End-to-end compressed sensing imaging method and device, storage medium and program product

Abstract

The embodiment of the invention discloses an end-to-end compressed sensing imaging method and device, a storage medium and a program product. The method comprises the steps that an image sample set is acquired, a compressed sensing imaging model is trained based on the image sample set, and trained network parameters are packaged into a global network model; receiving an image size specified by a user, and constructing a zero vector and a plurality of standard basis vectors based on the image size; inputting the zero vector into a global network model, and obtaining an offset vector output by the global network model; inputting the plurality of standard basis vectors into a global network model, and obtaining a plurality of basis response vectors output by the global network model; constructing a global matrix equivalent to the global network model based on the bias vector and the plurality of base response vectors; and deploying the global matrix to the target device. According to the method, a global matrix equivalent to a global network model is constructed and deployed to target equipment, so that the technology is promoted to be industrialized and commercialized, and wide application of the method in a real world scene is realized.

Description

Technical Field

[0001] This disclosure relates to the field of compressed sensing technology, and in particular to an end-to-end compressed sensing imaging method, apparatus, storage medium, and program product. Background Technology

[0002] In the field of image reconstruction, compressed sensing technology aims to acquire signals at a sampling rate far lower than that of the traditional sampling theorem and achieve high-quality image reconstruction. In recent years, deep learning (DL), with its powerful feature learning capabilities, has been widely applied to compressed sensing image reconstruction and has demonstrated performance that surpasses traditional optimization algorithms in academic metrics, bringing new development opportunities to image reconstruction.

[0003] Existing deep learning compressed sensing technology is difficult to industrialize and commercialize. This technology is computationally intensive, which is far from meeting the hardware deployment requirements and makes it difficult to run efficiently on real hardware. In addition, its lack of generalization ability also greatly limits its application scenarios. These problems seriously hinder the widespread application of this technology in real-world scenarios. Summary of the Invention

[0004] In view of this, the present disclosure provides an end-to-end compressed sensing imaging method, apparatus, storage medium, and program product, which can promote the industrialization and commercialization of the technology by constructing a global matrix equivalent to a global network model and deploying it to the target device, and realize its wide application in real-world scenarios.

[0005] In a first aspect, embodiments of this disclosure provide an end-to-end compressed sensing imaging method, employing the following technical solution: Obtain an image sample set, train the compressed sensing imaging model based on the image sample set, and encapsulate the trained network parameters into a global network model; Receive the image size specified by the user, and construct a zero vector and multiple standard basis vectors based on the image size; The zero vector is input into the global network model to obtain the bias vector output by the global network model; The multiple standard basis vectors are input into the global network model to obtain multiple basis response vectors output by the global network model; Based on the bias vector and the multiple basis response vectors, a global matrix equivalent to the global network model is constructed; Deploy the global matrix to the target device.

[0006] Optionally, the image sample set includes multiple image samples; The compressed sensing imaging model includes a multilayer linear convolutional neural network model, a dual-branch structure model, and an image reconstruction model.

[0007] Optionally, the multi-layer linear convolutional neural network model includes multiple convolutional layers; During the training process of the multi-layer linear convolutional neural network model based on the image samples, each convolutional layer performs a linear convolution operation on the corresponding input image and outputs a feature map, where the image samples are the initial input images. If the current convolutional layer belongs to the preset modulation level, then a modulation vector adapted to the current convolutional layer is generated based on the compression ratio vector specified by the user. The modulation vector is multiplied element-wise with the feature map output by the current convolutional layer, and the modulated feature map is output as the input image of the next convolutional layer. If the current convolutional layer does not belong to the preset modulation level, then there is no need to modulate the feature map output by the current convolutional layer and use it as the input image of the next convolutional layer. After all convolutional layers have completed the above process, the feature-enhanced image is obtained.

[0008] Optionally, the dual-branch structure model includes a global sampling branch module and a local sampling branch module; During the training process of the dual-branch structure model based on the image samples, the global sampling branch module uses the global discrete cosine transform method to perform global feature sampling on the feature-enhanced image to obtain global features; The local sampling branch module performs a scrambling operation on the feature enhancement image to generate a scrambling image; The scrambled image is block-sampled using a block diagonal Gaussian matrix to obtain local features; The global and local features constitute the compressed measurement data.

