optimize the reconstructed network in the direction that reduces the gradient of the overall loss function Loss: Loss=α·L1+β·L ADV +γ·L img ;L img =||f class (X)-f class (GT)||1; L ADV =E[logD(GT)]+E[logD(1-D(X))]; Where α, β, and γ represent the first, second, and third parameters, respectively; L1 represents the absolute value loss function; L ADV Let L represent the adversarial loss function; D represents the adversarial discriminator; D(·) represents the output value after processing with the adversarial discriminator; the bases of logD(GT) and logD(1-D(X)) are 10; E(·) represents the expected value; L img f represents the second consistency loss function; class (X) represents the recognition result vector of the reconstructed image X using a pre-trained recognition model; f class (GT) represents the recognition result vector of the sharp image GT corresponding to the reconstructed image X using a pre-trained recognition model; the adversarial discriminator D uses a VGG network.
6. The single-pixel image restoration method according to claim 5, characterized in that: Rebuild network f rec For the U-Net convolutional neural network, reconstruct the network f rec The first sampling block consists of six cascaded downsampling blocks and six cascaded upsampling blocks; each sampling block contains four cascaded fully connected layers FC1, FC2, FC3, and FC4; the input to the first downsampling block is the blurred single-pixel image Y and the latent feature vector e. Y The inputs of the other sampling blocks, excluding the latent feature vector e Y In addition, it includes the modulated image output from the previous sampling block; the modulated image output from the last upsampling block is the reconstructed image X.
7. The single-pixel image restoration method according to claim 3, characterized in that: The sampling rate of a blurred single-pixel image is less than 5%.
8. A single-pixel image recognition method, characterized in that: After obtaining a reconstructed image using a single-pixel image restoration method as described in any one of claims 1-7, the content in the reconstructed image is identified.
9. A single-pixel image restoration system, characterized in that, include: The hidden feature encoder module and the reconstruction network module; After the first single-pixel image is acquired, it is fed into the latent feature encoder to obtain the corresponding latent feature vector. The latent feature encoder module then sends the first single-pixel image and the corresponding latent feature vector into the reconstruction network module. After reconstructing the first single-pixel image based on the latent feature vector, the reconstruction network module outputs the corresponding reconstructed image; Each module is programmed or configured to perform the steps of a single-pixel image restoration method as described in any one of claims 1-8.
10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program that is programmed or configured to perform a single-pixel image restoration method as described in any one of claims 1-8.