Single-pixel non-vision field imaging system and super-resolution reconstruction method thereof

Abstract

The invention discloses a single-pixel non-vision-field imaging system and a super-resolution reconstruction method thereof, belongs to the field of optical systems, and aims to solve the problems in the prior art. The system comprises a light source, a DMD (Digital Micromirror Device), a plane mirror, a plano-concave lens, an object to be measured, a shielding object, a plano-convex lens, a plane mirror, a single-pixel detector, a data collector and a computer; the DMD, the plane mirror, the plano-concave lens, the plano-convex lens, the plane mirror and the single-pixel detector are coaxially arranged; light beams emitted by the light source are modulated by the DMD, illumination speckles preset by a computer are projected, the speckles are reflected to the plano-concave lens through the plane mirror and then irradiate the surface of an object to be measured through the plano-concave lens, and transmitted light is converged by the plano-convex lens after being reflected by the shielding object, then reflected by the plane mirror and finally captured by the single-pixel detector. The single-pixel detector transmits collected data to the data collector, the data collector transmits the data to the computer for processing, and finally a target image is reconstructed.

Description

Technical Field

[0001] The present invention belongs to the technical field of optical systems, and in particular relates to a single-pixel non-field-of-view imaging system and a super-resolution reconstruction method thereof. Background Art

[0002] Imaging technology is a vital branch of modern science and technology. Traditional imaging systems, such as digital cameras and video cameras, have achieved tremendous success over the past few decades and are widely used in a wide range of applications. While capable of providing high-resolution images, they also suffer from several technical limitations, such as photon efficiency, resolution limitations, noise, dynamic range, cost, and real-time imaging. Due to the physical limitations of optical systems, imaging of non-field-of-view areas is not possible.

[0003] Non-line-of-sight imaging technology transcends the limitations of traditional line-of-sight imaging. It uses the effective signal carried by indirect scattered light to reconstruct images of objects obscured by obstacles. Traditional non-line-of-sight imaging systems typically require complex optical equipment and extensive data acquisition, resulting in high system costs, bulk, and complex operation. Furthermore, the imaging resolution of traditional systems is often limited, making it difficult to meet the demand for high-resolution imaging.

[0004] Single-pixel imaging technology, as an emerging imaging method, offers unique advantages. It encodes the target by controlling the intensity fluctuations of the light field at the transmitter. At the receiver, the reflected echo energy is collected and the spatial information of the object is reconstructed using the principle of spatial intensity correlation. This is an indirect imaging method. It can significantly reduce system cost and complexity, while also achieving higher sensitivity in some cases. However, single-pixel imaging technology is relatively slow, and due to the limitations of a single detector, its imaging resolution is generally low.

[0005] Super-resolution reconstruction technology can restore high-resolution images from low-resolution images. This technology uses prior information and algorithm optimization to recover high-resolution images from low-resolution measurement data. However, existing super-resolution reconstruction techniques often face problems such as poor reconstruction results, high computational complexity, and sensitivity to noise when applied to single-pixel imaging data.

[0006] See the arXiv journal archive for the paper "SPI-GAN: Towards Single-Pixel Imaging through Generative Adversarial Network" by Nazmul Karim et al. This method proposes a reconstruction network based on a generative adversarial network. This network uses a generator to reconstruct images from noisy inputs and a discriminator to learn to distinguish between real and reconstructed images. However, achieving good reconstruction results requires a large amount of training data, and the reconstruction also loses edge and texture details. Summary of the Invention

[0007] In order to solve the problem that the existing single-pixel non-line-of-sight imaging technology loses edge and texture detail information when reconstructing an image, the present invention proposes a single-pixel non-line-of-sight imaging system and a super-resolution reconstruction method.

[0008] The present invention solves the above problems and adopts the following technical solutions:

[0009] A single-pixel non-line-of-sight imaging system, comprising: a light source, a DMD, a plane reflector, a plano-concave lens, an object to be measured, an obstruction, a plano-convex lens, a plane reflector, a single-pixel detector, a data collector, and a computer;

[0010] The DMD, plane reflector, plano-concave lens, plano-convex lens, plane reflector and single-pixel detector are coaxially arranged;

[0011] The light beam emitted by the light source is modulated by the DMD to project an illumination pattern preset by the computer. The speckle is reflected by a plane mirror to a plano-concave lens, and then passes through the plano-concave lens to illuminate the surface of the object to be measured. After being reflected by an obstruction, the transmitted light is converged by a plano-convex lens, and then reflected by a plane mirror, and finally captured by a single-pixel detector; the single-pixel detector transmits the collected data to a data collector, which transmits the data to a computer for processing, and finally reconstructs the target image.