[0009] Optionally, during each iteration of training the image reconstruction model based on the image samples, the image reconstruction model performs gradient descent on the compressed measurement data using a data fidelity term to generate intermediate variables; The intermediate variables are mapped using a proximal operator to generate the reconstructed image after the current iteration. Once the iterative training is complete, the globally reconstructed image is obtained.

[0010] Optionally, constructing a global matrix equivalent to the global network model based on the bias vector and the plurality of basis response vectors includes: The multiple basis response vectors are subtracted from the bias vector one by one, and the results are arranged column by column to form a linear global matrix.

[0011] Optionally, the target device is the static random access memory of a digital micromirror device.

[0012] Secondly, this disclosure also provides an end-to-end compressed sensing imaging system, employing the following technical solution: The model training module is used to acquire an image sample set, train the compressed sensing imaging model based on the image sample set, and encapsulate the trained network parameters into a global network model. The vector construction module is used to receive the image size specified by the user and construct a zero vector and multiple standard basis vectors based on the image size; The first acquisition module is used to input the zero vector into the global network model and acquire the bias vector output by the global network model; The second acquisition module is used to input the multiple standard basis vectors into the global network model and acquire the multiple basis response vectors output by the global network model; A matrix construction module is used to construct a global matrix equivalent to the global network model based on the bias vector and the multiple basis response vectors. The matrix deployment module is used to deploy the global matrix to the target device.

[0013] Thirdly, this disclosure also provides a computer device, which adopts the following technical solution: The computer device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform any of the end-to-end compressed sensing imaging methods described above.

[0014] Fourthly, embodiments of this disclosure also provide a computer-readable storage medium storing computer instructions for causing a computer to perform any of the end-to-end compressed sensing imaging methods described above.

[0015] Fifthly, embodiments of this disclosure also provide a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of any of the methods described above.

[0016] The end-to-end compressed sensing imaging method provided in this disclosure trains a compressed sensing imaging model by acquiring a set of image samples and encapsulates the trained network parameters into a global network model. This allows for the optimization of model performance using a large amount of data, enabling the model to learn the features and patterns of the image, laying the foundation for accurate and efficient imaging. A zero vector and multiple standard basis vectors are constructed based on the user-specified image size. The zero vector is input into the global network model to obtain a bias vector, and the standard basis vectors are input to obtain multiple basis response vectors. A global matrix equivalent to the global network model is then constructed based on these vectors. This series of operations transforms the complex network model into a standard global matrix, facilitating subsequent application and deployment, and making the model's operation more intuitive and interpretable. Deploying the global matrix to the target device allows the target device to directly utilize the matrix for efficient imaging, solving the problem from software simulation to physical implementation. Similar to the matrix extraction algorithm mentioned in the references, this provides an effective solution for the development of learning-based computational imaging hardware systems and has strong industrialization value.

[0017] The above description is merely an overview of the technical solution disclosed herein. In order to better understand the technical means of this disclosure and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this disclosure more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic flowchart of the end-to-end compressed sensing imaging method provided in the embodiments of this disclosure; Figure 2 A flowchart illustrating the dynamic conditionalization method based on compression ratio vector provided in this embodiment of the disclosure; Figure 3 A flowchart illustrating the sampling methods of the global sampling branch module and the local sampling branch module provided in this embodiment of the disclosure; Figure 4 A schematic flowchart illustrating the image reconstruction method of the image reconstruction model provided in this embodiment of the disclosure; Figure 5 This is a schematic block diagram of an end-to-end compressed sensing imaging system provided in an embodiment of the present disclosure; Figure 6This is a schematic diagram of the structure of a computer device provided in an embodiment of the present disclosure. Detailed Implementation

[0020] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0021] It should be understood that the following specific examples illustrate the implementation of this disclosure, and those skilled in the art can easily understand other advantages and effects of this disclosure from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. This disclosure can also be implemented or applied through other different specific implementation methods, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0022] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0023] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this disclosure. The drawings only show the components related to this disclosure and are not drawn according to the number, shape and size of the components in actual implementation. In actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0024] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0025] Reference Figure 1 This disclosure provides an end-to-end compressed sensing imaging method, comprising the following steps: S1: Obtain an image sample set, train the compressed sensing imaging model based on the image sample set, and encapsulate the trained network parameters into a global network model. S2: Receive the image size specified by the user, and construct a zero vector and multiple standard basis vectors based on the image size; S3: Input the zero vector into the global network model and obtain the bias vector output by the global network model; S4: Input multiple standard basis vectors into the global network model to obtain multiple basis response vectors output by the global network model; S5: Construct a global matrix equivalent to the global network model based on the bias vector and multiple basis response vectors; S6: Deploy the global matrix to the target device.

