Single-pixel imaging reconstruction method based on compressed sensing and related device

Abstract

The invention belongs to the technical field of imaging, and particularly relates to a single-pixel imaging reconstruction method and device based on compressed sensing. The core of the method is to construct an FRSPINet network architecture based on deep learning. The method comprises the following steps: firstly, constructing a cascade network model composed of a data preprocessing module, a deep separable convolution module, a full connection module, a shallow feature extraction module, an attention mechanism module and an image reconstruction module; carrying out gradient optimization on network parameters through multi-stage iterative training, dynamically adjusting a learning rate based on convergence of an SSIM loss function, and storing optimal weight parameters; in an imaging stage, a compression measurement value of a target scene is obtained through a single-pixel detector in a preset sampling mode, and after the compression measurement value is transmitted to a computing terminal through a data acquisition card, the compression measurement value is input into a pre-trained FRSPINet network for end-to-end nonlinear mapping, and finally a high-fidelity reconstructed image is output. And the image reconstruction quality and the system robustness at a low sampling rate are effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of computational imaging and image processing, and particularly relates to a single-pixel imaging reconstruction method based on compressed sensing and related devices. Background Art

[0002] As an emerging imaging method, single-pixel imaging technology uses a single detector to collect image information row by row and column by column, thereby significantly reducing the complexity and cost of imaging equipment. Due to its advantages in spatial resolution and sampling efficiency, this technology is particularly suitable for low-cost, portable devices and special scenarios where traditional imaging technologies are difficult to apply. However, traditional single-pixel imaging methods face problems such as high-dimensional data processing and noise interference during image reconstruction. As a result, the reconstruction results are easily affected significantly in low-light and noisy environments, and the reconstruction quality is difficult to guarantee. Therefore, how to improve the reconstruction quality of single-pixel imaging systems has become a technical challenge that needs to be solved urgently.

[0003] Currently, inventions in the field of single-pixel imaging rely heavily on the use of a fixed measurement matrix. However, this approach often leads to an increase in redundant information, which in turn reduces imaging efficiency and affects the quality of image reconstruction. Although this technology has made significant progress in recent years, significant challenges remain in the design of the measurement matrix and the optimization of the reconstruction algorithm. Commonly used measurement matrices in the early days included the Random Matrix (RM), the Random Gaussian Matrix (RGM), the Random Bernoulli Matrix (RBM), and the Random Fourier Matrix (RFM). While these matrices can reduce redundant information to a certain extent, their random nature still requires a large number of samples to achieve good reconstruction results. To improve sampling efficiency, subsequent inventions proposed novel measurement matrices, such as the Walsh-Ordered Standard Hadamard Matrix (WSHM) based on the Standard Hadamard Matrix (STHM) and the Walsh-Ordered Zigzag-Scanned Standard Hadamard Matrix (ZZWSHM). Compared to random matrices, these matrices have better symmetry and orthogonality, enabling better imaging with fewer samples. However, with the growing demand for efficient imaging, existing methods still have room for further improvement.

[0004] With the gradual promotion of single-pixel imaging technology in various applications, traditional reconstruction algorithms have gradually exposed certain shortcomings. First, when dealing with high-dimensional, low-sampling-rate image reconstruction, traditional algorithms are often unable to effectively restore high-quality images. Especially when faced with complex imaging scenes, traditional methods have poor reconstruction effects and weak generalization capabilities. Secondly, in practical applications where computing power, storage resources, and energy are limited, traditional reconstruction algorithms usually rely on matrix operations with high computational costs, making it difficult to run efficiently in real-time or large-scale data processing scenarios. With the continuous development of data volume and application scenarios, traditional single-pixel imaging reconstruction algorithms are facing increasingly serious performance bottlenecks.

[0005] Given this background, the present invention proposes a single-pixel imaging reconstruction method based on compressed sensing, aiming to address the challenges faced by single-pixel imaging at low sampling rates. This method combines matrix optimization techniques with an improved reconstruction algorithm and introduces an efficient multi-scale attention mechanism, enabling more accurate capture of important image features. In this way, the present invention improves the effectiveness and accuracy of the image reconstruction algorithm and optimizes imaging quality at low sampling rates. Summary of the Invention

[0006] In this study, a single-pixel imaging reconstruction method and related devices based on compressed sensing are proposed, aiming to improve the reconstruction accuracy and computational efficiency in single-pixel imaging.

[0007] A first aspect of an embodiment of the present application provides a single-pixel imaging reconstruction method based on compressed sensing, the method comprising:

[0008] Build a single-pixel reconstruction model based on deep learning (Feature Reconstruction Single-Pixel Imaging Network, FRSPINet);

[0009] The public datasets MNIST and Fashion MNIST are used as training and test sets to train the single-pixel reconstruction model and obtain the optimized single-pixel reconstruction model.

