A compressed sensing imaging system and method based on convolutional reconstruction network

Abstract

The compressed sensing imaging system and method based on the convolutional reconstruction network provided by the present invention include a light source, an optical shaping unit, a light modulation unit, a projection unit and a photoelectric detection unit arranged in sequence along the optical path, and also include a data acquisition processing unit and a display unit connected to the photoelectric detection unit signal in sequence, and a light modulation control unit connected to the light modulation unit and the data acquisition processing unit signal, and the data acquisition processing unit includes a data acquisition unit, a storage unit, a calculation control unit and a multi-core convolution operation unit connected to each other. The present invention can not only realize parallel optimization calculation, greatly improve the reconstruction speed, save hardware development costs, improve system operation efficiency, facilitate system engineering and miniaturization development, but also retain more image structure information and realize higher precision image reconstruction.

Description

Technical Field

[0001] The present invention relates to the technical field of signal reconstruction, and in particular to a compressed sensing imaging system and method based on a convolutional reconstruction network. Background Art

[0002] Computational ghost imaging is a novel imaging technology developed in recent years. It is also called single-pixel imaging because it often uses a single-pixel camera to measure the target. However, traditional computational ghost imaging has problems such as a large number of sampling times, long imaging time, and complex system structure. These problems limit the application of computational ghost imaging. Thanks to the introduction of compressed sensing theory, computational ghost imaging technology based on compressed sensing has developed rapidly.

[0003] Compressed sensing is an emerging technology applied to signal sampling and processing, which can effectively reduce the number of sampling times and reduce the amount of storage. The ghost imaging method based on compressed sensing uses a measurement matrix to observe the target object, reconstructs the observation results using a reconstruction algorithm, and then inverts the image of the object. However, the construction of the measurement matrix needs to meet strict finite isometry conditions, and the currently commonly used random measurement matrix hardware implementation is complex and cannot be widely used. In addition, in terms of signal reconstruction, traditional compressed sensing algorithms still face problems such as high algorithm complexity, long reconstruction time, and many iterations, which also limit their practical application in the field of computational ghost imaging. Summary of the invention

[0004] In view of the problems existing in the prior art, the present invention provides a compressed sensing imaging system and method based on a convolutional reconstruction network. The imaging system has a simple structure, low computational complexity and high reconstruction accuracy.

[0005] In order to solve the above technical problems, the technical solution of the present invention is: a compressed sensing imaging system based on a convolutional reconstruction network, comprising a light source, an optical shaping unit, a light modulation unit, a projection unit and a photoelectric detection unit arranged in sequence along the optical path, and also comprising a data acquisition processing unit and a display unit connected to the photoelectric detection unit signal in sequence, and a light modulation control unit connected to the light modulation unit and the data acquisition processing unit signal, the data acquisition processing unit comprising a data acquisition unit, a storage unit, a calculation control unit and a multi-core convolution operation unit connected to each other.

[0006] Furthermore, the multi-core convolution operation unit is a hardware circuit module based on FPGA or DSP or multi-core CPU or GPU or multi-core ARM, which is used to reconstruct the convolution network and perform convolution parallel operations.

[0007] Furthermore, the multi-core convolution operation unit includes a data processing circuit of at least two single-core convolution operation modules.

[0008] Furthermore, the storage unit includes a network parameter storage unit and an operation storage unit, and adopts a ROM or RAM memory.

[0009] Furthermore, the photoelectric detection unit includes a photoelectric detector and an optical receiving system, the optical receiving system is used to collect light energy reflected by the imaging target, and the photoelectric detector is a single-pixel detector or an array detector, which is used to convert the light signal collected by the optical receiving system into an analog electrical signal.

[0010] Furthermore, the data acquisition unit is an oscilloscope or a data acquisition card, which is used to perform analog-to-digital conversion, data acquisition and preprocessing on the electrical signal output by the photoelectric detection unit.

[0011] Furthermore, it also includes a planar array imaging system connected to the data acquisition processing unit by signal, which is used to acquire images of the imaging target and provide them to the data acquisition unit.

