CN114219843B - Method and system for constructing terahertz spectrum image reconstruction model and application - Google Patents
Method and system for constructing terahertz spectrum image reconstruction model and application Download PDFInfo
- Publication number
- CN114219843B CN114219843B CN202111546891.7A CN202111546891A CN114219843B CN 114219843 B CN114219843 B CN 114219843B CN 202111546891 A CN202111546891 A CN 202111546891A CN 114219843 B CN114219843 B CN 114219843B
- Authority
- CN
- China
- Prior art keywords
- image
- terahertz
- reconstruction
- model
- reconstructed
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000001228 spectrum Methods 0.000 title claims abstract description 53
- 238000000034 method Methods 0.000 title claims abstract description 36
- 238000012549 training Methods 0.000 claims abstract description 36
- 238000005070 sampling Methods 0.000 claims abstract description 33
- 238000013507 mapping Methods 0.000 claims abstract description 29
- 238000007781 pre-processing Methods 0.000 claims abstract description 10
- 230000003595 spectral effect Effects 0.000 claims description 28
- 239000011159 matrix material Substances 0.000 claims description 19
- 238000005259 measurement Methods 0.000 claims description 10
- 241000209140 Triticum Species 0.000 claims description 9
- 235000021307 Triticum Nutrition 0.000 claims description 9
- 230000006870 function Effects 0.000 claims description 7
- 240000008042 Zea mays Species 0.000 claims description 4
- 235000005824 Zea mays ssp. parviglumis Nutrition 0.000 claims description 4
- 235000002017 Zea mays subsp mays Nutrition 0.000 claims description 4
- 235000005822 corn Nutrition 0.000 claims description 4
- 230000004913 activation Effects 0.000 claims description 3
- 238000004590 computer program Methods 0.000 claims description 3
- 230000007246 mechanism Effects 0.000 claims description 3
- 235000013339 cereals Nutrition 0.000 claims description 2
- 230000000694 effects Effects 0.000 description 6
- 238000012545 processing Methods 0.000 description 5
- 238000003384 imaging method Methods 0.000 description 4
- 230000008901 benefit Effects 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 238000002474 experimental method Methods 0.000 description 3
- 230000008569 process Effects 0.000 description 3
- 230000000903 blocking effect Effects 0.000 description 2
- 238000013527 convolutional neural network Methods 0.000 description 2
- 238000002372 labelling Methods 0.000 description 2
- 239000000463 material Substances 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000013139 quantization Methods 0.000 description 2
- 239000000126 substance Substances 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 238000000862 absorption spectrum Methods 0.000 description 1
- 230000003044 adaptive effect Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 239000000919 ceramic Substances 0.000 description 1
- 238000000701 chemical imaging Methods 0.000 description 1
- 239000003153 chemical reaction reagent Substances 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007850 degeneration Effects 0.000 description 1
- 230000005670 electromagnetic radiation Effects 0.000 description 1
- 239000004744 fabric Substances 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 230000001537 neural effect Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000002601 radiography Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 238000004611 spectroscopical analysis Methods 0.000 description 1
- 238000001328 terahertz time-domain spectroscopy Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 239000002023 wood Substances 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/50—Depth or shape recovery
- G06T7/55—Depth or shape recovery from multiple images
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/70—Denoising; Smoothing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20021—Dividing image into blocks, subimages or windows
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Investigating Or Analysing Materials By Optical Means (AREA)
Abstract
The invention discloses a method for constructing a terahertz spectrum image reconstruction model, which comprises the following steps: s1, acquiring a terahertz spectrum image; s2, preprocessing the terahertz spectrum image; s3, performing undersampling observation on the preprocessed image; s4, sequentially inputting the undersampled images into a linear mapping network and a residual error network for network training to obtain a reconstructed model; s5, judging whether the number of times of training of the reconstructed model is not less than the preset number of times of training, and if so, outputting the reconstructed model; otherwise, entering S6; and S6, repeating the steps S4 to S5, and adding 1 to the training times each time. Aiming at terahertz spectrum images, the method aims at DR2On the basis of Net, a self-adaptive sampling step (preprocessing of a terahertz spectrogram and undersampling observation of a preprocessed image) is added, a residual error network thought is introduced, the image reconstruction quality is greatly improved, the image reconstruction time is shortened, and the image reconstruction efficiency of the terahertz system is improved to a certain extent.
