High-quality single-pixel imaging reconstruction method based on untrained convolutional neural network

Abstract

The application relates to the field of optical single-pixel imaging, and particularly relates to a high-quality single-pixel imaging reconstruction method based on a non-training convolutional neural network, which comprises the following steps: loading a coding pattern by using a digital micromirror array, coding a to-be-detected scene, and collecting single-pixel imaging of the coded to-be-detected scene to obtain a first one-dimensional signal; constructing a non-training convolutional neural network; taking the one-dimensional signal as the input of the non-training convolutional neural network, outputting a preliminary estimation of the to-be-detected scene, and performing inner product operation on the preliminary estimation and the coding pattern to obtain a second one-dimensional signal; calculating the absolute error of the first one-dimensional signal and the second one-dimensional signal as a loss value, and optimizing the parameters of the non-training convolutional neural network; and finally, outputting a single-pixel imaging reconstruction result corresponding to the first one-dimensional signal by the non-training convolutional neural network. The neural network constructed by the application does not need to train a large amount of data set, has strong generalization ability, and has high single-pixel reconstruction imaging quality.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of optical single-pixel imaging, and in particular to a high-quality single-pixel imaging reconstruction method based on a non-training convolutional neural network. BACKGROUND

[0002] Single-pixel imaging technology is a new imaging technology that first uses a single-pixel detector to obtain the spatial information of a target scene, and then reproduces the target scene through some reconstruction methods. Compared with traditional imaging methods, the single-pixel imaging method has stronger anti-noise capability, higher signal-to-noise ratio and lower cost in some special wavebands or low-illumination fields, and in particular provides a low-cost and effective means for imaging applications in the invisible spectral region. Therefore, single-pixel imaging technology has been widely used in underwater imaging, three-dimensional imaging, terahertz imaging, and phase imaging fields.

[0003] However, in the single-pixel imaging method, a large amount of data needs to be collected to obtain high-quality images, which results in a time-consuming data collection process and is not conducive to real-time imaging in some scenarios. Therefore, there is a trade-off between reconstruction quality and collection time. Therefore, the prior art proposes many methods to solve this problem, which are mainly divided into two categories: one is to design a coding matrix to obtain more information when collecting data; the other is to optimize the reconstruction algorithm to obtain higher reconstruction quality. The unique advantages of deep learning technology in solving inverse problems and nonlinear problems have attracted widespread attention, such as single-pixel imaging, digital holography, phase imaging and unfolding, fringe projection imaging, information security, etc. However, the data-driven deep learning reconstruction method needs to impose potential constraints on the reconstruction problem by using a large-scale data set, and the acquisition of the data set is not only time-consuming but also limited by experimental conditions; moreover, the deep learning implemented by data-driven has poor interpretability and limited generalization ability. SUMMARY

[0004] The present application aims to solve the technical problem that the prior art method of using deep learning technology for single-pixel reconstruction relies on a data set and has poor generalization ability.

[0005] To solve the above technical problems, the present application provides a high-quality single-pixel imaging reconstruction method based on a non-training convolutional neural network, which comprises the following steps:

[0006] S1, loading a coding pattern using a digital micromirror array to code a to-be-measured scene, and performing single-pixel imaging collection on the coded to-be-measured scene to obtain a first one-dimensional signal;

[0007] S2, constructing a non-training convolutional neural network;

[0008] S3, taking the one-dimensional signal as an input of a non-training convolutional neural network, outputting a preliminary estimation of the scene to be measured, and performing an inner product operation on the preliminary estimation and the encoding pattern, to output a second one-dimensional signal;

[0009] S4, calculating an absolute error of the first one-dimensional signal and the second one-dimensional signal as a loss value, and optimizing parameters of the non-training convolutional neural network according to the loss value, and finally outputting a single-pixel imaging reconstruction result corresponding to the first one-dimensional signal by the non-training convolutional neural network.

[0010] Further, in step S1, the encoding pattern is defined as P, the number of single-pixel imaging acquisition is M, and the first one-dimensional signal is s, which satisfies:

[0011] S(m) = <P m , O > m = {1, 2,..., M};

[0012] Wherein, O is the original scene to be measured, and <> is an inner product operation.

[0013] Further, in step S2, the non-training convolutional neural network is URNet.