[0012] The present application also provides an imaging method of a compressed sensing imaging system based on a convolutional reconstruction network, comprising the following steps:

[0013] S1: The data acquisition unit obtains sample data and expands it through data augmentation technology to form training samples, divides the obtained training samples into training sets and test sets, and stores them in the storage unit;

[0014] S2: The computing control unit allocates computing tasks and hardware resources according to the scale of the training set images and the initially set convolutional reconstruction network parameters, and controls the convolutional reconstruction network parameters and the data to be processed to enter the multi-core convolutional computing unit to train the initially constructed convolutional reconstruction network;

[0015] S3: After the training is completed, the first layer convolution parameters of the convolutional reconstruction network are used to construct a measurement matrix, and the constructed measurement matrix and the convolutional reconstruction network parameters are stored in a storage unit;

[0016] S4: The optical modulation control unit sequentially loads each row of data of the measurement matrix to control the modulation parameters of the optical modulation unit;

[0017] S5: The light source emits a light beam, which passes through the optical shaping unit and the light modulation unit in sequence to generate a light spot with a desired energy distribution, and then is projected onto the imaging target through the projection unit. The reflected light energy of the imaging target is received by the photoelectric detection unit and converted into an analog electrical signal;

[0018] S6: After the data acquisition unit acquires and preprocesses the analog electrical signal, it is stored in the storage unit. The calculation control unit transfers the convolution reconstruction network parameters in the storage unit and the data preprocessed by the data acquisition unit into the multi-core convolution operation unit for convolution parallel operation to obtain a reconstructed image of the imaging target, and displays it through the display unit.

[0019] Furthermore, the convolutional reconstruction network includes a sampling network and a reconstruction network connected in sequence, the sampling network includes a first convolutional layer and a second convolutional layer, the first convolutional layer is used to compress the input image, and the second convolutional layer is used to generate a plurality of feature maps; the reconstruction network consists of a plurality of convolutional layers. Modules The convolutional networks are stacked, each of which is composed of a plurality of convolutional layers. The reconstruction network outputs a feature map, which is added and fused with the output result of the sampling network to obtain the result of the convolutional reconstruction network.

[0020] Furthermore, in step S1, the sample data includes a number of randomly acquired images, and may also include a target image acquired by a surface imaging system.

[0021] The invention provides a method for a compressed sensing imaging system based on a convolutional reconstruction network, wherein the imaging system comprises a light source, an optical shaping unit, a light modulation unit, a projection unit and a photoelectric detection unit arranged in sequence along an optical path, and further comprises a data acquisition processing unit and a display unit connected to the photoelectric detection unit signal in sequence, and a light modulation control unit connected to the light modulation unit and the data acquisition processing unit signal, wherein the data acquisition processing unit comprises a data acquisition unit, a storage unit, a calculation control unit and a multi-core convolution operation unit connected to each other. The invention constructs a convolutional reconstruction network through a data acquisition processing unit and obtains an optimal measurement matrix after training it, modulates the light beam emitted by the light source through the light modulation control unit after loading the measurement matrix and projects it onto an imaging target, collects the light reflected by the imaging target through the photoelectric detection unit, and finally obtains a reconstructed image by performing convolution parallel operation through a multi-core convolution operation unit, and the reconstruction algorithm adopted by the imaging system is combined with a data processing system with a multi-core convolution operation unit as the core, which can not only realize parallel optimization calculation, greatly improve the reconstruction speed, save hardware development costs, improve system operation efficiency, and facilitate system engineering and miniaturization development, but also retain more image structure information and realize higher-precision image reconstruction. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 is a schematic structural diagram of a compressed sensing imaging system based on a convolutional reconstruction network in Example 1 of the present invention;

[0023] Figure 2Schematic diagram of the framework of the compressed sensing imaging system based on the convolutional reconstruction network in Embodiment 1 of the present invention;

[0024] Figure 3 is a schematic diagram of a convolutional reconstruction network in Example 1 of the present invention;

[0025] Figure 4 is a schematic structural diagram of a compressed sensing imaging system based on a convolutional reconstruction network in Embodiment 2 of the present invention;

[0026] Figure 5 is a schematic diagram of the framework of a compressed sensing imaging system based on a convolutional reconstruction network in Example 3 of the present invention;

[0027] As shown in the figure: 1. Sampling network; 11. First convolution layer; 12. Second convolution layer; 2. Reconstruction network; 10. Light source; 20. Optical shaping unit; 30. Light modulation unit; 40. Projection unit; 50. Photoelectric detection unit; 60. Data acquisition and processing unit; 610. Data acquisition unit; 620. Storage unit; 630. Computation control unit; 640. Multi-core convolution operation unit; 70. Light modulation control unit; 80. Imaging target; 90. Display unit; 100. Area array imaging system. DETAILED DESCRIPTION

[0028] The present invention is described in detail below in conjunction with the accompanying drawings.