[0010] The optimized single-pixel reconstruction model is deployed in the single-pixel imaging system to reconstruct the original image under low sampling rate conditions to obtain a high-quality target image.

[0011] In one possible implementation, the single-pixel reconstruction model includes a data preprocessing module, a depth-separable convolution module, a fully connected module, a shallow feature extraction module, a multi-scale attention mechanism module, and an image reconstruction module, wherein:

[0012] Depthwise separable convolution module: This module generates an optimized measurement matrix through a depthwise separable convolution layer and compresses the input raw data;

[0013] Fully connected module: This module converts low-dimensional sampled data into higher-dimensional data representation through feature mapping and nonlinear transformation;

[0014] Shallow feature extraction module: This module extracts shallow features from the sampled data of the fully connected module to capture the basic texture and structural information of the image;

[0015] Multi-scale attention mechanism module: This module dynamically adjusts the weights of different features by introducing multi-scale feature information;

[0016] Image reconstruction module: The feature map weighted by the attention mechanism is input into the image reconstruction module, which converts it into the final two-dimensional image output.

[0017] In one possible implementation, the data preprocessing module is specifically configured to:

[0018] The Min-Max Normalization method is used to normalize the original image collected by the detector to obtain a normalized original image. The normalization process is implemented by the following formula:

[0019]

[0020] Where I represents a value in the original data, min(I) and max(I) represent the minimum and maximum values ​​in the data set, respectively;

[0021] Perform rotation, translation, and scaling operations on the normalized original image to obtain the preprocessed original image.

[0022] In one possible implementation, the input image is normalized and data enhanced, and then convolved in a depthwise separable convolution module to achieve a nonlinear mapping of image features to one-dimensional compressed measurements. The depthwise separable convolution consists of depthwise convolution and pointwise convolution, which group convolutions in the feature dimension, independently perform depthwise convolution on each channel, and aggregate all channels using a 1×1 convolution before output.

[0023] In one possible implementation, the fully connected module adopts a three-level fully connected network architecture. The fully connected module is composed of three fully connected submodules connected in series. Each fully connected submodule includes a fully connected layer (Fully Connected Layer), a parametric rectified linear unit (PReLU), and a random dropout layer (Dropout Layer) with a probability of 0.2. Its mathematical expression can be described as:

[0024] F out =Dropout 0.2 (PReLU(WgF in +b))

[0025] Among them, the input feature vector F in ∈R 1024×1 After the fully connected layer, it is mapped to F in ∈R 4096×1 , through repeated "FC+PReLU+Dropout" structure to further enhance the feature representation ability, and finally output F out ∈R 4096×1 The compact feature vector of .

[0026] In one possible implementation, in the shallow feature extraction module, the output F out ∈R 4096×1 The feature vector of is converted into a two-dimensional feature map F through the tensor reshaping operation map ∈R 64×64 , then a large-size convolution kernel (5×5) is used to refine the feature map. After feature refinement, the feature map is downsampled using an average pooling layer with a kernel size of 5 and a stride of 2, compressing the feature map size from 64×64 to 32×32.

[0027] In one possible implementation, the multi-scale attention mechanism module first splits the input feature map X (with dimensions (C×H×W)) into G sub-feature groups in the channel dimension, i.e., X = [X0, X1, ..., X G-1 ], secondly, the EMA module designs three parallel paths, each path has different convolution kernels and processing methods. The first path uses a 1×1 convolution kernel combined with an average pooling operation; the second path also uses a 1×1 convolution kernel, but focuses on extracting high-level information of features through nonlinear transformations; the third path uses a 3×3 convolution kernel, focusing on extracting multi-scale features.

[0028] In one possible implementation, after the image reconstruction module is processed by the attention mechanism, a 1×1 convolution is used to fuse information between channels while maintaining the spatial structure, and finally a deep feature map of size 1×32×32 is generated; this deep feature map is normalized by the Sigmoid activation function, and the output value is mapped to between 0 and 1. The final 32×32 reconstructed image is generated as the output result of the single-pixel reconstruction model.

[0029] In one possible implementation, the images of the MNIST and Fashion MNIST datasets are divided into training and test sets in a ratio of 3:1, where the training set contains 6000 images and the test set contains 2000 images. During the model training process, the initial learning rate is set to 1×10 -4 , and adopt the ReduceLROnPlateau learning rate scheduling strategy to improve training efficiency. For network optimization, the structural similarity loss function (SSIM) is used. The SSIM expression is as follows:

[0030]

[0031] Among them, α>0, β>0, γ>0, x, y represent two pictures respectively, μ x and μ y Represent the average values ​​of x and y, σ x and σ y Represents the standard deviation of x and y, σ xy represents the covariance of x and y, while c1 c2c3 are constants respectively.