[0014] Further, the non-training convolutional neural network is divided into an input layer, a deconvolution layer, a convolution layer and an output layer according to the logical sequence of data processing, wherein:

[0015] The input layer includes a convolution kernel with a size of 4x4 and a step of 1;

[0016] The deconvolution layer includes five layers, and each deconvolution layer is provided with a batch normalization layer and a RELU activation function, and each convolution kernel in the deconvolution layer has a size of 4x4 and a step of 1;

[0017] The convolution layer includes four layers, and each convolution kernel in each convolution layer has a size of 3x3 and a step of 1;

[0018] The output layer is the same as the last layer in the convolution layer.

[0019] Further, the non-training convolutional neural network uses an Adam optimizer with a learning rate of 0.08 during training, the training step is 5000, and the learning rate is double decayed every 3000 training steps with a decay rate of 0.8.

[0020] Further, in step S3, the preliminary estimation is defined as The second one-dimensional information is Which respectively satisfy:

[0021]

[0022]

[0023] wherein, Delta URNet represents the processing procedure of the training-free convolutional neural network.

[0024] Further, in step S4, the single-pixel imaging reconstruction result is defined as The single-pixel imaging reconstruction result satisfies:

[0025]

[0026] wherein, represents the processing procedure of the training-free convolutional neural network after optimization according to the loss value, and ||·||L1 represents the least absolute deviation loss.

[0027] The present application has the advantages that the training-free convolutional neural network and the reconstruction method for image reconstruction in single-pixel imaging are provided, the physical-driven training-free network constructed by the present application can automatically optimize and finally reconstruct a two-dimensional image, the generalization ability of the neural network is stronger without training a large amount of data set, and the quality of the final single-pixel reconstruction imaging is higher. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 is a step flow chart of the high-quality single-pixel imaging reconstruction method based on the training-free convolutional neural network provided by the embodiment of the present application;

[0029] Figure 2 is a structural schematic diagram of the training-free convolutional neural network provided by the embodiment of the present application;

[0030] Figure 3 is a qualitative and quantitative result schematic diagram of the simple and complex natural images reconstructed by a 50% sampling rate provided by the embodiment of the present application;

[0031] Figure 4 is an interference pattern quality effect comparison schematic diagram provided by the embodiment of the present application;

[0032] Figure 5 is a PSNR and SSIM value schematic diagram compared with the prior art provided by the embodiment of the present application. DETAILED DESCRIPTION

[0033] In order to make the purpose, technical scheme and advantages of the present application more clear, the present application is further described in detail below by combining with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0034] Please refer to Figure 1 , Figure 1 is a step flow chart of a high-quality single-pixel imaging reconstruction method based on a non-training convolutional neural network provided by the embodiment of the application, and the high-quality single-pixel imaging reconstruction method comprises the following steps:

[0035] S1, a digital micromirror array is used to load an encoding pattern to encode a to-be-measured scene, and single-pixel imaging collection is performed on the encoded to-be-measured scene to obtain a first one-dimensional signal.

[0036] Further, in step S1, the encoding pattern is defined as P, the number of times of performing the single-pixel imaging collection is M, and the first one-dimensional signal is s, and the first one-dimensional signal s satisfies:

[0037] S(m)=<P m , O>,m={1,2,...M};

[0038] wherein O is the original to-be-measured scene, and <> is an inner product operation.

[0039] S2, a non-training convolutional neural network is constructed.

[0040] Further, in step S2, the non-training convolutional neural network is URNet.

[0041] Please refer to Figure 2 , Figure 2 is a structural schematic diagram of a non-training convolutional neural network provided by the embodiment of the application, and further, the non-training convolutional neural network is divided into an input layer, a deconvolution layer, a convolution layer and an output layer according to the logical order of data processing, wherein:

[0042] the input layer comprises a convolution kernel with a size of 4x4 and a step length of 1;

[0043] the deconvolution layer comprises five layers, and a batch normalization layer and a RELU activation function are arranged after each deconvolution layer, and each convolution kernel in the deconvolution layer has a size of 4x4 and a step length of 1;

[0044] the convolution layer comprises four layers, and each convolution kernel in each convolution layer has a size of 3x3 and a step length of 1;

[0045] the output layer is the same as the last layer in the convolution layer.