[0026] The end-to-end compressed sensing imaging method disclosed herein trains a compressed sensing imaging model by acquiring a set of image samples and encapsulates the trained network parameters into a global network model. This method can utilize a large amount of data to optimize model performance, enabling the model to learn the features and patterns of the images, thus laying the foundation for accurate and efficient imaging in the future.

[0027] Based on the user-specified image size, a zero vector and multiple standard basis vectors are constructed. The zero vector is input into the global network model to obtain a bias vector, and the standard basis vectors are input to obtain multiple basis response vectors. Then, a global matrix equivalent to the global network model is constructed based on these vectors. This series of operations transforms the complex network model into a standard global matrix, facilitating subsequent applications and deployments, while also making the model's operation more intuitive and interpretable.

[0028] Deploying a global matrix to the target device allows the device to directly utilize the matrix for efficient imaging, solving the problem of transitioning from software simulation to physical implementation. Just like the matrix extraction algorithm mentioned in the references, it provides an effective solution for the development of learning-based computational imaging hardware systems and has strong industrialization value.

[0029] In S1, a set of image samples input by the user is received. This set contains multiple image samples, which are used to train a compressed sensing imaging model that combines multiple network structures. This compressed sensing imaging model includes a multi-layer linear convolutional neural network model, a two-branch structure model, and an image reconstruction model. The multi-layer linear convolutional neural network model and the two-branch structure model are collectively referred to as the Cooperative Sampling Operator (COSO). In order to overcome the limitations of traditional sampling methods in information capture, the Cooperative Sampling Operator does not simply perform random projection or local convolution on the image. Instead, it performs intelligent preprocessing and comprehensive multi-band capture of image information through a two-stage, collaborative process, like an efficient "intelligent information funnel".

[0030] A multi-layer linear convolutional neural network (LCN) model consists of multiple convolutional layers (e.g., a neural network structure composed of seven stacked convolutional layers). During the training process of the LCN model based on image samples, each convolutional layer performs a linear convolution operation on the corresponding input image, outputting a feature map. The image sample is the initial input image. From the first to the seventh layer, each layer contains multiple convolutional kernels (filters). The convolutional kernels slide on the input data (the original image during the first convolutional layer, and intermediate feature maps during subsequent layers) to perform convolution operations, mapping the input data onto a set of feature maps. Common convolutional kernel sizes include 3x3 and 5x5. Smaller convolutional kernels can capture local features, while larger kernels can capture a wider range of features. The stride controls the size of the convolutional kernel as it slides across the input data; a larger stride results in a smaller output feature map size. Padding refers to adding extra pixels around the boundaries of the input data to maintain the same size as the input data or meet specific requirements.

[0031] A multi-layer linear convolutional neural network model performs deep conditional filtering on the input image. Deep conditional filtering acts as an "intelligent gatekeeper" for information acquisition, aiming to adaptively preprocess the input image before sampling to enhance the features most important for the reconstruction task and generate a corresponding feature-enhanced image. Unlike existing techniques that use fixed filters or simple convolutional layers, this scheme employs a compact seven-layer linear convolutional neural network to achieve this function. It can learn an implicit representation of a complex linear transformation matrix and dynamically conditionalize it according to a user-specified compression ratio vector. This means that for different sampling rates, the filtering network can automatically adjust its filtering characteristics (e.g., focusing on smoothing to preserve structure at low sampling rates and enhancing details at high sampling rates), thereby providing the input image with optimal information content for subsequent sampling steps. This adaptability is not available in static sampling methods. (See reference...) Figure 2 The flowchart illustrates a dynamic conditionalization method based on compression ratio vectors. The method for modulating convolutional layers based on compression ratio vectors includes the following steps: S11: Determine whether the current convolutional layer belongs to the preset modulation level; if yes, execute S12; if no, execute S14. S12: Generate a modulation vector that is adapted to the current convolutional layer based on the compression ratio vector specified by the user; S13: Multiply the modulation vector element-wise with the feature map output by the current convolutional layer, and output the modulated feature map as the input image of the next convolutional layer. S14: There is no need to modulate the feature map output by the current convolutional layer and use it as the input image for the next convolutional layer.

