Computing correlation imaging reconstruction algorithms based on deep convolutional hybrid neural networks

Abstract

The application belongs to the technical field of quantum imaging of computational imaging, and specifically discloses a correlation imaging reconstruction algorithm based on a deep convolutional hybrid neural network, which uses a Hadamard matrix gray scale coding sequence to replace a Gaussian matrix speckle sequence of original correlation imaging, and adjusts a digital micromirror device with a fixed sequence of speckle sequences in data acquisition and processing of correlation imaging. Under the same sampling rate, the algorithm can reconstruct a high-quality target object image, and has higher PSNR and SSIM values. Deep learning realizes end-to-end image recovery of an image in the algorithm implementation of correlation imaging, and reduces the steps of image calculation.

Description

Technical Field

[0001] This invention belongs to the field of quantum imaging technology in computational imaging, and specifically relates to an algorithm for computational correlation imaging reconstruction based on deep convolutional hybrid neural networks. Background Technology

[0002] Computational correlation imaging, also known as computational ghost imaging, is a type of imaging method that uses the correlation properties of light fields to calculate images, exhibiting the characteristics of "creating something from nothing" and "imaging away from the object." Correlation imaging has gone through two stages: The first stage involves a dual-optical-path system, namely the object-optical-arm system and the reference optical-arm system. In the object-optical-arm system, light passes through an object to be imaged, and the total light intensity is collected by bucket detection. In the reference optical-arm system, the light intensity distribution of the light field is received by a detector with resolution measurement. By jointly measuring the outgoing light from these two optical-path systems, the imaging information can be obtained. The second stage only requires a single-pixel detector to receive light intensity information, eliminating the need for the reference optical-arm system; this is called computational ghosting (CGI). Computational correlation imaging has overcome the constraints of imaging systems and light field modulation devices, mainly due to the development of spatial optical modulation devices, such as spatial light modulators, digital micromirror devices, LED arrays, and laser arrays.

[0003] Two key factors affecting image quality in computational correlation imaging are the construction of the illumination speckle pattern during measurement and the correlation algorithm for image reconstruction. In constructing the illumination speckle, methods such as orthogonalized sinusoidal speckle, orthogonal Hada matrices, Walsh-Hadama transform (WT), and 4-connected Hadama transform (CR) have been proposed to optimize the measurement illumination speckle. Regarding the reconstruction algorithm for computational correlation images, differential CGI and pseudo-inverse CGI algorithms have improved image quality. However, under low sampling conditions, the reconstructed images using differential CGI and pseudo-inverse CGI algorithms remain blurry and unclear. Summary of the Invention

[0004] The purpose of this invention is to overcome the defects in the existing technology and provide an algorithm for computational correlation imaging reconstruction based on deep convolutional hybrid neural networks.

[0005] This invention provides an algorithm for calculating correlated imaging reconstruction based on deep convolutional hybrid neural networks, comprising the following steps:

[0006] Step 1: Data Collection;

[0007] Prepare a clean MNIST dataset and enhance and expand it. Generate scattering spots based on the Hadamard matrix and obtain the image dataset using the data collected from the scattering spot sampling rate.

[0008] Step 2: Establish a deep convolutional hybrid neural network model;

[0009] The deep convolutional hybrid neural network model is a DCMGI model, including:

[0010] Feature extraction module: Contains 5 convolutional modules, the convolutional modules are as follows:

[0011] c_block_n(in_channel,out_channel)

[0012] Where n is the nth module of c_block, and in_channel and out_channel are the number of input channels and output channels of the convolution module;

[0013] Feature enhancement module: Includes channel attention mechanism and spatial attention mechanism. The output channel of the feature extraction module is used as the input channel of the feature enhancement module, and the channel attention mechanism and spatial attention mechanism are processed respectively.

[0014] Output module: includes convolutional layer Conlayer1 and convolutional layer Conlayer2;

[0015] The convolutional layer Conlayer1 has 10 input channels, 5 output channels, a kernel size of 3*3, padding of 1, and stride of 1.

[0016] The convolutional layer Conlayer2 has 5 input channels, 1 output channel, a kernel size of 3*3, padding of 1, and stride of 1.

