CN113516620B - Convolutional neural network translation error detection method - Google Patents
Convolutional neural network translation error detection method Download PDFInfo
- Publication number
- CN113516620B CN113516620B CN202110413187.8A CN202110413187A CN113516620B CN 113516620 B CN113516620 B CN 113516620B CN 202110413187 A CN202110413187 A CN 202110413187A CN 113516620 B CN113516620 B CN 113516620B
- Authority
- CN
- China
- Prior art keywords
- lsr
- sequence
- value
- translation error
- channel
- 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
- 238000013519 translation Methods 0.000 title claims abstract description 120
- 238000001514 detection method Methods 0.000 title claims abstract description 57
- 238000013527 convolutional neural network Methods 0.000 title claims description 8
- 238000012549 training Methods 0.000 claims abstract description 86
- 238000000034 method Methods 0.000 claims abstract description 81
- 239000013598 vector Substances 0.000 claims abstract description 25
- 238000013528 artificial neural network Methods 0.000 claims abstract description 23
- 238000003384 imaging method Methods 0.000 claims description 51
- 239000000919 ceramic Substances 0.000 claims description 10
- 238000005259 measurement Methods 0.000 claims description 9
- 230000003287 optical effect Effects 0.000 claims description 6
- 239000006185 dispersion Substances 0.000 description 13
- 238000010586 diagram Methods 0.000 description 7
- 238000009825 accumulation Methods 0.000 description 4
- 238000003062 neural network model Methods 0.000 description 4
- 238000012545 processing Methods 0.000 description 4
- 238000004519 manufacturing process Methods 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000010200 validation analysis Methods 0.000 description 2
- 238000012795 verification Methods 0.000 description 2
- 238000013459 approach Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000012886 linear function Methods 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 230000007704 transition Effects 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
- G06T7/0002—Inspection of images, e.g. flaw detection
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/70—Determining position or orientation of objects or cameras
- G06T7/73—Determining position or orientation of objects or cameras using feature-based methods
-
- 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
-
- 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
- G06T2207/20084—Artificial neural networks [ANN]
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Quality & Reliability (AREA)
- Testing Of Optical Devices Or Fibers (AREA)
Abstract
The invention discloses a convolution neural network translation error detection method, which mainly comprises the following steps: 1. acquiring an initial image; 2. establishing a training set; 3. establishing a neural network; 4. training a neural network; 5. detecting a translation error; the method of the invention introduces a plurality of wavelength channels, utilizes the different cycle periods of LSR values in each wavelength channel to form LSR characteristic vectors, and then uses a neural network to identify the LSR characteristic vectors to finish the detection of the translation error. In the process, the invention also utilizes data in one wavelength to construct a large training data set, and the establishment of the training data set solves the problem that the large-range training set is difficult to obtain when the traditional neural network method detects the translation error. Finally, the invention realizes the translation error detection with large range and high precision, and the method has strong noise resistance.
Description
Technical Field
The invention relates to a translation error detection method, in particular to a convolution neural network translation error detection method.
Background
In recent years, with the higher resolution requirement of astronomical observation, the requirement of the aperture of a telescope reaches 10m level. The single-caliber telescope has the sharply increased requirements on the manufacturing process along with the increase of the caliber, and the manufacturing cost is very high. Therefore, since the seventies of the last century, researchers break through the design concept of single-aperture telescope systems, and propose to use synthetic aperture technology to meet the requirements of large-aperture astronomical telescopes, so as to ensure the imaging quality of the system under the conditions of effectively reducing the manufacturing cost and meeting the requirements on the process level. However, the synthetic aperture telescope has a Tilt error (Tilt) and a translation error (Piston) between each sub-mirror after the track expansion, which greatly degrades the imaging quality. The tilt error can be well corrected through a precise confocal phase and a precise phase-sharing phase, while the translation error often exceeds the detection range of a general precise phase-sharing method, and a necessary coarse phase-sharing phase is required to be added as a transition.
Based on this, scientific researchers at home and abroad put forward a plurality of methods for the coarse and fine phase-sharing stage. Such as a wide band method for the coarse Phase-co-phasing stage, a series of methods based on a Dispersive Fringe Sensor (DFS), and a narrow band method, a Phase difference method (PD), a Pyramid detection method (Pyramid), an interference method, etc. for the fine Phase-co-phasing stage. However, due to the detection accuracy and range, these methods often need to be used in combination, so different devices are needed to detect and correct the translation error, for example, a PD method is used to perform fine co-phase detection after coarse co-phase detection is performed by a dispersion fringe sensor, and large-range and high-accuracy detection cannot be performed.
In 2017, the photoelectric institute proposed a left peak minus right peak method (DFA-LSR) based on dispersion fringe image accumulation, which accumulates dispersion fringe images along the dispersion direction and detects the linear relationship between the difference between the left peak and the right peak of the accumulated signals and the translation error. The accumulation of the whole dispersion fringe results in large bandwidth and small coherence length, so that the method has high precision, but the detection range can only reach one wavelength.
Disclosure of Invention
Aiming at the problem that the detection range of the existing DFA-LSR method mentioned in the background technology can only reach one wavelength, the invention provides a convolutional neural network translation error detection method.