Description
Technical Field
The invention relates to the technical field of image processing. More specifically, the invention relates to a method, a system and an application for constructing a terahertz spectrum image reconstruction model.
Background
Terahertz (THz) electromagnetic radiation, also called THz wave, with a wavelength of 0.03-3 mm and an emitted frequency of 0.1-10 THz, analyzes and studies the components, the structure and the interaction relationship of substances by utilizing different characteristic absorption spectra of the substances to the THz frequency band. Terahertz Imaging Technology (Terahertz Spectroscopy) was developed based on Terahertz time-domain Spectroscopy. THz wave has received increasing attention in the imaging field because it can penetrate through non-polar materials such as cloth, wood, ceramic, etc., and can image the inside of an object. The terahertz spectral image not only contains geometric information of a target, but also has complete information of the THz impulse response intensity, phase, time and the like of the target, the obtained sample image has the characteristic of map integration, and spatial image information and abundant spectral information can be simultaneously obtained.
Therefore, how to improve the reconstruction speed of the terahertz image and ensure the image quality is a research problem to be solved urgently.
Disclosure of Invention
It is an object of the present invention to address at least the above problems and to provide at least the advantages described hereinafter.
The invention also aims to provide a method and a system for constructing a terahertz spectral image reconstruction model, which solve the problems of long image acquisition time, large data redundancy and the like of a terahertz time-domain spectral imaging system and the technical problem of reduced imaging reconstruction quality caused by reduced sampling rate.
To achieve these objects and other advantages in accordance with the present invention, there is provided a method of constructing a terahertz spectral image reconstruction model, including:
s1, acquiring a terahertz spectrum image;
s2, preprocessing the terahertz spectrum image to obtain a preprocessed image;
s3, performing undersampling observation on the preprocessed image through a Gaussian random sampling matrix to obtain an undersampled image;
s4, sequentially inputting the undersampled images into a linear mapping network and a residual error network for training to obtain a reconstructed model;
s5, judging whether the number of times of training of the reconstructed model is not less than the preset number of times of training, and if so, outputting the reconstructed model; otherwise, entering S6;
and S6, repeating the S4-S5, and adding 1 to the training times each time.
Preferably, the method for constructing the terahertz spectral image reconstruction model includes the following steps of S2: dividing the terahertz spectrum image into a plurality of plane spectrum images, and dividing each plane spectrum image into a plurality of n multiplied by n non-overlapping image blocks to obtain a preprocessed image.
Preferably, the method for constructing the terahertz spectral image reconstruction model includes the following steps of S3: undersampling the preprocessed image by adopting a Gaussian random sampling matrix to obtain an undersampled image signal yiAs a compressed perceptual measurement, the number of samples a = N × b, where b is the sampling rate and N is the total number of pixels per image block.
Preferably, in the method for constructing a terahertz spectral image reconstruction model, step S4 is specifically:
s4.1, undersampled image signal yiInputting the data into a linear mapping network together, and performing initial reconstruction on the undersampled image to obtain an initial reconstructed image signal xi；
Wherein the model of the linear mapping network is:
xi＝f(yi,{W}) (1)
in the formula (1), W is ∈ Rm×nIs a mapping matrix;
the loss function of a linear mapping network is:
x in the formula (2)iIs the ith image block, yiIs the measured value of the image block, { WfUsing a back propagation mechanism, f (y)iAnd { W }) is the ith reconstructed image output by the linear mapping network, and S is the total number of image blocks in the training set;
x is in the formula (3) and belongs to { X ∈ [)1,X2,…XN}，Y∈{Y1,Y2,…YN}；
S4.2, initially reconstructing an image signal xiInputting the image into a residual error network, and carrying out precision reconstruction on the image to obtain an n multiplied by n block reconstructed image, namely a reconstructed model;
the residual error network structure comprises 6 residual error learning blocks, each learning block comprises 3 convolutional layers, the first convolutional layer uses an 11 multiplied by 11 kernel, the second convolutional layer uses a 1 multiplied by 1 kernel, and the third convolutional layer uses a 7 multiplied by 7 kernel; each convolutional layer selects PReLU as an activation function,
a in formula (4)iAre parameters.
Preferably, in the method for constructing the terahertz spectral image reconstruction model, in step S4.2, the BM3D filter is used to denoise the n × n blocked reconstructed image, so as to obtain a reconstructed image with high resolution, that is, the reconstructed model.