[0017] Add a normalization layer and activation function to the convolutional layer Conlayer1;

[0018] Step 3: Model training;

[0019] The image dataset described in step 1 is divided into a training set and a test set according to the proportion. The training set is input into the DCMGI model, and the DCMGI model is trained based on the training optimizer and loss function.

[0020] Step 4: Save the trained DCMGI model as DCMGI.pth to provide a model for real-world scene loading and image reconstruction.

[0021] A further approach is to augment the MNIST dataset from step 1 into an image of size N*N.

[0022] The formation of an N*N speckle sequence includes the following steps:

[0023] Step 1.11: Hadamard matrix column vector sorting generation: The right node value of the root node is [1,1]. The right node value of the root node is formed by concatenating [1,1] with 1*[1,1], and its value is [1,1,1,1]. The left node value is formed by concatenating [1,1] with -1*[1,1], and its value is [1,1,-1,-1]. This continues until the length of the leaf node value is N.

[0024] [1,-1] is the value of the left node rooted at [1]. The value of the left node rooted at [1,-1] is formed by concatenating [1,-1] with 1*[1,-1], and its value is [1,-1,1,-1]. The value of the right node is formed by concatenating [1,-1] with -1*[1,-1], and its value is [1,-1,-1,1]. This continues until the length of the leaf node is N. The left and right leaf nodes rooted at [1] are sorted into column vectors from right to left.

[0025] Step 1.12: Hadamard matrix row vector sorting generation: The left node value of the root node is [1,1]. The left node value of the root node is formed by concatenating [1,1] with 1*[1,1], and its value is [1,1,1,1]. The right node value is formed by concatenating [1,1] with -1*[1,1], and its value is [1,1,-1,-1]. This continues until the length of the leaf node value is N.

[0026] [1,-1] is the value of the left and right nodes with [1] as the root node. The value of the left node with [1,-1] as the root node is formed by concatenating [1,-1] with -1*[1,-1], and its value is [1,-1,-1,1]. The value of the right node is formed by concatenating [1,-1] with 1*[1,-1], and its value is [1,-1,1,-1]. This continues until the length of the leaf node is N. The left and right leaf nodes with [1] as the root are sorted into row vectors from left to right.

[0027] Step 1.13: Generation of an N*N speckle matrix: Each element is a speckle, and the speckle size is N*N. Multiply each element of the column vector sorted by each element of the row vector sorted, and set the speckle value of -1 to 0. Finally, an N*N speckle matrix is ​​formed.

[0028] Step 1.14: Perform a zigzag scan sort on the illumination speckle matrix to form a fixed illumination speckle sequence.

[0029] A further proposed solution is that the data collection method described in step 1 includes the following steps:

[0030] Step 1.21: Calculate the light intensity collected by the barrel detector in each measurement of the correlated imaging. The formula for calculating the light intensity is as follows:

[0031] It=∫dx|t(x)| 2 *Bt(x)

[0032] Where t(x) represents the spatial distribution information of the object to be imaged, Bt(x) represents the light intensity distribution of the light field at the object, It is the total light intensity value collected, and the subscript t indicates the t-th sampling;

[0033] Step (1.22): Setting the sampling rate β: If the image size is N*N, then the total number of samplings T is: N*N, where M is the number of samplings required, and the sampling rate β = M / T;

[0034] Step (1.23): Determine the number of samplings M based on the sampling rate β in step 1.22, and take M speckle patterns from the beginning of the speckle sequence generated in step 1.14.

[0035] Step (1.24): M illumination speckles are sequentially loaded onto the digital micromirror device to control the light field. The light intensity of the illuminated image is collected using a barrel detector, and the total collected light intensity is denoted as It (t represents the t-th sampling). The image is illuminated by M measurements, and the collected light intensity sequence is as follows:

[0036] I = [I1, I2, I3, ..., I M ]

[0037] The light intensity sequence is multiplied by the corresponding M scattered light spot sequences to form M-channel data, which is stored in JSON format. The file name is consistent with the image number. The image and the M-channel data form an image dataset.