The specific technical scheme of the invention is as follows:
the method for detecting the translation error of the convolutional neural network comprises the following implementation steps:
step 1: initial image acquisition:
step 1.1: setting a translation error interval delta; the selection of the plurality of wavelength channels is related to the spacing Δ of the translational errors: when there are n wavelength channels, the distance delta between each wavelength channel and the translation error needs to satisfy lambda i ＝M i Δ, where 1 < i ≦ n, M i Is an integer;
step 1.2: at any wavelength channel lambda i Imaging by adjusting the translation error by the translation error interval delta at lambda i In the channel, collect N i ＝λ i Δ +1 images; the adjusting range is as follows: when N is present i Is odd number [ - λ ] i /2,λ i /2]When N is present i When it is even [ - (lambda) i +Δ)/2,(λ i -Δ)/2]；
Step 2: establishing a training set;
step 2.1: accumulating the images of each wavelength channel according to the superposition principle of DFA-LSR, thereby obtaining an LSR value sequence corresponding to the translation error in one wavelength of each wavelength channel as an original data set;
step 2.2: respectively establishing training sets corresponding to positive and negative translation error ranges according to the original data set;
step 2.3: connecting the training sets with the positive translation error range and the training sets with the negative translation error range end to end, and adding an LSR value corresponding to the translation error being zero between the two training sets to form a training set;
and step 3: establishing a neural network;
establishing a neural network comprising Net1 and Net2;
the Net1 network structure comprises an input layer, a hidden layer and an output layer, wherein the hidden layer only has three nodes; net1 is used to fit translation error values within one wavelength of a single channel to the corresponding LSR value, with the input being a singleLSR values corresponding to translation errors within a wavelength of the channel, in the form of
Outputting a translation error value corresponding to the LSR value;
the Net2 network structure adopts a part of a Resnet18 network, has 11 layers in total, inputs the LSR eigenvector of nx1x1, and outputs the translation error interval corresponding to the LSR eigenvector; wherein, the interval between adjacent nodes of the output layer of Net2 is one wavelength distance of the wavelength channel selected by Net1, namely lambda i ；
And 4, step 4: training a neural network;
step 4.1: net1 training process
Inputting an LSR value corresponding to the translation error in one wavelength of any single channel into Net1 to obtain the corresponding relation between the translation error in one wavelength and the LSR value;
and 4.2: net2 training process
Inputting the training set obtained in the step 2 into Net2 to obtain the corresponding relation between the LSR characteristic vector and the translation error interval;
and 5: translational error detection
Step 5.1: after setting a translation error value in an imaging system, carrying out multi-channel imaging to obtain n channel images; obtaining LSR value of each channel image by DFA-LSR method, forming n × 1 multichannel LSR feature vector, and recording as
Step 5.2:
inputting the LSR value of the channel adopted during the Net1 training into the Net1 to obtain a result O 1 (ii) a Simultaneously inputting the feature vectors obtained in the step 4.1 into the Net2
Obtaining the result O 2 ；
The final translational error obtained is:
piston_detected＝O 2 *λ i +O 1 。
further, the specific implementation process of step 2.2 is as follows:
a: establishing a training set of a positive translation error range;
a1: when any channel lambda i When the sequence length of the middle LSR value sequence is an odd number, removing the first LSR value in the LSR value sequence, and then moving the whole first half sequence of the rest LSR value sequences to the second half sequence to form a new LSR value sequence;
a2: when any channel lambda i When the sequence length of the middle LSR value sequence is an even number, removing a first LSR value in the LSR value sequence, and taking the median of the rest LSR value sequences and the sequence value before the median as a first half sequence to integrally move to a second half sequence to form a new LSR value sequence;
a3: circularly arranging the new LSR value sequence obtained in the step A1 or the step A2 in the forward direction;
a4: performing steps A1-A3 on all channels to obtain a training set of positive translation error ranges;
b: establishing a training set of a negative translation error range;
b1: when the sequence length of the LSR value sequence in any channel is an odd number, removing the last LSR value in the LSR value sequence, and then moving the whole rear half sequence of the rest LSR value sequences to the front half sequence to form a new LSR value sequence;
b2: when the sequence length of the LSR value sequence in any channel is an even number, removing the first LSR value in the LSR value sequence, and then moving the median of the rest LSR value sequences and the sequence value after the median as a second half sequence to the front half sequence integrally to form a new LSR value sequence;
b3: reversely and circularly arranging the new LSR value sequence obtained in the step B1 or the step B2;
b4: steps B1-B3 are performed for all channels to obtain a training set of negative translational error ranges.
Further, the imaging system in the step 5 includes a monochromator, a beam expander, a beam splitter prism a, a beam splitter prism B, a D-type reflector a, a D-type reflector B, an interference area selection diaphragm, a six-foot translation stage, a piezoelectric ceramic actuator, an imaging lens, and a CMOS image sensor;
the D-type reflector A, the D-type reflector B, the interference area selection diaphragm, the piezoelectric ceramic actuator and the six-foot translation stage are combined into a spliced mirror unit;
the monochromator, the beam expander, the beam splitter prism A, the beam splitter prism B, the imaging lens and the CMOS image sensor are combined to form an imaging measurement unit;
in the splicing mirror unit, a D-shaped reflector A is fixed, a D-shaped reflector B is fixed on a piezoelectric ceramic actuator and a six-foot translation table through a supporting structure, a front-back position optical path difference exists between the D-shaped reflector A and the D-shaped reflector B, and an interference area selection diaphragm is close to the two reflectors;
in the imaging measurement unit, light emitted by a monochromator passes through a beam expander and then is converted into parallel light to be incident, the parallel light is transferred into a splicing mirror system by a beam splitter prism A and a beam splitter prism B, and the parallel light enters a CMOS image sensor after a phase difference is obtained in the splicing mirror system according to the optical path difference between the front position and the rear position of a D-type reflector A and the D-type reflector B.
Further, the imaging system is a dispersive fringe sensor system.
The invention has the beneficial effects that:
1. the method introduces multi-wavelength channels to obtain LSR characteristic vectors on the basis of the principle of the DFA-LSR method, thereby widening the detection range of the DFA-LSR method to a great extent and inheriting the anti-noise property and high precision of the DFA-LSR method.
2. The method of the invention realizes the detection of the translation error by distinguishing the LSR characteristic vector through the neural network, and establishes a large training data set through data in one wavelength, thereby solving the problem that the large training data set is difficult to obtain in the traditional method for detecting the translation error by utilizing the neural network, and ensuring the detection precision when the translation error is detected through the neural network subsequently.
