CN112581412A - Atomic force microscope image restoration method based on long-term and short-term memory network - Google Patents
Atomic force microscope image restoration method based on long-term and short-term memory network Download PDFInfo
- Publication number
- CN112581412A CN112581412A CN202011575092.8A CN202011575092A CN112581412A CN 112581412 A CN112581412 A CN 112581412A CN 202011575092 A CN202011575092 A CN 202011575092A CN 112581412 A CN112581412 A CN 112581412A
- Authority
- CN
- China
- Prior art keywords
- afm
- sample
- image
- training
- simulation
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 81
- 238000000089 atomic force micrograph Methods 0.000 title claims abstract description 35
- 230000007787 long-term memory Effects 0.000 title claims abstract description 6
- 230000006403 short-term memory Effects 0.000 title claims abstract description 6
- 238000012549 training Methods 0.000 claims abstract description 48
- 238000004088 simulation Methods 0.000 claims abstract description 28
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 25
- 230000006870 function Effects 0.000 claims abstract description 25
- 238000010606 normalization Methods 0.000 claims abstract description 18
- 238000012545 processing Methods 0.000 claims abstract description 14
- 238000005259 measurement Methods 0.000 claims abstract description 13
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims abstract description 11
- 238000005457 optimization Methods 0.000 claims abstract description 7
- 239000000523 sample Substances 0.000 claims description 82
- 230000008569 process Effects 0.000 claims description 39
- 230000015654 memory Effects 0.000 claims description 23
- 239000013598 vector Substances 0.000 claims description 19
- 239000011159 matrix material Substances 0.000 claims description 18
- 230000003044 adaptive effect Effects 0.000 claims description 6
- 229910052581 Si3N4 Inorganic materials 0.000 claims description 5
- HQVNEWCFYHHQES-UHFFFAOYSA-N silicon nitride Chemical compound N12[Si]34N5[Si]62N3[Si]51N64 HQVNEWCFYHHQES-UHFFFAOYSA-N 0.000 claims description 5
- 230000009467 reduction Effects 0.000 claims description 4
- IEDXPSOJFSVCKU-HOKPPMCLSA-N [4-[[(2S)-5-(carbamoylamino)-2-[[(2S)-2-[6-(2,5-dioxopyrrolidin-1-yl)hexanoylamino]-3-methylbutanoyl]amino]pentanoyl]amino]phenyl]methyl N-[(2S)-1-[[(2S)-1-[[(3R,4S,5S)-1-[(2S)-2-[(1R,2R)-3-[[(1S,2R)-1-hydroxy-1-phenylpropan-2-yl]amino]-1-methoxy-2-methyl-3-oxopropyl]pyrrolidin-1-yl]-3-methoxy-5-methyl-1-oxoheptan-4-yl]-methylamino]-3-methyl-1-oxobutan-2-yl]amino]-3-methyl-1-oxobutan-2-yl]-N-methylcarbamate Chemical compound CC[C@H](C)[C@@H]([C@@H](CC(=O)N1CCC[C@H]1[C@H](OC)[C@@H](C)C(=O)N[C@H](C)[C@@H](O)c1ccccc1)OC)N(C)C(=O)[C@@H](NC(=O)[C@H](C(C)C)N(C)C(=O)OCc1ccc(NC(=O)[C@H](CCCNC(N)=O)NC(=O)[C@@H](NC(=O)CCCCCN2C(=O)CCC2=O)C(C)C)cc1)C(C)C IEDXPSOJFSVCKU-HOKPPMCLSA-N 0.000 claims 2
- 238000013528 artificial neural network Methods 0.000 description 10
- 230000000694 effects Effects 0.000 description 8
- 238000012360 testing method Methods 0.000 description 6
- 230000008859 change Effects 0.000 description 5
- 230000004913 activation Effects 0.000 description 4
- 230000002596 correlated effect Effects 0.000 description 4
- 230000000875 corresponding effect Effects 0.000 description 4
- 238000010586 diagram Methods 0.000 description 4
- 238000003384 imaging method Methods 0.000 description 4
- 230000005540 biological transmission Effects 0.000 description 2
- 230000010339 dilation Effects 0.000 description 2
- 238000004880 explosion Methods 0.000 description 2
- 230000003993 interaction Effects 0.000 description 2
- 238000013507 mapping Methods 0.000 description 2
- 238000011084 recovery Methods 0.000 description 2
- 238000005299 abrasion Methods 0.000 description 1
- 230000009471 action Effects 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 238000011478 gradient descent method Methods 0.000 description 1
- 238000012804 iterative process Methods 0.000 description 1
- 210000001503 joint Anatomy 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 230000008439 repair process Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/20—Image enhancement or restoration using local operators
- G06T5/30—Erosion or dilatation, e.g. thinning
-
- 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