[0038] A further embodiment is that the convolutional module includes a convolutional layer Conv1 and a convolutional layer Conv2;

[0039] The Conv1 convolutional layer has a kernel size of 3*3, padding set to 1, stride set to 1, and both the input and output channels are in_channel.

[0040] The Conv2 convolutional layer has a kernel size of 3*3, padding of 1, stride of 1, input channel of in_channel, and output channel of out_channel.

[0041] A further solution is that the channel attention mechanism includes:

[0042] Global average pooling and global max pooling are performed on the input data of the input channel;

[0043] The results of average pooling and max pooling are processed using a shared multilayer perceptron. The two processed results are added together and then processed by an activation function to obtain the weight of each channel of the input data.

[0044] The weight of each channel is multiplied by the input data of the input channel to obtain the output data;

[0045] The spatial attention mechanism includes:

[0046] For each feature point of the input data in the input channel, the maximum value and average value are taken in the input channel, and the maximum value and average value are stacked to form 2-channel data;

[0047] The 2-channel data is passed through a convolutional layer with 1 output channel to adjust the number of channels, and then processed by an activation function to obtain the weight of each feature point.

[0048] Multiply the weight data by the input data;

[0049] The convolutional layer has a kernel size of 3*3, padding of 1, and stride of 1.

[0050] A further proposed solution is that, in step 3, the loss function is the mean squared loss function, and the formula for the loss function is:

[0051] Loss(x i y i )=(x i y i ) 2

[0052] Where loss is the name of the loss function, x i The original image, y i To reconstruct the image;

[0053] The training optimizer uses the Adam optimizer, and the learning rate parameter of the Adam optimizer is set to 0.0001.

[0054] Compared with existing technologies, the advantages of this invention are as follows: This invention discloses a correlation imaging reconstruction algorithm based on a deep convolutional hybrid neural network. At the same sampling rate, this algorithm can reconstruct high-quality target object images with higher PSNR and SSIM values. At high sampling rates, this algorithm provides superior reconstruction quality and retains richer image information. Even at low sampling rates, it can retain key image information. Furthermore, this invention optimizes the illumination speckle sequence during data acquisition, enabling the reconstruction of high-quality images with limited sampled data, reducing computational imaging time. Using the light intensity sequence and the corresponding speckle sequence as data sources, and a deep neural network model as the correlation function for calculating correlation imaging, end-to-end image reconstruction is achieved. Attached Figure Description

[0055] The following figures are for illustrative purposes only and are not intended to limit the scope of the invention, wherein:

[0056] Figure 1 : Block diagram of deep convolutional hybrid neural network structure;

[0057] Figure 2 : Block diagram of the convolution module c_block_n;

[0058] Figure 3 : Block diagram of feature extraction module;

[0059] Figure 4 Feature enhancement module structure diagram;

[0060] Figure 5 Shared MLP structure block diagram;

[0061] Figure 6 Channel attention structure diagram;

[0062] Figure 7 Spatial attention structure diagram;

[0063] Figure 8 Input module structure diagram;

[0064] Figure 9 Comparison of the effects of three computational imaging methods at a sampling rate of 6.25%. Detailed Implementation

[0065] To make the objectives, technical solutions, design methods, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0066] This invention provides a deep convolutional hybrid neural network-based algorithm for correlated imaging reconstruction. Specifically, it uses the gray-level encoding and sorting of Hadamard matrix scattering spots as a measurement, and uses intensity sequences and scattering spot sequences as data sources for deep neural network image reconstruction. The algorithm includes the following steps:

[0067] Step (1): Enhance the 70,000 data points of the MNIST dataset by performing horizontal, vertical and rotation operations, and uniformly expand them into N*N size images;

[0068] Step (2): Generate scattered light spots based on the Hadamard matrix, and sort the scattered light spots by grayscale encoding to form a fixed sequence of light scattered light spots. The specific steps include the following:

[0069] Step (2.1): The Hadamard matrix column vectors are sorted and generated. The right node value of the root node [1] is [1,1]. The right node value of the root node [1,1] is formed by concatenating [1,1] with 1*[1,1], and its value is [1,1,1,1]. The left node value is formed by concatenating [1,1] with -1*[1,1], and its value is [1,1,-1,-1]. The same pattern is followed for generating the right and left node values ​​of the subtree until the length of the leaf node value is N.