[0029] Example 1

[0030] like Figure 1-2As shown, the present invention provides a compressed sensing imaging system based on a convolutional reconstruction network, including a light source 10, an optical shaping unit 20, a light modulation unit 30, a projection unit 40 and a photoelectric detection unit 50 arranged in sequence along the optical path, and also including a data acquisition processing unit 60 and a display unit 90 connected to the photoelectric detection unit 50 in sequence, and a light modulation control unit 70 connected to the light modulation unit 30 and the data acquisition processing unit 60, wherein the data acquisition processing unit 60 includes a data acquisition unit 610, a storage unit 620, a calculation control unit 630 and a multi-core convolution operation unit 640 connected to each other. Specifically, the light source 10 can be any light source, such as an LED light source or a laser light source, or a pulse light source or a continuous light source. The optical shaping unit 20 is used to control the divergence angle, shape, and energy distribution of the outgoing light beam, and may adopt optical devices such as an optical lens group or a diffractive optical element or an aperture. The light modulation unit 30 is a spatial light modulator, such as a digital micromirror device (DMD), which is used to modulate the distribution of light intensity. The light beam emitted by the light source 10 forms a light spot with the required energy distribution after passing through the optical shaping unit 20 and the light modulation unit 30. The projection unit 40 projects the shaped and modulated light beam onto the imaging target 80 at a certain divergence angle, the photoelectric detection unit 50 receives the light energy reflected by the imaging target 80 and converts the light signal into an analog electrical signal, the data acquisition unit 610 is used to collect and preprocess the analog electrical signal output by the photoelectric detection unit 50, and store the preprocessed data in the storage unit 620, the calculation control unit 630 allocates the calculation tasks and hardware resources according to the parameters such as the calculation amount, and controls the convolution reconstruction network parameters and the preprocessed data of the data acquisition unit 610 to enter the multi-core convolution operation unit 640 for convolution parallel operation, and finally obtains the high-quality reconstructed image of the imaging target through multi-layer convolution reconstruction operation. The convolution reconstruction network provided by the present invention can not only solve the design problem of the measurement matrix in the computational ghost imaging system, but also provide a fast and effective reconstruction algorithm, which can retain more image structure information and achieve better reconstruction. In addition, the network model proposed by the present invention has a simple structure, can learn the mapping relationship between the measurement signal and the reconstructed image end-to-end, and compared with the traditional compressed sensing method, the present invention can obtain better reconstruction results at a lower sampling rate.

[0031] Preferably, the multi-core convolution operation unit 640 can perform convolution parallel operation, and can be a hardware circuit module based on FPGA, DSP, multi-core CPU, GPU, multi-core ARM or a chip that can realize the above functions. Preferably, the multi-core convolution operation unit 640 includes a data processing circuit of ≥2 single-core convolution operation modules, which can be developed based on FPGA (Field Programmable Gate Array, Field Programmable Logic Gate Array), which has the characteristics of real-time performance, high operating efficiency, low system integration difficulty, etc., and is suitable for small volume and low power consumption systems. It should be noted that a module that can independently complete a convolution operation task is called a single-core convolution operation unit, which is equivalent to a core of a CPU or a single-core ARM chip or a stream processing unit of a GPU or a convolution operation module composed of a few multipliers and adders in FPGA development. The multi-core convolution operation unit 640 includes an operation unit array composed of ≥2 single-core convolution operation units, which can independently complete the convolution operation tasks allocated by the calculation control unit 630. The function of parallel operation of multi-convolution tasks of greater than or equal to 2 is realized, and the operation tasks are arranged by the calculation control unit 630. Generally, the number of single-core convolution operation units in the multi-core convolution operation unit 640 can range from 2 to hundreds, and the maximum number is determined by hardware resources and flexibly adjusted according to the operation scale.

[0032] Preferably, the photoelectric detection unit 50 includes a photoelectric detection device and an optical receiving system. The optical receiving system is used to collect light energy reflected by the imaging target. The photoelectric detection device can be a single-pixel detector, such as APD, PMT, PIN, etc., which is used to convert light signals into electrical signals. Of course, it can also be an array detection device such as CCD, CMOS, APD array, etc. When using an array detection device, it is only necessary to superimpose the signals collected by all pixels as the final usable data, that is, use it as a single-pixel detector.