[0012] A super-resolution reconstruction method for a single-pixel non-line-of-sight imaging system, the method comprising the steps of:

[0013] Step 1: Collect single-pixel non-viewing-area imaging data:

[0014] First, an imaging system is built, which consists of a light source, a DMD, a plane reflector, a plano-concave lens, an object to be measured, an obstruction, a plano-convex lens, a plane reflector, a single-pixel detector, a data collector, and a computer. Light emitted from the light source projects a computer-defined illumination pattern through the DMD. The illumination pattern is reflected by the plane reflector onto the plano-concave lens, then passes through the plano-concave lens to illuminate the surface of the imaging object. After passing through the object surface, it is reflected by the obstruction, converged by the plano-convex lens, and reflected by the plane reflector before being collected by the single-pixel detector. The single-pixel detector transmits the single-pixel non-line-of-sight imaging data to the data collector. The data collector transmits the collected single-pixel non-line-of-sight imaging data to the computer for processing. This process is repeated multiple times, using a different illumination pattern each time, and ultimately a series of measurement values and corresponding illumination patterns are obtained.

[0015] Step 2: Build a super-resolution dataset: Preprocess the measurements and illumination speckles from step 1, number and label each data point to ensure the accuracy and consistency of all data; remove high-noise and erroneous data samples; and divide all prepared data into a training set and a validation set in an 8:2 ratio.

[0016] Step 3: Construct a super-resolution reconstruction network model. The super-resolution reconstruction network model consists of two parts: a generator and a discriminator. The input measurement value and illumination speckle are passed through the generator to output a high-resolution reconstruction. The real image and the high-resolution reconstruction are passed through the discriminator to obtain the probability value of the input being real data.

[0017] Step 4: Train the network model: Input the measurement values and illumination speckles prepared in step 2 into the super-resolution reconstruction network generator for training;

[0018] Step 5: Stop training and save the model: Based on the model's performance on the validation set, determine whether to end training early. Save the finalized model parameters as a data file. During model inference, load the model weight parameters and input values to obtain the final high-resolution image.

[0019] The single-pixel non-line-of-sight imaging system proposed in this invention utilizes a unique optical transmission structure in which illumination speckles projected by a DMD are reflected by a plane mirror, transmitted by a plano-concave lens, reflected by the obstruction surface, and converged by a plano-convex lens. This allows light to bypass obstructions and transmit information about hidden objects, enabling rapid imaging of objects obstructed by line of sight. Furthermore, the system performs exceptionally well in low-light environments, making it suitable for non-destructive imaging of objects and for use in concealed environments and hazardous areas.

[0020] The super-resolution reconstruction method proposed in the present invention simultaneously extracts global features and local details through an enhanced residual dense module, realizes detail enhancement within the generator of the super-resolution network, can restore high-resolution images from low-resolution images, improves the clarity of imaging details, and helps to more accurately identify and analyze the characteristics of target objects.

[0021] The present invention utilizes compressed sensing technology to reduce the amount of data collected and storage requirements, further reducing system operating costs and maintenance difficulty. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is a schematic diagram of a single-pixel non-line-of-sight imaging system of the present invention.

[0023] Figure 2 The figure is a flow chart of a super-resolution reconstruction method for a single-pixel non-field-of-view imaging system.

[0024] Figure 3 This is a super-resolution network structure diagram of a single-pixel non-field-of-view imaging system of the present invention.

[0025] Figure 4 This is a structural diagram of the super-resolution reconstruction network generator described in the present invention.

[0026] Figure 5 This is a structural diagram of the super-resolution reconstruction network TV module described in the present invention.

[0027] Figure 6 This is a diagram of the enhanced dense residual block structure of the super-resolution reconstruction network described in the present invention.

[0028] Figure 7 This is a dense block structure diagram of the super-resolution reconstruction network described in the present invention.

[0029] Figure 8 This is a structural diagram of the detail enhancement block of the super-resolution reconstruction network described in the present invention.

[0030] Figure 9 This is a structural diagram of the super-resolution reconstruction network discriminator described in the present invention. DETAILED DESCRIPTION

[0031] The present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0032] like Figure 1 As shown, a single-pixel non-line-of-sight imaging system includes a light source 1, a DMD 2, a plane reflector 3, a plano-concave lens 4, an object to be measured 5, an obstruction 6, a plano-convex lens 7, a plane reflector 8, a single-pixel detector 9, a data collector 10, and a computer 11. The DMD is a digital micromirror device.