The invention also provides a system for constructing the terahertz spectrum image reconstruction model, which comprises the following steps:
an acquisition module for acquiring a terahertz spectral image;
the self-adaptive sampling module is used for preprocessing the terahertz spectrum image and then performing undersampling observation on the preprocessed image through a Gaussian random sampling matrix to obtain an undersampled image;
and the reconstruction module is used for sequentially inputting the undersampled images into the linear mapping network and the residual error network for network training to obtain a reconstruction model.
The invention provides application of a terahertz spectrum image reconstruction model, which is applied to reconstruction of a terahertz spectrum image of wheat seeds or corn grains.
The present invention also provides an electronic device, comprising: the system comprises at least one processor and a memory communicatively connected with the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to cause the at least one processor to execute the method.
The invention also provides a storage medium on which a computer program is stored which, when executed by a processor, implements the method described above.
The invention at least comprises the following beneficial effects: aiming at terahertz spectrum images, the method aims at DR2On the basis of Net, a self-adaptive sampling step (preprocessing of the terahertz spectrogram and under-sampling observation of the preprocessed image) is added, a residual network idea is introduced, the image reconstruction quality is greatly improved, the image reconstruction time is shortened, and the image reconstruction efficiency of the terahertz system is improved to a certain extent.
Additional advantages, objects, and features of the invention will be set forth in part in the description which follows and in part will become apparent to those having ordinary skill in the art upon examination of the following or may be learned from practice of the invention.
Drawings
FIG. 1 is a schematic structural diagram of a terahertz spectral image reconstruction model according to the present invention;
FIG. 2 is a flow chart of the step reconstruction model training of the present invention;
fig. 3 is a diagram illustrating the reconstruction effect of the present invention.
Detailed Description
The present invention is further described in detail below with reference to the drawings and examples so that those skilled in the art can practice the invention with reference to the description.
It will be understood that terms such as "having," "including," and "comprising," as used herein, do not preclude the presence or addition of one or more other elements or groups thereof.
It should be noted that the experimental methods described in the following embodiments are all conventional methods unless otherwise specified, and the reagents and materials described therein are commercially available unless otherwise specified.
In the description of the present invention, the terms "lateral", "longitudinal", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate orientations or positional relationships based on the orientations or positional relationships shown in the drawings, are only for convenience in describing the present invention and simplifying the description, and do not indicate or imply that the device or element being referred to must have a particular orientation, be constructed and operated in a particular orientation, and thus, should not be construed as limiting the present invention.
The invention provides a method for constructing a terahertz spectral image reconstruction model, which comprises the following steps:
s1, acquiring a terahertz spectrum image; labeling the obtained terahertz spectrum image;
s2, preprocessing the terahertz spectrum image to obtain a preprocessed image;
s3, performing undersampling observation on the preprocessed image through a Gaussian random sampling matrix to obtain an undersampled image;
s4, sequentially inputting the undersampled images into a linear mapping network and a residual error network for network training to obtain a reconstructed model;
s5, judging whether the number of times of training of the reconstructed model is not less than the preset number of times of training, and if so, outputting the reconstructed model; otherwise, entering S6;
and S6, repeating the steps S4 to S5, and adding 1 to the training times each time.
In a traditional compressed sensing image reconstruction algorithm, the image needs to be subjected to blocking processing to realize special processing of complex information blocks, and reconstruction accuracy can be improved while complexity of image reconstruction is reduced. In order to solve the application bottleneck of the traditional compressed sensing reconstruction algorithm, reduce the calculated amount in the reconstruction stage and improve the reconstruction precision, a plurality of scholars propose to adopt a convolutional neural network to replace the optimization process and accelerate the reconstruction stage.
ReconNet is a non-iterative, fast block compressed perceptual reconstruction algorithm developed by Kulkarni et al. ReconNet based convolutional neural networkTo realize non-iterative compressed sensing image reconstruction, which is a representative work of applying convolutional neural network to reconstruct low-resolution mixed image measured by gaussian random matrix. The original image is divided into non-overlapping sub-blocks for reconstruction, so that not only can complex information blocks be processed, but also the purposes of reducing the reconstruction complexity and improving the reconstruction quality can be achieved. DR (digital radiography)2Net is an improvement over ReconNet, and the idea is also applied that divides the image into non-overlapping blocks, reconstructs each block by feeding the corresponding compressed perceptual measurements into its network, and combines the reconstructed blocks to form the entire reconstructed image. However, the reconstructed image has block artifacts, so the reconstructed image is input into a BM3D denoiser to remove the blocking artifacts and obtain a final output image.