3. The method of the invention can use a simple device to detect the translation error, can also be integrated on a Dispersion Fringe Sensor (DFS) to detect, and has strong applicability.
Drawings
Fig. 1 is a schematic diagram of the structure of an imaging system.
The reference numbers of fig. 1 are as follows:
the device comprises a 1-monochromator, a 2-beam expander, a 3-beam splitter A, a 4-beam splitter B, a 5-D reflector A, a 6-D reflector B, a 7-interference area selection diaphragm, an 8-hexapod translation stage, a 9-piezoelectric ceramic actuator, a 10-imaging lens and an 11-CMOS image sensor.
FIG. 2 is the image plane intensity distribution of the 660nm channel with a translational error of 50 μm.
FIG. 3 shows the image stacking process for 5 channels with a translation error of 50 μm.
FIG. 4 is an architecture diagram of a neural network; wherein (a) is a Net1 architecture diagram, and (b) is a Net2 architecture diagram;
FIG. 5 is a diagram showing the structure of an apparatus in which the CNN-Multi-LSR method is used in combination with a DFS.
Fig. 6 shows the extraction of LSR feature vector on the dispersed fringe pattern corresponding to-146.4953 μm translational error.
FIG. 7 shows the results of-165 μm translational error imaging when probed in conjunction with DFS, for each channel with no noise and SNR of 10, 15, and 20, respectively.
Detailed Description
Since the linear relationship between the difference between the left peak minus the right peak and the translational error within one wavelength still exists when monochromatic light is incident, and when the monochromatic light is incident, the difference continuously cycles with the increase of the translational error, which is also called as ambiguity of 2 pi. Based on this, the method of the invention introduces a plurality of wavelength channels, and utilizes the different cycle periods of LSR values in each wavelength channel to form LSR characteristic vector. And then, the neural network is used for identifying the LSR feature vector, and the detection of the translation error is completed. In the process, the invention also utilizes data in one wavelength to construct a large training data set, and the establishment of the training data set solves the problem that the large-range training set is difficult to obtain when the traditional neural network method detects the translation error. Finally, the method realizes the large-range and high-precision translation error detection, and has strong noise resistance.
The method mainly comprises four processes: establishing a training set, establishing a neural network model, training a neural network and detecting a translation error.
1. Establishing training set Process
The invention provides multi-channel monochromatic light imaging to break through the 2 pi fuzzy problem in the common precise phase-sharing method, obtain a large detection range and ensure the detection precision. In the method of detecting the translational error by using the neural network, the size of the detection range is often determined by the translational error range covered by the training set, and further determined by the size of the training set. However, it is very difficult to obtain a large training set by actually acquiring images, so the invention provides a method for acquiring images corresponding to translation errors in a wavelength range of each wavelength channel to manually construct a training set corresponding to a large translation error range.
1: acquiring an initial image:
1.1: setting a translation error interval delta; the selection of the plurality of wavelength channels is related to the spacing Δ of the translational errors: when there are n wavelength channels, the interval delta between each wavelength channel and the translation error needs to satisfy lambda i ＝M i Δ, where 1 < i ≦ n, M i Is an integer;
step 1.2: at any wavelength channel lambda i Imaging by adjusting the translation error by the translation error interval delta at lambda i In the channel, collect N i ＝λ i Δ +1 images; the adjusting range is as follows: when N is present i Is odd number [ - λ ] i /2,λ i /2]When N is present i When it is even [ - (lambda) i +Δ)/2,(λ i -Δ)/2]；
The specific imaging process is as follows: referring to fig. 1, an imaging system includes: the device comprises a monochromator 1, a beam expander 2, a beam splitter A3, a beam splitter B4, a D-type reflector A5, a D-type reflector B6, an interference area selection diaphragm 7, a six-foot translation stage 8, a piezoelectric ceramic actuator 9, an imaging lens 10 and a CMOS image sensor 11;
the D-type reflector A5, the D-type reflector B6, the interference area selection diaphragm 7, the piezoelectric ceramic actuator 9 and the six-foot translation stage 8 are combined to form a splicing mirror unit;
the monochromator 1, the beam expander 2, the beam splitter A3, the beam splitter B4, the imaging lens 10 and the CMOS image sensor 11 are combined to form an imaging measurement unit;
in the splicing mirror unit, a D-shaped reflector A5 is fixed, a D-shaped reflector B6 is fixed on a piezoelectric ceramic actuator 9 and a six-foot translation stage 8 through a supporting structure, a front-back position optical path difference (namely translation error) exists between the D-shaped reflector A5 and the D-shaped reflector B6, and an interference area selection diaphragm 7 is close to the two reflectors;
in the imaging measurement unit, light emitted by the monochromator 1 passes through the beam expander 2 and then becomes parallel light to be incident, the parallel light is converted into a splicing mirror unit by the beam splitter A3 and the beam splitter B4, and the splicing mirror unit obtains a phase difference according to a front-back position optical path difference between the D-type reflector A5 and the D-type reflector B6 and then enters the CMOS image sensor 11.