[0070] [1,-1] is the value of the left node rooted at [1]. The value of the left node rooted at [1,-1] is formed by concatenating [1,-1] with 1*[1,-1], and its value is [1,-1,1,-1]. The value of the right node is formed by concatenating [1,-1] with -1*[1,-1], and its value is [1,-1,-1,1]. The same pattern is followed when generating the values ​​of the left and right child nodes of the subtree until the length of the leaf node is N. The left and right leaf nodes rooted at [1] are sorted into column vectors from right to left.

[0071] Step (2.2): The Hadamard matrix row vectors are sorted and generated. The left node value of the root node [1] is [1,1]. The left node value of the root node [1,1] is formed by concatenating [1,1] with 1*[1,1], and its value is [1,1,1,1]. The right node value is formed by concatenating [1,1] with -1*[1,1], and its value is [1,1,-1,-1]. The same pattern is followed for generating the right and left node values ​​of the subtree until the length of the leaf node value is N.

[0072] [1,-1] is the value of the left and right nodes of the root node [1]. The value of the left node of the root node [1,-1] is formed by concatenating [1,-1] with -1*[1,-1], and its value is [1,-1,-1,1]. The value of the right node is formed by concatenating [1,-1] with 1*[1,-1], and its value is [1,-1,1,-1]. The same pattern is followed when generating the values ​​of the left and right child nodes of the subtree until the length of the leaf node is N. The left and right leaf nodes rooted at [1] are sorted into row vectors from left to right.

[0073] Step (2.3): Generate an N*N speckle matrix, where each element is a speckle and the speckle size is N*N. Multiply each element of the column vector sorted by each element of the row vector sorted, and reset the speckle values ​​from -1 to 0, finally forming an N*N speckle matrix.

[0074] Step (2.4): Perform a zigzag scan sort on the illumination speckle matrix to form a fixed speckle sequence;

[0075] Step (3): Image data acquisition is performed on 70,000 data points from the MNIST dataset, including the following steps:

[0076] Step (3.1): Calculate the light intensity collected by the barrel detector in each measurement of the correlated imaging. The formula for calculating the light intensity is as follows:

[0077] It=∫dxt(x) 2 *Bt(x)

[0078] Its t(x) is the spatial distribution information of the object to be imaged, Bt(x) represents the light intensity distribution of the light field at the object, It is the total light intensity value collected, and the subscript t indicates the t-th sampling.

[0079] Step (3.2): Set the sampling rate β. If the image size is N*N, then the total number of samplings T is: N*N, where M is the number of samplings required, and the sampling rate β = M / T.

[0080] Step (3.3): Determine the number of samplings M based on the sampling rate β in step (3.2), and take M light speckles from the beginning of the light speckle sequence generated in step 2.

[0081] Step (3.4): Load the M light speckles from step (2) onto the digital micromirror device in sequence to control the light field. Use a barrel detector to collect the light intensity of the illuminated image. Record the total light intensity corresponding to each light speckle as It. Illuminate the image with M measurements. Collect the light intensity sequence I = [I1, I2, I3...., IM]. Multiply the light intensity sequence with the corresponding M light speckle sequences to form M-channel data. Store the data in JSON format. The file name is consistent with the image number. The image and the M-channel data form an image dataset.

[0082] Step (4): The deep convolutional hybrid neural network is the construction of DCMGI, and its structural diagram is as follows. Figure 1 As shown, the neural network model includes a special object extraction module, a feature enhancement module, and an output module.

[0083] The feature extraction module consists of 5 convolutional modules, such as Figure 2 As shown, each convolutional module in the structure is c_block_n(in_channel, out_channel), where n is the nth convolutional module of the feature extraction module. The convolutional modules are as follows: Figure 3 As shown, `in_channel` and `out_channel` represent the number of input and output channels of the convolutional module. The convolutional module contains two convolutional layers: Conv1 (3x3 kernel), Conv2 ...