[0032] In S11-S14, the preset modulation layers are typically 2 to 6. Fully connected layers are constructed, each corresponding to a preset modulation convolutional layer. Each fully connected layer is used to multiply the modulation vector with the compression ratio vector to obtain a modulation vector adapted to the corresponding preset modulation convolutional layer. The formula for calculating the modulation vector is as follows: In Formula 1, Indicates the number of convolutional layers; Indicates the relationship with the first Modulation vectors adapted by the convolutional layers; Indicates the relationship with the first Fully connected layers adapted to convolutional layers; This represents a user-specified compression ratio vector. ,in, This represents the global structure sampling rate, i.e., the DCT branch sampling rate. (D-Branch) corresponds to the DCT branch (GlobalDCT, or D branch for short). DCT transform is usually used to capture global low-frequency information of an image (i.e., the general outline and structure of the image). This represents the local texture sampling rate, also known as the Gaussian branch sampling rate. (G-Branch) corresponds to the Scrambled Block-diagonal Gaussian branch (G-branch for short). This branch combines random permutation scrambling and block sampling to capture mid-to-high frequency information such as texture and detail. and The total compression ratio is obtained by adding them together.

[0033] The formula for calculating the modulated feature map is as follows: In Formula 2, Indicates the first Feature maps modulated by the convolutional layer; Indicates the first Each convolutional layer performs a linear convolution operation; Indicates the first The input image for each convolutional layer.

[0034] After all convolutional layers have completed the above steps, the feature-enhanced image is obtained. This scheme dynamically generates channel scaling factors through compression ratio vectors, enabling the filtering network to adaptively adjust feature weights according to different sampling rates. At low sampling rates, it focuses more on smoothing and preserving structure, while at high sampling rates, it enhances details.

[0035] After deep conditional filtering, the feature-enhanced image is fed into a parallel dual-branch structure for final undersampling measurement. This design diverts information with different characteristics to dedicated channels for processing, ensuring comprehensive information capture. This dual-branch fast sampling structure is a dual-branch model, comprising a global sampling branch module (D branch) and a local sampling branch module (G branch). (Refer to...) Figure 3 The flowchart illustrates the sampling methods of the global sampling branch module and the local sampling branch module. During the training of the dual-branch structure model based on image samples, the sampling method of the two parallel sampling branch modules includes the following steps: S15: The global sampling branch module uses the global discrete cosine transform method to perform global feature sampling on the feature-enhanced image to obtain global features; S16: The local sampling branch module performs a scrambling operation on the feature enhancement image to generate a scrambling image; S17: Use a block diagonal Gaussian matrix to perform block sampling on the scrambled image to obtain local features; S18: Global features and local features constitute compressed measurement data.

[0036] In S15, the D branch is used for global structure capture. It utilizes the global Discrete Cosine Transform (DCT) for sampling. The DCT transforms the feature-enhanced image globally, converting it from the spatial domain to the frequency domain, thereby efficiently concentrating the energy of the natural image on a few low-frequency coefficients. According to preset rules, the low-frequency coefficients after the DCT transform are selected. By retaining these main low-frequency coefficients and discarding high-frequency coefficients, the D branch can capture the overall contour and macroscopic structural information of the image with extremely high efficiency, achieving global feature sampling of the feature-enhanced image.

[0037] In steps S16-S18, a global random permutation is performed on all pixels of the feature-enhanced image. This can be achieved by generating a random pixel index sequence and then rearranging the pixels of the feature-enhanced image according to this sequence. For example, consider a simple 4x4 image whose pixels can be numbered from left to right and from top to bottom, as shown in the pixel table below: Pixel table The image has 16 pixels, numbered from 0 to 15. The random pixel index sequence is obtained by randomly arranging these 16 numbers. For example, a possible random pixel index sequence is: [3,11,7,1,13,4,15,9,0,5,14,8,2,6,12,10]. During the "scrambling" operation, the pixels of the image are rearranged according to this random sequence. That is, the pixel numbered 0 in the original image will be placed in the 3rd position in the new image (because the first element of the random sequence is 3), the pixel numbered 1 in the original image will be placed in the 11th position in the new image, and so on. After this rearrangement, the spatial relationship of the image pixels is disrupted, and subsequent block diagonal sampling can better capture the high-frequency texture and fine details of the image, avoiding the block artifacts caused by traditional block sampling. If the image size is larger, such as an 8x8 image (64 pixels in total), the random pixel index sequence is a sequence of length 64 obtained by randomly arranging the 64 numbers from 0 to 63. The principle is the same as that for the 4x4 image above.