- G06N3/044—Recurrent networks, e.g. Hopfield 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- 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
- G06T2207/10—Image acquisition modality
- G06T2207/10056—Microscopic image
-
- 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
- G06T2207/20081—Training; Learning
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (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)
- Image Processing (AREA)
Abstract
The invention discloses an atomic force microscope image restoration method of a long-term and short-term memory network. The method comprises the steps of firstly, obtaining a simulation sample by using an expansion method in mathematical morphology, then, carrying out normalization operation on the simulation sample, inputting the simulation sample into an LSTM model for training, perfecting the LSTM model by adopting an improved MAE loss function and an Adam optimization algorithm, and then, carrying out regularization and inverse normalization processing on the training sample. And finally, inputting the actual AFM image into the trained model to obtain an AFM restoration image, and processing by using a self-adaptive algorithm with gradual interpolation, so that the position of the characteristic point in the sample is accurately obtained. The method has stability and better robustness, and can effectively recover the AFM image, thereby improving the measurement accuracy.
Description
Technical Field
The invention relates to the fixed value of standard instruments such as grids and steps in nano metering, in particular to an atomic force microscope image restoration method based on a long-short term memory (LSTM) network.
Background
Since an Atomic Force Microscope (AFM) performs imaging by using an interaction force between a minute probe and an object to be measured, it is widely used in the field of nano-science metrology. The AFM obtains relevant geometric quantity parameters through high-resolution imaging of the nano-metering standard instrument and calibrates the geometric quantity parameters, so that the quantity value transmission from the standard metering instrument to the working metering instrument is realized. In the scanning imaging process of the AFM, due to the influence of the needle tip, an image obtained by scanning is a result of the combined action of the atomic force probe and the sample instead of the real description of the appearance of the sample, and the phenomenon is the 'artifact effect' of the AFM. The phenomenon reduces the accuracy of the measurement result of the standard device, and affects the quantity value transmission and traceability of the nanometer scientific measurement.
From the viewpoint of improving the accuracy of the measuring instrument, the "artifact effect" can be reduced by using a probe with a smaller curvature radius, but such a probe tends to have poor durability and high cost, and the "artifact effect" cannot be eliminated in principle. Because the nanometer material has higher rigidity and hardness and is not easy to deform and wear, the AFM adopts contact measurement in the process of scanning a sample. Based on the measuring mode, the real surface of the sample can be obtained through a certain algorithm according to the imaging of the sample. However, this deconvolution process is non-deterministic and cannot restore an image with an accurate mathematical expression, so the process is highly demanding in algorithm selection.
At present, the main algorithms for restoring AFM images include a radial basis function neural network and multilayer perceptron-based restoration algorithm, a probe blind reconstruction algorithm based on mathematical morphology and a back propagation neural network algorithm. The probe blind reconstruction algorithm often depends excessively on AFM measurement images, so that the robustness of the reconstructed probe morphology is poor, and AFM image restoration is influenced. The radial basis function neural network and the back propagation neural network algorithm are limited by the upper depth limit of the neural network in the training process, and the problem of training failure caused by gradient attenuation or explosion is often accompanied.
Disclosure of Invention
The invention provides a method for restoring AFM images based on a long-short term memory (LSTM) recurrent neural network, aiming at the defects of the prior art.