[0032] A second aspect of an embodiment of the present application provides a single-pixel imaging reconstruction device based on compressed sensing, the device comprising:

[0033] A construction unit for building a single-pixel reconstruction model based on deep learning (Feature Reconstruction Single-Pixel Imaging Network, FRSPINet);

[0034] A training unit is used to train the single-pixel reconstruction model using the public datasets MNIST and Fashion MNIST as training sets and test sets to obtain an optimized single-pixel reconstruction model;

[0035] The reconstruction unit is used to deploy the optimized single-pixel reconstruction model to the single-pixel imaging system, reconstruct the original image under low sampling rate conditions, and obtain a high-quality target image.

[0036] A third aspect of an embodiment of the present application provides a terminal, comprising a processor, an input device, an output device, and a memory, wherein the processor, input device, output device, and memory are interconnected, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is configured to call the program instructions and execute the step instructions in the first aspect of the embodiment of the present application.

[0037] The fourth aspect of the embodiments of the present application provides a computer-readable storage medium, wherein the above-mentioned computer-readable storage medium stores a computer program for electronic data exchange, wherein the above-mentioned computer program enables a computer to execute some or all of the steps described in the first aspect of the embodiments of the present application.

[0038] A fifth aspect of the embodiments of the present application provides a computer program product, wherein the computer program product includes a non-transitory computer-readable storage medium storing a computer program, wherein the computer program is operable to cause a computer to perform some or all of the steps described in the first aspect of the embodiments of the present application. The computer program product may be a software installation package.

[0039] The present invention has the following beneficial effects:

[0040] By integrating multiple modules such as data preprocessing, depthwise separable convolution, fully connected layers, shallow feature extraction, and attention mechanisms, a method based on the FRSPINet deep learning network architecture was designed. This method effectively improves the reconstruction accuracy and computational efficiency of single-pixel imaging, enabling high-quality image reconstruction.

[0041] The depthwise separable convolutional layer combined with the matrix optimization strategy improves the perception accuracy of the target image information and can learn adaptively, thereby enhancing the accuracy and reliability of single-pixel imaging;

[0042] The optimization and synergy of multiple modules further enhance the feature extraction, generalization and nonlinear expression capabilities of this method;

[0043] By introducing the attention mechanism module, the network can automatically learn and fuse multi-scale features, while generating dynamic attention weights, thereby effectively improving image quality during image reconstruction. In particular, in the reconstruction of complex images, it demonstrates better robustness and detail recovery capabilities.

[0044] By building a single-pixel imaging system, this method has application potential and also provides new technical means for related fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 A schematic diagram of a flow chart of a single-pixel imaging reconstruction method based on compressed sensing provided for the implementation of the present invention;

[0046] Figure 2 Schematic diagram of the depthwise separable convolution module structure of the present invention;

[0047] Figure 3 Schematic diagram of the EMA multi-scale attention module of the present invention;

[0048] Figure 4 The dataset samples (MNIST, Fashion MNIST) used for training in the present invention;

[0049] Figure 5 This is a schematic diagram of the loss function structure used in the present invention;

[0050] Figure 6 This is a schematic diagram of the principle of the single-pixel imaging system of the present invention;

[0051] Figure 7 A schematic diagram of the process flow of a single-pixel imaging reconstruction device based on compressed sensing provided for the implementation of the present invention. DETAILED DESCRIPTION

[0052] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings involved in the embodiments of the present invention will be briefly described below. Obviously, the drawings only represent some embodiments of the present invention. For those skilled in the art, other relevant drawings can be derived based on these drawings without creative work. This article will clearly and completely describe the technical solutions of the present invention in conjunction with the drawings in the embodiments of the present invention. It should be emphasized that the embodiments are only some embodiments of the present invention and do not mean that the present invention is limited to this. Any other embodiments based on the embodiments of the present invention that have not undergone creative work should be deemed to fall within the scope of protection of the present invention.

[0053] See also Figure 1 As shown, Figure 1 A schematic flow chart of a single-pixel imaging reconstruction method based on compressed sensing provided for the implementation of the present invention includes the following steps:

[0054] 101. Build a single-pixel reconstruction model based on deep learning (Feature Reconstruction Single-Pixel Imaging Network, FRSPINet).

