A highly sensitive anti-scattering imaging method based on coding and network

Abstract

The present invention discloses a high-sensitivity anti-scattering imaging method based on coding and network, comprising the following steps: 1) establishing a speckle acquisition optical system, and obtaining a target speckle image carrying coding information by using the speckle acquisition optical system; 2) constructing an imaging target reconstruction model based on a convolutional neural network; 3) inputting corresponding target data sets and speckle signal data sets into the imaging target reconstruction model for training to obtain a trained imaging target reconstruction model; 4) inputting the acquired target speckle image into the trained imaging target reconstruction model to obtain the restored original target. The present invention proposes a high-sensitivity anti-scattering imaging target reconstruction model based on a convolutional neural network, which can realize the restoration of the original target hidden behind a scattering medium. By introducing coding into the front-end optical system, the structure is simple. After the front-end optical system collects the output signal of the optical system, it is directly used for the reconstruction of the back-end algorithm. This method can significantly improve the sensitivity of the optical system.

Description

Technical Field

[0001] The present invention belongs to the field of machine learning and image reconstruction, and particularly relates to a high-sensitivity anti-scattering imaging method based on coding and network. Background Art

[0002] The presence of scattering media seriously affects the imaging quality. Although traditional anti-scattering imaging techniques such as speckle correlation imaging techniques have achieved certain effects, due to their limitations such as being inapplicable to low-light environments and relying on high-quality imaging devices, it has posed certain obstacles to the practical application of this technology. In a low-light environment, the intensity of the scattered light signal is limited, and it is difficult for high-quality imaging devices to collect speckle images with high signal-to-noise ratio and high contrast. Existing reconstruction algorithms such as phase retrieval algorithms have limited noise robustness and it is difficult to reconstruct the original target from such low-quality speckle signals. Currently, the methods that can achieve high-sensitivity anti-scattering imaging generally face problems such as the need to improve the algorithm reconstruction ability and noise robustness, the time-consuming optimization solution, and the need to improve the reconstruction speed, and none of them can achieve high-sensitivity target imaging in a low-light environment.

[0003] Therefore, a new high-sensitivity anti-scattering imaging method is needed to solve the above problems. Summary of the Invention

[0004] The purpose of the present invention is to provide a high-sensitivity anti-scattering imaging method based on coding and network to solve the defect that in a low-light environment, the intensity of the scattered light signal is limited and it is difficult for imaging devices to collect speckle images with high signal-to-noise ratio and high contrast.

[0005] The technical solution for the purpose of the present invention is as follows:

[0006] A high-sensitivity anti-scattering imaging method based on coding and network, comprising the following steps:

[0007] 1), Establish a speckle acquisition optical system, the speckle acquisition optical system is used to collect the target speckle image carrying coding information hidden behind the scattering medium, and the target speckle image carrying coding information is obtained by using the speckle acquisition optical system;

[0008] 2), Construct an imaging target reconstruction model based on a convolutional neural network;

[0009] 3), Input the corresponding target data set and speckle signal data set into the imaging target reconstruction model in step 2) for training to obtain a trained imaging target reconstruction model;

[0010] 4), Input the target speckle image carrying coding information collected in step 1) into the trained imaging target reconstruction model obtained in step 3) to obtain the restored original target.

[0011] Furthermore, the speckle collection optical system includes a first digital micromirror array, a scattering medium and a second digital micromirror array, wherein the first digital micromirror array and the second digital micromirror array are respectively located on both sides of the scattering medium. Two digital micromirror arrays are used to present the target and the code respectively, and after being modulated by the scattering medium, a speckle signal carrying the code information is finally formed on the detector.

[0012] Furthermore, the speckle collection optical system also includes a light source and a single-pixel detector, and the first digital micromirror array, the scattering medium, and the second digital micromirror array are sequentially arranged on an optical path between the light source and the single-pixel detector. The present invention uses a single-pixel detector instead of an area array camera, and after collecting the output signal of the optical system, it is directly used for back-end algorithm reconstruction.