The technical scheme adopted by the invention is that,
an atomic force microscope image restoration method based on a long-term and short-term memory network comprises the following steps:
step 1: obtaining a simulation sample by using an expansion method in mathematical morphology;
assuming that the expression of the probe cross-sectional profile shape when the tip is at the home position in the coordinate system is s (x), the expression of the sample cross-sectional profile shape is f (x). The position of the probe in the scan line is defined by the abscissa of the tip, assuming that the abscissa of the tip is x0The ordinate h of the tip is then indicated at x0The measured height value of the sample scanning point. At this time, assuming that the expression of the cross-sectional profile shape of the probe is t (x), the relationship between s (x) and t (x) satisfies the expression
t(x)=s(x)+h (1)
The relationship between f (x) and t (x) satisfies the expression
AFM contact measurement is equivalent to a dilation operation in mathematical morphology, the mathematical expression of which is
Thus, according to equation (3), the expression for h can be written as
By combining formula (1) and formula (2), formula (3) can be converted into
Since h and x satisfy a one-to-one mapping relationship, h can be regarded as a function of x, and h (x) is used for representing a height result change curve measured by the probe on a certain horizontal scanning line of the sample, namely results of each line of the simulation image. h (x) includes the relationship among the sample profile, the tip profile and the measurement image.
In the simulation process, a nano-grid standard device is selected as a tested sample, and a silicon nitride needle tip is selected as a scanning device. The simulated scanning process may acquire 2000 sets of sample images, with 1600 sets of images used as a training set and 400 sets of images used as a test set. The sample selection for the training set need not contain all probe types, but not all nanogrid types.
Step 2: carrying out normalization operation on the simulation sample image obtained in the step 1;
and decomposing the AFM image line by line to realize dimension reduction, so that the line vector becomes a training sample of the network.
Since the AFM simulation image is a 140 × 140 matrix, both the input data and the output result are 1 × 140 row vectors. In order to inhibit the phenomena of incapability of convergence and overfitting in the training process and further improve the stability and the performance of the neural network, normalization processing is adopted for input data, namely
In formula (6), x is a numerical value in the input matrix, xmaxAnd xminRespectively the maximum and minimum values, x, in the input matrix*The result is normalized for the input data.
And step 3: inputting the data obtained in the step (2) into an LSTM model for training;
the LSTM model comprises a forgetting gate FtInput gate ItOutput gate OtAnd candidate memory cellsAnd fourthly, the method comprises the following steps. The gate input of LSTM is the current time step input XtHidden state H with last time stept-1This is matched by the AFM scan line point gradient determined by the height values of adjacent points. The output is calculated by a full connection layer of a sigmoid activation function, and the expression is
To obtain candidate memory cells with larger gradient near 0 AFM scan line input and faster model convergence, tanh activation function is used, expressed as tan h
The LSTM can store AFM scanning line vector information with higher dimension through memory cells, so that the correlation degree between scanning points is obtained. For example, points at flat areas of the scan line are less correlated, while points at areas with greater variations in scan line height are more correlated. Setting a time step memory cell Ct-1Current time step memory cell CtAt the time of obtaining CtWhen passing through FtAnd ItAnd the Hadamard product of the matrix controls the flow of information between each point of the scanning line:
if FtIs always approximately 1 and ItThe element in (1) is always approximately 0, and past memory cells will always be saved by time and passed to the current time step. The process can store effective information in the AFM scanning line for a long time, solves the problem of gradient attenuation caused by the increase of scanning line vector dimensions, and improves the training reliability. Meanwhile, the dependency relationship among points with larger distance among scanning lines can be obtained, and the image restoration of a region with serious artifact effect is realized.
At the acquisition of CtThen, through OtTo control the slave HtFlow of information to the current time step hidden state:
Ht=Ot⊙tanh(Ct) (10)
for the nano-grid scan line, when OtWhen the element in (A) is approximately 0 or 0, CtThe information in the method is usually only the information of a flat area of the nano grid, is not the required characteristic information, and is generally reserved for the user; when O is presenttWhen the element in (B) is far from 0, then C is presenttThe characteristic information such as the position of a step and the like of the nano-grid is contained in the H-shaped magnetic field, and the H-shaped magnetic field is transmitted to the HtFor use by the output layer.
And 4, step 4: continuously improving an LSTM model by adopting an MSE loss function and an Adam optimization algorithm;
the recovery effect of AFM images was evaluated using Modified Mean Absolute Error (MMAE) as a loss function and using a 1-norm in the summation process. The loss function is expressed as
In equation (11), n represents the number of training samples, l represents the vector dimension of the network output,estimated value, y, representing each line of AFM restored imageiThe real values of the respective lines of the AFM restored image are shown.
The learning rate parameter adaptive formula of the Adam algorithm is
In the formula (12), β0Denotes the initial given hyper-parameter, βiRepresenting the hyperparameter after i iterations.