[0033] The data acquisition unit 610 can be an oscilloscope or a data acquisition card, or any other device with data acquisition function, for collecting and preprocessing the electrical signal output by the photoelectric detection unit 50, and has data acquisition, data preprocessing and analog-to-digital conversion functions.

[0034] Preferably, the storage unit 620 includes a network parameter storage unit and an operation storage unit, which can be a ROM or RAM memory, or other devices and equipment for data storage. The network parameter storage unit is used to store the parameters initially set during the training of the convolution reconstruction network, continuously updated during the training process, and obtained after the training, while the operation storage unit is used to store the intermediate training quantity, the data preprocessed by the data acquisition unit, the final reconstructed image, the observation value vector and other related data. The network parameter storage unit determines the required storage space based on the selected network characteristics, and the operation storage unit determines the required storage space based on the maximum amount of data generated during the operation.

[0035] The present invention also provides an imaging method of a compressed sensing imaging system based on a convolutional reconstruction network, comprising the following steps:

[0036] S1: The data acquisition unit 610 obtains sample data and expands it through data augmentation technology to form training samples, and divides the obtained training samples into training sets and test sets, and stores them in the storage unit 620; specifically, in this embodiment, the data acquisition unit 610 can obtain an image set consisting of a random number of images, such as from any image database, and randomly cut out multiple image blocks with a resolution of 128*128 from each image, and obtain a total of 200,000 images with a resolution of 128*128 through data augmentation technologies such as rotation, translation, and folding, and divide them into training sets and test sets according to 8:2, and send the images in the training set to the storage unit 620 in sequence. It should be noted that the imaging target here does not refer to a specific imaging target, but refers to various different imaging targets.

[0037] S2: The calculation control unit 630 allocates computing tasks and hardware resources according to the scale of the images in the training set and the initially set convolutional reconstruction network parameters, and controls the convolutional reconstruction network parameters and the images in the training set to enter the multi-core convolutional operation unit to train the initially constructed convolutional reconstruction network; specifically, the convolutional reconstruction network to be trained includes two parts, one is the sampling network 1, which is responsible for compressing the input image at a certain sampling rate; the other is the reconstruction network 2, which reconstructs the compressed signal to restore the original signal. In this embodiment, the sampling network 1 includes two convolutional layers, namely the first convolutional layer 11 and the second convolutional layer 12, wherein the first convolutional layer 11 uses a convolution kernel of 64*64 size to compress the input image signal according to a step size of 64; when the sampling rate is 0.1, the number of convolutional channels is set to 410; when the sampling rate is 0.2, the number of convolutional channels is set to 819. The second convolution layer 12 of the sampling network 1 uses a convolution kernel of size 3*3, and 4096 feature maps will be obtained. The number of pixels contained in the feature map is exactly equal to the number of pixels of the input image, so the feature map is converted to the image resolution size here.

[0038] like Figure 3 As shown, the reconstruction network 2 is connected to the entire sampling network to reconstruct the sampling results. In this embodiment, the reconstruction network uses a structure of 20 residual blocks stacked, each residual block contains three convolution layers, and the convolution layers in the residual blocks all use 3*3 convolution kernels, with a step size of 3 and a padding of 1. Each convolution layer will generate 128 feature maps, and each convolution layer will be followed by a ReLU activation function. The last layer of the reconstruction network uses a 5*5 convolution to output a feature map, which is additively fused with the output of the sampling network as the reconstruction result of the convolution reconstruction network.

[0039] Then, the error between the reconstruction result and the corresponding input image is calculated using the loss function. When the error between the two is less than the set value, the training ends. Otherwise, the above training process is repeated. In this embodiment, the mean square error (MSE) function can be used, as shown in formula (1):

[0040]

[0041] Where N represents the number of pixels of the input image, X and Y represent the image input by the sampling network and the reconstruction result output by the convolutional reconstruction network, respectively.