[0055] The present invention first constructs an image reconstruction network FRSPINet, which includes multiple key modules: a data preprocessing module, a depthwise separable convolution module, a fully connected module, a shallow feature extraction module, a multi-scale attention mechanism module, and an image reconstruction module. The design of each module aims to improve reconstruction performance and image quality, ensuring effective processing and reconstruction of image data at different levels and dimensions. Specifically, the data preprocessing module consists of a batch normalization layer, a data enhancement layer, and a compressed sampling layer; the depthwise separable convolution module consists of a compressed depthwise convolution layer and a pointwise convolution layer; the fully connected module consists of three fully connected layers, each followed by a PreLU activation function layer and a Dropout layer; the shallow feature extraction module consists of a cascaded convolution layer, a normalization layer, a pooling layer, and a Relu layer; the multi-scale attention mechanism module consists of a multi-branch parallel convolution layer, a cross-scale interaction layer, a dynamic attention generation layer, and a residual connection layer; and the image reconstruction module consists of a convolution layer and an activation function layer.

[0056] Furthermore, the data preprocessing module first converts the image resolution of 28×28 pixels to 32×32 pixels. We then use the Min-Max Normalization method to normalize the data collected by the detector. This method aims to map the numerical range of the input data to a uniform interval, thereby eliminating the dimensionality differences between different features and ensuring that each feature contributes consistently to the results during model training. Specifically, the normalization process is implemented using the following formula:

[0057]

[0058] Here, I represents a value in the original data, and min(I) and max(I) represent the minimum and maximum values ​​in the dataset, respectively. This conversion process linearly maps the original data to the interval ([0,1]). This method effectively eliminates scale differences between features and prevents certain features from dominating the training process, thereby accelerating the convergence of the optimization algorithm, improving training efficiency and stability, and enhancing the model's learning ability and generalization performance.

[0059] At the same time, the present invention enhances the data by performing a series of random transformations on the original training data, such as rotation, translation, and flipping. These transformations simulate the diversity of the real environment, can improve the robustness of the model, and increase the diversity of the data, thereby enhancing the generalization ability of the model and improving the training effect.

[0060] exist Figure 2In the depthwise separable convolution (DSC) module shown, the present invention innovatively applies it as a measurement matrix to single-pixel imaging tasks. This measurement matrix captures different spatial feature patterns in a more flexible manner. DSC consists of depthwise convolution and pointwise convolution, with depthwise convolution used to extract spatial features and pointwise convolution used to extract channel features. By decomposing the convolution operation into two independent steps along the spatial and channel dimensions, the model can effectively extract finer-grained spatial information, thereby optimizing the processing of image data. This design not only helps preserve key image information but also enables effective optimization during processing. Compared to traditional convolution operations, traditional two-dimensional convolution requires a convolution operation for each pixel, while depthwise separable convolution decomposes it into two one-dimensional convolution steps, significantly reducing computational complexity and memory requirements. This decomposition not only reduces the amount of computation but also improves the processing efficiency of the model. Specifically, the independent convolution operations in the spatial and channel dimensions reduce the number of convolution kernel parameters and improve computational parallelism. Furthermore, traditional measurement matrices are typically fixed and rely on prior knowledge, making them difficult to adapt to actual data. They often fail to fully preserve the spatial structure of the input data, leading to information loss. Using depthwise separable convolution as the measurement matrix, however, dynamically adapts to the characteristics of the data, improving the accuracy of information capture and avoiding the information loss that can occur in traditional methods. Consequently, the measurement matrix based on depthwise separable convolution demonstrates superior performance in single-pixel imaging tasks.

[0061] In this invention, we first preprocess the target image and then compressively sample it using an innovative measurement matrix. In the single-pixel imaging problem, the target image I is usually represented as an M×1 column vector, i.e., I=[I1,I2,...,I M ] T , where M represents the number of pixels in the image. For a given target image, single-pixel imaging uses N measurement matrices A1, A2, ..., A N , where each matrix A i Represents a specific perceptual mode. By performing inner product operations on these measurement matrices and the target image, a set of scalar values ​​y1, y2, ..., y N These scalar values ​​form an N×1 sampling vector y=[y1,y2,...,y N ] T ,Combining the above steps, we can get the following complete mathematical expression:

[0062] y=A·I,

[0063] Where A is an N×M measurement matrix, y is an N×1 sampling vector, and I is an M×1 target image column vector. Each scalar y i It represents the inner product result of the i-th measurement matrix and the target image. The inner product operation is essentially a process of compressing or sampling the target image, aiming to obtain a compressed representation of the image through these sampling values.