[0013] Furthermore, the imaging target reconstruction model in step 2) includes an encoding module and a decoding module, wherein the encoding module uses a convolutional layer to extract features from the input information with the target light intensity amplitude to obtain high-dimensional, low-sensory recognition feature information; and the decoding module projects the feature information obtained by the encoding module into a low-dimensional, high-sensory recognition pixel space. Through this dense clustering expression of global information, the conversion from local semantic information to global pixel grayscale information is realized, and the reconstructed original target image is output.

[0014] Furthermore, the imaging target reconstruction model in step 2) is expressed by the following formula:

[0015] I=O*S

[0016] Where I is the target speckle image, S is the system point spread function, and O is the target image;

[0017]

[0018]

[0019] Among them, I i (x, y) is the i-th target speckle image, F i (x, y) is the i-th two-dimensional speckle image, T i (x, y) is the random Gaussian coding matrix, and E(i) is the final output signal of the speckle collection optical system.

[0020] The present invention encodes the speckle signal multiple times, uses a detector to collect the total intensity of the speckle signal after each encoding, and finally forms a one-dimensional output signal. This method can significantly improve the sensitivity of the optical system.

[0021] Furthermore, the imaging target reconstruction model described in step 2) adopts a skip connection strategy to fuse features of different dimensions, enhance the flow of gradients, avoid the problem of gradient disappearance during the encoding stage, and achieve full fusion and efficient extraction of feature information.

[0022] Furthermore, the imaging target reconstruction model described in step 2) adopts a dropout strategy to avoid overfitting during the training process by using the dropout strategy.

[0023] Furthermore, in step 3), the imaging target reconstruction model is trained using the mean square error function as the loss function.

[0024] Furthermore, in step 4), a two-step reconstruction method or an end-to-end reconstruction method is used for reconstruction. The two-step reconstruction method first restores the speckle image before encoding based on the compressed sensing theory, and then uses the speckle image reconstructed in the first step as the input of the constructed imaging target reconstruction model to obtain the original structure distribution of the target to be measured. The end-to-end reconstruction method regards each process as a whole, does not require secondary reconstruction of the collected data, directly uses the original information as the input of the network, and uses a network to achieve information extraction and target reconstruction, that is, directly uses the output signal of the optical system as the input of the network to obtain the original structure distribution of the target to be measured. The reconstruction speed of the present invention is relatively fast. The average reconstruction speed of the end-to-end reconstruction algorithm is 95 frames per second, providing an effective technical approach to solve the problem of anti-scattering imaging of weak signals.

[0025] Furthermore, the two-step reconstruction method includes the following steps:

[0026] 1. Based on the compressed sensing theory, first restore the target speckle image before encoding;

[0027] 2. Use the target speckle image obtained in step 1 as the input of the imaging target reconstruction model to obtain the restored original target

[0028] Beneficial effects:

[0029] The present invention proposes a high-sensitivity anti-scattering imaging target reconstruction model based on a convolutional neural network, which can realize the restoration of the original target hidden behind the scattering medium. Encoding is introduced into the front-end optical system, and the structure is simple. After the front-end optical system collects the output signal of the optical system, it is directly used for the reconstruction of the back-end algorithm. This method can significantly improve the sensitivity of the optical system. Description of the Drawings

[0030] Figure 1 is a schematic diagram of a high-sensitivity anti-scattering imaging optical system based on an area array detector used in the present invention.

[0031] Figure 2It is the process diagram of data acquisition and target reconstruction of the present invention.

[0032] Figure 3 It is the structural block diagram of the HSSRNet network of the present invention.

[0033] Figure 4 It is the effect diagram of the HSSRNet of the present invention reconstructing data under different training set data volumes.

[0034] Figure 5 It is the reconstruction effect diagram when the width of the HSSRNet network of the present invention is different.