Aiming at terahertz spectrum images, the method aims at DR2On the basis of Net, a self-adaptive sampling step (preprocessing of a terahertz spectrogram and undersampling observation of a preprocessed image) is added, a residual error network thought is introduced, the image reconstruction quality is greatly improved, the image reconstruction time is shortened, and the image reconstruction efficiency of the terahertz system is improved to a certain extent.
In another technical scheme, the method for constructing the terahertz spectrum image reconstruction model specifically comprises the following steps of S2: dividing the terahertz spectrum image into a plurality of plane spectrum images, and dividing each plane spectrum image into a plurality of n multiplied by n non-overlapping subblocks to obtain a preprocessed image.
In another technical scheme, the method for constructing the terahertz spectrum image reconstruction model specifically comprises the following steps of S3: adopting a Gaussian random sampling matrix to carry out undersampling on the preprocessed image to obtain an undersampled image signal yiAs a compressed perceptual measurement value, the number of sampling times a = N × b, where b is a sampling rate, N is a total number of pixels per image block, and N =1089.
The method comprises the steps that a group of U multiplied by V multiplied by Y terahertz spectral image data is provided, each 1 multiplied by Y spectral image data is a complete THz frequency domain or time domain spectrum, the images are divided into Y planar THz spectral images according to the frequency domain, then the images are divided into 33 multiplied by 33 non-overlapping sub-blocks, a Gaussian random measurement matrix is adopted to observe the sub-blocks, and an M-dimensional vector is obtained and serves as the input of a linear mapping network and is recorded as phi X. Where Φ is an observation matrix of dimension M × H, M is measurement data, and X is a vectorized image block. The network divides the image into 33 x 33 non-overlapping sub-blocks for reconstruction.
In another technical scheme, the method for constructing the terahertz spectrum image reconstruction model specifically comprises the following steps of S4:
s4.1, undersampled image signal yiInputting the data into a linear mapping network, and performing initial reconstruction on the undersampled image to obtain an initial reconstruction image signal xi；
Wherein the model of the linear mapping network is:
xi＝f(yi,{W}) (1)
the formula (1) is an over-determined equation, W belongs to Rm×nIs a mapping matrix;
the loss function (MSE) for a linear mapping network is:
wherein x isiIs the ith image block, yiIs the measured value of the image block, { WfUsing a back propagation mechanism, f (y)iAnd W) is the i-th reconstructed image output by the linear mapping network, and S is the total number of image blocks in the training set. MSE is the squared and summed average of the difference between the real and predicted values. Since MSE is in the form of a square, it is easy to derive, and the gradient of the loss of MSE increases as the loss increases, and decreases as the loss approaches 0.
Suppose a mapping matrix WfSo that
There is a minimum error. y isi∈RmRepresenting a compressed perceptual measurement, xi∈RnRepresenting its corresponding source signal, the training data comprising n training samples may then be represented as { (y)1,x1),(y2,x2),…(yN,xN) }. From these training data, W can be obtained by solving the following equationf：
X is in the formula (3) and belongs to { X ∈ [)1,X2,…XN}，Y∈{Y1,Y2,…YN}；
Equation (3) provides a linear mapping function that can be modeled efficiently in deep learning through fully connected layers. Therefore, the optimal mapping matrix W corresponding to the minimum equation can be obtained on all training samplesf；
S4.2, initially reconstructing an image signal xiInputting the image into a residual error network, and carrying out precision reconstruction on the image to obtain an n multiplied by n partitioned reconstructed image, namely a reconstructed model; the residual error network can improve the reconstruction precision of the image and obtain a high-resolution output;
the residual error network structure comprises 6 residual error learning blocks, each learning block comprises 3 convolution layers, the first convolution layer uses an 11 multiplied by 11 kernel, the second convolution layer uses a 1 multiplied by 1 kernel, and the third convolution layer uses a 7 multiplied by 7 kernel; each convolution layer selects PReLU (parametric Rectified Linear Unit) as an activation function, and the specific expression is as follows:
a in formula (4)iIs a parameter, in particular a coefficient controlling the slope of the negative part. If a isi0, the degeneration is Relu; if a isiIs a very small fixed value, then PRelu degenerates to Leaky Relu (LRelu). In this application aiThe initialization was 0.2.