[0038] A block diagonal Gaussian matrix is ​​a special matrix structure that divides a scrambled image into multiple non-overlapping blocks. The sampling operations within each block are independent, and the sampling coefficients follow a Gaussian distribution. In this way, the scrambled image is sampled to capture local features of the image.

[0039] The core function of scrambling during sampling is to break spatial locality constraints, enabling the block diagonal sampling matrix to gain global perception of the entire image. Through random permutation, block sampling, which could only capture local information, can now collect information from the entire image. This significantly improves information preservation quality and robustness to spatial changes while maintaining computational efficiency. In other words, pixels at any location in the original image can potentially be assigned to any sampling block, effectively capturing high-frequency textures and fine details while avoiding block artifacts caused by traditional block sampling. This dual-branch design works collaboratively to ensure that, within a limited measurement budget, both the macroscopic structure and microscopic details of the image are preserved to the greatest extent possible.

[0040] After the entire sampling phase is completed, the acquired global and local features are combined to form compressed measurement data (also known as undersampled measurement data). Next, the reconstruction phase begins, where the image reconstruction model performs the task of restoring the image. To solve the "black box" problem of deep learning models, a new approach is taken in the design of the reconstruction network. It is not a deep network with arbitrarily stacked structures, but is constructed by "unrolling" the computation steps of a classic, mathematically transparent iterative optimization algorithm—Proximal Gradient Descent (PGD).

[0041] Reference Figure 4 The flowchart illustrating the image reconstruction method of the image reconstruction model demonstrates that, during the training process of the image reconstruction model based on image samples, the method for image reconstruction includes the following steps: S19: The image reconstruction model uses gradient descent on compressed measurement data through a data fidelity term to generate intermediate variables; S20: Use the proximal operator to perform proximal mapping on the intermediate variables to generate the reconstructed image after the current iteration; S21: After the iterative training is completed, the globally reconstructed image is obtained.

[0042] In S19-S21, the formulas for calculating intermediate variables are as follows: In formula 3, Indicates the first Intermediate variables in the next iteration; Indicates the number of iterations; Indicates the first The reconstructed image from the next iteration, also known as the image estimate, ; Indicates the learning step size (a learnable parameter). ; Represents the sampling matrix, ; This represents the transpose of the sampling matrix; This indicates compressed measurement data. .in, Represents the set of real numbers. and Representing dimensions, compressing the dimensions of measurement data Much smaller than the dimensions of the original image ; , Indicates residual connection, Generated by initializing the subnet; Equation 3 refers to the gradient descent step, obtained through the data fidelity term. Gradient update estimation.

[0043] The formula for reconstructing the image is as follows: In formula 4, express; This represents the proximal operator, implemented by a neural network in the PCNet case. Equation 4 refers to the proximal mapping step, which applies prior image constraints through a learnable neural network module.

[0044] Formulas 3 and 4 constitute one iteration of the proximal gradient descent algorithm. The first step updates the gradient by constraining the fidelity of the measurement data, and the second step applies image prior regularization through a learnable neural network proximal operator. These steps are executed alternately to gradually refine the reconstruction result. After the iteration is complete, the final reconstructed image is the globally reconstructed image. Each "layer" or "stage" in the reconstruction network of this scheme strictly corresponds to one complete iteration of the PGD algorithm. This makes every step of the entire reconstruction process—from gradient calculation to regularization constraints—have clear mathematical meaning, and its internal information flow is completely traceable and analyzable, thus transforming the "black box" into a "transparent box." Instead of using the manually designed, fixed regularization terms (such as the L1 norm) in the traditional PGD algorithm, it replaces the proximal mapping operator with a learnable module composed of a small neural network. This module can automatically learn more complex and accurate image intrinsic structure and texture priors from large-scale data than traditional sparse priors. Ultimately, the reconstruction network perfectly combines the logical clarity and interpretability of traditional optimization algorithms with the powerful feature representation capabilities of deep learning, achieving industry-leading reconstruction performance while ensuring model transparency.

[0045] In S2, the image dimensions are transformed into a one-dimensional vector, and the length N of this vector is obtained. For example, if the image dimensions are H×W, then N=H×W. An N-dimensional zero vector (i.e., a completely black image) and N standard basis vectors are constructed, where each standard basis vector contains only the first... One element is 1, and the rest are 0, and the positions of the elements with a value of 1 in each standard basis vector are distinct. .