[0035] Figure 6 It is the effect diagram of the HSSRNet of the present invention reconstructing data under different data dimensions.

[0036] Figure 7 It is the effect diagram of the HSSRNet of the present invention reconstructing data under different coding matrix sizes.

[0037] Figure 8 It is the effect diagram of the HSSRNet of the present invention reconstructing different noise data.

[0038] Figure 9 It is the schematic diagram of the high-sensitivity anti-scattering imaging optical system based on a single-pixel detector used in the present invention.

[0039] Figure 10 It is the effect diagram of the HSSRNet of the present invention reconstructing a digital target based on a single-pixel detector. Detailed implementation manners

[0040] A high-sensitivity anti-scattering imaging method based on coding and a network includes the following steps:

[0041] Step 1: Establish an optical system for speckle acquisition. The optical system acquires speckle images hidden behind a scattering medium to obtain a speckle data set of several targets;

[0042] As Figure 1 shown, the optical system for speckle acquisition used in the present invention has a simple structure. Two digital micromirror arrays are used to present the target and the code respectively. After being modulated by the scattering medium, a speckle signal carrying the coding information is finally formed on the detector, which is the input signal of the HSSRNet, Figure 2 which is the process diagram of data acquisition and target reconstruction.

[0043] Step 2: Construct a high-sensitivity anti-scattering imaging target reconstruction model based on a convolutional neural network;

[0044] Construct as Figure 3The HSSRNet shown below. In the encoding stage, a convolutional layer is used to extract features from the information with the target light intensity amplitude input to the network. In the decoding stage, the high-dimensional and low-sensory-identifiability feature information obtained in the encoding stage is projected into the low-dimensional and high-sensory-identifiability pixel space. Through this dense clustering expression of global information, the conversion from local semantic information to global pixel grayscale information is achieved, and the reconstructed original target image is output. A skip connection strategy is used in the network to fuse features of different dimensions and enhance the flow of gradients to avoid the problem of gradient disappearance in the encoding stage, realizing the full fusion and efficient extraction of feature information. A random dropout strategy is used in the network to avoid overfitting during training. Mean squared error (MSE) is used as the loss function during training.

[0045] Step 3: Input the corresponding target dataset and the speckle dataset carrying encoding information into the high-sensitivity anti-scattering imaging target reconstruction model constructed in Step 2 for training;

[0046] It is known that within the optical memory range, the speckle image I can be described as the convolution of the object O hidden behind the scattering medium and the system point spread function S, which can be expressed as

[0047] I = O * S (1)

[0048] where I is the speckle image, S is the system point spread function, and O is the target image. The speckle signal I i (x, y) is modulated by a series of random Gaussian encoding matrices T i (x, y) to obtain a series of two-dimensional speckle images F i (x, y). Through a focusing lens, it is converged onto the detector, and the detector can collect the final optical system output signal E(i). This process can be expressed as

[0049]

[0050]

[0051] where I i (x, y) is the speckle signal, F i (x, y) is the two-dimensional speckle image, T i (x, y) is the random Gaussian encoding matrix, and E(i) is the final output signal of the optical system.

[0052] Step 4: Input the collected target speckle image into the two trained high-sensitivity anti-scattering imaging target reconstruction models to obtain the restored original target.

[0053] Use the optimal model saved by HSSRNet in Step 3 to reconstruct 1874 groups, 1406 groups, and 938 groups of training set data. The results are asFigure 4 As shown, when the amount of training set data is only 938 groups, the target reconstruction can still be well completed.