1.3: establishing a training set;
1.3.1: accumulating the images of each wavelength channel according to the superposition principle of DFA-LSR, thereby obtaining an LSR value sequence corresponding to the translation error in one wavelength of each channel as an original data set;
the specific accumulation process in this step is:
will N this i Each image of the images is accumulated according to columns to obtain N i A result of addition, i.e. N i Each column vector is processed according to the formula (1) to obtain N i Individual LSR value, noted:
wherein L is p-aver Is the average of three nodes centered on the left peak, R p-aver 、M p-aver The same is the average value of three nodes with the right peak value and the middle peak value as the centers. This process of superimposing and calculating LSR values stems from the DFA-LSR approach. However, here the results of monochromatic light imaging are superimposed, rather than the dispersive stripsA fringe image, and calculating an LSR value using an average of each peak value and a neighbor value to improve resistance to noise;
1.3.2: respectively establishing training sets corresponding to the positive and negative translation error ranges according to the original data set, wherein the training sets are established in the following specific steps:
a: establishing a training set of a positive translation error range;
a1: when any channel lambda i When the sequence length of the middle LSR value sequence is an odd number, removing the first LSR value in the LSR value sequence, and then moving the whole first half sequence of the rest LSR value sequences to the second half sequence to form a new LSR value sequence;
a2: when any channel lambda i When the sequence length of the middle LSR value sequence is an even number, removing a first LSR value in the LSR value sequence, and taking the median of the rest LSR value sequences and the sequence value before the median as a first half sequence to integrally move to a second half sequence to form a new LSR value sequence;
a3: the new LSR value sequences obtained in the step A1 or the step A2 are arranged in a forward circulation mode;
a4: performing the steps A1-A3 on all channels to obtain a training set of a positive translation error range;
b: establishing a training set of a negative translation error range;
b1: when the sequence length of the LSR value sequence in any channel is an odd number, removing the last LSR value in the LSR value sequence, and then moving the whole rear half sequence of the rest LSR value sequences to the front half sequence to form a new LSR value sequence;
b2: when the sequence length of the LSR value sequence in any channel is an even number, removing the first LSR value in the LSR value sequence, and then moving the median of the rest LSR value sequences and the sequence value after the median as a second half sequence to the front half sequence integrally to form a new LSR value sequence;
b3: reversely and circularly arranging the new LSR value sequence obtained in the step B1 or the step B2;
b4: steps B1-B3 are performed for all channels to obtain a training set of negative translational error ranges.
1.3.3: connecting the training sets with positive translation error ranges and the training sets with negative translation error ranges end to end, and adding a corresponding LSR value with zero translation error between the two training sets to form a training set;
for ease of understanding, the above steps 1.3.2 and 1.3.3 are now described by way of example:
a, establishing a training set of a positive translation error range;
a1: for channels with an odd sequence length of the sequence of LSR values, for example 660nm channels, the shift error interval Δ in step 1.1 is set to 30nm. The length of the LSR value sequence obtained by acquisition and processing is N i =660/30+1=23, as:
the first LSR value in the sequence of LSR values is removed and then the first half sequence (part a in the table above) of the remaining sequence of LSR values is moved in its entirety to the second half sequence before the new sequence of LSR values is composed:
| 0.0585 | …… | 0.6344 | 0.6881 | -0.6344 | …… | -0.0585 | 0 |
a2: for channels with an even sequence length of the sequence of LSR values, for example 690nm channels, the interval Δ of the translation errors in step 1.1 is set to 30nm. The length of the LSR value sequence obtained by acquisition and processing is N i 690/30+1=24, as:
the first LSR value in the LSR value sequence is removed, and then the middle and the sequence values before the middle of the remaining LSR value sequence are moved as the whole of the first half sequence (part B in the above table) to the second half sequence to form a new LSR value sequence:
| 0.0457 | …… | 0.5828 | 0.6643 | -0.6643 | -0.5828 | …… | -0.0457 | 0 |
a3: circularly arranging the new LSR value sequence obtained in the step A1 or the step A2 in the forward direction;
and B, circularly arranging the new LSR value sequence obtained in the step A2 in the forward direction:
| 0.0457 | …… | 0 | 0.0457 | …… | 0 | 0.0457 | …… | 0 | …… |
a4: performing the steps A1-A3 on all channels to obtain a training set of a positive translation error range;
when the number n of channels in step 1.1 is 5:
| 660nm channel | 0.0585 | …… | 0 | 0.0585 | …… | 0 | 0.0585 | …… | …… |
| 690nm channel | 0.0457 | …… | -0.0457 | 0 | 0.0457 | …… | -0.0457 | 0 | …… |
| 720nm channel | 0.0672 | …… | -0.0962 | -0.0672 | 0 | 0.0672 | …… | -0.0962 | …… |
| 750nm channel | 0.0509 | …… | -0.1534 | -0.1086 | -0.0509 | 0 | 0.0509 | …… | …… |
| 780nm channel | 0.0419 | …… | -0.1939 | -0.1516 | -0.1095 | -0.0419 | 0 | 0.0419 | …… |
B: establishing a training set of a negative translation error range;
b1: for channels with an odd sequence length of the sequence of LSR values, for example 660nm channels, the interval Δ of the translational errors in step 1.1 is set to 30nm. The length of the sequence of LSR values obtained by acquisition and processing is N i =660/30+1=23, as:
removing the last LSR value in the LSR value sequence, and then integrally moving the rear half sequence (part C in the table) of the rest LSR value sequences to the front half sequence to form a new LSR value sequence;
| 0 | 0.0585 | …… | 0.6344 | -0.6881 | -0.6344 | …… | -0.0585 |
b2: for channels with an even sequence length of the sequence of LSR values, for example 690nm channels, the spacing Δ of the translation errors in step 1.1 is set to 30nm. The obtained LSR value sequence is collected and processed
Column length of N i =690/30+1=24, such as:
the first LSR value in the LSR value sequence is removed and then the median of the remaining LSR value sequences and the sequence values after the median (part D in the above table) are moved as a whole of the second half sequence before the first half sequence to form a new LSR value sequence:
| 0 | 0.0457 | …… | 0.5828 | 0.6643 | -0.6643 | -0.5828 | …… | -0.0457 |
b3: reversely and circularly arranging the new LSR value sequence obtained in the step B1 or the step B2;
and B1, reversely and circularly arranging the new LSR value sequence obtained in the step B1:
| …… | 0 | …… | -0.0457 | 0 | …… | -0.0457 | 0 | …… | -0.0457 |
b4: steps B1-B3 are performed for all channels to obtain a training set of negative translational error ranges.