[0046] In S3-S6, any linear operator F:RB→RD can be uniquely represented in matrix form. This scheme utilizes this fundamental principle, employing a series of carefully designed "probe" inputs to determine each element of the measurement matrix. The entire compressed sensing imaging model is treated as a linear "black box" function GA(⋅); then a zero vector is input. , The output of the compressed sensing imaging model at this point is the bias vector. ; sequentially convert each basis response vector Input the compressed sensing imaging model and obtain the corresponding output. The global matrix of the first The column is By traversing the N basis response vectors, the entire M×N global sampling matrix can be constructed. In other words, by performing subtraction operations between each basis response vector and the bias vector, and arranging the results column-wise, a linear global matrix can be formed. This operation ensures that each result becomes a column vector of the global matrix. Here, M is obtained by multiplying N by the user-specified global compression rate, where N is the total number of image pixels and M is the total number of measurements.

[0047] This global matrix deployment scheme is a core enabling technology that allows the technology to move from the laboratory to the factory and achieve commercial applications. It provides a specific and feasible algorithm that transforms the Cooperative Sampling Operator (COSO) and image reconstruction model, trained in software as a neural network, into explicit matrices and bias vectors that can be loaded and executed by physical hardware. The extracted global matrix contains M row vectors, each of which can be viewed as an independent sampling pattern. In a single-pixel imaging system, these patterns can be digitized and sequentially loaded into the SRAM (Static Random Access Memory) of a Digital Micromirror Device (DMD), which is the target device for global matrix deployment. The DMD controls the deflection of its micromirror array according to the loaded pattern, thereby physically achieving spatial modulation and sampling of the light field. This scheme is not only a technology specific to this invention but can also be seen as a "universal adapter" connecting artificial intelligence models and physical sensing hardware. It provides a physical implementation path for all learning-based linear sensing models, solving a core bottleneck that has long plagued the industry-academia-research transformation in this field, and possesses broad platform-level application value.

[0048] In summary, this solution demonstrates significant advantages in several aspects, including medical imaging. Firstly, it combines superior performance with complete interpretability. Through an algorithmic unfolding mechanism, the reconstruction process is built upon a PGD optimization framework, ensuring a clear physical explanation for each step of image reconstruction. This eliminates the uncertainty and trust issues inherent in "black box" models, providing traceable evidence for medical diagnosis and facilitating compliant product launches. Simultaneously, while maintaining interpretability, it achieves industry-leading reconstruction accuracy, outperforming existing methods when processing high-resolution medical images. Secondly, this solution enables seamless deployment from software models to physical hardware. Its unique matrix extraction scheme provides a practical hardware deployment path for complex learning-based sampling operators, transforming AI models into executable instructions for physical hardware. This allows advanced AI-based computational imaging technology to be truly implemented, applicable to fields such as industrial non-destructive testing where traditional cameras struggle. Furthermore, the solution's network model is compact and efficient, achieving superior performance with fewer parameters and computational complexity. This reduces the demands on hardware computing power, allowing deployment on conventional computing platforms and resource-constrained devices, thus lowering hardware costs. Furthermore, it possesses excellent generalization ability, exhibits outstanding performance and robustness when processing high-resolution images, and a single model can adapt to imaging tasks at any sampling rate, enhancing its practical value in various scenarios such as industrial inspection and clinical medical imaging.

[0049] To more intuitively demonstrate the advantages of this solution, the following performance comparison table compares it with traditional compressed sensing optimization algorithms and mainstream black-box deep learning models from multiple key dimensions. The performance comparison table is as follows: Effect Comparison Table Reference Figure 5 This disclosure provides an end-to-end compressed sensing imaging system, comprising: The model training module 101 is used to acquire an image sample set, train the compressed sensing imaging model based on the image sample set, and encapsulate the trained network parameters into a global network model. Vector construction module 102 is used to receive the image size specified by the user and construct a zero vector and multiple standard basis vectors based on the image size; The first acquisition module 103 is used to input the zero vector into the global network model and acquire the bias vector output by the global network model. The second acquisition module 104 is used to input multiple standard basis vectors into the global network model and acquire multiple basis response vectors output by the global network model; Matrix construction module 105 is used to construct a global matrix equivalent to the global network model based on the bias vector and multiple basis response vectors; The matrix deployment module 106 is used to deploy the global matrix to the target device.