[0084] c_block_1(M,50)-->c_block_2(50,40)-->c_block_3(40,30)-->c_b lock_4(30,20)-->c_block_5(20,10)

[0085] In its convolution module c_block_1, M represents the number of input channels, and --> represents the data flow direction.

[0086] The feature enhancement module includes channel attention and spatial attention mechanisms, as shown in Figure 4. The feature enhancement module uses the output channels of the feature extraction module as input channels, and performs channel attention and spatial attention processing respectively. The structure of the channel attention mechanism is as follows: Figure 5 As shown, this structure can be divided into three parts. The first part performs global average pooling and global max pooling on the input data. The second part uses a shared multilayer perceptron (e.g., ...) to process the results of average pooling and max pooling. Figure 6 The process (as shown) involves adding the two processed results, then passing them through an activation function to obtain the weights of each channel of the input data, which are between 0 and 1. The third step multiplies the weights of each channel obtained in the second step by the original channel input data to obtain the output data. The spatial attention mechanism structure is as follows: Figure 7 As shown, the structure consists of three parts. The first part takes the maximum and average values ​​of each feature point in the input data across channels. These maximum and average values ​​are then stacked to form a 2-channel dataset. The second part processes the 2-channel output from the first part through a convolutional layer with a 3x3 kernel, 1 unit of padding, and a stride of 1. After activation, the weights of each feature point are set between 0 and 1. The third part multiplies the spatial weights with the input data. This module outputs 10 channels, and the output data is the same size as the input data.

[0087] The output module consists of two convolutional layers, such as Figure 8 As shown, its convolutional layer Conlayer1 has 10 input channels, 5 output channels, a kernel size of 3*3, padding of 1, and stride of 1; convolutional layer Conlayer2 has 5 input channels, 1 output channel, a kernel size of 3*3, padding of 1, and stride of 1. The biggest difference between convolutional layers Conlayer1 and Conlayer2 is that Conlayer1 adds a normalization layer, namely BatchNorm2d, and an activation function after it.

[0088] Step (5): Training a deep convolutional hybrid neural network, i.e., the DCMGI model, includes the following steps:

[0089] Step (5.1): Divide the image dataset from step (3) into a training set and a test set, with the training set and test set divided in a 6:1 ratio.

[0090] Step (5.2): Determining the optimizer and loss function for training the deep convolutional hybrid neural network. The mean squared loss function is adopted, and its formula is as follows:

[0091] Loss(x i y i )=(x i y i ) 2

[0092] Its Loss is the name of the loss function, x i The original image, y i To reconstruct the image;

[0093] The training optimizer uses the Adam optimizer, which combines the advantages of both AdaGrad and RMSProp optimization algorithms. The learning rate parameter of the optimizer is set to 0.0001.

[0094] Step (5.3): Set the number of epochs for model training parameters to 200.

[0095] Step (5.4): The quality assessment of the reconstructed image uses Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM). SSIM refers to the degree of similarity between the generated image and the real image, and is calculated as follows:

[0096]

[0097] In the above formula, C1 = (k1l)², C2 = (k2l)², l refers to the dynamic range of pixel values. Typically, k1 is 0.01, k2 is 0.03, μ is the average value, and σ is the variance. For covariance.

[0098] Peak signal-to-noise ratio (PSNR) is the ratio between the maximum signal strength and the noise intensity. A higher PSNR value indicates less noise interference to the image. The mathematical expression is as follows:

[0099]

[0100] Where MAXi is the maximum pixel value of the original image, and MSE is the mean square error between the original image and the reconstructed image.

[0101] Step (6): The deep convolutional hybrid neural network is the model saved as DCMGI. The deep convolutional hybrid neural network model is saved as DCMGI.pth. The saved model is used to load the model of the actual scene for image reconstruction.

[0102] Example 1

[0103] This embodiment provides an algorithm for calculating correlated imaging reconstruction based on a deep convolutional hybrid neural network, including the following steps:

[0104] Step (1): Expand the images in the training set and test machine of the MNIST dataset into grayscale images of size 64*64;

[0105] Step (2): Generate illumination speckle based on the Hadamard matrix, and sort the illumination speckle by grayscale encoding to form a fixed sequence of illumination speckle sequences with a sequence length of 64*64.