And 5: carrying out regularization and inverse normalization processing on the training samples;
for the output layer passing through the LSTM model, a Dropout regularization method is adopted, and some output characteristics of the layer are abandoned randomly in the training process, so that the overfitting phenomenon is reduced, and the model obtained through training is suitable for more types of AFM scanning sample images.
All sample data in the LSTM model training process are in the [0,1] interval, and the finally obtained restored image needs to be close to the actual sample, so that all data need to be subjected to inverse normalization operation before an output result is obtained.
Step 6: inputting the actual AFM image into the trained model to obtain an AFM restoration image;
and for the AFM restored image, processing by using an adaptive algorithm with gradual interpolation, so that the positions of the characteristic points in the sample can be more accurately obtained.
The invention has the beneficial effects that: after AFM simulation images are acquired by a simulation scanning sample, the AFM image deconvolution process is obtained through LSTM model training, and finally the model is used for restoring the actual AFM images. The method can solve the problems of cost consumption and image quality reduction caused by needle point abrasion in the process of obtaining AFM images in a large scale. The method has the advantages that the LSTM model is adopted for deconvolution operation, so that the repair of the position with serious artifact effect in the AFM image is facilitated, and meanwhile, the problem of gradient disappearance or explosion in the training process is solved.
Drawings
FIG. 1 is a schematic diagram of AFM simulation images acquired in the present invention.
FIG. 2 is a three-dimensional view of a simulated needle tip according to the method of the present invention.
FIG. 3 is a three-dimensional diagram of a simulated nano-grating according to the method of the present invention.
FIG. 4 is a three-dimensional graph of the results of the simulated nano-raster scan of the present invention.
FIG. 5 is an algorithmic flow chart of the method of the present invention.
FIG. 6 is a diagram of the LSTM model architecture of the method of the present invention.
FIG. 7 is a graph of loss value versus iteration number in neural network training for the method of the present invention.
FIG. 8 is a top view of a nano-grating in the test results of the method of the present invention.
Detailed Description
The method obtains the neural network training sample by simulating the process of scanning the nano-grid of the nano-metering standard instrument by the Matlab. And then, equating each line scanning height curve in the measurement image of the nano grid by AFM as a time sequence sample, inputting the time sequence sample into a constructed LSTM frame for training to obtain a prediction model, inputting the test data into the model to obtain a prediction result, and comparing the prediction result with an actual result, thereby proving that the model can effectively restore the actual appearance of the sample, has practical significance for processing AFM scanning data and improving measurement accuracy, and can realize the butt joint with related AFM imaging software.
The invention comprises the following steps:
step 1: obtaining a simulation sample by using an expansion method in mathematical morphology;
the scanning process of AFM contact measurements does not need to take into account the small deformations due to the interaction forces between the probe and the sample. Assuming that the expression of the probe cross-sectional profile shape when the tip is at the home position in the coordinate system is s (x), the expression of the sample cross-sectional profile shape is f (x). The position of the probe in the scan line is defined by the abscissa of the tip, assuming that the abscissa of the tip is x0The ordinate h of the tip is then indicated at x0The measured height value of the sample scanning point. At this time, assuming that the expression of the cross-sectional profile shape of the probe is t (x), the relationship between s (x) and t (x) satisfies the expression
t(x)=s(x)+h (1)
The relationship between f (x) and t (x) satisfies the expression
AFM contact measurement is equivalent to a dilation operation in mathematical morphology, the mathematical expression of which is
Thus, according to equation (3), the expression for h can be written as
By combining formula (1) and formula (2), formula (3) can be converted into
Since h and x satisfy a one-to-one mapping relationship, h can be regarded as a function of x, and h (x) is used for representing a height result change curve measured by the probe on a certain horizontal scanning line of the sample, namely results of each line of the simulation image. h (x) includes the relationship among the sample profile, the tip profile and the measurement image.
In the simulation process, a nano-grid standard device is selected as a tested sample, and a silicon nitride needle tip is selected as a scanning device.
