• 专利附录
  • 专利说明
  • 权利要求
  • 类似技术
  • 同族专利
  • 引用文献
  • 法律状态
  • 优先权
    • 专利摘要:
    This disclosure proposes a method for denoising ESPI images based on deep learning, which consists of building a training dataset; build a convolutional nerve network model based on the self-adaptive BM3D-TV algorithm; training said convolutional nerve network model with the training dataset; inputting a speckle interference fringe image to be denoised into the trained model, and denoising the speckle interference fringe image; the method provided by the present disclosure can effectively perform denoising on ESPI images, which not only guarantees low noise residual, but also provides effective protection on the edge information of fringed images, moreover, ESPI images can have good adaptability, which facilitates batch image processing and builds a basis for subsequent operations such as phase extraction.
    公开号:BE1027609B1
    申请号:E20215108
    申请日:2021-02-17
    公开日:2022-01-17
    发明作者:Huamei Xin
    申请人:Univ Shandong；
    IPC主号:
    专利说明:

    ESPI image denoising method and system based on deep learning
    TECHNICAL FIELD This disclosure relates to the technical field of electron speckle interferometry, and in particular to an ESPI image denoising method and system based on deep learning.
    TECHNICAL BACKGROUND The description in this section provides only technical background information relating to this disclosure and does not necessarily constitute prior art.
    Electronic speckle interferometry technology realizes non-contact high-precision full-field detection and at the same time has advantages such as wide frequency range, low level of environmental impacts, wide scene range applicable, simplicity and ease of operation. With the development of electronic speckle interference technology research, this technology plays an important role in detecting surface damage, flaw, deformation and displacement of objects, which has unique advantages for detections in many fields, such as automotive industry, transportation, electronics industry, archaeology, art research, bridges, medicine, materials science, space science and others for detecting vibration, detecting profiles of objects, surveying displacements, inspecting deformations of physical machines and equipment, evaluating surface roughness and strength of objects, etc. Additionally, ESPI image research has also become a hot spot, bringing groundbreaking advancements to optical measurements.
    An ESPI image can accurately and efficiently express the deformation of an object and can be created by various operation modes, including addition, subtraction and multiplication modes. The subtraction mode will not be affected by the light intensity of the background, so the final image will have contrast advantages and produce relatively little noise, but noise production will still be unavoidable. Speckle interference fringed images obtained by any mode present problems such as image blur, low resolution and high level of inherent noise, which disadvantage further research. Therefore, the exploration of denoising methods is very significant.
    By denoising, we mean filtering out noise that is unfavorable to the collection and use of target information contained in images. During denoising, the degree of protection on useful information, the level of residual noise and the image clarity are all indicators for evaluating the effects of denoising, and also standards for testing the benefits and
    ; BE2021/5108 the disadvantages of denoising methods. With the increasingly extensive research on ESPI image denoising algorithms in China and abroad in recent years, commonly used methods include spatial domain filtering and transformation domain filtering, including the mean filtering, median filtering, Wiener filtering, Fourier transform and wavelet transform. These methods will reduce image contrast while filtering out noise, leading to image blurring, and image fringe structure will be destroyed by some levels, which will affect subsequent phase extraction and calculation. .
    With the advancement in technology, the requirements on the denoising of speckle interference fringed images becoming higher and higher, the inventor found that the existing denoising methods cannot achieve desired effects and the level of ESPI image noise is relatively high and the fringes show great diversity. Existing methods fail to achieve denoising by guaranteeing contrast and edge detail information of fringed images, although denoising and edge information protection are two aspects that need to be balanced and coordinated when researching methods. denoising of ESPI images; With the mature of current deep learning technology, the nerve network have also been applied to image research, including image recognition, image classification and others, and the structure of the nerve network related to denoising imagery is also constantly evolving and innovating, but the structure of the convolutional nerve network used for speckle interference fringed image denoising remains relatively simple, and due to training data limitations, processing on ESPI images by a denoising network model also has some limitations; due to the influences of noise residual and information loss, the results of corresponding phase information extraction, residual prediction, edge detection, object deformation measurement and d 'other operations are not ideal.
    DESCRIPTION OF THE DISCLOSURE To solve the above problems, the present disclosure proposes a deep learning-based ESPI image denoising method and system, which first consists of training a new proposed convolutional nerve network model on the basis of self-adaptive BM3D-TV, then denoise real speckle interference fringed images with the trained model, thereby protecting the fringe information while ensuring the denoising effect. The above solution has good universality of ESPI images. According to a first aspect of the example of the present disclosure, a method for denoising ESPI images based on deep learning is provided, which comprises: constructing a training data set: constructing a convolutional nerve network model based on self-adaptive BM3D-TV algorithm
    ; BE2021/5108 train said convolutional nerve network model with the training dataset; inputting a speckle interference fringe image to be denoised into the trained model, and denoising the speckle interference fringe image.
