A curve envelope fitting method based on VGG16 network

Abstract

The invention discloses a curve envelope fitting method based on a VGG16 network. The algorithm uses a data set sample with a label established by CCD acquisition to train a neural network, and applies the neural network algorithm to the acquisition data set to verify the accuracy and infer the microscopic morphological characteristics of the optical fiber surface; a reading data set is created through a tenserflow framework, and a gray value change sequence of the group of images at a certain pixel point (a, b) is recorded as X(a, b)(t); a cubic Hermite interpolation method is used to supplement the sequence X(a, b)(t) into a one-dimensional sequence X2(a, b)(t) with a size of 224*224, and then the sequence X2(a, b)(t) is converted into a two-dimensional image matrix X2(a, b)(m, n); the actual height of the pixel point predicted by the specially designed convolutional neural network is processed and output, and compared with the sample label to make the error within a set threshold range; the application of the neural network enables the algorithm to have better self-learning, self-organizing ability and fault tolerance and excellent nonlinear approximation ability, which can improve the accuracy and fault tolerance of the envelope algorithm and has certain reference significance.

Description

Technical Field

[0001] The present invention relates to the field of deep learning data processing, and in particular to an envelope fitting algorithm for a surface characteristic curve of an optical fiber ferrule based on a deep learning tool. Background Art

[0002] Neural network algorithms are a type of algorithm that learns from data based on neural networks, and they still fall into the category of machine learning. With the improvement of computing power and the advent of the big data era, highly parallel GPUs and massive data make it possible to train large-scale neural networks. In the field of image classification, convolutional neural networks are very useful in deep learning. Compared with traditional image classification methods, it no longer requires manual feature description and extraction of target images, but instead uses neural networks to autonomously learn features from training samples, and these features are closely related to classifiers, which effectively solves the problem of manual feature extraction and classifier selection.

[0003] In white light interference, when the optical path difference between two coherent light beams is zero, the interference fringes have extreme values, and as the optical path difference gradually increases, the peak value of the interference fringes gradually decreases. Therefore, white light scanning interferometry is achieved by determining the zero optical path difference position of the collected white light interference signal to achieve surface morphology measurement. In addition, the accuracy of the white light scanning interferometry measurement system is directly affected by the positioning accuracy of the zero optical path difference position. The positioning of the zero optical path difference position can be converted into the positioning of the peak position of the interference light intensity. At present, the commonly used peak positioning methods include extreme value method, centroid method, Fourier transform method, wavelet transform method, Hilbert transform method, spatial frequency domain algorithm, and white light phase shift method, but the positioning accuracy and accuracy of these methods are poor. In contrast, the powerful learning ability, adaptive ability, and feature extraction ability of deep learning may well make up for the measurement accuracy and precision problems of the above-mentioned classical algorithms. Summary of the invention

[0004] In order to solve the shortcomings of the above-mentioned prior art, the purpose of the present invention is to provide a curve envelope fitting algorithm based on the VGG16 network. The algorithm adopts a deep learning tool and takes the VGG16 network as the basic framework. The one-dimensional discrete signal is complemented by the cubic Hermite interpolation method and converted into two-dimensional image data. The cosine signal modulated by the Gaussian function and the feature extraction of the white light interference signal are realized through convolution with the convolution kernel.

[0005] The technical solution adopted by the present invention is: to establish the required data set through CCD acquisition, the sample should contain clear interference pictures and accurate three-dimensional microscopic data of the object. The total number of samples is required to be greater than 6000 groups, and the sample pictures contained in each group depend on the scanning step length, and the peak position of white light interference is required to be included, and the sample picture specifications are unified.

[0006] After reading a group of images, record the gray value change sequence of the group of images at a certain pixel point (a, b) as X(a, b) (t) , use the cubic Hermite interpolation method to convert the sequence X(a,b) (t) Supplemented to a one-dimensional sequence X of size 224*224 2 (a,b) (t) , the sequence X 2 (a,b) (t) Convert to a two-dimensional image matrix X 2 (a,b) (m,n) It is convenient for subsequent deep learning algorithm processing.

[0007] Since the white light interference signal is a cosine signal modulated by a Gaussian function, the light intensity value at the zero-order fringe in the center reaches the highest peak. Therefore, a corresponding convolution kernel G(9,9) of size 9*9 is designed, and the data of the one-dimensional sequence G(t) constructed according to the same rule of G(9,9) is required to be Gaussian distributed.

[0008] The two-dimensional image matrix X of a certain pixel 2 (a,b) (m,n) As the model input, it is first convolved twice with 64 specially designed convolution kernels, and then maxpooled for dimensionality reduction;

[0009] After that, it is convolved twice again with 128 specially designed convolution kernels and the dimension is reduced using maxpool.