[0046] Further, the non-training convolutional neural network uses an Adam optimizer with a learning rate of 0.08 during training, the training step length is 5000, the learning rate is double attenuated every 3000 training step lengths, and the attenuation rate is 0.8.

[0047] S3. The one-dimensional signal is used as the input of an untrained convolutional neural network to obtain a preliminary estimate of the scene to be tested. The preliminary estimate is then used to perform an inner product operation with the encoded pattern to obtain a second one-dimensional signal.

[0048] Furthermore, in step S3, the preliminary estimate is defined as... The second one-dimensional information is They respectively satisfy:

[0049]

[0050]

[0051] Where, Δ URNet This describes the processing procedure of the untrained convolutional neural network.

[0052] S4. Calculate the absolute error between the first one-dimensional signal and the second one-dimensional signal as the loss value, optimize the parameters of the untrained convolutional neural network based on the loss value, and finally output the single-pixel imaging reconstruction result corresponding to the first one-dimensional signal through the untrained convolutional neural network.

[0053] Furthermore, in step S4, the single-pixel imaging reconstruction result is defined as... The single-pixel imaging reconstruction result satisfy:

[0054]

[0055] in, This represents the processing procedure of the untrained convolutional neural network after optimization based on the loss value, where ||·||L1 represents the minimum absolute bias loss.

[0056] For example, the high-quality single-pixel imaging reconstruction method based on untrained convolutional neural networks provided in this embodiment of the invention is compared with the prior art as follows:

[0057] like Figure 3 As shown, Figure 3 This presents qualitative and quantitative results of reconstructed simple and complex natural images at a 50% sampling rate. Prior techniques for comparison include Total Variational Minimization (TVAL3) implemented with enhanced Lagrangian and alternating orientation algorithms, Differential Ghost Imaging (DGI), Conjugate Gradient Descent (CGD), and Alternating Projection (AP). Evaluation metrics include peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) for quantitative analysis of the reconstruction results.Figure 3 The reconstruction result and the quantitative result of Table 1 can show that the high-quality single-pixel imaging reconstruction method based on the untrained convolutional neural network has the best performance in terms of visual quality and quantitative metrics (SSIM and PSNR).

[0058] Table 1: PSNR and SSIM values of reconstruction results of different reconstruction methods

[0059]

[0060] Figure 3 The results show that the reconstruction quality of TVAL3 is only lower than that of the URNet method. On this basis, the performance of the proposed untrained reconstruction method based on URNet is analyzed compared with TVAL3 under different sampling rates, and the method is applied to the phase shift interferograms collected in the actual experiment to verify the feasibility in the experiment. Please refer to Figure 4 It can be seen that the quality of the interferogram reconstructed by the method proposed in the embodiment of the application is higher, Figure 5 Figs. 1 (a) and (b) are respectively the PSNR and SSIM values of the untrained reconstruction method based on URNet compared with TVAL3 under different sampling rates, and the results show that when the sampling rate is 5%, the PSNR value of the method proposed in the embodiment of the application exceeds 22 dB, compared with the TVAL3 algorithm, the PSNR value is increased by 7-12 dB, and the SSIM value can be increased by 0.1-0.2 on average, which further proves the feasibility and advantages of the embodiment of the application in the experiment.

[0061] The application has the beneficial effects that the untrained convolutional neural network and the reconstruction method for image reconstruction in single-pixel imaging are proposed, the physically driven untrained network constructed by the application can automatically optimize and finally reconstruct a two-dimensional image, compared with the prior art, a large number of data sets do not need to be trained, the generalization ability of the neural network is stronger, and the final single-pixel reconstruction imaging quality is higher.

[0062] The embodiment of the application further provides a high-quality single-pixel imaging reconstruction based on an untrained convolutional neural network, which comprises:

[0063] The imaging acquisition module is configured to load a coded pattern by using a digital micromirror array to code a scene to be measured, and to acquire a first one-dimensional signal by performing single-pixel imaging acquisition on the coded scene to be measured.

[0064] The network construction module is configured to construct an untrained convolutional neural network.

[0065] a network training module configured to input the one-dimensional signal into a non-training convolutional neural network, output a preliminary estimation of the scene to be measured, and perform an inner product operation between the preliminary estimation and the coded pattern to output a second one-dimensional signal;

[0066] an optimization output module configured to calculate an absolute error of the first one-dimensional signal and the second one-dimensional signal as a loss value, optimize parameters of the non-training convolutional neural network according to the loss value, and finally output a single-pixel imaging reconstruction result corresponding to the first one-dimensional signal by the non-training convolutional neural network.