[0042] S3: After the training is completed, the first convolution layer parameters of the convolution reconstruction network are used to construct a measurement matrix, and the constructed measurement matrix and the convolution reconstruction network parameters are stored in the storage unit 620; specifically, the first convolution layer of the network is used to compress the input image signal, and a mapping relationship between the input image and the compressed signal can be obtained, which is expressed as: Where X is the input image signal, is the compressed signal, and F represents the convolution operation. The above convolution process can be converted into matrix multiplication, which is mathematically expressed as The dimension of C is M×N, N is the dimension of the input image signal, and C is the constructed measurement matrix. The calculated measurement matrix and the remaining network parameters after removing the first convolution layer parameters of the sampling network are stored in the corresponding positions of the storage unit 620. Since the first convolution layer parameters have been converted into the measurement matrix, no additional storage is required. It should be noted that the measurement matrix constructed by the present invention is a 0-1 binary matrix, which can be effectively applied and stored by the computational ghost imaging system.

[0043] S4: The light modulation control unit 70 loads each row of data of the measurement matrix in sequence to control the modulation parameters of the light modulation unit 30.

[0044] S5: The light source 10 emits a light beam, which passes through the optical shaping unit 20 and the light modulation unit 30 in sequence to generate a light spot with a desired energy distribution, and then is projected onto the imaging target 80 through the projection unit 40. The reflected light energy of the imaging target 80 is received by the photoelectric detection unit 50 and converted into an analog electrical signal;

[0045] S6: After the data acquisition unit 610 acquires and preprocesses the analog electrical signal, it is stored in the storage unit 620. The calculation control unit 630 transfers the convolution reconstruction network parameters in the storage unit 620 and the data preprocessed by the data acquisition unit 610 into the multi-core convolution operation unit 640 for convolution parallel operation to obtain a reconstructed image of the imaging target. Specifically, in the process of convolution parallel operation, the result obtained after each convolution layer operation is completed is returned to the storage unit 620 for temporary storage, waiting for the next convolution layer operation until the preset number of reconstruction layers is completed. After being sent to the multi-core convolution operation unit 640 for parallel calculation, high-quality imaging of the imaging target 80 can be obtained. For an image with a resolution of 128*128, a GPU with 192 stream processing units is used for convolution parallel calculation, which takes only 0.015ms, and the fastest reconstruction efficiency can reach more than 60fps.

[0046] Example 2

[0047] like Figure 4 As shown, unlike Example 1, the imaging system in this embodiment also includes a planar array imaging system 100 connected to the data acquisition processing unit 60 by signal, which is used to acquire images of the imaging target 80 and provide them to the data acquisition unit 610 for training the convolutional reconstruction network. The planar array imaging system 100 is used to receive high-resolution images of the imaging target 80. It is possible to acquire sample data in real time under the system's own lighting conditions, and obtain optimal parameters by training the convolutional reconstruction network parameters. After obtaining the optimal network parameters, the imaging system is started to work normally. This method can be carried on any vehicle or movable platform to collect target images in a specific scene in real time for convolutional reconstruction network training. This makes the training process more suitable for a specific application scenario, the reconstruction accuracy is higher, and the environmental adaptability of the entire compressed sensing imaging system is improved.

[0048] Example 3

[0049] like Figure 5As shown, different from Example 1, in this embodiment, the data acquisition processing unit 60 and the light modulation control unit 70 are integrated into the overall control and processing unit of the compressed sensing system. The computing control unit 630 is an overall control unit developed with a single or multiple FPGAs as the core, which realizes the unified resource allocation and control of the light source 10, the light modulation unit 30, the data acquisition unit 610, the storage unit 620, and the multi-core convolution operation unit 640. The storage unit 620 realizes the storage function of the network parameters, the light modulation matrix, and the optoelectronic data. The overall control and processing unit of the compressed sensing system is obtained by integration, which reduces the system complexity, saves the hardware development cost, and reduces the system volume and power consumption.

[0050] Although the embodiments of the present invention are described in the specification, these embodiments are only for reference and should not limit the protection scope of the present invention. Various omissions, substitutions and changes within the scope of the present invention should be included in the protection scope of the present invention.

Claims

1. A compressed sensing imaging system based on a convolutional reconstruction network, characterized in that: The invention comprises a light source, an optical shaping unit, a light modulation unit, a projection unit and a photoelectric detection unit arranged in sequence along the optical path, and also comprises a data acquisition processing unit and a display unit connected to the photoelectric detection unit signal in sequence, and a light modulation control unit connected to the light modulation unit and the data acquisition processing unit signal. The data acquisition processing unit comprises a data acquisition unit, a storage unit, a calculation control unit and a multi-core convolution operation unit connected to each other. The data acquisition processing unit constructs a convolution reconstruction network and obtains an optimal measurement matrix after training the network. The light modulation control unit loads the measurement matrix to modulate the light beam emitted by the light source and then projects the modulated light beam onto an imaging target. The light reflected by the imaging target is collected by the photoelectric detection unit. The calculation control unit controls the parameters of the convolution reconstruction network and the data preprocessed by the data acquisition unit to enter the multi-core convolution operation unit for convolution parallel operation, and obtains a reconstructed image through multi-layer convolution reconstruction operation.