[0064] When obtaining an N×1-dimensional observation vector, in order to achieve efficient feature extraction, the present invention designs a cascaded feature extraction network to design a fully connected module. This architecture adopts a three-level fully connected topology to achieve step-by-step mapping and dimensional conversion of the feature space through hierarchical nonlinear transformation and regularization mechanism. Specifically, the fully connected module is composed of three fully connected submodules in series, each of which contains a fully connected layer (Fully Connected Layer), a parametric linear correction unit (Parametric Rectified Linear Unit, PReLU) and a random dropout layer (DropoutLayer) with a probability of 0.2. Its mathematical expression can be described as:

[0065] F out =Dropout 0.2 (PReLU(WgF in +b))

[0066] Among them, the input feature vector F in ∈R 1024×1 First, the number of features is expanded from 1024 to 4096 through the first-level fully connected layer, and adaptive nonlinear transformation is introduced using parameterized rectified linear units. Subsequently, a dropout layer with probability p = 0.2 is used to randomly inactivate neurons to suppress overfitting. While maintaining the number of features at 4096, the second and third-level networks further optimize the feature representation through a cascaded "fully connected + PReLU + Dropout" structure, and finally output a compact feature vector F out ∈R 4096×1 .

[0067] This architecture's dimensionality expansion strategy first enhances the model's representational capabilities by expanding the number of features from 1024 to 4096, providing ample space for subsequent feature interactions. Compared to the traditional ReLU function, the PReLU activation function features a learnable negative slope parameter, enabling more flexible adjustment of feature distribution, thereby improving model performance. The Dropout layer after each layer effectively suppresses overfitting during training by randomly inactivating neurons, enhancing the model's generalization capabilities. Finally, the three-level network cascade progressively removes the influence of input noise, generating highly discriminative feature representations.

[0068] Furthermore, in the shallow feature extraction module, the output F out ∈R 4096×1 The compact feature vector of is converted into a two-dimensional feature map F through the tensor reshaping operation map ∈R 64×64 , using large receptive field convolution kernel for feature refinement:

[0069] F conv =Conv2D(F map :W)

[0070] Among them, F conv Represents the output feature map after the convolution operation. Conv2D is a two-dimensional convolution layer that performs convolution operations. map Denotes the input feature map, W is the convolution kernel weight matrix, and the convolution kernel size is designed to be 5×5 to capture contextual features. In addition, a zero padding (padding = 2) strategy is used to ensure that the spatial resolution is maintained after the convolution operation. The subsequent average pooling layer performs a downsampling operation to reduce the size of the feature map:

[0071] F pool =AvgPool2D(F conv ; kernel = 5, stride = 2)

[0072] Among them, F pool Represents the output feature map after the pooling operation. AvgPool2D represents the operation of the two-dimensional average pooling layer, which calculates the average value of the local area through the sliding window. conv is the convolution output from the previous layer. The pooling window size is set to 5×5 (i.e., kernel=5), and the sliding step parameter is 2 (i.e., stride=2), which means that the window slides with a step size of 2 each time. This operation reduces the size of the feature map from 64×64 to 32×32, thereby reducing computational complexity and providing stable gradient propagation for the subsequent reconstruction stage by preserving the smoothness of the feature distribution.

[0073] Furthermore, after obtaining F map ∈R 32×32 After the feature map, we introduced the efficient multi-scale attention mechanism (EMA), such as Figure 3 As shown in the figure, the multi-scale attention mechanism module first divides the input feature map X (with dimensions (C×H×W)) into G sub-feature groups (such as Figure 3 Each C / / G feature group in X=[X0,X1,...,X G-1], each subgroup is processed independently to capture diverse feature representations. Through this grouping strategy, EMA can not only process feature information of different scales more flexibly, but also effectively reduce computational complexity. The EMA module is designed with three parallel paths, each with different convolution kernels and processing methods. The first path uses a 1×1 convolution kernel (concat+1×1) combined with average pooling operations (X Avg Pool and Y Avg Pool, where "X Avg Pool" represents 1D horizontal global average pooling and "Y AvgPool" represents 1D vertical global average pooling), aiming to measure channel information and capture global context. The second path also uses a 1×1 convolution kernel, but focuses on the high-level feature extraction through nonlinear transformations.

[0074] The EMA module is designed with three parallel paths, each with different convolution kernels and processing methods. The first path uses a 1×1 convolution kernel (concat+1×1) combined with an average pooling operation (where "X Avg Pool" represents 1D global average pooling in the horizontal direction, and "Y Avg Pool" represents 1D global average pooling in the vertical direction), aiming to capture channel information and extract global context. The second path also uses a 1×1 convolution kernel, but its design focuses on extracting high-level information of features through nonlinear transformations. The third path uses a 3×3 convolution kernel, focusing on extracting multi-scale features, avoiding conventional normalization operations, thereby better capturing local information and retaining spatial details. Through the collaborative work of these three paths, the EMA module not only improves the model's ability to process information of different scales, but also effectively avoids information loss or redundant calculations.