[0050] The various variations and specific examples of the end-to-end compressed sensing imaging method provided above are also applicable to the end-to-end compressed sensing imaging system provided in this disclosure. Through the foregoing detailed description of the end-to-end compressed sensing imaging method, those skilled in the art can clearly understand the implementation method of the end-to-end compressed sensing imaging system. For the sake of brevity, they will not be described in detail here.

[0051] A computer device according to embodiments of the present disclosure includes a memory and a processor. The memory is used to store non-transitory computer-readable instructions. Specifically, the memory may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may, for example, include random access memory (RAM) and / or cache memory. The non-volatile memory may, for example, include read-only memory (ROM), hard disk, flash memory, etc.

[0052] The processor may be a central processing unit (CPU) or other form of processing unit with data processing capabilities and / or instruction execution capabilities, and may control other components in the computer device to perform desired functions. In one embodiment of this disclosure, the processor is used to execute computer-readable instructions stored in the memory, causing the computer device to perform all or part of the steps of the end-to-end compressed sensing imaging methods of the foregoing embodiments of this disclosure.

[0053] Those skilled in the art will understand that, in order to solve the technical problem of how to achieve a good user experience, this embodiment may also include well-known structures such as communication buses and interfaces, and these well-known structures should also be included within the protection scope of this disclosure.

[0054] like Figure 6 This is a schematic diagram of a computer device provided for an embodiment of the present disclosure. It illustrates a structural schematic diagram suitable for implementing the computer device in the embodiments of the present disclosure. Figure 6 The computer device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein.

[0055] like Figure 6 As shown, a computer device may include a processor (such as a central processing unit, graphics processing unit, etc.), which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) or programs loaded from storage devices into random access memory (RAM). The RAM also stores various programs and data required for the operation of the computer device. The processor, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0056] Typically, the following devices can be connected to the I / O interface: input devices, such as sensors or visual information acquisition devices; output devices, such as displays; storage devices, such as magnetic tapes or hard drives; and communication devices. Communication devices allow the computer device to communicate wirelessly or wiredly with other devices (such as edge computing devices) to exchange data. Although Figure 6 A computer apparatus with various devices is shown, but it should be understood that it is not required to implement or have all of the devices shown. More or fewer devices may be implemented or included alternatively.

[0057] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device, or installed from a ROM. When the computer program is executed by a processor, all or part of the steps of the end-to-end compressed sensing imaging method of embodiments of this disclosure are performed.

[0058] For a detailed description of this embodiment, please refer to the corresponding descriptions in the foregoing embodiments, which will not be repeated here.

[0059] A computer-readable storage medium according to embodiments of the present disclosure stores non-transitory computer-readable instructions. When these non-transitory computer-readable instructions are executed by a processor, all or part of the steps of the end-to-end compressed sensing imaging methods of the foregoing embodiments of the present disclosure are performed.

[0060] The aforementioned computer-readable storage media include, but are not limited to: optical storage media (e.g., CD-ROM and DVD), magneto-optical storage media (e.g., MO), magnetic storage media (e.g., magnetic tape or portable hard drive), media with built-in rewritable non-volatile memory (e.g., memory card), and media with built-in ROM (e.g., ROM cartridge).

[0061] For a detailed description of this embodiment, please refer to the corresponding descriptions in the foregoing embodiments, which will not be repeated here.

[0062] The basic principles of this disclosure have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this disclosure are merely examples and not limitations, and should not be considered as essential features of each embodiment of this disclosure. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the scope of this disclosure to the necessity of employing the aforementioned specific details for implementation.

[0063] In this disclosure, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The block diagrams of devices, apparatuses, devices, and systems involved in this disclosure are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as "comprising," "including," "having," etc., are open-ended terms meaning "including but not limited to," and are used interchangeably with them. The terms "or" and "and" as used herein refer to the terms "and / or," and are used interchangeably with them unless the context clearly indicates otherwise. The term "such as" as used herein refers to the phrase "such as but not limited to," and is used interchangeably with it.

[0064] Additionally, as used herein, the "or" used in a list of items beginning with "at least one" indicates a separate list, such that a list of, for example, "at least one of A, B, or C" means A or B or C, or AB or AC or BC, or ABC (i.e., A and B and C). Furthermore, the word "exemplary" does not imply that the described example is preferred or better than other examples.