[0106] Step (3): Set the sampling rate to 6.25%, extract 256 light speckles from the light speckle sequence from front to back, load them onto the DMD in a fixed order, collect the light intensity sequence, and form the training dataset and the test dataset respectively;

[0107] Step (4): The deep convolutional hybrid neural network, namely DCMGI, is trained using Adam as the training optimizer, with a learning rate of 0.0001 and mean squared error as the loss function. The model is trained for 200 rounds.

[0108] At a sampling rate of 6.25%, the traditional computational correlation imaging algorithm, the pseudo-inverse computational correlation imaging algorithm, and the computational correlation imaging algorithm based on deep convolutional hybrid neural networks reconstructed images from handwritten digits 0 to 9. The imaging effect of the computational correlation imaging algorithm based on deep convolutional hybrid neural networks was significantly better than the other two algorithms. The imaging results are as follows: Figure 9 As shown, the first row of images is the original image of the handwritten digits, the second row is the result of traditional computational correlation imaging, the third row is the imaging effect of the pseudo-inverse computational correlation imaging algorithm, and the fourth row is the imaging result of computational correlation imaging based on deep convolutional hybrid neural networks.

[0109] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. An algorithm for computational correlation imaging reconstruction based on deep convolutional hybrid neural networks, characterized in that, Includes the following steps: Step 1: Data Collection; Prepare a clean MNIST dataset and augment it. Scattered light spots are generated based on the Hadamard matrix, and the scattered light spots are sorted by grayscale encoding to form a fixed sequence of light speckle sequences; Data sampling is performed based on the illumination speckle sequence; the light intensity sequence is multiplied by the measured illumination speckle sequence and then combined with the original image to form an image dataset. Step 2: Establish a deep convolutional hybrid neural network model; Step 3: Model training; The image dataset described in step 1 is divided into a training set and a test set according to the proportion. The training set is input into the DCMGI model, and the DCMGI model is trained based on the training optimizer and loss function. Step 4: Use the trained DCMGI model to calculate associated imaging and complete the image reconstruction; The deep convolutional hybrid neural network model is a DCMGI model, including: Feature extraction module: Contains n convolutional modules, wherein the convolutional modules are: c_block_n(in_channel,out_channel) Where n is the nth module of c_block, and in_channel and out_channel are the number of input channels and output channels of the convolution module; Feature enhancement module: Includes channel attention mechanism and spatial attention mechanism. The output channel of the feature extraction module is used as the input channel of the feature enhancement module, and the channel attention mechanism and spatial attention mechanism are processed respectively. Output module: includes convolutional layer Conlayer1 and convolutional layer Conlayer2; The convolutional layer Conlayer1 has 10 input channels, 5 output channels, a kernel size of 3*3, padding of 1, and stride of 1. The convolutional layer Conlayer2 has 5 input channels, 1 output channel, a kernel size of 3*3, padding of 1, and stride of 1. Add a normalization layer and activation function to the convolutional layer Conlayer1.

2. The algorithm for computational correlation imaging reconstruction based on deep convolutional hybrid neural networks according to claim 1, characterized in that, The MNIST dataset from step 1 is augmented and expanded to an image of size N*N; The formation of an N*N speckle sequence includes the following steps: Step 1.11: Hadamard matrix column vector sorting generation: The right node value of the root node [1] is [1,1]. The right node value of the root node [1,1] is formed by concatenating [1,1] with 1*[1,1], and its value is [1,1,1,1]. The left node value is formed by concatenating [1,1] with -1*[1,1], and its value is [1,1,-1,-1]. This continues until the length of the leaf node value is N. [1,-1] is the value of the left node rooted at [1]. The value of the left node rooted at [1,-1] is formed by concatenating [1,-1] with 1*[1,-1], and its value is [1,-1,1,-1]. The value of the right node is formed by concatenating [1,-1] with -1*[1,-1], and its value is [1,-1,-1,1]. This continues until the length of the leaf node is N. The left and right leaf nodes rooted at [1] are sorted into column vectors from right to left. Step 1.12: Hadamard matrix row vector sorting generation: The left node value of the root node [1] is [1,1]. The left node value of the root node [1,1] is formed by concatenating [1,1] with 1*[1,1], and its value is [1,1,1,1]. The right node value is formed by concatenating [1,1] with -1*[1,1], and its value is [1,1,-1,-1]. This continues until the length of the leaf node value is N. [1,-1] is the value of the left and right nodes with [1] as the root node. The value of the left node with [1,-1] as the root node is formed by concatenating [1,-1] with -1*[1,-1], and its value is [1,-1,-1,1]. The value of the right node is formed by concatenating [1,-1] with 1*[1,-1], and its value is [1,-1,1,-1]. This continues until the length of the leaf node is N. The left and right leaf nodes with [1] as the root are sorted into row vectors from left to right. Step 1.13: Generation of an N*N speckle matrix: Each element is a speckle, and the speckle size is N*N. Multiply each element of the column vector sorted by each element of the row vector sorted, and set the speckle value of -1 to 0. Finally, an N*N speckle matrix is ​​formed. Step 1.14: Perform a zigzag scan sort on the illumination speckle matrix to form a fixed illumination speckle sequence.