In order to better acquire the characteristic vectors of the nano-grid and the probe in training, the following scheme can be used for setting the parameters of the nano-grid and the probe: the line width W of the nano-grid is 20-40nm, the distance is 5nm, the height H is 10-30nm, the distance is 5nm, 25 nano-grid types are arranged in total, the size of a corresponding matrix is 140 multiplied by 140, and the resolution is 1 nm/pixel; the curvature radius R of the silicon nitride needle tips is 10-30nm, the pitch is 5nm, the cone angle theta is 5-80 degrees, the pitch is 5nm, 80 needle tip types are arranged, the size of a corresponding matrix is 30 multiplied by 30, and the resolution is 1 nm/pixel. The simulated scanning process may acquire 2000 sets of sample images, where 1600 sets of images are used as a training set and 400 sets of images are used as a testing set. Meanwhile, in the actual scanning process, the probe types are often considered to be limited, and the nano-grid morphology is infinite, so that the sample selection of the training set needs to include all the probe types, but does not need to include all the nano-grid types.
Step 2: carrying out normalization operation on the simulation sample image obtained in the step 1;
because the AFM image obtained in the step 1 is equivalent to a set of row vectors, the AFM image can be decomposed line by line to realize dimension reduction, so that the row vectors become training samples of the network. In the process, as the AFM simulation image is a 140 × 140 matrix, the input data and the output result are both row vectors of 1 × 140. In order to inhibit the phenomena of incapability of convergence and overfitting in the training process and further improve the stability and the performance of the neural network, normalization processing is required to be adopted on input data, namely
In formula (6), x is a numerical value in the input matrix, xmaxAnd xminRespectively the maximum and minimum values, x, in the input matrix*The result is normalized for the input data.
And step 3: inputting the data obtained in the step (2) into an LSTM model for training;
the convolution process in the AFM scanning sample is generated by the gradient change of the sample appearance and is not related to the height value of the sample, so the process of deconvolution of the AFM image focuses on researching the change condition of the sample scanning height curve. For an input matrix, if the position of each data in the matrix is changed, the data in the output matrix itself will change, not just spatially. Thus, the data in the input matrix is ordered and can be viewed as a time series, while the LSTM model is suitable for training of the process since the input and output matrices are the same size.
The LSTM model comprises a forgetting gate FtInput gate ItOutput gate OtAnd candidate memory cellsAnd fourthly, the method comprises the following steps. The gate input of LSTM is the current time step input XtHidden state H with last time stept-1This is matched by the AFM scan line point gradient determined by the height values of adjacent points. The output is given by sigmoThe full connection layer calculation of the id activation function is obtained, and the expression is
To obtain candidate memory cells with larger gradient near 0 AFM scan line input and faster model convergence, tanh activation function is used, expressed as tan h
The LSTM can store AFM scanning line vector information with higher dimension through memory cells, so that the correlation degree between scanning points is obtained. For example, points at flat areas of the scan line are less correlated, while points at areas with greater variations in scan line height are more correlated. Setting a time step memory cell Ct-1Current time step memory cell CtAt the time of obtaining CtWhen passing through FtAnd ItAnd the Hadamard product of the matrix controls the flow of information between each point of the scanning line:
if FtIs always approximately 1 and ItThe element in (1) is always approximately 0, and past memory cells will always be saved by time and passed to the current time step. The process can store effective information in the AFM scanning line for a long time, solves the problem of gradient attenuation caused by the increase of scanning line vector dimensions, and improves the training reliability. Meanwhile, the dependency relationship among points with larger distance among scanning lines can be obtained, and the image restoration of a region with serious artifact effect is realized.
At the acquisition of CtThen, it is necessary to pass through OtTo control the slave HtFlow of information to the current time step hidden state:
Ht=Ot⊙tanh(Ct) (10)
for the nano-grid scan line, when OtWhen the element in (A) is approximately 0 or 0, CtThe information in the method is usually only the information of a flat area of the nano grid, is not the required characteristic information, and is generally reserved for the user; when O is presenttWhen the element in (B) is far from 0, then C is presenttThe characteristic information such as the position of a step and the like of the nano-grid is contained in the H-shaped magnetic field, and the H-shaped magnetic field is transmitted to the HtFor use by the output layer.
And 4, step 4: continuously improving an LSTM model by adopting an MSE loss function and an Adam optimization algorithm;
in the process of adjusting the LSTM model to realize gradient descent, due to the output AFM scanning behavior vector form, the recovery effect of the AFM image is evaluated by adopting Modified Mean Absolute Error (MMAE) as a loss function, and a 1-norm is used in the summation process. The loss function is expressed as
In equation (11), n represents the number of training samples, l represents the vector dimension of the network output,estimated value, y, representing each line of AFM restored imageiThe real values of the respective lines of the AFM restored image are shown.