2. The compressed sensing imaging system based on convolutional reconstruction network according to claim 1, characterized in that: The multi-core convolution operation unit is a hardware circuit module based on FPGA or DSP or multi-core CPU or GPU or multi-core ARM, which is used to reconstruct the convolution network and perform convolution parallel operations.

3. The compressed sensing imaging system based on convolutional reconstruction network according to claim 1, characterized in that: The multi-core convolution operation unit includes a data processing circuit of at least two single-core convolution operation modules.

4. The compressed sensing imaging system based on convolutional reconstruction network according to claim 1, characterized in that: The storage unit includes a network parameter storage unit and an operation storage unit, and adopts a ROM or RAM memory.

5. The compressed sensing imaging system based on convolutional reconstruction network according to claim 1, characterized in that: The photoelectric detection unit includes a photoelectric detector and an optical receiving system. The optical receiving system is used to collect light energy reflected by the imaging target. The photoelectric detector is a single-pixel detector or an array detector, which is used to convert the light signal collected by the optical receiving system into an analog electrical signal.

6. The compressed sensing imaging system based on convolutional reconstruction network according to claim 1, characterized in that: The data acquisition unit is an oscilloscope or a data acquisition card, which is used to perform analog-to-digital conversion, data acquisition and preprocessing on the electrical signal output by the photoelectric detection unit.

7. The compressed sensing imaging system based on convolutional reconstruction network according to claim 1, characterized in that: It also includes a planar array imaging system connected to the data acquisition processing unit for signal acquisition and processing, and is used to acquire images of the imaging target and provide the images to the data acquisition unit.

8. An imaging method of a compressed sensing imaging system based on a convolutional reconstruction network as claimed in claim 1, characterized in that: The following steps are involved: S1: The data acquisition unit obtains sample data and expands it through data augmentation technology to form training samples, divides the obtained training samples into training sets and test sets, and stores them in the storage unit; S2: The computing control unit allocates computing tasks and hardware resources according to the scale of the training set images and the initially set convolutional reconstruction network parameters, and controls the convolutional reconstruction network parameters and the data to be processed to enter the multi-core convolutional computing unit to train the initially constructed convolutional reconstruction network; S3: After the training is completed, the first layer convolution parameters of the convolutional reconstruction network are used to construct a measurement matrix, and the constructed measurement matrix and the convolutional reconstruction network parameters are stored in a storage unit; S4: The optical modulation control unit sequentially loads each row of data of the measurement matrix to control the modulation parameters of the optical modulation unit; S5: The light source emits a light beam, which passes through the optical shaping unit and the light modulation unit in sequence to generate a light spot with a desired energy distribution, and then is projected onto the imaging target through the projection unit. The reflected light energy of the imaging target is received by the photoelectric detection unit and converted into an analog electrical signal; S6: After the data acquisition unit acquires and preprocesses the analog electrical signal, it is stored in the storage unit. The calculation control unit transfers the convolution reconstruction network parameters in the storage unit and the data preprocessed by the data acquisition unit into the multi-core convolution operation unit for convolution parallel operation to obtain a reconstructed image of the imaging target, and displays it through the display unit.

9. The imaging method of the compressed sensing imaging system based on the convolutional reconstruction network according to claim 8, characterized in that: The convolutional reconstruction network includes a sampling network and a reconstruction network connected in sequence, the sampling network includes a first convolutional layer and a second convolutional layer, the first convolutional layer is used to compress the input image, and the second convolutional layer is used to generate a plurality of feature maps; the reconstruction network is formed by stacking a plurality of convolutional modules, each of which is composed of a plurality of convolutional layers, and the reconstruction network outputs a feature map, which is additively fused with the output result of the sampling network to obtain the result of the convolutional reconstruction network.

10. The imaging method of the compressed sensing imaging system based on the convolutional reconstruction network according to claim 8, characterized in that: In step S1, the sample data includes a number of randomly acquired images, and may also include a target image acquired by a surface imaging system.

Images