A self-supervised dual-domain dual-path single-pixel imaging method

Abstract

The application provides a self-supervised dual-domain dual-path single-pixel imaging method, and the specific steps are as follows: S1, acquiring a single-pixel measurement value; S2, constructing a neural network SDD-Net, inputting the single-pixel measurement value I into the neural network to obtain a preliminary reconstructed image T1; the preliminary reconstructed image T1 is subjected to a single-pixel forward imaging process to generate corresponding measurement value PT1; the measurement value PT1 is input into the same neural network again to obtain a final reconstructed image T2; S3, loss function design, designing a dual-domain loss function and adopting a total variation method to guide the neural network to separate and reconstruct structure and texture information; S4, generation of a reconstructed image, a structure decoder outputs T s and a texture decoder outputs T t The final reconstructed image is obtained by adding the structure and the texture at the pixel level; and the method improves the imaging image quality and the imaging rate.

Description

Technical Field

[0001] This invention belongs to the field of single-pixel imaging, and particularly relates to a self-supervised dual-domain dual-path single-pixel imaging method. Background Technology

[0002] Single-pixel imaging (SPI) is a novel computational imaging technique. SPI utilizes spatial light modulators such as digital micromirrors (DMDs) to modulate a light source, projecting different imaging patterns onto a target object. A single-pixel detector collects the total light intensity reflected or refracted by the target object. Finally, the target image is reconstructed by calculating the correlation between the imaging pattern and the obtained light intensity signal. Unlike traditional cameras that use pixel arrays to collect image information, SPI only requires a single photodetector unit with no spatial resolution to collect total light intensity information and reconstruct the image. Furthermore, current SPI technology can operate under visible, ultraviolet, infrared, and terahertz lighting conditions. Due to its unique imaging principle, SPI can solve imaging problems that are unsuitable for traditional camera imaging, such as terahertz imaging, hyperspectral imaging, microscopic imaging, and underwater imaging. However, because the single-pixel detector detects a globally compressed measurement signal of the entire target object each time, it is difficult to reconstruct high-quality target images at low sampling rates.

[0003] Existing literature, such as Bian et al., proposed a self-supervised dual-domain constrained SPI imaging model. First, a preliminary image is reconstructed using the traditional pseudo-inverse matrix method. This preliminary image is then input into a neural network to obtain an output image. After undergoing a reversible transformation, the output image is used as a label for the SPI imaging process to obtain corresponding measurement values, which are then input back into the network. This method not only considers the constraints of the measurement domain but also adds image space constraints through reversible transformations, enabling the network to reconstruct images more accurately and with higher quality. However, Bian's method employs a two-step reconstruction process. The first step, using traditional algorithms for preliminary reconstruction, often loses a significant amount of feature information in the SPI measurement space, and the quality and speed of the reconstruction largely depend on the traditional algorithm used. Furthermore, because Bian adds spatial constraints to the transformed image, the more transformation operations are performed on the image, the more single-pixel imaging processes are required, resulting in increased time consumption. Summary of the Invention

[0004] The purpose of this invention is to solve the problem of feature information loss during the initial reconstruction process of traditional algorithms, as well as the quality and efficiency issues of single-pixel imaging.

[0005] To achieve the above objectives, the present invention provides a self-supervised dual-domain dual-path single-pixel imaging method, characterized by comprising the following steps:

[0006] S1. Obtain single pixel measurement value

[0007] The single-pixel measurement value I is obtained by calculating the target image T and the imaging mode P; the specific calculation formula is as follows:

[0008] I i =∫P i (x, y)T(x, y)dxdy

[0009] In the formula I i I represents the measurement value of the i-th single pixel; P represents the modulation pattern, i.e., the imaging mode; P i This represents the i-th modulation pattern; the subscripts i = 1, 2, ..., M represents the i-th measurement, T is the target image, and (x, y) are the pixel coordinates.