| …… | …… | -0.0585 | 0 | …… | -0.0585 | 0 | …… | -0.0585 | 660nm channel |
| …… | 0 | 0.0457 | …… | -0.0457 | 0 | 0.0457 | …… | -0.0457 | 690nm channel |
| …… | 0.0962 | …… | -0.0672 | 0 | 0.0672 | 0.0962 | …… | -0.0672 | 720nm channel |
| …… | 0.1534 | -0.0509 | 0 | 0.0509 | 0.1086 | 0.1534 | …… | -0.0509 | 750nm channel |
| …… | -0.0419 | 0 | 0.0419 | 0.1095 | 0.1516 | 0.1939 | …… | -0.0419 | 780nm channel |
1.3.3: connecting the training sets with the positive translation error range and the training sets with the negative translation error range end to end, and adding an LSR value corresponding to the translation error being zero between the two training sets to form a training set;
| …… | …… | -0.0585 | 0 | …… | -0.0585 | 0 | 0.0585 | …… | 0 | 0.0585 | …… | …… |
| …… | -0.0457 | 0 | 0.0457 | …… | -0.0457 | 0 | 0.0457 | …… | -0.0457 | 0 | 0.0457 | …… |
| …… | 0 | 0.0672 | 0.0962 | …… | -0.0672 | 0 | 0.0672 | …… | -0.0962 | -0.0672 | 0 | …… |
| …… | 0.0509 | 0.1086 | 0.1534 | …… | -0.0509 | 0 | 0.0509 | …… | -0.1534 | -0.1086 | -0.0509 | …… |
| …… | 0.1095 | 0.1516 | 0.1939 | …… | -0.0419 | 0 | 0.0419 | …… | -0.1939 | -0.1516 | -0.1095 | …… |
2. creating neural network models
The neural network model comprises two neural network models, namely Net1 and Net2;
the Net1 network structure comprises an input layer, a hidden layer and an output layer, wherein the hidden layer only has three nodes. Net1 is used to fit the shift error value in one wavelength of a single channel to the corresponding LSR value, which acts as a perfect co-phasing. During training, the input is LSR value in one wavelength of single channel in the form of LSR λi And outputting the translation error value corresponding to the LSR value.
The Net2 network structure adopts a part of a Resnet18 network, and has 11 layers in total, wherein the input format and some node parameters are changed, the LSR characteristic vector with the shape of nx1x1 is input, and the translation error interval corresponding to the LSR characteristic vector is output; the spacing between adjacent nodes of the output layer being one wavelength distance of the selected wavelength channel of Net1, i.e. λ i . Net2 is equivalent to a coarse co-phasing process, unknown translation errors are defined in a small range, and accurate detection is carried out by utilizing the result of Net 1. In the training process, the input of Net2 is a set formed by m eigenvectors with the form of n × 1 established by us, and the label is a translation error value corresponding to the training set.
3. Training neural networks
3.1: net1 training process
Inputting an LSR value corresponding to the translation error in one wavelength of any single channel into Net1 to obtain the corresponding relation between the translation error in one wavelength and the LSR value;
3.2: net2 training process
Inputting the training set obtained in the step 1 into Net2 to obtain the corresponding relation between the LSR characteristic vector and the translation error interval;
4. translational error detection
4.1: after setting a translation error value in an imaging system, carrying out multi-channel imaging to obtain n channel images; obtaining LSR value of each channel image by DFA-LSR method, forming n × 1 multichannel LSR feature vector, and recording as
4.2: inputting the LSR value of the channel adopted during the Net1 training into the Net1 to obtain the resultO 1 (ii) a Simultaneously inputting the feature vectors obtained in the step 4.1 into the Net2
Obtaining the result O 2 (ii) a The final translational error obtained is:
piston_detected＝O 2 *λ i +O 1 。
at this point, the translational error detection is completed.
The index specification of the translation error detection method provided by the invention is as follows:
a: the detection range boundary condition (maximum detection range) of the translational error detection method provided by the invention depends on the monochromaticity of incident light, and the calculation formula is as shown in formula (3):
wherein λ is min At the center wavelength of the smallest wavelength channel used, Δ 1/2 λ, Δ λ are the full width at half maximum of the minimum wavelength channel and the total bandwidth of that channel, respectively.
B: detection precision: the detection accuracy of the method mainly depends on the fitting accuracy of Net1, and the detection residual error is basically stabilized below the fitting error of Net 1.
The method of the present invention is further explained below with reference to specific simulation examples:
1. parameter selection for imaging systems
Selecting long wavelength as possible within the bandwidth range of a monochromator to obtain a larger coherence length; this example takes n =5, Δ =30nm, the center wavelength channel is 720nm, and the 5 operating wavelength channels are: 660nm,690nm,720nm,750nm and 780nm;
(II) full width at half maximum Δ of each wavelength channel 1/2 λ and bandwidth Δ λ (total bandwidth of each wavelength channel), the smaller the values of these two parameters, the higher monochromaticity of monochromatic light of each wavelength channel is, the larger the corresponding detection range is, the values of this example are: 2.5nm and 4nm, the values being selectedThe monochromaticity of the actual monochromator is considered;
(III) the size of the image element of the image sensor is 1.67 mu m; in the process of independently using the CNN-Multi-LSR method, the pixel size of the image sensor can be not a unique value, or an image sensor with the pixel size of 4.6 mu m can be used, but the pixel size is not too large, so that the resolution of an imaging area is ensured to be at least 15 multiplied by 15, and a margin is reserved for extracting a superposed signal peak value;
(IV) interference area selection diaphragm: the length of the rectangular hole is 2mm, and the width of the rectangular hole is 2mm; the center distance between the two rectangular holes is 5mm;
(V) the focal length of the imaging lens is 10cm; the parameters In (IV) and (V) jointly determine that the two-hole diffraction image is in the imaging area mentioned in (III) and three diffraction peaks are obvious, and the values simultaneously consider the stripe contrast limit when the two-hole diffraction image is combined with DFS.