[0075] In the feature fusion stage, EMA effectively combines the attention weights extracted from these three paths through cross-spatial information aggregation technology. In particular, for the 1×1 convolution path, global average pooling is used to compress the spatial dimension to extract global context information. Its mathematical expression is shown as follows:

[0076]

[0077] where Z c Represents the feature map after pooling, H and W are the height and width of the feature map, x c (i, j) is the pixel value of the c-th channel and position (i, j) in the original feature map.

[0078] Furthermore, the feature map after the EMA module is reconstructed using the image reconstruction module, using 1×1 convolution to fuse information between channels while maintaining the spatial structure, ultimately obtaining a deep feature map of size 1×32×32. Finally, the convolutional feature map is normalized using the Sigmoid activation function, mapping the output value to between 0 and 1, generating the final 32×32 reconstructed image as the network output.

[0079] In this paper, the public MNIST and Fashion MNIST datasets are used as experimental datasets, such as Figure 4 As shown in the figure, the two datasets used in the experiment have different characteristics. The MNIST dataset contains images of handwritten digits with large contrast variations between images, while the Fashion MNIST dataset contains images of clothing with moderate detail and complex backgrounds. Both datasets contain 10 categories, each with multiple samples, and the image resolution is uniformly 28×28 pixels. To ensure the validity of the training and testing data, the images in the MNIST and Fashion MNIST datasets are divided into training and test sets in a 3:1 ratio, with the training set containing 6,000 images and the test set containing 2,000 images.

[0080] Finally, for model optimization, we adopted the Structural Similarity Index (SSIM) loss function to achieve a more comprehensive and refined image quality assessment. SSIM constructs a multi-dimensional image quality assessment framework based on three key dimensions: luminance comparison, contrast comparison, and structure comparison. This framework aims to quantify the differences between the original image and the reconstructed image to assess their similarity. Figure 5 A schematic diagram of the structural similarity (SSIM) measurement system is shown. Specifically, the SSIM expression is as follows:

[0081]

[0082] Among them, α>0, β>0, γ>0, μ x and μ y Represent the average values ​​of x and y, σ x and σ y Represents the standard deviation of x and y respectively. xy represents the covariance of x and y. c1, c2, and c3 are constants to avoid system errors caused by denominators of zero.

[0083] Similarly, to comprehensively evaluate image restoration results, this paper employs a key evaluation metric: the Structural Similarity Index (SSIM). SSIM is an advanced method widely used in image quality assessment that quantitatively analyzes image restoration quality. Its value ranges between 0 and 1, with values ​​closer to 1 indicating a higher visual similarity between the restored image and the original. As an indicator that combines subjective perception with objective assessment, SSIM demonstrates greater credibility and effectiveness in evaluating image restoration results than traditional error metrics.

[0084] In the model training and verification, in order to evaluate the effectiveness of the depthwise separable convolution matrix and the FRSPINet model, this paper conducted a series of simulation experiments using the Pytorch framework on the NVIDIA GeForce RTX 2070 hardware platform. During the model training process, we set the initial learning rate to 1×10 -4 , and adopts the ReduceLROnPlateau learning rate scheduling strategy to improve training efficiency. Specifically, if the model does not significantly improve performance within 20 consecutive epochs, the learning rate will automatically decrease by 50% until the end of training, for a total of 150 epochs. This strategy aims to dynamically adjust the learning rate to avoid training stagnation and optimize model parameters.

[0085] In model deployment and application, the present invention also proposes a single-pixel imaging reconstruction device based on matrix optimization, which is used to implement a single-pixel imaging reconstruction method based on matrix optimization. The device includes a customized projector (Project) and a single-pixel detector (Single-pixel detector). The projector has a refresh rate of 120Hz, a wavelength range of 380-750nm, and is loaded with a measurement matrix (Measurement matrix) with a resolution of 800×800 pixels. During the simulation stage, the pixel size of the measurement matrix and the target image (Target) is set to 32×32. In actual testing, the measurement matrix is ​​expanded into 32×32 sub-blocks, and each sub-block is magnified 25 times to ensure the projection accuracy requirements. The light signal reflected from the target surface is reflected by a frosted glass plate and collected by a single-pixel detector model Ksdtek DT-30-V. The collected data is converted by the ART USB3200 data acquisition board (Data Acquisition board). The sampling rate is 500kS / s, and the processed data is transmitted to a computer and image reconstruction analysis is performed using Python and PyTorch frameworks. Figure 6 As shown in the figure, the schematic diagram and physical diagram of the whole system.