In another technical scheme, in the method for constructing the terahertz spectrum image reconstruction model, in step S4.2, a BM3D filter is used to denoise an nxn blocked reconstructed image, so as to obtain a reconstructed image with high resolution, that is, the reconstructed model. And after passing through a residual error network, 33 × 33 block reconstructed images are output, a block effect generated by quantization errors of block quantization of the intermediate reconstructed image is removed through a BM3D filter, and a final high-resolution reconstructed image is output.
The invention also provides a system for constructing the terahertz spectrum image reconstruction model, which comprises the following steps:
an acquisition module for acquiring a terahertz spectral image;
the self-adaptive sampling module is used for preprocessing the terahertz spectrum image and then performing undersampling observation on the preprocessed image through a Gaussian random sampling matrix to obtain an undersampled image;
and the reconstruction module is used for sequentially inputting the undersampled images into the linear mapping network and the residual error network for network training to obtain a reconstruction model. The reconstruction module comprises a linear mapping network and a residual error network;
and the denoising module is used for carrying out denoising processing on the reconstructed model.
As shown in fig. 1, the terahertz spectrum image reconstruction model structure of the present invention includes an image acquisition module, an adaptive sampling module, a linear mapping network, a residual error network, and a denoising module.
The invention also provides an application of the terahertz spectrum image reconstruction model, and the terahertz spectrum image reconstruction model is applied to reconstruction of the terahertz spectrum image of the wheat seeds or the corn seeds.
The method specifically comprises the following steps:
in order to verify the performance of the model proposed by the inventor, the THz instrument used in the experiment is a Z3 series THz time domain spectrometer of Zomega company, the effective spectral range is 0.1-3.5 THz, the peak dynamic range is more than 1000 (70 dB), and the system signal-to-noise ratio is more than 3000. The wheat seeds are used as an experimental object, a sample is placed on a mobile platform in the THz-TDS system, reflection imaging measurement is carried out, the maximum scanning area of the system is 50mm multiplied by 50mm, and the spatial resolution is set to be 0.1mm. After THz spectral image data of a sample are obtained, image reconstruction is carried out by using the terahertz spectral image reconstruction model constructed by the invention. In order to meet daily requirements, all experimental processes are carried out in a normal-temperature normal-humidity environment. The image reconstruction includes the steps of:
the method comprises the following steps: extracting THz spectrums from all pixel points in a THz image of the wheat seeds, and averaging at all frequency points to obtain an average spectrum signal, wherein each spectrum comprises 500 points;
step two: as shown in fig. 2, the reconstruction model trains:
s1, obtaining a terahertz spectrum image training set: selecting 12 wheat seeds and corn kernels respectively, using the same sample and 500 THz spectral images in different frequency domains as a group, and using 12000 THz spectral images in total of twelve groups as a training set;
s2, entering a self-adaptive sampling module for processing: labeling the spectral images in the training set, and performing dimension reduction on the spectral images in the training set, dividing the spectral images into a plurality of N × N (33 × 33) of image blocks which are not overlapped with each other to obtain a preprocessed image, wherein the total pixel number of each image block is N =1089; performing undersampling on the preprocessed image signals by adopting a Gaussian random sampling matrix phi to obtain an undersampled image, wherein four undersampled images are obtained by sequentially performing four undersampling on the preprocessed image at 4 different sampling rates b (0.8,0.5,0.25 and 0.1), and the measurement times A (A = N multiplied by b) of the four undersampled images through the Gaussian matrix are 871, 546, 272 and 109 respectively;
s3, sequentially inputting the four undersampled images into a linear mapping network and a residual error network for network model training to obtain four image reconstruction models (ATResCS), wherein the number of times of model training is not less than the preset number of times of training, the trained image reconstruction models are subjected to denoising treatment, specifically, a BM3D filter is adopted to denoise n multiplied by n partitioned reconstructed images to obtain high-resolution reconstructed images, namely the image reconstruction models;
step three, based on the four sampling rates in the step two, adopting the training set in the step two to sequentially carry out comparison and Matching on OMP (organic Matching Pursuit) IRLS (Iterative weighted Least square), IHT (Iterative Hard threshold) DR (Iterative weighted threshold)2Training by using a Net to respectively obtain four contrast image reconstruction model groups, namely a group A, a group B, a group C and a group D, wherein all the four contrast image reconstruction model groupsFour contrast image reconstruction models are included; parameters (including iteration times) of the four algorithms are set as default values, and due to the fact that block effects can be generated in reconstruction, a BM3D denoising device is adopted to denoise a reconstructed image model; OMP (organic Matching Pursuit), IRLS (Iterative weighted Least Squares), IHT (Iterative Hard Thresholding) are three traditional compressed sensing image reconstruction algorithms;
and step four, selecting the frequency domain spectrum at the strongest signal (with the most obvious characteristics) in the step one, sequentially inputting the frequency domain spectrum into four image reconstruction models, four A group comparison image reconstruction models, four B group comparison image reconstruction models, four C group comparison image reconstruction models and four D group comparison image reconstruction models, and finally reconstructing to obtain the effect diagram shown in the figure 3.