In the iterative process of gradient descent, an Adam optimization algorithm is adopted. The Adam algorithm performs exponential weighted moving average on the small-batch random gradient, performs deviation correction on partial variables, and simultaneously adopts self-adaptive learning rate to continuously optimize network parameters. Compared with a random gradient descent method, the algorithm can more effectively update the network weight and finally accelerate the convergence speed. The learning rate parameter adaptive formula of the Adam algorithm is
In the formula (12), β0Denotes the initial given hyper-parameter, βiRepresenting the hyperparameter after i iterations.
And 5: carrying out regularization and inverse normalization processing on the training samples;
for the output layer passing through the LSTM model, a Dropout regularization method is adopted, and some output characteristics of the layer are abandoned randomly in the training process, so that the overfitting phenomenon is reduced, and the model obtained through training is suitable for more types of AFM scanning sample images.
All sample data in the LSTM model training process are in the [0,1] interval, and the finally obtained restored image needs to be close to the actual sample, so that all data need to be subjected to inverse normalization operation before an output result is obtained.
Step 6: inputting the actual AFM image into the trained model to obtain an AFM restoration image;
and for the AFM restored image, processing by using an adaptive algorithm with gradual interpolation, so that the positions of the characteristic points in the sample can be more accurately obtained.
The invention is further illustrated by the following figures and examples.
Example (b):
step 1: obtaining a simulation sample by using an expansion method in mathematical morphology;
according to the AFM simulation image schematic diagram shown in FIG. 1, the following scheme is adopted to set parameters: the line width W of the nano grid is 20-40nm, the spacing is 5nm, the height H is 10-30nm, the spacing is 5nm, the size of a corresponding matrix is 140 multiplied by 140, and the resolution is 1 nm/pixel; the curvature radius R of the silicon nitride needle tip is 10-30nm, the cone angle theta is 5-80 degrees, the interval is 5nm, the size of a corresponding matrix is 30 multiplied by 30, and the resolution is 1 nm/pixel. In 2000 groups of sample images obtained in the simulation scanning process, 1600 groups of images are used as a training set, 400 groups of images are used as a testing set, and the sample selection of the training set needs to include all probe types. The simulated tip model, the simulated nano-grating model and the simulated AFM image are respectively shown in FIG. 2, FIG. 3 and FIG. 4.
After obtaining the AFM simulation image, the sample needs to be trained, and the flow of the adopted algorithm is shown in fig. 5.
Step 2: and (3) carrying out normalization operation on the simulation sample image obtained in the step (1).
And step 3: and (3) inputting the data obtained in the step (2) into an LSTM model for training.
The LSTM model comprises four parts, namely a forgetting gate, an input gate, an output gate and a candidate memory cell. The gate inputs of the LSTM are the current time step input and the previous time step hidden state, and the output is obtained by a full connection layer of a sigmoid function. In the memory layer, the memory cells at the current time step are obtained by the Hadamard product of the matrix, and the hidden state at the current time step is obtained by the tanh function, as shown in FIG. 6.
And 4, step 4: and continuously perfecting the LSTM model by adopting an improved MAE loss function and an Adam optimization algorithm.
A graph of loss values versus number of iterations in the training is shown in fig. 7.
And 5: carrying out regularization and inverse normalization processing on the training samples;
for the output layer passing through the LSTM model, a Dropout regularization method is adopted, and the output characteristics of the output layer are abandoned randomly in the training process, so that the overfitting phenomenon is reduced, and the model obtained through training is suitable for more types of AFM images.
Step 6: inputting the actual AFM image into the trained model to obtain an AFM restoration image;
and for an output result obtained by inputting the actual AFM image into the trained model, processing by using a step-by-step interpolation adaptive algorithm, so that the positions of the feature points in the sample can be more accurately obtained.
The test results of the method of the invention are shown in fig. 8, (a) is the top view of the actual nano-grating sample, (b) is the top view of the actual nano-grating sample after AFM scanning, and (c) is the top view of the AFM restored image obtained by the trained model. Therefore, the atomic force microscope image restoration is realized, namely.
The above description is only a preferred embodiment of the method of the present invention, but the scope of the method of the present invention is not limited thereto, and any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope of the method of the present invention are included in the scope of the method of the present invention. Therefore, the protection scope of the method of the present invention shall be subject to the protection scope of the claims.