[0086] During the system verification phase, this study set three different sampling times: 10, 51, and 102, corresponding to sampling rates of 1%, 5%, and 10%. Furthermore, regarding the selection of inventive matrices, the present invention employed five different matrices: RM, STHM, WSHM, ZZWSHM, and the innovative depthwise separable convolutional layer matrix proposed in this paper. Regarding reconstruction algorithms, we selected four methods for comparison: the traditional computer ghost imaging algorithm (CGI), the total variation (TV) algorithm, the feature learning ghost imaging network (FLGINet), and the FRSPINet proposed in this paper.

[0087] The above mainly introduces the scheme of the embodiment of the present application from the perspective of the execution process on the method side. It is understandable that, in order to implement the above functions, the terminal includes hardware structures and / or software modules corresponding to the execution of each function. Those skilled in the art should easily realize that, in combination with the units and algorithm steps of the various examples described in the embodiments provided herein, the present application can be implemented in the form of hardware or a combination of hardware and computer software. Whether a function is executed in a hardware or computer software driven hardware manner depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0088] The embodiment of the present application can divide the terminal into functional units according to the above method example. For example, each functional unit can be divided according to each function, or two or more functions can be integrated into one processing unit. The above integrated unit can be implemented in the form of hardware or in the form of software functional units. It should be noted that the division of units in the embodiment of the present application is schematic and is only a logical functional division. In actual implementation, there may be other division methods.

[0089] In line with the above, please see Figure 7 , Figure 7 The present invention provides a schematic diagram of a single-pixel imaging reconstruction device based on compressed sensing. Figure 7 As shown, the device includes:

[0090] A construction unit 701 is used to construct a single-pixel reconstruction model based on deep learning (Feature Reconstruction Single-Pixel Imaging Network, FRSPINet);

[0091] A training unit 702 is configured to train the single-pixel reconstruction model using the public datasets MNIST and Fashion MNIST as training and test sets to obtain an optimized single-pixel reconstruction model.

[0092] The reconstruction unit 703 is used to deploy the optimized single-pixel reconstruction model to the single-pixel imaging system to reconstruct the original image under low sampling rate conditions to obtain a high-quality target image.

[0093] An embodiment of the present application also provides a computer storage medium, wherein the computer storage medium stores a computer program for electronic data exchange, and the computer program enables a computer to execute part or all of the steps of any one of the compressed sensing-based single-pixel imaging reconstruction methods described in the above method embodiments.

[0094] An embodiment of the present application also provides a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program, and the computer program enables a computer to execute part or all of the steps of any one of the compressed sensing-based single-pixel imaging reconstruction methods recorded in the above method embodiments.

[0095] It should be noted that for the aforementioned method embodiments, for the sake of simplicity, they are all expressed as a series of action combinations, but those skilled in the art should be aware that this application is not limited by the order of the actions described, because according to this application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by this application.

[0096] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0097] In the several embodiments provided in this application, it should be understood that the disclosed devices can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, and the indirect coupling or communication connection of devices or units can be electrical or other forms.

[0098] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0099] In addition, the functional units in the various embodiments of the application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software program modules.

[0100] If the integrated unit is implemented in the form of a software program module and sold or used as an independent product, it can be stored in a computer-readable memory. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a memory, including a number of instructions for enabling a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned memory includes: various media that can store program codes, such as a U disk, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk.

[0101] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable memory, which may include: a flash drive, a read-only memory, a random access memory, a magnetic disk or an optical disk, etc.

[0102] The above is a detailed introduction to the embodiments of the present application. Specific examples are used herein to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method and core idea of ​​the present application. At the same time, for those skilled in the art, according to the idea of ​​the present application, there may be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A single-pixel imaging reconstruction method based on compressed sensing, characterized in that: The method comprises: Build a single-pixel reconstruction model based on deep learning (Feature Reconstruction Single-Pixel Imaging Network, FRSPINet); The public datasets MNIST and Fashion MNIST are used as training and test sets to train the single-pixel reconstruction model and obtain the optimized single-pixel reconstruction model. The optimized single-pixel reconstruction model is deployed in the single-pixel imaging system to reconstruct the original image under low sampling rate conditions to obtain a high-quality target image.