In order to reduce errors caused by random measurement matrixes, PSNR, SSIM and reconstruction time of five (the application and four comparison groups) reconstruction methods are averaged for 100 test results, and the effect of reconstructing wheat seeds is shown in figure 3 when sampling rates are respectively 0.8,0.5,0.25 and 0.1 by adopting five algorithms.
It can be seen from fig. 3 that, when the sampling rate is 0.8, the original image can be clearly reconstructed by the five algorithms, and when the sampling rate is lower than 0.5, the image reconstructed by using the OMP algorithm is severely distorted and is basically submerged by noise, and at this time, IRLS, IHT, DR2-Net and ATResCS still have high reconstruction quality. When the sampling rate is 0.1, the image reconstructed by the IHT algorithm is basically submerged by noise, the characteristics of the image reconstructed by the IRLS algorithm are also blurred, and the characteristics of the DR2-Net and ATResCS reconstructed images are relatively clear. Experiments prove that the reconstruction effect of the ATResCS reconstruction network provided by the application exceeds that of the traditional reconstruction algorithm, and the image reconstruction quality under the low sampling rate is greatly improved.
Compared with three traditional compressed sensing methods and DR2-Net, the ATResCS not only improves the reconstruction quality of the image to a certain extent, but also improves the reconstruction speed. The THz spectrum image reconstruction time comparison of the wheat seeds at different sampling rates is shown in table 1. From table 1, the reconstruction of the ATResCS reconstruction network is lower than the reconstruction time required by other reconstruction algorithms, and the image reconstruction efficiency of the terahertz system is improved to a certain extent.
TABLE 1 comparison of THz spectrum image reconstruction time of wheat seeds with different sampling rates
The present invention also provides an electronic device, comprising: the system comprises at least one processor and a memory communicatively connected with the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to cause the at least one processor to execute the method.
The invention also provides a storage medium on which a computer program is stored which, when executed by a processor, carries out the method described above.
The number of apparatuses and the scale of the process described herein are intended to simplify the description of the present invention. Applications, modifications and variations of the present invention will be apparent to those skilled in the art.
While embodiments of the invention have been described above, it is not limited to the applications set forth in the description and the embodiments, which are fully applicable in various fields of endeavor to which the invention pertains, and further modifications may readily be made by those skilled in the art, it being understood that the invention is not limited to the details shown and described herein without departing from the general concept defined by the appended claims and their equivalents.