[0010] The convolution is repeated three times with 256 specially designed convolution kernels, and the dimension is reduced by maxpool.

[0011] The network is convolved three times with 512 specially designed convolution kernels and the dimension is reduced by maxpool.

[0012] Once again, 512 specially designed convolution kernels are used for convolution three times, followed by dimensionality reduction through maxpool.

[0013] The two-dimensional sequence is stretched into a one-dimensional sequence through three fully connected layers, and the values ​​are normalized through soft-max to select the value X with the highest probability. The value is converted into the peak height of this pixel using proportional normalization, and then multiplied by the scanning step to get the actual height of the point.

[0014] The actual height output by the neural network is compared with the actual height label to obtain the error. If the error is within the artificially specified threshold, the output neural network model is saved; otherwise, the neural network model parameters are continuously adjusted and the network is trained again until the error meets the expectation and the output neural network model is saved.

[0015] Then, by repeating the above operation for all pixels, the three-dimensional features of the microscopic object can be obtained through the deep learning algorithm.

[0016] The convolution operation padding = "same" means adding 0s around the image matrix to pad the image to the original image size.

[0017] The BatchNormalization layer is used after the convolution operation to speed up the model training and prevent overfitting of the model training. At the same time, the activation function adopts the "Relu" function.

[0018] The maxpool operation is a typical pooling matrix of size 2*2, with stride=2, that is, the sliding step size is 2.

[0019] The beneficial effects of the present invention are: the algorithm improves the accuracy of envelope algorithm processing. At the same time, the use of neural network enables the algorithm to have good self-learning, self-organizing ability and fault tolerance and excellent nonlinear approximation ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a schematic diagram of the neural network data preprocessing process of the present invention;

[0021] Figure 2 is a schematic diagram of the neural network structure of the present invention;

[0022] Figure 3 It is the algorithm flow chart of the present invention; DETAILED DESCRIPTION

[0023] Figure 1-3 In the process, the required data set is established through CCD acquisition. The sample should contain clear interference images and accurate three-dimensional microscopic data of the object. The total number of samples is required to be greater than 6,000 groups. The number of sample images in each group depends on the scanning step length. It is required to include the peak position of white light interference, and the sample image specifications are unified.

[0024] After reading a group of images, record the gray value change sequence of the group of images at a certain pixel point (a, b) as X(a, b) (t) , use the cubic Hermite interpolation method to convert the sequence X(a,b) (t) Supplemented to a one-dimensional sequence X of size 224*224 2 (a,b) (t) , the sequence X 2 (a,b) (t) Convert to a two-dimensional image matrix X 2 (a,b) (m,n) It is convenient for subsequent deep learning algorithm processing.

[0025] Since the white light interference signal is a cosine signal modulated by a Gaussian function, the light intensity value at the zero-order fringe in the center reaches the highest peak. Therefore, a corresponding convolution kernel G(9,9) of size 9*9 is designed, and the data of the one-dimensional sequence G(t) constructed according to the same rule of G(9,9) is required to be Gaussian distributed.

[0026] The two-dimensional image matrix X of a certain pixel 2 (a,b) (m,n) The data input, i.e. a single-channel image with a length and width of 224, is used as the model input. It is first convolved twice with 64 convolution kernels G(9,9) of size 9*9, and then the maxpool is used to reduce the dimension to 112*112*64.

[0027] After that, it is convolved twice again with 128 convolution kernels G(9,9) and the dimension is reduced to 56*56*128 using maxpool, that is, an image with a size of 56*56 and 128 channels.

[0028] Repeat the convolution three times with 256 convolution kernels G(9,9), reduce the dimension to 28*28*256 through maxpool, convolve three times with 512 convolution kernels G(9,9), reduce the dimension to 14*14*512 through maxpool again, convolve three times with 512 convolution kernels G(9,9), and reduce the dimension to 7*7*512 through maxpool.

[0029] Through three fully connected layers with sizes of 4096, 4096, and 1000 nodes, (7, 7, 512) is stretched to a sequence of size (1, 1, 1000). The values ​​are normalized by soft-max to select the value X with the highest probability. The value is converted to the peak height of this pixel using proportional normalization, that is, X / 1000*the number of sample image groups is the peak height of this pixel, and then multiplied by the scanning step size is the actual height of the point.

[0030] The error is calculated between the actual height output by the neural network and the actual height label. If the error is within the artificially specified threshold, the output neural network model is saved; otherwise, the neural network model parameters are continued to be adjusted, including the convolution kernel size, number, data distribution, and number of neural network layers, and the network is trained again until the error meets the expectation and the output neural network model is saved.