[0065] It should also be noted that in the systems and methods of this disclosure, the components or steps can be decomposed and / or recombined. These decompositions and / or recombinations should be considered as equivalent solutions to this disclosure.

[0066] Various changes, substitutions, and modifications can be made to the technology described herein without departing from the teachings defined by the appended claims. Furthermore, the scope of the claims of this disclosure is not limited to the specific aspects of the processes, machines, manufactures, events, means, methods, and actions described above. Currently existing or later-developed processes, machines, manufactures, events, means, methods, or actions that perform substantially the same function or achieve substantially the same result as the corresponding aspects described herein can be utilized. Therefore, the appended claims include such processes, machines, manufactures, events, means, methods, or actions within their scope.

[0067] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this disclosure. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects without departing from the scope of this disclosure. Therefore, this disclosure is not intended to be limited to the aspects shown herein, but rather to be carried out within the widest scope consistent with the principles and novel features disclosed herein.

[0068] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this disclosure to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations therein.

Claims

1. An end-to-end compressive sensing imaging method, characterized in that, The method comprises: obtaining an image sample set, training a compressed sensing imaging model based on the image sample set, and encapsulating trained network parameters as a global network model; receiving a user-specified image size, and constructing a zero vector and a plurality of standard basis vectors based on the image size; inputting the zero vector into the global network model to obtain a bias vector output by the global network model; inputting the plurality of standard basis vectors into the global network model to obtain a plurality of basis response vectors output by the global network model; constructing a global matrix equivalent to the global network model based on the bias vector and the plurality of basis response vectors; and deploying the global matrix to a target device.

2. The end-to-end compressive sensing imaging method of claim 1, wherein, The image sample set comprises a plurality of image samples; The compressed sensing imaging model comprises a multi-layer linear convolutional neural network model, a double-branch structure model, and an image reconstruction model.

3. The end-to-end compressive sensing imaging method of claim 2, wherein, The multi-layer linear convolutional neural network model comprises a plurality of convolutional layers; In the training process of the multi-layer linear convolutional neural network model based on the image sample, each convolutional layer performs linear convolution operation on a corresponding input image to output a feature map, and the image sample is an initial input image; if the current convolutional layer belongs to a preset modulation level, a modulation vector adapted to the current convolutional layer is generated based on a user-specified compression ratio vector; the modulation vector is multiplied by the feature map output by the current convolutional layer element by element to output a modulated feature map as an input image of a next convolutional layer; if the current convolutional layer does not belong to the preset modulation level, the feature map output by the current convolutional layer is not modulated and is used as an input image of a next convolutional layer; after all convolutional layers are executed according to the above process, a feature-enhanced image is obtained.

4. The end-to-end compressive sensing imaging method of claim 3, wherein, The double-branch structure model comprises a global sampling branch module and a local sampling branch module; in the training process of the double-branch structure model based on the image sample, the global sampling branch module uses a global discrete cosine transform method to perform global feature sampling on the feature-enhanced image to obtain global features; the local sampling branch module performs scrambling operation on the feature-enhanced image to generate a scrambled image; a block-diagonal Gaussian matrix is used to perform block sampling on the scrambled image to obtain local features; the global features and the local features constitute compressed measurement data.

5. The end-to-end compressed sensing imaging method of claim 4, wherein in each iterative training process of the image reconstruction model based on the image sample, the image reconstruction model performs gradient descent on the compressed measurement data through a data fidelity term to generate an intermediate variable; a proximal operator is used to perform proximal mapping on the intermediate variable to generate a reconstructed image after current iteration; after iterative training is completed, a global reconstructed image is obtained.

6. The end-to-end compressive sensing imaging method of claim 1, wherein, The method of constructing a global matrix equivalent to the global network model based on the bias vector and the plurality of basis response vectors comprises: performing subtraction operation on the plurality of basis response vectors and the bias vector one by one, and arranging the obtained results in columns to form a linear global matrix.

7. The end-to-end compressive sensing imaging method of claim 1, wherein, The target device is a static random access memory of a digital micromirror device.

8. A computer apparatus, comprising: The computer device comprises: at least one processor; and a memory connected to the at least one processor in communication; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the end-to-end compressive sensing imaging method of any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer instructions for causing a computer to perform the end-to-end compressive sensing imaging method of any one of claims 1-7.

10. A computer program product comprising computer instructions, characterized in that, The computer instructions, when executed by a processor, implement the steps of the end-to-end compressive sensing imaging method of any one of claims 1-7.

Images