3. The algorithm for computational correlation imaging reconstruction based on deep convolutional hybrid neural networks according to claim 2, characterized in that, The method for collecting data described in step 1 includes the following steps: Step 1.21: Calculate the total light intensity collected by the barrel detector in each measurement of the correlated imaging; The calculation formula is as follows: ； Where t(x) represents the spatial distribution information of the object to be imaged. This represents the intensity distribution of the light field at the object. This is the total light intensity value collected, and the subscript t indicates the t-th sampling. Step 1.22: Setting the sampling rate β: If the image size is N*N, then the total number of samplings T is: N*N, where M is the number of samplings required, and the sampling rate β = M / T; Step 1.23: Determine the number of samplings M based on the sampling rate β in Step 1.22, and take M speckle patterns from the beginning of the speckle sequence generated in Step 1.14; Step 1.24: M scattered light spots are sequentially loaded onto the digital micromirror device to control the light field. A barrel detector is used to collect the light intensity of the illuminated image. The total light intensity corresponding to each scattered light spot is denoted as... The image was illuminated M times, and the collected light intensity sequence is as follows: I=[I1,I2,I3....,I M ] The light intensity sequence and the corresponding M scattered light spots are stored, and the image and the corresponding light intensity sequence form an image dataset.

4. The algorithm for computational correlation imaging reconstruction based on deep convolutional hybrid neural networks according to claim 2, characterized in that, The convolutional module includes a convolutional layer Conv1 and a convolutional layer Conv2; The Conv1 convolutional layer has a kernel size of 3*3, padding is set to 1, stride is set to 1, and both the input and output channels are in_channel. The Conv2 convolutional layer has a kernel size of 3*3, padding of 1, stride of 1, input channel of in_channel, and output channel of out_channel.

5. The algorithm for computational correlation imaging reconstruction based on deep convolutional hybrid neural networks according to claim 2, characterized in that, The channel attention mechanism includes: Global average pooling and global max pooling are performed on the input data of the input channel; The results of average pooling and max pooling are processed using a shared multilayer perceptron. The two processed results are added together and then processed by an activation function to obtain the weight of each channel of the input data. The weight of each channel is multiplied by the input data of the input channel to obtain the output data; The spatial attention mechanism includes: For each feature point of the input data in the input channel, the maximum value and average value are taken in the input channel, and the maximum value and average value are stacked to form 2-channel data; The 2-channel data is passed through a convolutional layer with 1 output channel to adjust the number of channels, and then processed by an activation function to obtain the weights of each feature point. Multiply the weight data by the input data; The convolutional layer has a kernel size of 3*3, padding of 1, and stride of 1.

6. The algorithm for computational correlation imaging reconstruction based on deep convolutional hybrid neural networks according to claim 1, characterized in that, In step 3, the loss function adopted is the mean square loss function, and the formula for the loss function is: ； in, The loss function name. The original image. To reconstruct the image; The training optimizer uses the Adam optimizer, and the learning rate parameter of the Adam optimizer is set to 0.0001.

Images