2. Single wavelength channel imaging process
Referring to fig. 1, light emitted by the monochromator passes through the beam expander and then becomes parallel light to be incident, the parallel light is transferred into the splicing mirror unit by the beam splitter prism a and the beam splitter prism B, the light enters the imaging measurement unit after phase difference is obtained in the splicing mirror unit, and the light is imaged on the CMOS image sensor by the imaging measurement unit. FIG. 2 shows the imaging results for the 660nm wavelength channel at a translation error of 50 μm.
3. Image overlay processing procedure
For a specific translation error, each of the 5 images obtained is accumulated column by column to obtain 5 accumulation results, i.e. 5 column vectors. Finally, each column vector is processed according to the formula (1) to obtain 5 LSR values, and the five LSR values can form a characteristic vector [ LSR ] 660 ,LSR 690 ...LSR 780 ]After the superposition process shown in FIG. 3, the eigenvector formed by LSR values of 5 channels with a translational error of 50 μm [ -0.3095,0.6354,0.5866, -0.4367,0.1390 ] can be obtained by using the formula (1)]。
4. Training set establishment process:
the training set is established in the manner of (1.3) above. In this example, a training set corresponding to a translational error range of (-180.9 μm,108.9 μm) was established for a total of 7260 sets of data.
5. A neural network training process:
a block diagram of two neural networks is shown in fig. 4. And training the two networks, wherein the selected wavelength channel of Net1 is a 720nm channel, and the adjacent nodes corresponding to the output nodes of Net2 are 720nm apart. The training set and validation set for Net1 have 720/30=24 groups of data. The Net2 training set has 7260 groups of data, and the validation set has 4357 groups of data. Because the training of Net1 is similar to the fitting process of linear function, the training process time is short (within 5 s), the training result precision is high, and the optimal performance is achieved at epoch 38. The verification precision of Net2 reaches 96.44% after 400epochs, and the whole training process is about 69mins.
6. Detecting a translation error: and distinguishing the LSR feature vector by using the trained Net1 and Net 2. When Net1 and Net2 are respectively input into LSR 720 (LSR value of 720nm channel) and LSR eigenvector, let the output results of Net1 and Net2 be O 1 And O 2 Then the corresponding translation error detection result is given by equation (2). Taking a translation error detection process of 50 μm as an example, O 2 ＝69，O 1 =0.3095, and data of 720nm adopted by Net1 at this time, so λ i =0.72 μm, the detected translation error is obtained according to equation (2):
piston detected ＝69×0.72+0.3095＝49.9895μm
the detection error is 10.5nm. The method can realize large-range and high-precision detection and has strong resistance to noise.
Expanding applications
The translational error detection method provided by the present invention has device applicability, and can be used with the imaging system shown in fig. 1, or in combination with a Dispersive Fringe Sensor (DFS), and fig. 5 shows a device structure diagram of the DFS. In fig. 5, the dispersion coefficient of the gridlines is:
the dispersion coefficient of the grism was 5.5X 10 5 ；
The process of using the method in combination with the dispersion fringe sensor is similar to the process of using the imaging system shown in fig. 1, and the difference is that a broadband light source is used for incidence, and a prism grating is added into the imaging measurement system for dispersion to obtain a dispersion fringe image, so that superposed images under two conditions are different. When the system is used jointly, different areas need to be selected on the dispersion fringe image for superposition and calculation to form LSR characteristic vectors, and the superposition of imaging results of each channel when the imaging system in FIG. 1 is used is replaced. The specific selection area and the superposition flow of each channel are given in fig. 6.
The LSR feature vector on the dispersed fringe pattern can be obtained through the process of fig. 6, and then the training set and the verification set are obtained by combining the (1.3) data set establishing method. Then, a network having the same structure as Net1 and Net2 is continuously used, except that the equivalent bandwidth of the image region of each selected channel is about 2.42nm due to the dispersion effect of the edge grating, and the maximum detection range is (-180 μm,180 μm) as can be seen from equation (2), so that the number of output nodes of the network 2 is 500, and the node interval is still 720nm.
Fig. 7 shows five channel region images corresponding to-165 μm translational error when jointly detecting with DFS, where (a) represents no noise, (b) represents SNR =10, (c) represents SNR =15, and (b) represents each channel region image when SNR =20, according to the proposed method. The detection result is as follows: when no noise exists: -165.0024 μm; SNR = 10: -164.9999 μm. Therefore, whether noise exists or not, the error of the detection result combined with DFS is less than 5nm, and the adaptability of the method to the device is further verified on the basis of large range, high precision and strong noise resistance.
Description of the Properties
Translation error detection range: using detectable translation errors P detected Showing, when the imaging system of fig. 1 is used in conjunction with the method of the present invention: p is less than or equal to-100 um detected ≤100um；
When a Dispersive Fringe Sensor (DFS) is used in conjunction with the method of the invention: -169 um.ltoreq.P detected ≤169um；
Average detection precision of translational error: detection error P using translation error error Showing that the imaging system of figure 1 is used in conjunction with the method of the present invention and when P is noise free error <20nm, using the imaging of FIG. 1The system incorporates the method of the invention and SNR = 15: p error <20nm；
The use of a Dispersive Fringe Sensor (DFS) in conjunction with the method of the invention and without noise: p error <40nm, using a Dispersive Fringe Sensor (DFS) in combination with the method of the invention and SNR = 15: p error <50nm；
When the imaging system of fig. 1 is used in conjunction with the method of the present invention, translation errors of 100um are detected. In the absence of noise, the detection result is: 99.9974um, error less than 5nm; when SNR =15, the detection result is: 99.9985um, and the error is less than 20nm.