[0054] The effects of the present invention can be further illustrated by the following results:

[0055] First, verify the reconstruction capabilities of the two reconstruction networks under different amounts of training set data. Reconstruction is performed when the amounts of training set data are 1874 groups, 1406 groups, and 938 groups respectively, to test the reconstruction capabilities of the network when the amount of data decreases. Generally, when training a network, the larger the amount of data provided for the network to learn, the richer the information that can be extracted, and the more beneficial it is to the recovery of the original target. The two reconstruction algorithms proposed in this paper can achieve the reconstruction of the original target when the amount of training set data is only 938 groups. To test the reconstruction capabilities of the networks with different widths, the width coefficient of the original HSSRNet network is set to 1.0. On this basis, the number of feature channels in each layer is adjusted to form networks with width coefficients of 2.0 and 0.5 respectively, to explore the influence of different network widths on the reconstruction algorithm. The network width is the number of feature channels. Sufficient network width can enable each layer in the feature extraction process to obtain effective features. Too narrow a width will result in insufficient feature extraction, and the performance of the network model will also be limited. However, blindly increasing the network width will also cause problems such as repeated feature extraction and increased computational burden. After the width reaches a certain magnitude, the performance of the algorithm may show a downward trend. The reconstruction effects of the test set are as Figure 5 shown. It can be seen that appropriately increasing the network width has a certain effect on improving the reconstruction capabilities of the algorithm, but as the network becomes wider, the reconstruction effect does not necessarily become better.

[0056] To explore the influence of different input data dimensions on the reconstruction performance of the algorithm, compare the reconstruction effects of the algorithm when the dimension parameters are 32×32, 28×28, 24×24, and 20×20 respectively. The reconstruction effects of the test set are as Figure 6 shown. It can be concluded that increasing the input data dimension, that is, increasing the sampling number of the speckle optical signal formed by the target, is beneficial to target recovery. Increasing the coding matrix, at this time, more original speckles are encoded, and the representation of the speckle field is closer to the system average. Taking the size of different coding matrices as a variable, explore the reconstruction effects of the original target when the sizes of the coding matrices are different. Experimental data with coding matrix sizes of 32×32, 64×64, and 80×80 are collected respectively. The reconstruction effects of the test set are as Figure 7As shown in the figure, increasing the size of the encoding matrix is beneficial to the reconstruction result. To explore the noise resistance ability of the reconstruction algorithm, in this subsection, Gaussian noise with a normal distribution is added during the data processing to interfere with the image. During the experiment, Gaussian noise with standard deviations of 0.01, 0.03, 0.05, and 0.07 is added to the two-step and end-to-end data respectively to simulate the noise interference in the actual experiment process, so as to explore the reconstruction results of the two reconstruction algorithms when there is noise in the experimental environment. The reconstruction effect of the test set is as Figure 8 shown. It can be seen that compared with the noise-free data, as the standard deviation of the noise increases, the reconstruction quality drops significantly.

[0057] From the above several comparative experiments, it can be seen that the reconstruction ability of the end-to-end reconstruction method is stronger than that of the two-step reconstruction method. The reason for this difference in ability is the reconstruction process of the first compressive sensing algorithm in the two-step method. Since the compressive sensing reconstruction process is not a completely lossless process, it is not excluded that effective information for speckle reconstruction will be lost during this process. This makes the data input to the subsequent network have some information loss compared with the original data, while the data input to the end-to-end network contains all the information in the original input data, which can enable the network to maximize the mining of the original information and obtain the global optimal solution. Through experiments, it is proved that both methods can complete the task of reconstructing hidden targets from weak scattering signals, but the reconstruction ability of the end-to-end method is better than that of the two-step reconstruction method.