2. The single-pixel imaging reconstruction method based on compressed sensing according to claim 1, characterized in that: The single-pixel reconstruction model includes a data preprocessing module, a depth-separable convolution module, a fully connected module, a shallow feature extraction module, a multi-scale attention mechanism module and an image reconstruction module, wherein: Depthwise separable convolution module: This module generates an optimized measurement matrix through a depthwise separable convolution layer and compresses the input raw data; Fully connected module: This module converts low-dimensional sampled data into higher-dimensional data representation through feature mapping and nonlinear transformation; Shallow feature extraction module: This module extracts shallow features from the sampled data of the fully connected module to capture the basic texture and structural information of the image; Multi-scale attention mechanism module: This module dynamically adjusts the weights of different features by introducing multi-scale feature information; Image reconstruction module: The feature map weighted by the attention mechanism is input into the image reconstruction module, which converts it into the final two-dimensional image output.

3. The single-pixel imaging reconstruction method based on compressed sensing according to claim 2, characterized in that: The data preprocessing module is specifically used for: The Min-Max Normalization method is used to normalize the original image collected by the detector to obtain a normalized original image. The normalization process is implemented by the following formula: Where I represents a value in the original data, min(I) and max(I) represent the minimum and maximum values ​​in the data set, respectively; Perform rotation, translation, and scaling operations on the normalized original image to obtain the preprocessed original image.

4. The single-pixel imaging reconstruction method based on compressed sensing according to claim 2, characterized in that: After the input image is normalized and data enhanced, it is convolved in the depthwise separable convolution module to achieve a nonlinear mapping of image features to one-dimensional compressed measurements. The depthwise separable convolution consists of depthwise convolution and pointwise convolution, which groups convolutions in the feature dimension, performs depthwise convolution on each channel independently, and aggregates all channels using a 1×1 convolution before output.

5. The single-pixel imaging reconstruction method based on compressed sensing according to claim 2, characterized in that: The fully connected module adopts a three-level fully connected network architecture. The fully connected module consists of three fully connected submodules in series. Each fully connected submodule contains a fully connected layer (Fully Connected Layer), a parametric rectified linear unit (PReLU) and a random dropout layer (Dropout Layer) with a probability of 0.

2. Its mathematical expression can be described as: F out JDropout 0.2 (PReLU(WgF). in +b)) Among them, the input feature vector F in ∈R 1024×1 After the fully connected layer, it is mapped to F in ∈R 4096×1 , through repeated "FC+PReLU+Dropout" structure to further enhance the feature representation ability, and finally output F out ∈R 4096×1 The compact feature vector of .

6. The single-pixel imaging reconstruction method based on compressed sensing according to claim 5, characterized in that: In the shallow feature extraction module, the output F out ∈R 4096×1 The feature vector of is converted into a two-dimensional feature map F through the tensor reshaping operation map ∈R 64 ×64 , then a large-size convolution kernel (5×5) is used to refine the feature map. After feature refinement, the feature map is downsampled using an average pooling layer with a kernel size of 5 and a stride of 2, compressing the feature map size from 64×64 to 32×32.

7. The single-pixel imaging reconstruction method based on compressed sensing according to claim 6, characterized in that: The multi-scale attention mechanism module first divides the input feature map X (dimension is (C×H×W)) into G sub-feature groups in the channel dimension, that is, X=[X0,X1,...,X G-1 ], secondly, the EMA module designs three parallel paths, each path has different convolution kernels and processing methods. The first path uses a 1×1 convolution kernel combined with an average pooling operation; the second path also uses a 1×1 convolution kernel, but focuses on extracting high-level information of features through nonlinear transformations; the third path uses a 3×3 convolution kernel, focusing on extracting multi-scale features.

8. The single-pixel imaging reconstruction method based on compressed sensing according to claim 7, characterized in that: After the image reconstruction module is processed by the attention mechanism, 1×1 convolution is used to fuse information between channels while maintaining the spatial structure, and finally a deep feature map of size 1×32×32 is generated; the deep feature map is normalized by the Sigmoid activation function, the output value is mapped to between 0 and 1, and the final 32×32 reconstructed image is generated as the output result of the single-pixel reconstruction model.

9. The single-pixel imaging reconstruction method based on compressed sensing according to any one of claims 1 to 8, characterized in that: The images of the MNIST and Fashion MNIST datasets are divided into training and test sets in a ratio of 3:

1. The training set contains 6000 images and the test set contains 2000 images. During the model training process, the initial learning rate is set to 1×10 -4 , and adopt the ReduceLROnPlateau learning rate scheduling strategy to improve training efficiency. For network optimization, the structural similarity loss function (SSIM) is used. The SSIM expression is as follows: Among them, α>0, β>0, γ>0, x, y represent two pictures respectively, μ x and μ y Represent the average values ​​of x and y, σ x and σ y Represents the standard deviation of x and y, σ xy represents the covariance of x and y, while c1 c2c3 are constants respectively.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, the processor is caused to perform the method according to any one of claims 1 to 9.

Images