[0033] Light emitted from light source 1 passes through DMD 2 and projects an illumination pattern predefined by computer 11. This illumination pattern is then reflected by plane mirror 3 onto plano-concave lens 4, which then passes through that lens onto the surface of imaging object 5. After passing through object 5, the light is reflected by obstruction 6, converged by plano-convex lens 7, and reflected by plane mirror 8 before being collected by single-pixel detector 9. This single-pixel detector 9 transmits the data to data collector 10. From collector 10, the data is transmitted to computer 11 for processing. The entire system can acquire target information through indirect optical signals even when the imaging object 5 is obscured.

[0034] like Figure 2 As shown, a super-resolution reconstruction method for a single-pixel non-line-of-sight imaging system specifically includes the following steps:

[0035] Step 1: Collect single-pixel non-viewing-area imaging data:

[0036] First, an imaging system is constructed, consisting of a light source 1, a DMD 2, a plane mirror 3, a plano-concave lens 4, an object to be measured 5, an obstruction 6, a plano-convex lens 7, a plane mirror 8, a single-pixel detector 9, a data acquisition unit 10, and a computer 11. Light emitted from light source 1 passes through DMD 2 and projects an illumination pattern predefined by computer 11. The illumination pattern is then reflected by plane mirror 3 onto plano-concave lens 4, then passes through plano-concave lens 4 to illuminate the surface of the imaging object 5. After passing through the surface of object 5, it is reflected by obstruction 6, converged by plano-convex lens 7, and reflected by plane mirror 8 before being collected by single-pixel detector 9. The single-pixel detector 9 transmits the single-pixel non-line-of-sight imaging data to data acquisition unit 10. Data acquisition unit 10 then transfers the collected single-pixel non-line-of-sight imaging data to computer 11 for processing. This process is repeated multiple times, each time using a different illumination pattern, ultimately obtaining a series of measurement values and corresponding illumination patterns.

[0037] Step 2: Build a super-resolution dataset: Preprocess the measurements and illumination speckle patterns from Step 1, label each data point, and ensure all data is accurate. Remove high-noise and erroneous data samples. Split all prepared data into a training set and a validation set in an 8:2 ratio.

[0038] Step 3: Build a super-resolution reconstruction network model:

[0039] like Figure 3 As shown in Figure 2, the super-resolution network model consists of two parts: a generator and a discriminator. The generator takes the input measurements and illumination speckles, and outputs a high-resolution reconstruction. The discriminator then passes the real image and the high-resolution reconstruction through the discriminator to determine the probability that the input is real data.

[0040] like Figure 4As shown in the figure, a generator is constructed, which mainly includes a total variation (TV) module, multiple convolutional activation layers, an enhanced dense residual block (ERRDB), and an upsampling layer. The generator is responsible for generating high-resolution images. The input measurement values and illumination speckle are passed through the TV module to output a low-resolution image. The low-resolution image is passed through the convolutional activation layer to output a low-frequency feature image. The low-frequency feature image is passed through convolution layer 1 and activation layer 1, an ERRDB block, convolution layer 2 and activation layer 2 to output a high-frequency feature image. The high-frequency feature image is passed through an upsampling layer, convolution layer 3 and activation layer 3, and convolution layer 4 to output a high-resolution image.

[0041] like Figure 5 As shown in the figure, the TV module consists of an optimization problem solver and a judgment block. During operation, the TV module first generates an initialization image with all elements set to zero. This image, along with the input measurement values and illumination speckle pattern, is then fed into the optimization problem solver to update the initialization image. The updated initialization image then enters the judgment block, which determines whether its L2 norm is less than 0.001. If so, the initialization image is again fed into the optimization problem solver for an update. If so, the initialization image is output as a low-resolution image.

[0042] like Figure 6 As shown in the figure, the ERRDB block consists of a dense block, residual factors, convolutional layer 6, activation layer 6, convolutional layer 7, and a detail enhancement block. The input image is multiplied by the dense block and the residual factors to produce a feature map. The output feature map is added to the input image and then added to the result of convolutional layer 6, activation layer 6, and convolutional layer 7 to produce a depth feature map. The depth feature map is multiplied by the result of the detail enhancement block and then added to the feature map to return the output of the ERRDB block.