[0067] The high-quality single-pixel imaging reconstruction system based on the non-training convolutional neural network can realize the steps in the high-quality single-pixel imaging reconstruction method based on the non-training convolutional neural network in the above embodiments, and achieve the same technical effects. For details, refer to the description in the above embodiments, which will not be repeated here.

[0068] The embodiment of the present application also provides a computer device, which comprises a memory, a processor and a computer program stored in the memory and executable on the processor.

[0069] The processor invokes the computer program stored in the memory to execute the steps in the high-quality single-pixel imaging reconstruction method based on the non-training convolutional neural network provided by the embodiment of the present application. For details, refer to the description in the above embodiments. Figure 1 , and specifically comprises:

[0070] S1, loading a coded pattern by using a digital micromirror array to code a scene to be measured, and performing single-pixel imaging collection on the coded scene to be measured to obtain a first one-dimensional signal.

[0071] Further, in step S1, the coded pattern is defined as P, the number of times of performing the single-pixel imaging collection is M, the first one-dimensional signal is s, and the first one-dimensional signal s satisfies:

[0072] S(m) = <P m , O >, m = {1, 2,..., M};

[0073] wherein O is the original scene to be measured, and <> is an inner product operation.

[0074] S2, constructing a non-training convolutional neural network.

[0075] Further, in step S2, the non-training convolutional neural network is URNet.

[0076] Further, the non-training convolutional neural network is divided into an input layer, a deconvolution layer, a convolution layer and an output layer according to the logical sequence of data processing.

[0077] The input layer comprises a convolution kernel with a size of 4x4 and a step of 1;

[0078] The deconvolution layer comprises five layers, and each of the deconvolution layers is provided with a batch normalization layer and a RELU activation function after the deconvolution layer, each of the convolution kernels in the deconvolution layer has a size of 4x4 and a step of 1;

[0079] The convolution layer comprises four layers, and each of the convolution kernels in each of the convolution layers has a size of 3x3 and a step of 1;

[0080] The output layer is the same as the last layer in the convolution layer.

[0081] Further, the training-free convolutional neural network uses an Adam optimizer with a learning rate of 0.08 during training, the training step is 5000, the learning rate is double-decayed every 3000 training steps, and the decay rate is 0.8.

[0082] S3, taking the one-dimensional signal as an input of a training-free convolutional neural network, outputting a preliminary estimation of the scene to be measured, and performing an inner product operation on the preliminary estimation and the encoding pattern to output a second one-dimensional signal.

[0083] Further, in step S3, the preliminary estimation is defined as The second one-dimensional information is which respectively satisfy:

[0084]

[0085]

[0086] wherein, Δ URNet represents a processing process of the training-free convolutional neural network.

[0087] S4, calculating the absolute error of the first one-dimensional signal and the second one-dimensional signal as a loss value, and optimizing the parameters of the training-free convolutional neural network according to the loss value, and finally outputting a single-pixel imaging reconstruction result corresponding to the first one-dimensional signal by the training-free convolutional neural network.

[0088] Further, in step S4, the single-pixel imaging reconstruction result is defined as The single-pixel imaging reconstruction result satisfies:

[0089]

[0090] wherein, A processing procedure of the untrained convolutional neural network optimized according to the loss value, ||·||L1 represents a least absolute deviation loss.

[0091] The computer device provided in the embodiments of the present application can realize the steps in the high-quality single-pixel imaging reconstruction based on the untrained convolutional neural network in the above-described embodiments, and can realize the same technical effects. For details, refer to the description in the above-described embodiments, which will not be described here again.

[0092] The computer device provided in the embodiments of the present application can realize the steps in the high-quality single-pixel imaging reconstruction based on the untrained convolutional neural network in the above-described embodiments, and can realize the same technical effects. For details, refer to the description in the above-described embodiments, which will not be described here again.

[0093] Those skilled in the art can understand that all or part of the processes in the above-described embodiments can be completed by a computer program instructing related hardware, and the program can be stored in a computer readable storage medium. When the program is executed, the processes of the above-described embodiments can be included. The storage medium can be a magnetic disc, an optical disc, a read-only memory (ROM) or a random access memory (RAM).

[0094] It should be noted that in this document, the terms "comprising", "including", or any other variant thereof are intended to cover non-exclusive inclusions, so that processes, methods, articles or devices that include a series of elements not only include those elements, but also include other elements not explicitly listed, or include elements inherent to such processes, methods, articles or devices. Without more limitations, the element defined by the statement "including a" does not exclude the presence of another identical element in the process, method, article or device that includes the element.

[0095] From the above description of the embodiments, those skilled in the art can clearly understand that the above-described embodiment methods can be realized by means of software and a general hardware platform, and of course, they can also be realized by hardware, but in many cases, the former is a better embodiment. Based on such understanding, the technical solutions of the present application can be embodied in the form of a software product, which is stored in a storage medium (such as a ROM / RAM, a magnetic disc, an optical disc), and includes a plurality of instructions for making a terminal (which can be a mobile phone, a computer, a server, an air conditioner, or a network device) execute the methods described in the embodiments of the present application.

[0096] The embodiments of the present application are described above with reference to the accompanying drawings, and the disclosed are only the preferred embodiments of the present application, but the present application is not limited to the above-mentioned specific embodiments, and the above-mentioned specific embodiments are only illustrative, but not restrictive, and those skilled in the art can make many equivalent changes without departing from the spirit of the present application and the scope of the claims, and all the equivalent changes are within the protection scope of the present application.

Claims

1. A high-quality single-pixel imaging reconstruction method based on a training-free convolutional neural network, characterized in that, The high-quality single-pixel imaging reconstruction method comprises the following steps: S1, load a coding pattern by using a digital micromirror array to code a scene to be measured, and collect single-pixel imaging of the coded scene to be measured to obtain a first one-dimensional signal; S2, construct an untrained convolutional neural network; S3, input the one-dimensional signal into the untrained convolutional neural network as an input, output a preliminary estimation of the scene to be measured, and perform inner product operation on the preliminary estimation and the coding pattern to output a second one-dimensional signal; S4, calculate the absolute error of the first one-dimensional signal and the second one-dimensional signal as a loss value, optimize the parameters of the untrained convolutional neural network according to the loss value, and finally output a single-pixel imaging reconstruction result corresponding to the first one-dimensional signal by the untrained convolutional neural network; In step S2, the untrained convolutional neural network is URNet; The untrained convolutional neural network is divided into an input layer, a deconvolution layer, a convolution layer and an output layer according to the logical sequence of data processing, wherein: The input layer includes a convolution kernel with a size of 4x4 and a step of 1; The deconvolution layer includes five layers, and each deconvolution layer is provided with a batch normalization layer and a RELU activation function, and each convolution kernel in the deconvolution layer has a size of 4x4 and a step of 1; The convolution layer includes four layers, and each convolution kernel in each convolution layer has a size of 3x3 and a step of 1; The output layer is the same as the last layer in the convolution layer.

2. The high-quality single-pixel imaging reconstruction method based on the training-free convolutional neural network according to claim 1, wherein, In step S1, the encoding pattern is defined as P , the number of times of the single-pixel imaging acquisition is M , the first one-dimensional signal is s , the first one-dimensional signal s satisfies: ， ； wherein, O is the original scene to be measured, is an inner product operation.

3. The high-quality single-pixel imaging reconstruction method based on the training-free convolutional neural network according to claim 1, wherein, The untrained convolutional neural network uses an Adam optimizer with a learning rate of 0.08 during training, the training step is 5000, and the learning rate is double decayed every 3000 training steps with a decay rate of 0.

8.

4. The high-quality single-pixel imaging reconstruction method based on the training-free convolutional neural network according to claim 2, wherein, In step S3, the preliminary estimate is defined as , the second one-dimensional signal being , respectively, ； ， ； wherein, represents a processing procedure of the untrained convolutional neural network.

5. The high-quality single-pixel imaging reconstruction method based on the training-free convolutional neural network according to claim 4, wherein, In step S4, the single-pixel imaging reconstruction result is defined as , the single-pixel imaging reconstruction result satisfies: ， ； wherein, represents a processing procedure of the untrained convolutional neural network optimized according to the loss value, represents an L1 norm calculation.

Images