[0010] S2. Construction of the neural network SDD-Net

[0011] The single-pixel measurement value I is input into the neural network SDD-Net to obtain the preliminary reconstructed image T1; T1 then undergoes a single-pixel forward imaging process to generate the corresponding measurement value PT1; the measurement value PT1 is input into the same neural network SDD-Net again to obtain the final reconstructed image T2; where I and PT1 constitute the measurement domain loss, and T1 and T2 constitute the image domain loss constraint; the neural network SDD-Net consists of an encoder and two decoders, the decoders are composed of a structure decoder and a texture decoder, the decoders and encoder are connected by a special skip connection that includes channel spatial attention, and the two decoders are connected by an interactive attention module;

[0012] S3. Loss Function Design

[0013] The total variation method is used to guide the network in separating and reconstructing structural and texture information; the SDD-Net neural network is trained under constraints using the following loss function:

[0014]

[0015] T o =f θ =T s +T t =T gn(g＝1，2)

[0016] In the formula T o T represents the output image. s Represents a structural image, T t P represents the texture image, and I represents the pattern image. iLet n be the input measurement value of the network, and i be the index of the image domain and the measurement domain, respectively. g is the output sequence number of the network in which iteration. μ is the weight of the image domain constraint term in the loss function. σ is the weight of the third part of the regularization in the loss function, and λ is the weight of the fourth part of the regularization in the loss function. N represents the total number of pixels in the image. L1 is the loss function. T 1n This represents the position of the nth pixel in the first output of the network; TV represents the total variation loss function; f θ Represents the parameters of the neural network;

[0017] The total variation loss function is calculated as follows:

[0018]

[0019] This represents the pixel value at coordinates (x+1, y) in the structured image, which is the value of the pixel adjacent to the right. This represents the pixel value at coordinates (x, y) in the structured image; This represents the pixel value at coordinates (x, y+1) in the structured image, i.e., the value of the adjacent pixel below. In order to preserve the edge details of the image and obtain better convergence, the L1 loss function is used to constrain the two domains.

[0020] S4. Generation of reconstructed images

[0021] The output T of the structure path s and the output T of the texture path t The final reconstructed image is obtained by adding the pixels together.

[0022] Furthermore, the interactive attention module in S2 consists of a frequency domain branch and a spatial domain branch. The frequency domain uses Fourier convolutional layers to extract global frequency information from the input feature map; the spatial domain uses cascaded channel spatial attention to extract local spatial information. Finally, the frequency domain branch and the spatial domain branch perform feature fusion on the channels.

[0023] Furthermore, the loss function in S3 consists of four parts. The first part is the measurement domain constraint term, which consists of I and PT1, to ensure that the reconstructed image accurately reflects the original measurement value. The second part is the image domain constraint term, which consists of T1 and T2, to ensure the consistency and quality of the network output image. These two parts are used to generate a correct and unique formal reconstructed image. The third and fourth parts are regularization terms used to guide the two paths of the neural network to focus on their respective reconstruction tasks.

[0024] Beneficial effects:

[0025] This invention addresses the problems of non-unique imaging results and poor reconstructed image quality caused by constraining the neural network using only the measurement domain, by combining the measurement and image domains. The self-supervised deep learning strategy avoids the problems of excessive pre-training time and poor generalization ability of trained networks found in supervised learning methods. To address the issue of poor reconstructed image quality due to the large amount of feature information in the initial reconstructed measurement space, this application utilizes the inherent structure and texture of the image, designing the network for dual-task image reconstruction. Finally, the images are summed at the pixel level to obtain a high-quality reconstructed image, while also improving the imaging speed. Attached Figure Description

[0026] Figure 1 This is a flowchart of a single-pixel imaging process provided in an embodiment of the present invention;

[0027] Figure 2a This is a flowchart of the SDD-Net reconstruction method provided in an embodiment of the present invention;

[0028] Figure 2b This is the dual-path reconstruction network provided in the embodiments of the present invention;

[0029] Figure 2c This is the network module structure provided in the embodiments of the present invention;

[0030] Figure 3 This invention provides a structure-texture reconstruction image output by the network at a sampling rate of 10%.

[0031] Figure 4a These are the simulation results of the five algorithms provided by this invention on the Urban100 dataset;

[0032] Figure 4b This is a magnified view of the details of a portion of the reconstructed image at a 30% sampling rate provided by the present invention;

[0033] Figure 4c This invention provides a magnified view of the low-contrast region on the right side of the reconstructed image at a 30% sampling rate and the SSIM index.

[0034] Figure 4d This invention provides a quantitative evaluation of the PSNR and SSIM of five methods for reconstructing images.

[0035] Figure 5a These are magnified images and spectrograms of the single-path and dual-path image reconstruction results at a 10% sampling rate provided by this invention.

[0036] Figure 5b This invention provides a comparison of SSIM and PSNR results for single-path and dual-path reconstruction.