[0043] like Figure 7 As shown in Figure 1, the dense block consists of the eighth to twelfth convolutional activation layers. The output of each convolutional activation layer is concatenated with the outputs of all the previous convolutional activation layers to form high-dimensional features for super-resolution reconstruction.

[0044] like Figure 8 As shown in the figure, the detail enhancement block consists of global max pooling, global min pooling, global average pooling, and multiple activation layers consisting of a 3×3 convolutional layer, a 5×5 convolutional layer, a 7×7 convolutional layer, a 3×3 convolutional layer (1), a 3×3 convolutional layer (2), and a 1×1 convolutional layer. The input image undergoes global max pooling, global min pooling, and global average pooling to output feature 1. Feature 1 is concatenated and passed through the 3×3 convolutional layer (1) before being input into feature 2. Simultaneously, the input image also passes through the 3×3 convolutional layer, a 5×5 convolutional layer, and a 7×7 convolutional layer to output feature 3. Feature 3 is concatenated and passed through a 1×1 convolutional layer to output feature 4. Features 3 and 4 are concatenated and passed through the 3×3 convolutional layer (2) before returning to the output of the detail enhancement block.

[0045] like Figure 9 As shown in the figure, a discriminator is constructed, which consists of multiple convolutional layers, activation layers, global average pooling, linear layers 1, 5, and 2. The discriminator receives two types of data: real data samples or data synthesized by the generator. The discriminator's input image first passes through multiple convolutional layers 5 and 4 activation layers to generate feature maps. The feature maps are then subjected to global average pooling to output a one-dimensional feature vector. The feature vector then passes through linear layers 1, 5, and 2 activation layers to output a scalar value representing the probability that the input is real data.

[0046] Step 4: Train the network model: Using the input and output data from Step 2, train the network model by minimizing the loss functions of the generator and discriminator. The generator loss consists of adversarial loss, content loss, and perceptual loss. The adversarial loss is based on the discriminator output and makes the generated images visually closer to real images.

[0047] During model training, the performance of the model on the validation set is evaluated using three key metrics to determine the optimal model weights: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Normalized Root Mean Square Error (NRMSE). PSNR primarily assesses the quality loss caused by signal compression. SSIM provides a comprehensive assessment of structural similarity, focusing on aspects such as contrast, brightness, and structural integrity. NRMSE is highly sensitive to errors in image data, making it suitable for more consistent evaluation across different datasets. The calculation formulas for the above three evaluation metrics are as follows:

[0048]

[0049] Where x and y represent the real image and the reconstructed image respectively. μ represents the average pixel value of the image, σ represents the variance of the image, and σ xy Represents the covariance between the two images. In addition, c1 and c2 are introduced constants with a value of 0.01.

[0050] Step 5: Stop training and save the model: Monitor the loss and quality of the generated data during training. Stop training when the loss of the generator and discriminator on the validation set from step 2 is sufficiently small. Save the model parameters at this point in time as a file with the suffix ".pth". For subsequent super-resolution reconstruction, input the measured values and illumination speckle pattern into the network's generator to return the final image.

[0051] Example:

[0052] A single-pixel non-line-of-sight imaging system comprises: light source 1, a 200 mW laser with a wavelength of 660 nm; digital micromirror device 2, a Texas Instruments DLP7000 DMD with a resolution of 1024x768 pixels, projecting an illumination speckle pattern at a maximum modulation rate of 22,727 Hz; plane reflectors 3 and 8, JLGD mirrors with dimensions of 60 mm x 80 mm x 5 mm; plano-concave lens 4, a Hengyang Optics GLA13-025B-150A; A4-printed cardboard with a hollowed-out finish as the object to be measured 5; a black acrylic sheet as the obstruction 6; and a plano-convex lens 7, a Hengyang Optics GLA11-025-200B. Single-pixel detector 9, a Thorlabs PDA36A2 detector, captures the total reflected light intensity. This detector has an effective wavelength range of 350 nm to 1100 nm and a detection area of 13 square millimeters. A data acquisition card 10 (model HDM-DAQ-24128-U) with a maximum sampling rate of 128,000 Hz was used to convert the measurements into electrical signals. These measurements were then transferred to a computer 11 (Asus desktop) for image reconstruction.