Claims (5)
1. The atomic force microscope image restoration method based on the long-term and short-term memory network is characterized by comprising the following steps:
step 1: obtaining a simulation sample by using an expansion method in mathematical morphology;
step 2: carrying out normalization operation on the simulation sample;
and step 3: inputting the normalized simulation sample into an LSTM model for training;
the LSTM model comprises a forgetting gate, an input gate, an output gate and candidate memory cells; the gate input of the LSTM is the current time step input and the previous time step hidden state, and the output is obtained by a full connection layer of a sigmoid function;
in a memory layer, memory cells in the current time step are obtained through the Hadamard product of the matrix, so that the flow of information in the row vector of the AFM image is controlled, and the hidden state in the current time step is obtained through a tanh function;
and 4, step 4: continuously improving an LSTM model by adopting an MSE loss function and an Adam optimization algorithm;
and 5: carrying out regularization and inverse normalization processing on the training samples;
step 6: and inputting the actual AFM image into the trained model to obtain an AFM restoration image.
2. The atomic force microscope image restoration method based on the long-short term memory network as claimed in claim 1, wherein the process of obtaining the simulation sample in step 1 is as follows:
converting a three-dimensional process of AFM contact measurement into a two-dimensional process, and converting a binary function representing the process into a univariate function; using s (x) to represent the cross section outline shape of the probe when the needle point is at the original point position in the coordinate system, f (x) the cross section outline shape of the sample, and then obtaining the relation between the height h (x) of the sample obtained by AFM scanning and f (x) and s (x);
setting parameters of a simulation nano grid and a simulation silicon nitride needle tip;
and (4) simulating by using an expansion method in mathematical morphology to obtain a plurality of groups of training samples.
3. The atomic force microscope image restoration method based on the long-short term memory network as claimed in claim 1, wherein the normalization process in the step 2 is as follows:
the simulation sample obtained in the step 1 is equivalent to a set of row vectors, and the AFM image is decomposed line by line to realize dimension reduction, so that the row vectors become training samples of the network; in this process, both the input data and the output result are row vectors of 1 × 140, and normalization processing is applied to the input data.
4. The atomic force microscope image restoration method based on the long and short term memory network as claimed in claim 1, wherein in the step 4, in order to satisfy the output result in a vector form, the LSTM model is continuously perfected by using an improved MAE loss function MMAE and Adam optimization algorithm; the MMAE loss function is expressed as
Where n represents the number of training samples, l represents the vector dimension of the network output,estimated value, y, representing each line of AFM restored imageiRepresenting real values of each line of the AFM restored image;
the learning rate parameter adaptive formula of the Adam algorithm is
Wherein beta is0Denotes the initial given hyper-parameter, βiRepresenting the hyperparameter after i iterations.
5. The atomic force microscope image restoration method based on the long-short term memory network as claimed in claim 1, wherein the regularization and de-normalization processing of the training samples in step 5 is as follows:
for the output layer passing through the LSTM model, a Dropout regularization method is adopted, and some output characteristics of the layer are abandoned randomly in the training process, so that the overfitting phenomenon is reduced, and the model obtained through training is suitable for more types of scanning sample images;
in order to make the AFM restored image close to the actual sample, all data are subjected to an inverse normalization operation before an output result is obtained.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202011575092.8A CN112581412A (en) | 2020-12-28 | 2020-12-28 | Atomic force microscope image restoration method based on long-term and short-term memory network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202011575092.8A CN112581412A (en) | 2020-12-28 | 2020-12-28 | Atomic force microscope image restoration method based on long-term and short-term memory network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN112581412A true CN112581412A (en) | 2021-03-30 |
Family
ID=75140221