[0031] Then, by repeating the above operation for all pixel points, the three-dimensional features of the microscopic object can be obtained through the deep learning algorithm, thereby correctly constructing the three-dimensional morphology of the microscopic surface.

[0032] In summary, the algorithm processing process of the curve envelope fitting method based on the VGG16 network is as follows: first, the neural network is trained through labeled samples, then the required data set is established through CCD acquisition, and the neural network algorithm is applied to the acquired data set to verify the accuracy and infer the microscopic morphological characteristics of the optical fiber surface. Finally, the trained neural network can be deployed and applied to the image data collected each time, so as to obtain the microscopic morphological characteristics of the measured surface.

Claims

1. A curve envelope fitting method based on VGG16 network, characterized in that: include: Establish the required data set through CCD acquisition, unify the sample image specifications, and perform corresponding data preprocessing based on the feature information; The preprocessed two-dimensional image is trained by comparing the designed neural network with the actual height label, and the corresponding neural network is modified and optimized according to the training analysis results; Train the neural network through training data, obtain the target neural network that meets the requirements and deploy it to micro-morphology detection; The steps to train a neural network include: Since the white light interference signal is a cosine signal modulated by a Gaussian function, the light intensity value of the zero-order fringe at the center reaches the highest peak. Therefore, a corresponding convolution kernel G(9,9) with a size of 9*9 is designed. It is required that the data of the one-dimensional sequence G(t) constructed according to the same rule of G(9,9) should be Gaussian distributed. The two-dimensional image matrix X2(a,b) of a certain pixel (m,n) The data input, i.e. a single-channel image with a length and width of 224, is used as the model input. First, it is convolved twice with 64 convolution kernels G(9,9) of size 9*9, and then the maxpool is used to reduce the dimension to 112*112*64. Then, it is convolved twice again with 128 convolution kernels G(9,9), and the dimension is reduced to 56*56*128 using maxpool, that is, an image with a size of 56*56 and 128 channels; Repeat the convolution three times with 256 convolution kernels G(9,9), reduce the dimension to 28*28*256 through maxpool, convolve three times with 512 convolution kernels G(9,9), reduce the dimension to 14*14*512 through maxpool again, convolve three times with 512 convolution kernels G(9,9), and reduce the dimension to 7*7*512 through maxpool; Through three fully connected layers with sizes of 4096, 4096, and 1000 nodes, (7, 7, 512) is stretched into a sequence of size (1, 1, 1000). The values ​​are normalized by soft-max to select the value X with the highest probability. The value is converted into the peak height of this pixel point using proportional normalization, that is, X / 1000*the number of sample image groups is the peak height of this pixel point, and then multiplied by the scanning step distance to get the actual height of the point; The error is calculated by comparing the actual height output by the neural network with the actual height label. If the error is within the artificially specified threshold, Then save the output neural network model; otherwise, continue to adjust the neural network model parameters, including the size and number of convolution kernels, data distribution, and the number of neural network layers, train the network again until the error meets the expectation, and save the output neural network model; Then, by repeating the above operation for all pixel points, the three-dimensional features of the microscopic object can be obtained through the deep learning algorithm, thereby correctly constructing the three-dimensional morphology of the microscopic surface.

2. The method according to claim 1, characterized in that The data preprocessing steps include: The required data set is established through CCD acquisition. The sample should contain clear interference images and accurate three-dimensional microscopic data of the object. The total number of samples is required to be greater than 6,000 groups. The number of sample images in each group depends on the scanning step length. It is required to include the peak position of white light interference, and the sample image specifications should be unified. After reading a group of images, record the gray value change sequence of the group of images at a certain pixel point (a, b) as X(a, b) (t) , use the cubic Hermite interpolation method to convert the sequence X(a,b) (t) Supplemented to a one-dimensional sequence X2(a,b) of size 224*224 (t) , the sequence X2(a,b) (t) Convert to a two-dimensional image matrix X2(a,b) (m,n) It is convenient for subsequent deep learning algorithm processing.

3. The method according to claim 1, characterized in that: The steps for deploying a neural network application should include: The neural network that meets the requirements is exported and applied to the actual pixel height detection. The actual height of the point is estimated through the neural network to construct the entire microstructure appearance characteristics.

4. The method according to claim 1, characterized in that The convolution operation padding = "same", that is, add 0s around the image matrix to pad the image to the original image size.

5. The method according to claim 1, characterized in that The BatchNormalization layer is used after the convolution operation to speed up the model training and prevent overfitting of the model training. At the same time, the activation function adopts the "Relu" function.

6. The method according to claim 1, characterized in that The maxpool operation is a typical 2*2 pooling matrix with stride=2, that is, the sliding step size is 2.

Images