[0053] A super-resolution reconstruction method for a single-pixel non-line-of-sight imaging system comprises the following steps:

[0054] The imaging system described in step 1 primarily consists of a light source 1 (660 nm 200 mW laser), a DMD 2 (Texas Instruments DLP7000 DMD), plane mirrors 3 and 8 (JLGD 60 mm x 80 mm x 5 mm), a plano-concave lens 4 (Hengyang Optics GLA13-025B-150A), an object to be measured 5 (cardboard with a hollowed-out A4 print), an obstruction 6 (black acrylic sheet), a plano-convex lens 7 (Hengyang Optics GLA11-025-200B), a single-pixel detector 9 (Thorlabs PDA36A2), a data acquisition unit 10 (HDM-DAQ-24128-U), and a computer 11 (Asus desktop computer). The target object was imaged under 1200 illumination speckle modulations, and the corresponding single-pixel detector values and illumination speckle patterns were recorded to ensure data reproducibility.

[0055] As described in Step 2, a super-resolution dataset was constructed from the 1200 experimental data points. All collected data was labeled, processed, and partitioned. The quality and consistency of all data were ensured for subsequent processing and analysis. Raw data from the measurements and corresponding speckle images were first collected and labeled. This ensured the complete accuracy and consistency of the collected data. The processed measurements and illumination speckle patterns were then divided into training and validation sets for model training and evaluation.

[0056] Construct the super-resolution reconstruction network model as described in step 3, define 23 enhanced residual dense blocks, the default convolution kernel size is 3×3 and the ReLu activation function. Use the He method to initialize the network weights to ensure that the initial state of the network is reasonable. Introduce multiple convolution layers of sizes 3×3, 5×5, and 7×7 and a variety of global pooling to form a detail enhancement block, such as Figure 8 As shown in the figure, these convolutional layers process features in parallel, capturing information at different scales to enhance the model's ability to capture image detail. Spatial features are weighted using 3×3 convolutions, which are then concentrated on key image regions through global average pooling, global max pooling, and global min pooling. By combining multi-scale and spatial features, the processed data is further refined using 3×3 convolutions and mapped to a range between 0 and 1 using a sigmoid activation function. This approach amplifies key details while preserving the original information.

[0057] In the training network model described in Step 4, use the Adam optimizer with a learning rate of 0.0002, a β1 parameter of 0.5, and a β2 parameter of 0.999. Automatic mixed precision is employed to accelerate training. Network weights are updated using gradient descent. At each training iteration, the loss function is calculated and the network weights are updated. During training, monitor the loss and the quality of the generated images, and adjust the learning rate and network parameters to optimize performance.

[0058] As mentioned in step 5, stop training when the validation set loss stops decreasing significantly or reaches 1000 training rounds. At this point, the model's performance on the validation set meets expectations. Save the trained model weights to a local hard drive. Ensure model reproducibility by revisiting the model architecture and training parameters.

[0059] By comparing the results of the present invention with those of existing reconstruction methods, the feasibility and superiority of the proposed method are further verified. The comparison of relevant indicators between the existing technology and the proposed method is shown in Table 1:

[0060] Evaluation indicators Differential ghost imaging Compressed Sensing SPI-GAN The present invention Peak signal-to-noise ratio 9.57 9.67 4.86 11.81 Structural similarity 0.108 0.114 0.128 0.566 Normalized error 0.866 0.868 2.525 0.548

[0061] As can be seen from the table, the method proposed in the present invention has higher peak signal-to-noise ratio, structural similarity and normalized error. These indicators further illustrate that the present invention has better super-resolution reconstruction quality.

Claims

1. A single-pixel non-line-of-sight imaging system, characterized in that: The system comprises: a light source (1), a DMD (2), a plane reflector (3), a plano-concave lens (4), an object to be measured (5), an obstruction (6), a plano-convex lens (7), a plane reflector (8), a single-pixel detector (9), a data collector (10) and a computer (11); The DMD (2), the plane reflector (3), the plano-concave lens (4), the plano-convex lens (7), the plane reflector (8) and the single-pixel detector (9) are coaxially arranged; The light beam emitted by the light source (1) is modulated by the DMD (2) to project an illumination pattern preset by the computer (11). The illumination pattern is reflected by the plane reflector (3) to the plano-concave lens (4), and then irradiated onto the surface of the object to be measured (5) through the plano-concave lens (4). The light passing through the object to be measured (5) is reflected by the obstruction (6), converged by the plano-convex lens (7), and then reflected by the plane reflector (8), and finally captured by the single-pixel detector (9); the single-pixel detector (9) transmits the collected data to the data collector (10), and the data collector (10) transmits the data to the computer (11) for processing, and finally reconstructs the target image.