Claims (8)
1. The method for constructing the terahertz spectrum image reconstruction model is characterized by comprising the following steps of:
s1, acquiring a terahertz spectrum image;
s2, preprocessing the terahertz spectrum image to obtain a preprocessed image;
s3, carrying out undersampling observation on the preprocessed image through a Gaussian random sampling matrix to obtain an undersampled image;
s4, sequentially inputting the undersampled images into a linear mapping network and a residual error network for training to obtain a reconstructed model;
s5, judging whether the number of times of training of the reconstructed model is not less than the preset number of times of training, and if so, outputting the reconstructed model; otherwise, entering S6;
s6, repeating S4-S5, and adding 1 to the training times each time;
step S4 specifically includes:
s4.1, undersampled image signal yiInputting the data into a linear mapping network together, and performing initial reconstruction on the undersampled image to obtain an initial reconstructed image signal xi；
Wherein the model of the linear mapping network is:
xi＝f(yi,{W}) (1)
in the formula (1), W is ∈ Rm×nIs a mapping matrix;
the loss function of a linear mapping network is:
x in the formula (2)iIs the ith image block, yiIs the measured value of the image block, { WfUsing a back propagation mechanism, f (y)iAnd { W }) is the i-th reconstructed image output by the linear mapping network, and S is the total number of image blocks in the training set;
x is in the formula (3) and belongs to { X ∈ [)1,X2,…XN}，Y∈{Y1,Y2,…YN}；
S4.2, initially reconstructing an image signal xiInputting the image into a residual error network, and carrying out precision reconstruction on the image to obtain an n multiplied by n block reconstructed image, namely a reconstructed model;
the residual error network structure comprises 6 residual error learning blocks, each learning block comprises 3 convolutional layers, the first convolutional layer uses an 11 multiplied by 11 kernel, the second convolutional layer uses a 1 multiplied by 1 kernel, and the third convolutional layer uses a 7 multiplied by 7 kernel; each convolutional layer selects PReLU as an activation function,
a in formula (4)iAre parameters.
2. The method for constructing the terahertz spectral image reconstruction model according to claim 1, wherein the step S2 specifically comprises: dividing the terahertz spectrum image into a plurality of plane spectrum images, and dividing each plane spectrum image into a plurality of n multiplied by n non-overlapping image blocks to obtain a preprocessed image.
3. The method for constructing the terahertz spectral image reconstruction model according to claim 2, wherein the step S3 specifically comprises: undersampling the preprocessed image by adopting a Gaussian random sampling matrix to obtain an undersampled image signal yiAs a compressed perceptual measurement, the number of samples a = N × b, where b is the sampling rate and N is the total number of pixels per image block.
4. The method for constructing the terahertz spectral image reconstruction model as claimed in claim 3, wherein in step S4.2, the BM3D filter is adopted to denoise the n × n blocked reconstructed image, so as to obtain a reconstructed image with high resolution, namely the reconstructed model.
5. The system for constructing the terahertz spectral image reconstruction model is characterized by comprising:
an acquisition module for acquiring a terahertz spectral image;
the self-adaptive sampling module is used for preprocessing the terahertz spectrum image and then performing undersampling observation on the preprocessed image through a Gaussian random sampling matrix to obtain an undersampled image;
and the reconstruction module is used for sequentially inputting the undersampled images into the linear mapping network and the residual error network for network training to obtain a reconstruction model.
6. The application of the terahertz spectrum image reconstruction model is characterized in that the terahertz spectrum image reconstruction model of any one of claims 1 to 4 is applied to reconstruction of terahertz spectrum images of wheat seeds or corn grains.
7. An electronic device, comprising: at least one processor, and a memory communicatively coupled to the at least one processor, wherein the memory stores instructions executable by the at least one processor to cause the at least one processor to perform the method of any of claims 1-4.
8. Storage medium on which a computer program is stored which, when being executed by a processor, carries out the method of any one of claims 1-4.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111546891.7A CN114219843B (en) | 2021-12-16 | 2021-12-16 | Method and system for constructing terahertz spectrum image reconstruction model and application |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111546891.7A CN114219843B (en) | 2021-12-16 | 2021-12-16 | Method and system for constructing terahertz spectrum image reconstruction model and application |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114219843A CN114219843A (en) | 2022-03-22 |
| CN114219843B true CN114219843B (en) | 2022-11-01 |
Family
ID=80703306