A Dispersive Fringe Sensor (DFS) was used in conjunction with the method of the invention to detect 169um translational errors. In the absence of noise, the detection result is: 169.0269um, error less than 40nm; when SNR =15, the detection result is: 169.0228um, error less than 50nm.
Compared with the existing detection technical indexes:
Claims (4)
1. a convolutional neural network translation error detection method is characterized by comprising the following steps:
step 1: initial image acquisition:
step 1.1: setting a translation error interval delta; the selection of the plurality of wavelength channels is related to the spacing Δ of the translational errors: when there are n wavelength channels, the interval delta between each wavelength channel and the translation error needs to satisfy lambda i ＝M i Δ, where 1 < i ≦ n, M i Is an integer;
step 1.2: at any wavelength channel lambda i Imaging by adjusting the translation error by the translation error interval delta at lambda i In the channel, collect N i ＝λ i (Δ +1 images); the adjusting range is as follows: when N is present i Is odd number [ - λ ] i /2,λ i /2]When N is present i When it is even [ - (lambda) i +Δ)/2,(λ i -Δ)/2]；
And 2, step: establishing a training set;
step 2.1: accumulating the images of each wavelength channel according to the superposition principle of DFA-LSR, thereby obtaining an LSR value sequence corresponding to the translation error in one wavelength of each wavelength channel as an original data set;
step 2.2: respectively establishing training sets corresponding to the positive translation error range and the negative translation error range according to the original data set;
step 2.3: connecting the training sets with positive translation error ranges and the training sets with negative translation error ranges end to end, and adding a corresponding LSR value with zero translation error between the two training sets to form a training set;
and step 3: establishing a neural network;
establishing a neural network comprising Net1 and Net2;
the Net1 network structure comprises an input layer, a hidden layer and an output layer, wherein the hidden layer is only provided with three nodes; net1 is used for fitting the relation between the translation error value in one wavelength of a single channel and the corresponding LSR value, and the input of the translation error value in one wavelength of the single channel is the LSR value corresponding to the translation error in the form of LSR λi Outputting a translation error value corresponding to the LSR value;
the Net2 network structure adopts a part of a Resnet18 network, has 11 layers in total, inputs the LSR eigenvector of nx1x1, and outputs the translation error interval corresponding to the LSR eigenvector; wherein the interval between adjacent nodes of the output layer of Net2 is one wavelength distance, namely lambda, of the wavelength channel selected by Net1 i ；
And 4, step 4: training a neural network;
step 4.1: net1 training process
Inputting an LSR value corresponding to the translation error in one wavelength of any single channel into Net1 to obtain the corresponding relation between the translation error in one wavelength and the LSR value;
and 4.2: net2 training process
Inputting the training set obtained in the step 2 into Net2 to obtain the corresponding relation between the LSR characteristic vector and the translation error interval;
and 5: translational error detection
Step 5.1: after setting a translation error value in an imaging system, carrying out multi-channel imaging to obtain n channel images; obtaining LSR value of each channel image by DFA-LSR method, forming n × 1 multichannel LSR feature vector, and recording as
Step 5.2:
inputting the LSR value of the channel adopted during the Net1 training into the Net1 to obtain a result O 1 (ii) a Simultaneously inputting the feature vectors obtained in step 5.1 into Net2
Obtaining the result O 2 ；
The final translational error obtained is:
piston_detected＝O 2 *λ i +O 1 。
2. the convolutional neural network translational error detection method according to claim 1, wherein the specific implementation process of the step 2.2 is as follows:
a: establishing a training set of a positive translation error range;
a1: when any channel lambda i When the sequence length of the middle LSR value sequence is an odd number, removing the first LSR value in the LSR value sequence, and then moving the whole first half sequence of the rest LSR value sequences to the second half sequence to form a new LSR value sequence;
a2: when any channel lambda i When the sequence length of the middle LSR value sequence is an even number, removing a first LSR value in the LSR value sequence, and taking the median of the rest LSR value sequences and the sequence value before the median as a first half sequence to integrally move to a second half sequence to form a new LSR value sequence;
a3: the new LSR value sequences obtained in the step A1 or the step A2 are arranged in a forward circulation mode;
a4: performing the steps A1-A3 on all channels to obtain a training set of a positive translation error range;
b: establishing a training set of a negative translation error range;
b1: when the sequence length of the LSR value sequence in any channel is an odd number, removing the last LSR value in the LSR value sequence, and then moving the whole rear half sequence of the rest LSR value sequences to the front half sequence to form a new LSR value sequence;
b2: when the sequence length of the LSR value sequence in any channel is an even number, removing a first LSR value in the LSR value sequence, and then moving the median of the rest LSR value sequences and the sequence value after the median as a whole of a rear half sequence to the front half sequence to form a new LSR value sequence;
b3: reversely and circularly arranging the new LSR value sequence obtained in the step B1 or the step B2;
b4: steps B1-B3 are performed for all channels to obtain a training set of negative translational error ranges.