2. A super-resolution reconstruction method for a single-pixel non-line-of-sight imaging system, characterized by: The method comprises the steps of: Step 1: Collect single-pixel non-viewing-area imaging data: First, an imaging system is constructed, which consists of a light source (1), a DMD (2), a plane reflector (3), a plano-concave lens (4), an object to be measured (5), an obstruction (6), a plano-convex lens (7), a plane reflector (8), a single-pixel detector (9), a data collector (10) and a computer (11); light emitted from the light source (1) projects an illumination pattern predefined by the computer (11) through the DMD (2); the illumination pattern is reflected by the plane reflector (3) onto the plano-concave lens (4), and then irradiated onto the surface of the imaging object (5) through the plano-concave lens (4), and then passes through the surface of the object to be measured (5) and is reflected by the obstruction (6), converged by the plano-convex lens (7) and reflected by the plane reflector (8) before being collected by the single-pixel detector (9), and the single-pixel detector (9) transmits the single-pixel non-field-of-view imaging data to the data collector (10); the data collector (10) transmits the collected single-pixel non-field-of-view imaging data to the computer (11) for processing; This process is repeated multiple times, using different illumination patterns each time, and finally a series of measurement values and corresponding illumination patterns are obtained; Step 2: Build a super-resolution dataset: Preprocess the measurements and illumination speckles from step 1, number each data point and assign a category label to ensure the accuracy and consistency of all data. Remove high-noise and erroneous data samples; divide all prepared data into training set and validation set in an 8:2 ratio; Step 3: Construct a super-resolution reconstruction network model. The super-resolution reconstruction network model consists of two parts: a generator and a discriminator. The input measurement value and illumination speckle are passed through the generator to output a high-resolution reconstruction. The real image and the high-resolution reconstruction are passed through the discriminator to obtain the probability value of the input being real data. Step 4: Train the network model: Input the measurement values and illumination speckles prepared in step 2 into the super-resolution reconstruction network generator for training; Step 5: Stop training and save the model: Based on the model's performance on the validation set, determine whether to end training early. Save the finalized model parameters as a data file. During model inference, load the model weight parameters and input values to obtain the final high-resolution image.

3. The super-resolution reconstruction method of a single-pixel non-line-of-sight imaging system according to claim 2, characterized in that: The generator described in step 3 mainly includes a TV module, multiple convolutional activation layers, ERRDB and an upsampling layer; the generator is responsible for generating high-resolution images, and the input measurement values and illumination speckles are output as low-resolution images through the TV module, and the low-resolution images are output as low-frequency feature images through the convolutional activation layer, and the low-frequency feature images are output as high-frequency feature images through the convolutional layer 1 and activation layer 1, ERRDB block, convolutional layer 2 and activation layer 2; the high-frequency feature images are output as high-resolution images through the upsampling layer, convolutional layer 3 and activation layer 3, and convolutional layer 4.

4. The super-resolution reconstruction method of a single-pixel non-line-of-sight imaging system according to claim 3, characterized in that: The ERRDB block consists of a dense block, a residual factor, a convolutional layer six, an activation layer six, a convolutional layer seven and a detail enhancement block. The input image is multiplied by the dense block and the residual factor to output a feature map; the output feature map is added to the input image, and then added to the result of the convolutional layer six, the activation layer six, and the convolutional layer seven to output a depth feature map; the depth feature map is multiplied with the result of the detail enhancement block, and then added to the feature map to return to the output of the ERRDB block.

5. The super-resolution reconstruction method of a single-pixel non-line-of-sight imaging system according to claim 3, characterized in that: The detail enhancement block consists of global maximum pooling, global minimum pooling, global average pooling, and multiple activation layers consisting of a 3×3 convolutional layer, a 5×5 convolutional layer, a 7×7 convolutional layer, a 3×3 convolutional layer 1, a 3×3 convolutional layer 2, and a 1×1 convolutional layer; the input image is subjected to global maximum pooling, global minimum pooling, and global average pooling to output feature 1; Feature one is concatenated and passed through the first 3×3 convolution layer to output feature two. The input image also passes through the 3×3 convolution layer, the 5×5 convolution layer, and the 7×7 convolution layer to output feature three. Feature three is concatenated and passed through the 1×1 convolution layer to output feature four. Feature three and feature four are concatenated and passed through the second 3×3 convolution layer to return to the output of the detail enhancement block.

Images