[0058] The structural diagram of the high-sensitivity anti-scattering imaging optical system based on a single-pixel detector is as Figure 9 shown. The light source, scattering medium, and DMD device used are Figure 1 the same. The difference is that the area array detector is replaced by a single-pixel detector, and a data acquisition card is used to output the acquired signal intensity, which is the final output signal of the optical system. During the data acquisition process, the sampling rate of the data acquisition card is set to 1000, and the number of samples is set to 100. For the acquired data, the relatively better-performing HSSRNet-E network is used for reconstruction. The signal with a size of 1024×1 is directly used as the input of the network, and together with its corresponding target true distribution, a training set containing 1800 groups of data is formed. The remaining 50 groups of data that have not been trained are used to evaluate the reconstruction ability of the trained model for hidden targets. Figure 10The figure shows the results of the reconstruction test set. The average MSE of the reconstruction results of the test set obtained by the end-to-end reconstruction network HSSRNet-E is 36.2986, the average SSIM is 0.4111, and the average PSNR is 16.9337 dB. This indicates that when the data volume is 1800 groups, the end-to-end reconstruction network HSSRNet-E proposed in this paper can better perform target reconstruction on the data collected by the single-pixel detector. The reconstructed result figure and the target real image are highly similar in morphology and structure, and good reconstruction results can be obtained for each digital target. This shows that the end-to-end reconstruction algorithm has strong reconstruction ability in low-light environments, can achieve high-quality reconstruction of targets in low-light environments, and can be well applied to the high-sensitivity anti-scattering imaging task based on the single-pixel detector to reconstruct the targets hidden behind the scattering medium.

Claims

1. A highly sensitive anti-scatter imaging method based on coding and network, It is characterized in that The following steps are involved: 1) Establishing a speckle collection optical system, wherein the speckle collection optical system is used to collect a target speckle image carrying coded information hidden behind a scattering medium, and obtaining a target speckle image carrying coded information by using the speckle collection optical system; 2) Construct an imaging target reconstruction model based on convolutional neural network; In step 2), the imaging target reconstruction model includes an encoding module and a decoding module. The encoding module uses a convolutional layer to extract features from the input information with the target light intensity amplitude to obtain high-dimensional, low-sensory recognition feature information; the decoding module projects the feature information obtained by the encoding module into a low-dimensional, high-sensory recognition pixel space: The imaging target reconstruction model in step 2) is expressed by the following formula: ； Among them, is the target speckle image, is the system point spread function, is the target image; ； ； Among them, is the i-th target speckle image, is the i-th two-dimensional speckle image, is a random Gaussian coding matrix, is the final output signal of the speckle acquisition optical system; 3) Inputting the corresponding target data set and the speckle signal data set into the imaging target reconstruction model of step 2) for training to obtain a trained imaging target reconstruction model; 4) Inputting the target speckle image carrying the coded information collected in step 1) into the trained imaging target reconstruction model obtained in step 3) to obtain the restored original target.

2. A highly sensitive anti-scatter imaging method based on coding and network according to claim 1, It is characterized in that The speckle collection optical system comprises a first digital micromirror array, a scattering medium and a second digital micromirror array, wherein the first digital micromirror array and the second digital micromirror array are respectively located on two sides of the scattering medium.

3. A highly sensitive anti-scatter imaging method based on coding and network according to claim 2, It is characterized in that The speckle collection optical system also includes a light source and a single-pixel detector, and the first digital micromirror array, the scattering medium, and the second digital micromirror array are sequentially arranged on an optical path between the light source and the single-pixel detector.

4. The high-sensitivity anti-scatter imaging method based on coding and network according to claim 1, It is characterized in that The imaging target reconstruction model described in step 2) adopts a skip-layer connection strategy.

5. The high-sensitivity anti-scatter imaging method based on coding and network according to claim 1, It is characterized in that The imaging target reconstruction model described in step 2) adopts a random discarding strategy.

6. The highly sensitive anti-scatter imaging method based on coding and network according to claim 1, It is characterized in that In step 3), the imaging target reconstruction model is trained, and a mean square error function is used as a loss function.

7. The highly sensitive anti-scatter imaging method based on coding and network according to claim 1, It is characterized in that In step 4), reconstruction is performed using a two-step reconstruction method or an end-to-end reconstruction method.

8. The high-sensitivity anti-scatter imaging method based on coding and network according to claim 7, It is characterized in that The two-step reconstruction method comprises the following steps: First, based on the theory of compressed sensing, the target speckle image before encoding is restored; 2. Using the target speckle image obtained in step 1 as input of the imaging target reconstruction model to obtain the restored original target.

Images