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111546891.7A Active CN114219843B (en) | 2021-12-16 | 2021-12-16 | Method and system for constructing terahertz spectrum image reconstruction model and application |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114219843B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114894740A (en) * | 2022-04-06 | 2022-08-12 | 深圳先进技术研究院 | Terahertz single-pixel imaging method and system |
| CN115953334A (en) * | 2022-12-16 | 2023-04-11 | 哈尔滨工业大学 | Fourier single-pixel imaging reconstruction method based on time domain and frequency domain coordination |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111951344A (en) * | 2020-08-09 | 2020-11-17 | 昆明理工大学 | Magnetic resonance image reconstruction method based on cascade parallel convolution network |
| CN112435162A (en) * | 2020-11-13 | 2021-03-02 | 中国科学院沈阳自动化研究所 | Terahertz image super-resolution reconstruction method based on complex field neural network |
Family Cites Families (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106441575B (en) * | 2016-09-18 | 2017-11-21 | 河南工业大学 | A kind of sparse imaging method of terahertz time-domain spectroscopy |
| US11327137B2 (en) * | 2017-06-06 | 2022-05-10 | Shenzhen Institutes Of Advanced Technology | One-dimensional partial Fourier parallel magnetic resonance imaging method based on deep convolutional network |
| US10699163B1 (en) * | 2017-08-18 | 2020-06-30 | Massachusetts Institute Of Technology | Methods and apparatus for classification |
| CN110599558B (en) * | 2018-06-12 | 2021-06-29 | 清华大学 | Data reconstruction method, device and equipment for terahertz spectral imaging and storage medium |
| CN110378980B (en) * | 2019-07-16 | 2020-07-03 | 厦门大学 | Multichannel magnetic resonance image reconstruction method based on deep learning |
| CN112525851B (en) * | 2020-12-10 | 2024-08-06 | 深圳先进技术研究院 | Terahertz single-pixel imaging method and system thereof |
| CN113379598B (en) * | 2021-05-20 | 2023-07-14 | 山东省科学院自动化研究所 | Terahertz image reconstruction method and system based on residual channel attention network |
-
2021
- 2021-12-16 CN CN202111546891.7A patent/CN114219843B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111951344A (en) * | 2020-08-09 | 2020-11-17 | 昆明理工大学 | Magnetic resonance image reconstruction method based on cascade parallel convolution network |
| CN112435162A (en) * | 2020-11-13 | 2021-03-02 | 中国科学院沈阳自动化研究所 | Terahertz image super-resolution reconstruction method based on complex field neural network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114219843A (en) | 2022-03-22 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN114219843B (en) | Method and system for constructing terahertz spectrum image reconstruction model and application | |
| CN110599420B (en) | CT image block reconstruction method and system based on deep learning | |
| US10573031B2 (en) | Magnetic resonance image reconstruction with deep reinforcement learning | |
| Kang et al. | A deep convolutional neural network using directional wavelets for low‐dose X‐ray CT reconstruction | |
| CN112205993B (en) | System and method for processing data acquired using multi-energy computed tomography imaging | |
| KR20190133728A (en) | Tomography Reconstruction Based on Deep Learning | |
| CN111492406A (en) | Image generation using machine learning | |
| CN110047113A (en) | Neural network training method and equipment, image processing method and equipment and storage medium | |
| CN111081354A (en) | System and method for denoising medical images through deep learning network | |
| CN107480619B (en) | Noise-reduction method and system based on EEMD and the Ground Penetrating Radar B-scan image for arranging entropy | |
| RU2626184C2 (en) | Method, device and system for reconstructing magnetic resonance image | |
| JP2014520295A (en) | Method for determining PIXON mapping in iterative image reconstruction | |
| CN104574456B (en) | A kind of super lack sampling K data imaging method of magnetic resonance based on figure regularization sparse coding | |
| CN112102428B (en) | CT cone beam scanning image reconstruction method, scanning system and storage medium | |
| CN104200436B (en) | Multispectral image reconstruction method based on dual-tree complex wavelet transformation | |
| CN111784581B (en) | SAR image super-resolution reconstruction method based on self-normalization generation countermeasure network | |
| CN112819949B (en) | Magnetic resonance fingerprint image reconstruction method based on structured low-rank matrix | |
| CN112578471B (en) | Clutter noise removing method for ground penetrating radar | |
| CN108230280A (en) | Image speckle noise minimizing technology based on tensor model and compressive sensing theory | |
| CN114758031A (en) | Method and device for reconstructing capacitive coupling electrical impedance tomography image | |
| Zhang et al. | Spectral CT image-domain material decomposition via sparsity residual prior and dictionary learning | |
| CN117195960A (en) | Atmospheric turbulence wave front detection method based on shack Hartmann UT conversion model | |
| Zhang et al. | Projection domain denoising method based on dictionary learning for low-dose CT image reconstruction | |
| CN117523095A (en) | Sparse angle THz-CT image reconstruction method based on deep learning | |
| CN105976412B (en) | A kind of CT image rebuilding methods of the low tube current intensity scan based on the sparse regularization of offline dictionary |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |