Disclosure of Invention
In order to solve the technical problem that a scattering medium affects image reconstruction, the invention provides a double-channel ghost imaging reconstruction method and system based on a human eye imitation mechanism.
Therefore, the double-channel ghost imaging reconstruction method based on the human eye imitation mechanism provided by the invention specifically comprises the following steps of:
s1, constructing a two-channel image reconstruction model;
s2, a training data set is manufactured, the two-channel image reconstruction model is trained, and the two-channel image reconstruction model is trained to be converged based on the training data set;
s3, generating a measurement matrix based on a human-eye-simulated strategy, loading the measurement matrix to a micro-mirror array, synchronously controlling the micro-mirror array and a photoelectric detector, and acquiring a light intensity sequence by the photoelectric detector to finish data acquisition of a target sample;
s4, correcting the acquired data by adopting a nonlinear light intensity correction method;
s5, inputting the corrected data into the two-channel image reconstruction model, and returning the corrected data into a two-dimensional image after passing through two channels to obtain a deep learning ghost imaging clear image.
Further, the two-channel image reconstruction model adopts a two-channel enhanced neural network, two branches of the two-channel enhanced neural network are a ghost imaging image branch and a light intensity residual branch, and data sources of the two-channel enhanced neural network are ghost imaging reconstructed images and light intensity characteristic sequences respectively.
Further, the ghost imaging reconstructed image is calculated by the following formula:
I(x)=<(S i -<S i >)(P i -<P i >)>
wherein the method comprises the steps of<·>Representing the arithmetic mean of all measured values, S i Representing the sequence of light intensities input by the network, P i Representing a random measurement pattern employed to acquire the intensity data.
Further, the ghost image branches acquire preliminary feature images from six convolution layers having feature image dimensions of 128×128×1, 128×128×32, 128×128×16, 128×128×32, 128×128×1, respectively.
Further, the light intensity residual branch comprises a full connection layer and a two-dimensional residual block and two convolution layers, the full connection layer being located between the measurement value sequence and the convolution layers.
Further, in the step S2, a specific image set is generated by adopting a human eye simulation strategy to simulate the light intensity, a monte carlo simulation mode is adopted to simulate the scattering process, a degradation data set with the scattering characteristic of the transmission path is introduced, and a data set to be trained is created.
Further, in the step S4, the nonlinear light intensity correction method specifically includes that when the optical system receives interference of the scattering medium, the measured value is disturbed, assuming that the attenuation coefficient of the previous L frames is ζ, and after L measurements, the attenuation coefficient of the scattering medium is mutated, the following formula may be obtained:
wherein y is a light intensity measurement value, M represents the total number of measurements, N represents the length of a one-dimensional signal, and dividing the measurement average values before and after dynamic change according to the characteristics of ghost imaging and compressed sensing algorithms can obtain the following formula:
multiplying data with a sequence number greater than L by a correction factor ζ according to the above formula L And/ζ, a new light intensity sequence is formed.
Therefore, the two-channel ghost imaging reconstruction system based on the human-simulated eye mechanism comprises a central processing unit, a memory and a data acquisition part, wherein a two-channel image reconstruction model and a program which can be run by the central processing unit are stored in the memory, and the program can realize the two-channel ghost imaging reconstruction method based on the human-simulated eye mechanism in the process of being run by the central processing unit.
Further, the data acquisition portion includes a micromirror array, a photodetector, an imaging lens, and a laser.
Therefore, the computer readable storage medium provided by the invention stores a two-channel image reconstruction model and a program which can be run by a central processing unit, and the program can realize the two-channel ghost imaging reconstruction method based on the human eye imitation mechanism in the process of being run by the central processing unit.
Compared with the prior art, the invention has the following beneficial effects:
the light intensity data is acquired through the human eye imitation mechanism and used for model training and input, the parameterization of the imaging process of the imaging optical system is facilitated, a nonlinear light intensity correction method is provided, and nonlinear influence of chaotic media is eliminated.
In some embodiments of the invention, there are also the following benefits:
1) The dual-channel image reconstruction model based on the residual block convolutional neural network is provided, and the comprehensive extraction capacity and regression capacity of one-dimensional light intensity and two-dimensional image characteristics are improved from the model design level;
2) In order to eliminate the effect of the scattering medium in the transmit path on the imaging, the scattering process is simulated in a Monte Carlo simulation and a data set to be trained is created.
Detailed Description
For a clearer understanding of technical features, objects, and effects of the present invention, a specific embodiment of the present invention will be described with reference to the accompanying drawings.
The dual-channel ghost imaging reconstruction system based on the human eye imitation mechanism comprises a central processing unit, a memory and a data acquisition part. As shown in fig. 1, the data acquisition section includes a micromirror array (DMD) 2, a Photodetector (PD) 6, an imaging lens 5, and a laser 1, 3 in the drawing representing a scattering medium, 4 representing an object, a laser light source is modulated using the micromirror array (DMD), wherein the DMD uses a digital voltage signal to control the micromirrors to perform mechanical movement, intensity modulation of the light source is achieved by controlling deflection angles of the micromirrors, and a state of each of the micromirrors is controlled by 0 or 1 in the loaded measurement matrix data. The memory stores a two-channel image reconstruction model and a program related to the two-channel ghost imaging reconstruction. The central processing unit mainly realizes two functions: 1) Manufacturing a training data set and training a two-channel image reconstruction model; 2) In the actual imaging process, the nonlinear light intensity correction method is used for correcting input data, the corrected data is input into the two-channel image reconstruction model, and the deep learning ghost imaging clear image is obtained through operation.
As shown in fig. 2, the two-channel ghost imaging reconstruction method based on the human eye imitation mechanism in the embodiment of the invention specifically includes the following steps:
s1, constructing a two-channel image reconstruction model, wherein the two-channel image reconstruction model adopts a two-channel enhanced neural network, as shown in FIG. 3, two branches of the neural network are a ghost imaging image branch and a light intensity residual branch, and data sources of the two branches are a ghost imaging reconstruction image (GI reconstruction image) and a light intensity characteristic sequence respectively. Wherein the GI reconstructed image is calculated by the following formula:
I(x)=<(S i -<S i >)(P i -<P i >)> (1)
wherein the method comprises the steps of<·>Representing the arithmetic mean of all measured values, S i Representing the sequence of light intensities input by the network, P i Representing a random measurement pattern employed to acquire the intensity data.
The ghost imaging image branches obtain preliminary feature images by six convolution layers, wherein the feature image dimensions are 128×128×1, 128×128×32, 128×128×16, 128×128×32 and 128×128×1 respectively, and 32 feature images with different 128×128 sizes can be obtained after the 128×128×32 convolution layers, namely the convolution layers contain 32 channels.
The light intensity residual branch comprises a full connection layer, a two-dimensional residual Block (Res 2Net Block) and two convolution layers, wherein the full connection layer is positioned between the measured value sequence and the convolution layers and is used for extracting features of one-dimensional data and arranging the features into a two-dimensional image. Each residual block comprises a residual structure, the head and the tail of which are respectively connected by a convolution layer, the convolution kernel of the front convolution layer is 5×5×32, namely 128×128×64 is output after a feature map with 128×128×1 dimensions is input, and the convolution kernel of the rear convolution layer is 5×5×1, namely 128×128×1 is output after a feature map with 128×128×32 dimensions is input. In RM, with ReLU as activation function, each residual structure is followed by a batch normalization layer (Batch Normalization, BN). In the residual structure, the 128×128×32 feature map is equally divided into four parts, each part corresponds to a convolution kernel of 5×5×4 except for the first part, and the result after the convolution of the ith layer is added to the (i+1) th layer as a combined input.
The two characteristic branches are combined in an addition mode, namely, the two characteristic graphs with 128 multiplied by 1 dimensions are added and then sent to a second residual error module for characteristic induction, and finally, the product of the output of the second residual error and the GI reconstructed image is used as a predicted image, and the cost function is defined by the mean square error between the product and the truth image for training. Because the GI reconstructed image contains noise characteristics of ghost images, the channel can realize image denoising, and fluctuation in the light intensity sequence contains the most original signal acquisition characteristics of ghost images, and the introduction of the fluctuation in the light intensity sequence can enhance the detail reconstruction of the image.
S2, a training data set is manufactured, a two-channel image reconstruction model is trained, the key of the two-channel image reconstruction model training is the consistency of the training set and the light intensity data input by an actual imaging scene, for example, the light intensity data to be input during imaging are acquired under a strong scattering environment, and the network training data set also has strong scattering characteristics, namely, the input light intensity signal is required to be scattered and degraded. In order to realize key sampling, a human-simulated eye acquisition mode is adopted for data acquisition during imaging, so that a human-simulated eye strategy is adopted during training set manufacturing. The data set theory of deep learning ghost imaging is to be collected and manufactured in the field, but the time cost is high, in order to manufacture the data set more economically and conveniently, a human eye imitation strategy is adopted to simulate the generation of light intensity of a specific image set, a Monte Carlo simulation mode is adopted to simulate the scattering process, a degradation data set with the scattering characteristic of a transmission path is introduced into the self-manufactured data set, a data set to be trained is created, the training data set comprises the characteristics of an area of interest of the human eye imitation and the scattering characteristics, and a double-channel image reconstruction model is trained to be converged (mean square error is smaller than 0.5) based on the training data set.
As shown in fig. 4, the static scattering medium was simulated, and θ was calculated using the following equation, respectively
0 ,
θ 0 =sin -1 {1-ξ 1 [1-cos(β E /2)]} (2)
Wherein, xi 1 And xi 2 Are all random numbers with values between 0 and 1, beta E The value is 0.
The motion of the photons can be simulated using the random numbers generated by the program, and the linear distance between each two consecutive collision sites of a single photon is calculated using the following formula:
wherein, xi 3 To take on a random number between 0 and 1, sigma a And sigma (sigma) s The sum represents the linear extinction coefficient of the particles.
By pairs of
Representing the relative transmission direction of photons after undergoing the ith scattering, the probability distribution density function is:
the probability of a photon experiencing a scattered backward direction departure angle is specified for a scattering phase function for a particular scattering type, expressed as a Rayleigh scattering phase function:
wherein v is an atmospheric model parameter, and only the parameter is needed to pass through in the specific calculation process
Solving for θ from probability density distribution
i And set to a random number between 0 and 1 to determine θ in reverse
i Is a value of (a). The direction vector of photons in the global cartesian coordinate system has been updated as:
after the nth scattering, the energy remainder of the photon is:
setting the energy threshold of photons to epsilon=10 -5 I.e. when the current energy value of a photon is below a threshold, it is considered a non-viable photon and cannot be received. The scattered photons and the energy carried by the scattered photons comprehensively represent the intensity distribution of the degenerated light field, and the light field and the target gray level are subjected to dot multiplication operation, so that simulated light intensity data can be finally obtained.
S3, generating a measurement matrix based on a human eye imitation strategy, loading the measurement matrix into the DMD for light intensity data acquisition, taking a reconstruction target with the resolution of 64 multiplied by 64 as an example, setting the size of the basic measurement matrix to 1024 rows and 4096 columns under the condition that the sampling rate is 25%, wherein the acquisition mode is as follows: firstly generating a Hadamard matrix with the size of 4096×4096, then taking the first 1024 rows of the matrix as a basic measurement matrix, respectively taking out each row of the basic measurement matrix, recombining each 4096-number-containing sequence into a 64×64 image according to the rows to form a matrix group M 1 The Hadamard matrix with 1024×1024 size is reorganized into 1024 matrices with 32×32 size according to rows in the above way, and then the matrix is fully expanded into matrix group M with 64×64 size by bilinear interpolation method 2 A region of interest R is artificially selected in a 64×64 matrix, the unselected region is NR, a region control matrix C is set, wherein the pixel value of the R-containing region in C is 1, otherwise, 0, and thus the human eye simulated measurement pattern group is obtained: m is M F =M 1· *C+M 2· * And (I-C), wherein I is an identity matrix, a matrix group containing 1024 measurement patterns is loaded into an upper computer of the DMD, the DMD and a Photoelectric Detector (PD) are synchronously controlled, the PD can acquire 1024 frames of light intensity sequences at the same frame rate, and the process completes the data acquisition of one target sample.
S4, because the scattering medium in the training set is uniform, the change caused by the mutation or the non-uniformity of the scattering medium is needed to be considered in the actual collection, so that the influence of the non-uniformity characteristic on the data is required to be weakened, the nonlinear light intensity correction method is used for correcting the input data, when the optical system receives the interference of the scattering medium, the measured value is disturbed, the attenuation coefficient of the former L frames is assumed to be xi, and after L times of measurement, the attenuation coefficient of the scattering medium is mutated, and the following formula can be obtained:
where y is the intensity measurement, M represents the total number of measurements and N represents the length of the one-dimensional signal. The nonlinear correction method for the light intensity value is shown as a formula (10), wherein xi is L And/ζ is an accurate correction factor. According to the characteristics of ghost imaging and compressed sensing algorithms, the average value measured before and after dynamic change can be divided to obtain the following expression:
the method corrects the light intensity data once and finds out the mutation point with the most serious influence on the light intensity by the dynamic change of the scattering medium. Calculated from (10)
Where l=1, 2,3 …, M. Find to make->
The value of L with the largest value is taken as the abrupt point L of light intensity correction
0 ζ at this time
L And/ζ is a correction factor. Correcting the light intensity sequence according to equations (9) and (10), i.e. multiplying the data with sequence number greater than L by correction factor xi
L And/ζ, a new light intensity sequence is formed.
S5, inputting the corrected data into a two-channel image reconstruction model, and returning the corrected data into a two-dimensional image after passing through two channels to obtain a deep learning ghost imaging clear image.
The laser range finders are adopted to measure the distance in Shenzhen urban areas to ensure that the scattering distance is about 50m for imaging, the experiment is carried out in a thick fog scene at certain night, and a prototype can be verified in an actual more extreme haze environment. The target uses standard test boards printed with black and white stripes and a three-dimensional shaped ornament.
As can be seen from the reconstruction results and the contrast quantification values, the method has a significant effect on the contrast improvement of black and white stripes, as shown in fig. 5A-5F. When the compression ratio is not more than 20, the reconstruction contrast ratio is kept above 200, and a reconstruction result with the contrast ratio of 62 can still be obtained under the compression ratio of 40.
In order to verify the advantages of the PC-GI network (named MC-PC-GI) of Monte Carlo simulation data, as shown in FIGS. 6A-6D, the PC-GI network based on the compressed sensing reconstruction method CSGI and the PC-GI network based on non-scattering simulation data are adopted for experimental comparison, the result of training the PC-GI network by adopting the non-scattering simulation data is slightly better than that of the CSGI method, the detail information is more sufficient, but the noise is more, in contrast, the MC-PC-GI method drives the network training under more ideal data, the regression result is more accurate, and the robustness is stronger under the strong scattering imaging environment.
The double-channel ghost imaging reconstruction method and system based on the human eye imitation mechanism provided by the invention have the following beneficial effects:
1) The method comprises the steps of providing a dual-channel image reconstruction model based on a residual block convolutional neural network, considering data training modes of different dimensions, improving comprehensive extraction capacity and regression capacity of one-dimensional light intensity and two-dimensional image characteristics from a model design level, acquiring light intensity data for model training and input through a human eye imitation mechanism, enhancing the light intensity data characteristics and overall image quality through key acquisition of a region of interest, facilitating realization of parameterization of an imaging optical system imaging process, and further improving image accuracy in a deep neural network;
2) In order to recover the inter-frame light intensity correlation under the dynamic chaotic scattering medium, a nonlinear light intensity correction method is provided, and a compensation factor is designed to eliminate nonlinear influence of the chaotic medium;
3) Because the actual light intensity data is collected in a strong scattering environment, the training set data also has strong scattering characteristics, namely the input light intensity signal is scattered and degraded, and in order to eliminate the influence of scattering medium in a transmitting path on imaging, a Monte Carlo simulation mode is adopted to simulate the scattering process and create a data set to be trained.
The foregoing disclosure is illustrative of the present invention and is not to be construed as limiting the scope of the invention, which is defined by the appended claims.