    In addition, said training dataset includes fringed images without noise, simulated speckle interference fringed images with noise, and real speckle interference fringed images collected by ESPI technology; said simulated speckle interference fringed images with noise are obtained by adding additive noise and multiplicative noise to the fringed images.
    In addition, to ensure the number of real speckle interference fringed images in the training dataset, the collected real speckle interference fringed images will be clipped and rotated, in order to improve data. and to obtain an expansion of the number of real images.
    In addition, the self-adaptive BM3D-TV algorithm is used to denoise the real speckle interference fringed images, and the denoised real speckle interference fringed images are used as the output of the nerve net training. convolutional.
    Moreover, the target function of the self-adaptive BM3D-TV algorithm is expressed as: min, (7) = (/PV" ad +(2/2) | aso] dry where, is a fringed image d information with noise to be processed, P(X,Y) is a self-adaptive parameter, /zusp is an image of results obtained after denoising by a BM3D filter, ©) is an image area and À is a Lagrange multiplier.
    In addition, the convolutional nerve network includes a plurality of layers, and the ReLU activation function is adopted, except for the first and last layers of the convolutional nerve network, the middle layers will undergo batch normalization after the operation of convolution.
    In addition, the convolutional nerve network model based on the self-adaptive BM3D-TV algorithm can perform model training by taking the actual ESPI image (as input, and the signal from the input processed with the model being as output), the simulated fringed image with no noise (output) and the simulated fringed ESPI image with noise (input) as the training sample of the two channels in the network model, and taking the actual ESPI image denoised by the self-adaptive BM3D-TV algorithm as model output; the two channels are the real image drive channel and the simulated image drive channel, the input of the real image drive channel is the real ESPI image, and the output is l real processed image (self-adaptive BM3D-TV filtering and data reinforcement) inside the network model; the simulated image drive channel input is the simulated ESPI image,
    and the output is the simulated ESPI image without noise before adding noise. BE2021/5108 According to a second aspect of the example of the present disclosure, an ESPI image denoising system based on deep learning is provided, which comprises: a training data acquisition module, the assembly of acquired training data including fringed images without noise, simulated speckle interference fringed images with noise, and real speckle interference fringed images collected by ESPI technology: a model building module, for constructing the convolutional nerve network model based on the self-adaptive BM3D-TV algorithm and training said convolutional nerve network model with the training data set: a real ESPI image acquisition module, for acquiring ESPI fringe images of an object by electron speckle interferometry technology; a denoising processing module, for inputting the real ESPI fringed images to be processed into the trained model and outputting the denoised ESPI fringed images.
    According to a third aspect of the example of the present disclosure, electrical equipment is provided, which comprises a memory, a processor and an executable computer program stored in the memory, the processor implementing said ESPI image denoising method based on deep learning by running the program.
    According to a fourth aspect of the example of the present disclosure, a computer-readable storage medium is provided in which is stored a computer program, the processor implementing said ESPI image denoising method based on deep learning by running the program.
    Compared with existing techniques, the present disclosure has the following beneficial effects: The present disclosure proposes a convolutional nerve network model based on the self-adaptive BM3D-TV algorithm, which allows training of the model by taking the image real ESPI, the simulated fringed image without noise, and the simulated fringed ESPI image with noise as the training sample of the two channels in the network model, and taking the real ESPI image denoised by the BM3D-TV algorithm self-adaptive as model output; and — denoising the real speckle interference fringed image with the trained model. The method described in the present disclosure enables denoising to be performed on ESPI images, which not only guarantees low noise residual, but also provides effective protection on the edge information of fringed images, moreover, ESPI images can have good adaptability, which facilitates batch image processing and builds a basis for subsequent operations such as phase extraction.
    DESCRIPTION OF FIGURES BE2021/5108 This disclosure will be better understood with reference to the appended figures, and the examples given to explain this disclosure are given solely by way of illustrative and non-limiting example.
    Figure 1(a) is a simulated ESPI image of the noiseless fringed image generated by simulation in Example I of this disclosure; Figure 1(b) is a simulated ESPI image with noise generated by simulation in Example I of the present disclosure; Figure 2 is a real speckle interference fringed image in Example I of the present disclosure; Figure 3(a) is a sectional view of Figure 1(a) in Example I of the present disclosure; Figure 3(b) is a sectional view of Figure 1(b) in Example I of the present disclosure; Figure 3(c) is a sectional view of Figure 2 in Example I of the present disclosure; Figure 4 is a step diagram of the BM3D-based self-adaptive TV denoising algorithm in Example I of the present disclosure; Figure 5 shows the structure of the network model in Example I of this disclosure; Figure 6(a) to Figure 6(b) is a view and a cross-sectional view of the actual speckle interference fringed image in Example I of this disclosure after denoising with the PCNN model; Figure 6(c) to Figure 6(b) is a view and a cross-sectional view of the actual speckle interference fringed image in Example I of the present disclosure after denoising with the SRCNN model: Figure 6(e) to Figure 6(f) is a view and a sectional view of the actual speckle interference fringed image in Example I of the present disclosure after denoising with the DnCNN model; Figure 7(a) is a view of the actual speckle interference fringed image in Example I of this disclosure after denoising by the self-adaptive BM3D-TV based convolutional nerve network algorithm; Figure 7(b) is a cross-sectional view of the real speckle interference fringed image in Example I of the present disclosure after denoising by the BM3D-TV based convolutional nerve network algorithm auto- adaptive.