[0037] Figure 6 This invention provides a quantitative performance evaluation of the five methods provided under Gaussian noise levels with variances of 0.05, 0.25, and 0.5.

[0038] Figure 7a These are the reconstruction results of the five algorithms provided by this invention under five low sampling rates;

[0039] Figure 7b This invention provides a comparison of the reconstruction metrics of five algorithms at low sampling rates.

[0040] Figure 8 This refers to the performance of SDD-Net provided in this embodiment of the invention in actual experiments. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0042] The application principle of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0043] Example 1:

[0044] like Figure 1 As shown, a self-supervised dual-domain dual-path single-pixel imaging method is described, with the following specific steps:

[0045] S1. Obtain single pixel measurement value

[0046] The single-pixel measurement value I is obtained by calculating the target image T and the imaging mode P; the specific calculation formula is as follows:

[0047] I i =∫P i (x, y)T(x, y)dxdy (1)

[0048] In the formula I i I represents the measurement value of the i-th single pixel; P represents the modulation pattern, i.e., the imaging mode; P i This represents the i-th modulation pattern; the subscripts i = 1, 2, ..., M represents the i-th measurement, T is the target image, and (x, y) are the pixel coordinates.

[0049] S2. Construction of the neural network SDD-Net

[0050] The single-pixel measurement value I is input into the neural network SDD-Net to obtain the preliminary reconstructed image T1. T1 then undergoes a single-pixel forward imaging process to generate the corresponding measurement value PT1.

[0051] During self-supervised training, the constraint of only a single pixel measurement value I may lead to the same I value corresponding to multiple different target images. Therefore, the measurement value PT1 is input into the same neural network SDD-Net again to obtain the final reconstructed image T2. Here, I and PT1 constitute the measurement domain loss, and T1 and T2 constitute the image domain loss constraint.

[0052] Please refer to Figure 2a and Figure 2b The SDD-Net neural network comprises two branches: structure reconstruction and texture reconstruction, which operate independently. It consists of one encoder and two decoders. The decoder section comprises a structure decoder and a texture decoder. Each decoder is connected to the encoder using a special skip connection that incorporates channel-space attention. The two decoders are connected by an interactive attention module (IAM), which transfers structural information from the structure path image to the texture path, thereby guiding the generation of texture information. Figure 2c The specific structures of the encoder block, decoder block, channel spatial attention module, and interactive attention module are illustrated. The interactive attention module (IAM) consists of a frequency domain branch and a spatial domain branch. In the frequency domain, Fourier convolutional layers are used to extract global frequency information from the input feature map; the 1×1 convolutional layers aim to reduce the number of channels in the feature map to 1, thus reducing the complexity of the Fourier convolution. In the spatial domain, cascaded channel spatial attention is used to extract local spatial information. Finally, the outputs of the frequency domain branch and the spatial branch are used for channel-wise feature fusion.

[0053] S3. Loss Function Design

[0054] This paper employs the total variation method to guide the SDD-Net neural network in separating and reconstructing structural and texture information. The SDD-Net neural network is trained under constraints using the following loss function:

[0055]

[0056] T o =f θ =T s +T t =T gn’(g＝1，2) (4)

[0057] In the formula T o T represents the output image. s Represents a structural image, T t P represents the texture image, and I represents the pattern image. irepresents the input measurement value of the network, n and i are the indices of the image domain and measurement domain, respectively, and g is the output sequence number of the network in which iteration. μ is the weight of the image domain constraint term in the loss function; σ is the weight of the third part of the regularization in the loss function, and λ is the weight of the fourth part of the regularization in the loss function; N represents the total number of pixels in the image; L1 is the loss function; T 1n This represents the position of the nth pixel in the first output of the network. TV represents the total variation loss function. θ This represents the parameters of the neural network.

[0058] The entire loss function consists of four parts. The first part is the measurement domain constraint term, composed of I and PT1, ensuring that the reconstructed image accurately reflects the original measurement values. The second part is the image domain constraint term, composed of T1 and T2, ensuring the consistency and quality of the network output image. These two parts are used to generate a correct and unique formal reconstructed image. The third and fourth parts serve as regularization terms, guiding the neural network's dual paths to focus on their respective reconstruction tasks. The third part ensures that the output of the structural path is close to the true measurement value in the measurement domain, thus enabling the structural image to possess most of the subject structure of the real target. The fourth part uses the total variation loss function to smooth the detailed texture in the structural image, so that the structure image only reconstructs structural information, while the remaining texture information is supplemented by the texture path. Through the interactive attention module, the structural information in the structural path is fed to the texture path, guiding the generation of texture information, thus preserving both clear contour structure and delicate texture information in the reconstructed image. The total variation loss function is calculated as follows:

[0059]

[0060] This represents the pixel value at coordinates (x+1, y) in the structured image, which is the value of the pixel adjacent to the right. This represents the pixel value at coordinates (x, y) in the structured image. This represents the pixel value at coordinates (x, y+1) in the structured image, which is the value of the adjacent pixel below.