3. The convolutional neural network translation error detection method of claim 1, wherein: the imaging system in the step 5 comprises a monochromator, a beam expander, a beam splitter prism A, a beam splitter prism B, a D-shaped reflector A, a D-shaped reflector B, an interference area selection diaphragm, a six-foot translation table, a piezoelectric ceramic actuator, an imaging lens and a CMOS image sensor;
the D-type reflector A, the D-type reflector B, the interference area selection diaphragm, the piezoelectric ceramic actuator and the six-foot translation stage are combined into a spliced mirror unit;
the monochromator, the beam expander, the beam splitter prism A, the beam splitter prism B, the imaging lens and the CMOS image sensor are combined to form an imaging measurement unit;
in the splicing mirror unit, a D-shaped reflector A is fixed, a D-shaped reflector B is fixed on a piezoelectric ceramic actuator and a six-foot translation table through a supporting structure, a front-back position optical path difference exists between the D-shaped reflector A and the D-shaped reflector B, and an interference area selection diaphragm is close to the two reflectors;
in the imaging measurement unit, light emitted by a monochromator passes through a beam expanding mirror and then becomes parallel light to be incident, the parallel light is switched into a splicing mirror system by a beam splitting prism A and a beam splitting prism B, and the parallel light enters a CMOS image sensor after a phase difference is obtained in the splicing mirror system according to the optical path difference between the front position and the rear position of a D-type reflector A and the D-type reflector B.
4. The convolutional neural network translational error detection method of claim 1, wherein: the imaging system is a dispersive fringe sensor system.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110413187.8A CN113516620B (en) | 2021-04-16 | 2021-04-16 | Convolutional neural network translation error detection method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110413187.8A CN113516620B (en) | 2021-04-16 | 2021-04-16 | Convolutional neural network translation error detection method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113516620A CN113516620A (en) | 2021-10-19 |
| CN113516620B true CN113516620B (en) | 2023-01-06 |
Family
ID=78061514
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110413187.8A Active CN113516620B (en) | 2021-04-16 | 2021-04-16 | Convolutional neural network translation error detection method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113516620B (en) |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107945146A (en) * | 2017-11-23 | 2018-04-20 | 南京信息工程大学 | A kind of space-time Satellite Images Fusion method based on depth convolutional neural networks |
| CN111311666A (en) * | 2020-05-13 | 2020-06-19 | 南京晓庄学院 | Monocular vision odometer method integrating edge features and deep learning |
| CN112508904A (en) * | 2020-12-02 | 2021-03-16 | 长春理工大学 | Spliced telescope translation aberration detection method based on BP neural network |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6787747B2 (en) * | 2002-09-24 | 2004-09-07 | Lockheed Martin Corporation | Fast phase diversity wavefront correction using a neural network |
| CN105115427B (en) * | 2015-09-16 | 2017-12-26 | 中国科学院光电技术研究所 | Method for improving noise-resistant capacity of dispersion fringe sensor co-phase system |
-
2021
- 2021-04-16 CN CN202110413187.8A patent/CN113516620B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107945146A (en) * | 2017-11-23 | 2018-04-20 | 南京信息工程大学 | A kind of space-time Satellite Images Fusion method based on depth convolutional neural networks |
| CN111311666A (en) * | 2020-05-13 | 2020-06-19 | 南京晓庄学院 | Monocular vision odometer method integrating edge features and deep learning |
| CN112508904A (en) * | 2020-12-02 | 2021-03-16 | 长春理工大学 | Spliced telescope translation aberration detection method based on BP neural network |
Non-Patent Citations (4)
| Title |
|---|
| Dispersed-fringe-accumulation-based left-subtract-right method for fine co-phasing of a dispersed fringe sensor;Yang Li 等;《Applied Optics》;20171231;全文 * |
| Global piston restoration of segmented mirrors with recurrent neural networks;Dailos Guerra-Ramos 等;《OSA Continuum》;20201231;全文 * |
| Large-scale piston error detection technology for segmented optical mirrors via convolutional neural networks;Dequan Li 等;《Optics Letters》;20191231;全文 * |
| 采用神经网络提高EIT成像质量的方法研究;陈晓艳等;《天津科技大学学报》;20160825(第04期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113516620A (en) | 2021-10-19 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US9945721B2 (en) | Selective wavelength imaging systems and methods | |
| CN104344793A (en) | Single-frame light intensity transmission quantitative phase microscope system and method | |
| CN111220361B (en) | Method for measuring focal length of micro-lens array | |
| US20140374573A1 (en) | Spectral imaging sensors and methods | |
| CN111238397B (en) | Transient digital Moire phase-shifting interferometry device and method for optical element surface shape | |
| US9091534B2 (en) | Measuring apparatus, measuring method, and method of manufacturing an optical component | |
| JP5144841B1 (en) | Imaging device | |
| WO1999056096A1 (en) | Corrected concentric spectrometer | |
| CN106500589A (en) | A kind of measuring method of multi-wavelength tunable micro-interference and its device | |
| JP6223734B2 (en) | Calibrable beamforming system and method | |
| CN206311061U (en) | A kind of measurement apparatus of multi-wavelength tunable micro-interference | |
| CN112508904A (en) | Spliced telescope translation aberration detection method based on BP neural network | |
| CN113516620B (en) | Convolutional neural network translation error detection method | |
| CN110849591A (en) | Multispectral common-phase error detection method based on pupil space modulation | |
| Park et al. | Miniaturization of a Fresnel spectrometer | |
| CN107631687A (en) | Point source dystopy expands simultaneous phase-shifting fizeau interferometer and its measuring method | |
| CN110307781B (en) | Interferometric imaging system | |
| CN112097904A (en) | Spectral imaging system and method based on diffraction lens/zoom lens array | |
| DE102014215952A1 (en) | Illumination module for an optical sensor and optical sensor with such an illumination module for a coordinate measuring machine for measuring internal threads or boreholes of a workpiece | |
| CN105446111B (en) | A kind of focusing method applied to digital hologram restructuring procedure | |
| US10002275B2 (en) | Optical element having two phase masks | |
| CN111238396B (en) | Transient digital moire phase-shifting interferometry device and method | |
| CN113029526B (en) | Synthetic aperture co-phasing error estimation method and device | |
| CN110672305A (en) | Array type telescope multispectral common-phase error detection method for spatial modulation based on electronic shutter | |
| Wang et al. | Large-range piston error detection technology based on dispersed fringe sensor |
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 |