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202011575092.8A Pending CN112581412A (en) | 2020-12-28 | 2020-12-28 | Atomic force microscope image restoration method based on long-term and short-term memory network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN112581412A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114236181A (en) * | 2021-12-02 | 2022-03-25 | 中国电子科技集团公司第十三研究所 | AFM probe measuring method, device, control equipment and storage medium |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20180033144A1 (en) * | 2016-09-21 | 2018-02-01 | Realize, Inc. | Anomaly detection in volumetric images |
CN110706173A (en) * | 2019-09-27 | 2020-01-17 | 中国计量大学 | Atomic force microscope image blind restoration method based on convolutional neural network |
CN111833266A (en) * | 2020-06-18 | 2020-10-27 | 杭州电子科技大学 | Motion blur restoration method for recurrent neural network |
-
2020
- 2020-12-28 CN CN202011575092.8A patent/CN112581412A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20180033144A1 (en) * | 2016-09-21 | 2018-02-01 | Realize, Inc. | Anomaly detection in volumetric images |
CN110706173A (en) * | 2019-09-27 | 2020-01-17 | 中国计量大学 | Atomic force microscope image blind restoration method based on convolutional neural network |
CN111833266A (en) * | 2020-06-18 | 2020-10-27 | 杭州电子科技大学 | Motion blur restoration method for recurrent neural network |
Non-Patent Citations (5)
Title |
---|
武玉伟 等: "深度学习基础与应用", vol. 2020, 30 November 2020, 北京理工大学出版社, pages: 147 - 99 * |
范春奇;任坤;孟丽莎;黄泷;: "基于深度学习的数字图像修复算法最新进展", 信号处理, no. 01, 25 January 2020 (2020-01-25) * |
许海燕, 王琛: "纳米生物医学技术", vol. 2009, 30 June 2009, 中国协和医科大学出版社, pages: 195 * |
钟跃崎: "人工智能技术原理与应用", vol. 2020, 30 September 2020, 东华大学出版社, pages: 254 * |
陈敏: "人工智能通信理论与方法", vol. 2020, 31 January 2020, 华中科技大学出版社, pages: 268 * |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114236181A (en) * | 2021-12-02 | 2022-03-25 | 中国电子科技集团公司第十三研究所 | AFM probe measuring method, device, control equipment and storage medium |
CN114236181B (en) * | 2021-12-02 | 2023-10-20 | 中国电子科技集团公司第十三研究所 | AFM probe measuring method, device, control equipment and storage medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Narwaria et al. | Objective image quality assessment based on support vector regression | |
CN111539132B (en) | Dynamic load time domain identification method based on convolutional neural network | |
US9092737B2 (en) | Systems, methods, and apparatus for 3-D surface mapping, compliance mapping, and spatial registration with an array of cantilevered tactile hair or whisker sensors | |
Miller et al. | Statistical behavior of retrospective patterns and their effects on estimation of stock and harvest status | |
Zheng et al. | A CNN-based image reconstruction for electrical capacitance tomography | |
Memarzadeh et al. | Hierarchical modeling of systems with similar components: A framework for adaptive monitoring and control | |
Wu et al. | Gated recurrent unit based frequency-dependent hysteresis modeling and end-to-end compensation | |
CN112581412A (en) | Atomic force microscope image restoration method based on long-term and short-term memory network | |
CN114065919A (en) | Deficiency value completion method and medium based on generation countermeasure network | |
Fablet et al. | Joint learning of variational representations and solvers for inverse problems with partially-observed data | |
JP7097541B2 (en) | Information processing equipment, information processing system, information processing method and program | |
CN115990875A (en) | Flexible cable state prediction and control system based on hidden space interpolation | |
CN112388628A (en) | Apparatus and method for training a gaussian process regression model | |
CN114511025A (en) | Fan fault diagnosis method and device based on weighted multi-sensor fusion filtering | |
CN116957422B (en) | Ecological environment evaluation method based on convolution self-coding and sharp point mutation model | |
CN109191503A (en) | Remote sensing image variation detection method and system based on condition random field | |
Nakamura-Zimmerer et al. | Structured Covariance Gaussian Networks for Orion Crew Module Aerodynamic Uncertainty Quantification | |
CN117370771A (en) | Knowledge embedding filling soft measurement method based on conditional fractional diffusion | |
CN113783186B (en) | Voltage prediction method considering topological structure change of power distribution network | |
CN115760603A (en) | Interference array broadband imaging method based on big data technology | |
JP7404989B2 (en) | Observation value estimation method and device, and image reconstruction method and device | |
CN110852451B (en) | Recursive kernel self-adaptive filtering method based on kernel function | |
CN114048762A (en) | Double-attention-guided rotating machine health assessment method | |
CN114155354A (en) | Capacitance tomography reconstruction method and device based on graph convolution network | |
Kewalramani et al. | Estimation of Remaining Useful Life of Electric Motor using supervised deep learning methods |
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 |