[0061] In order to preserve the edge details of the image and achieve better convergence, the L1 loss function is used to constrain the two domains.

[0062] S4. Generation of reconstructed images

[0063] The output T of the structure path s and the output T of the texture path t The final reconstructed image is obtained by adding the pixels together. The image structure and texture paradigm output by the network is as follows: Figure 3 As shown, T can be seen s The images in the text mainly consist of the outline structure of the images, while T...t The image's oscillating and repetitive texture information was then reconstructed.

[0064] Example 2:

[0065] The neural network training process consists of 1000 epochs with a batch size of 1 and an initial learning rate of 0.0005. Cosine annealing is used to reduce the learning rate. All experiments used the Urban100 and Set11 datasets. Images from Urban100 were preprocessed to 64×64 pixels as real images. The loss function parameters μ, σ, and λ were set to 5, 0.1, and 0.2, respectively. In the Set11 dataset, 128×128 images were used to verify the reconstruction effect of this method on larger resolution images. The loss function parameters μ, σ, and λ were set to 1, 0.1, and 0.2, respectively. The entire network structure was built on the PyTorch framework and deployed on an NVIDIA RTX 3090 GPU (24GB). To facilitate result comparison, structural similarity (SSIM) and peak signal-to-noise ratio (PSNR) were used for quantification. The specific techniques are as follows:

[0066]

[0067] Where μ and σ represent the mean and variance of the image, respectively, σ xv Let C1 and C2 be the covariance between images x and y, and C1 and C2 be constants. X and Y represent the width and height of the image, and B is the bit depth of a single-channel pixel value. SSIM can measure the similarity between two images based on metrics such as brightness and contrast. The closer the value is to 1, the closer the two images are, and the better the image reconstruction effect. PSNR represents the overall average error between the images; a larger value indicates better image quality. MSE represents the mean squared error loss, which calculates the squared mean of the pixel differences between the predicted image and the original image.

[0068] In comparative simulation experiments, the performance of traditional algorithm DGI, pseudo-inverse matrix method Pse-Inv, total variation compressed sensing CS-TV, and the deep learning method proposed by Bian are compared on the Urban100 dataset. The experimental results for Pse-Inv, CS-TV, and Bian are derived from Bian's work. The network proposed by Bian is named SD-Net. Figure 4aThe simulation experiment of reconstructing a 64×64 image demonstrates the reconstruction effects of five algorithms at sampling rates of 20%, 30%, and 40%. It can be seen that at a sampling rate of 20%, the reconstructed images from the three traditional algorithms all exhibit significant noise. SD-Net, due to its use of the MSE loss function to constrain the reconstruction results, produces an overly smooth image, resulting in insufficient reconstruction details. The neural network SDD-Net, due to its unique structure, recovers the striped structure of buildings in the image relatively well. Figure 4b and Figure 4c The images shown are magnified versions of local regions and the right-side shadow area in the reconstruction results of SD-Net and SDD-Net at a 30% sampling rate. It can be seen that due to the reduced contrast in the shadow area, SD-Net cannot capture the line texture under the shadow, thus losing some image detail. Visually and based on the SSIM metric, this method can significantly reconstruct image details under low contrast. Figure 4d This paper presents a quantitative evaluation of five reconstruction methods under three sampling rates. It can be seen that as the sampling rate increases, the PSNR and SSIM indices of the image also increase, indicating better image reconstruction results. Under all three sampling rates, the PSNR and SSIM of the neural network SDD-Net are higher than the other four methods. When the sampling rate is 20%, the PSNR and SSIM of the neural network SDD-Net are improved by 3.934 dB and 0.2505 respectively compared to the traditional algorithm CS-TV, and by 1.984 dB and 0.0705 respectively compared to SD-Net.

[0069] Example 3:

[0070] Table 1 presents the ablation study of the interactive attention module and dual-domain constraint strategy in the proposed SSD-Net network on 128×128 images in the Set11 dataset. Please refer to... Figure 5a and Figure 5b All ablation experiments were conducted at a 10% sampling rate. It can be seen that both module designs improve image reconstruction performance. Adding an image domain constraint to the network improved PSNR by 4.03 dB and SSIM by 0.0714 in quantitative image metrics. The network model with the interactive attention module increased PSNR and SSIM by 0.46 dB and 0.0045, respectively. The reconstruction performance was optimal when the network model incorporated both modules.

[0071] Table 1 shows the module ablation experiments on the Set11 dataset.

[0072]

[0073] To investigate the effectiveness of dual-path reconstruction in a network, a single-path version of the network, S-Net, was implemented, removing the texture path so that the entire image's structure and texture are restored using only one path. The original dual-path network was named DP-Net. From magnified local detail images and the overall image spectrum, DP-Net shows sharper and clearer texture edges and contains more high-frequency texture details than S-Net. Quantitatively, DP-Net outperforms S-Net in both SSIM and PSNR. This demonstrates the limitation of the network using only one path to simultaneously restore image structure and texture information; the structural and texture components interfere with each other during reconstruction, leading to unclear detail imaging. Using dual-path structure-texture reconstruction can effectively solve this problem.

[0074] Example 4:

[0075] To investigate the robustness of this method, Gaussian noise with variances of 0.05, 0.25, and 0.5 was added to the forward process of SPI. Figure 6 The reconstruction metrics of five methods are shown at a 64×64 image size and a 50% sampling rate. Because the added Gaussian noise is random, the noise addition process was repeated 10 times, and error bars for the proposed method were generated. It can be seen that the short error bars indicate strong noise resistance of the proposed method. As the noise variance increases, the performance of all five reconstruction algorithms decreases. Our method exhibits better performance than other methods at relatively high noise levels. Especially when the noise variance is 0.5, SD-Net and the other three traditional algorithms show significant performance degradation, while our method maintains good performance even at high noise levels.

[0076] Example 5:

[0077] Table 2 shows the reconstruction time and SSIM values ​​for five methods reconstructing a 64×64 image. The sampling rate is 0.5, and CS-TV has 300 iterations. SD-Net and SSD-Net converge at 200 and 550 iterations, respectively, achieving the best reconstruction results. DGI, Pse-Inv, and CS-TV have shorter reconstruction times but poorer reconstruction results. SD-Net, due to its two-step reconstruction method, suffers from significantly reduced reconstruction efficiency due to the initial reconstruction step of traditional algorithms, requiring 5.13 seconds per iteration. SSD-Net, using a one-step reconstruction, only requires 0.46 seconds per iteration. SDD-Net shows a significant improvement in imaging speed, reducing reconstruction time to one-fifth of SD-Net, and also achieves the best reconstruction results among the five algorithms.

[0078] Table 2 Comparison of Reconstruction Time and Quality of Five Algorithms

[0079]

[0080] Example 6:

[0081] Reconstruction performance at low sampling rates: Previous experiments were conducted at relatively high sampling rates. To investigate the performance of this method at low sampling rates, we used the Set11 dataset with image size of 128×128 and conducted experiments at sampling rates of 2%, 5.45%, 6.75%, 8%, and 10%. Figure 7a and Figure 7b The reconstructed images and quantitative results of five algorithms are presented. It can be seen that DGI and Pse-Inv are almost unable to reconstruct the target image at low sampling rates. CS-TV performs well at sampling rates above 6.75%, but exhibits significant noise and a substantial decrease in image quality at lower sampling rates. SD-Net, due to its two-step reconstruction method, relies heavily on the quality of the reconstructed image from traditional algorithms. SD-Net uses the Pse-Inv algorithm as its first step. Since Pse-Inv cannot reconstruct a good image at low sampling rates, the network cannot learn image features, resulting in poor reconstruction performance for SD-Net, even worse than the traditional CS-TV algorithm. In contrast, our method uses a one-step reconstruction approach; the neural network SDD-Net can directly extract image features from the entire measurement space, thereby reconstructing a high-quality image. SDD-Net demonstrates a significant advantage in PSNR compared to the other four algorithms.

[0082] To further verify the effectiveness of this method in real-world scenarios at low sampling rates, a system was established... Figure 8 The SPI experimental system shown uses a DLP LightCrafter4500 projector with a resolution of 912×1140 and a MER2-502-79U3M CMOS camera with a resolution of 2448×2048. The camera lens is a DTCA230-37C object-image telecentric lens. The DLP LightCrafter4500 projects a patterned light field onto the object, and the camera takes pictures of the object. The sum of all pixel values ​​in the obtained image is used as the total light intensity signal. 410 measurement signals were collected, with an SR of 10%. It can be seen that the DGI and Pse-Inv methods almost fail to reconstruct the target object. Although CS-TV can roughly reconstruct the image, the smoothness of the algorithm leads to the loss of image edge contour details. SD-Net suffers from poor image quality due to the poor image quality reconstructed by the Pse-Inv algorithm in the first step, resulting in low-quality output images. Our proposed method achieves the best reconstruction results, with less noise and clearer structural edges in the generated images.

[0083] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A self-supervised dual-domain dual-path single-pixel imaging method, characterized in that, Includes the following steps: S1. Obtain single pixel measurement value The single-pixel measurement value I is obtained by calculating the target image T and the imaging mode P; the specific calculation formula is as follows: Yo i =∫P i (x,y)T(x,y)dxdy In the formula I i I represents the measurement value of the i-th single pixel; P represents the modulation pattern, i.e., the imaging mode; P i This represents the i-th modulation pattern; the subscripts i = 1, 2, ..., M indicate the i-th measurement, T is the target image, and (x, y) are the pixel coordinates. S2. Construction of the neural network SDD-Net The single-pixel measurement value I is input into the neural network SDD-Net to obtain the preliminary reconstructed image T1; T1 then undergoes a single-pixel forward imaging process to generate the corresponding measurement value PT1; the measurement value PT1 is input into the same neural network SDD-Net again to obtain the final reconstructed image T2; where I and PT1 constitute the measurement domain loss, and T1 and T2 constitute the image domain loss constraint; the neural network SDD-Net consists of an encoder and two decoders, the decoders are composed of a structure decoder and a texture decoder, the decoders and encoder are connected by a special skip connection that includes channel spatial attention, and the two decoders are connected by an interactive attention module; S3. Loss Function Design The total variation method is used to guide the network in separating and reconstructing structural and texture information; the SDD-Net neural network is trained under constraints using the following loss function: T o =f θ =T s +T t =T gn(g＝1,2) In the formula T o T represents the output image. s Represents a structural image, T t P represents the texture image, and I represents the pattern image. i Let n be the input measurement value of the network, and i be the index of the image domain and the measurement domain, respectively. g is the output sequence number of the network in which iteration. μ is the weight of the image domain constraint term in the loss function. σ is the weight of the third part of the regularization in the loss function, and λ is the weight of the fourth part of the regularization in the loss function. N represents the total number of pixels in the image. L1 is the loss function. T 1n This indicates the position of the nth pixel in the first output of the network; TV represents the total variation loss function; f θ Represents the parameters of the neural network; The total variation loss function is calculated as follows: This represents the pixel value at coordinates (x+1, y) in the structured image, which is the value of the pixel adjacent to the right. This represents the pixel value at coordinates (x, y) in the structured image; This represents the pixel value at coordinates (x, y+1) in the structured image, i.e., the value of the adjacent pixel below. In order to preserve the edge details of the image and obtain better convergence, the L1 loss function is used to constrain the two domains. S4. Generation of reconstructed images The output T of the structure path s and the output T of the texture path t The final reconstructed image is obtained by adding the pixels together.

2. The self-supervised dual-domain dual-path single-pixel imaging method according to claim 1, characterized in that, In S2, the interactive attention module consists of a frequency domain branch and a spatial domain branch. The frequency domain uses Fourier convolutional layers to extract global frequency information from the input feature map; the spatial domain uses cascaded channel spatial attention to extract local spatial information. Finally, the frequency domain branch and the spatial domain branch perform feature fusion on the channel.

3. The self-supervised dual-domain dual-path single-pixel imaging method according to claim 1, characterized in that, The loss function in S3 consists of four parts. The first part is the measurement domain constraint term, which consists of I and PT1, to ensure that the reconstructed image accurately reflects the original measurement value. The second part is the image domain constraint term, which consists of T1 and T2, to ensure the consistency and quality of the network output image. These two parts are used to generate a correct and unique formal reconstructed image. The third and fourth parts are regularization terms used to guide the two paths of the neural network to focus on their respective reconstruction tasks.

Images