    DETAILED PRESENTATION OF EMBODIMENT This disclosure can be described below in more detail via the exemplary embodiments and the appended figures.
    It should be noted that the detailed description below is only illustrative to better understand the
    ° BE2021/5108 this disclosure. It should be noted that, unless otherwise indicated, all technical and scientific terms used in this disclosure have the same meaning as those well known to those skilled in the art.
    It should be noted that the terms used herein are intended to describe embodiments, rather than limit the illustrative embodiment according to the present application. Unless otherwise indicated, the singular form includes the plural, further, the words used herein "understand" and/or "include", indicate the presence of the characteristic, step, work, device, component and/or the combination of these. Example I: A preferred embodiment of the present disclosure is described below. Figure 1(a) shows an example of a simulated fringed image, which consists of firstly generating 300 noiseless simulated fringed images by a computer and saving them as an output sample for model training, whose size is 180*180. Figure 1(b) shows an example of simulated fringed image, which consists of adding simulated noise to the fringed image, including additive noise and multiplicative noise, and the details of the added noise are as follows: I. is a fringed image with no noise, M and M, are random Gaussian noises, / is a simulated speckle interference fringed image with noise thus one can obtain an expression of the relationship between the three: Ii, fn +n, where, fl and M respect a random Gaussian distribution and can be expressed as follows (based on python): n = np.randon.normal(loa = a, scale= b/3.5æ = 180*180) n, = np.randon.normal(loa = a,scale= b, Sze= 180*180) where, &@ is an average value and À is the standard deviation. A speckle interference fringed image is thus generated and recorded. Figure 2 shows an example of the real speckle interference fringed image. By collecting an actual ESPI image and then expanding it to 100 images by the data enhancement method (it is achieved mainly by cutting and rotating here), then adjusting the images to an identical size of 180*180, the images will be used with the simulated images as a training sample of the model, providing supporting data for subsequent trials. Figure 3 (a) - Figure 3 (c) are the cross-sectional views of Figure 1 (a), Figure 1 (b) and Figure 2 above, the cross-sectional figure is a line obtained by connecting the gray values of the points corresponding to all x values, in which the gray level value is the y axis and the
    | BE2021/5108 continuous pixel position in the image is the x axis. The cut figure helps to better analyze the noise and fringing information of the image, which is one of the effective ways to evaluate the image quality.
    Figure 4 is a step diagram of the self-adaptive TV denoising algorithm based on BM3D; In particular, the BM3D method consists of two steps: the initial estimate and the final estimate, the details of which are as follows: (a) Initial estimate Retrieve approximate blocks from the image, overlay the results to form a matrix three-dimensional and divide into groups; making the hard threshold values on the three-dimensional coefficients in the three-dimensional frequency domain to eliminate noise, and performing a hard threshold value filtering operation by three-dimensional inverse conversion; weighting all the estimated values at the same position and taking the average value, the data obtained being the average value of the pixels at this position, so as to carry out the initial estimation of the real image and carry out the aggregation.
    (b) Final estimation Perform block adaptation on the initial estimation result image, and build an initial estimation result image and a 3D matrix corresponding to the noise image according to the corresponding area coordinates similar blocks; perform a 3D conversion on the two 3D matrices constructed respectively in the step above, perform a Wiener filtering on the conversion coefficients of the 3D matrix composed of the image with noise by taking an energy spectrum of the image real as energy spectrum of the image of the initial estimate, then perform a 3D conversion to obtain denoised image blocks and return them to its positions; acquiring a final estimated value of each pixel by taking the weighted average value of all the estimated values at the same position to perform an aggregation.
    We take / ij To represent at the gray level value at the center of the image x =ihy,= jh , 1,j=0,1,2--N, Nh=L and L are each the length of l 'original image, À is the spatial step, / (x,%t,) is the result of #7 iterations of the algorithm, and the data of this result is denoted as / L p where, 1 =n- At ‚ At is the time step.
    The diffusion term of the self-adaptive TV denoising algorithm based on BM3D can be expressed as follows:
    v. pvI |_ Il, Add +1 Ts vz I;+1, Substitute the difference quotient for the partial derivative to obtain: (I, X, = Ti, JI (, ), 7 1 LA (In X, = Li 15 —21/, Le (1, X, 2 + U, X, = Tae Ia ne +1 F5 Consequently, the discrete iteration form in the Euler-Lagrange equation of the self-adaptive TV denoising model can be expressed as follows: VI.
    FRI — AT”, —1° + A V.
    PV 1,7 1,J 1,7 1,J n 2-p vr] where, M is the number of iterations, À is a Lagrange multiplier, ;,j=0,1,2---N, and boundary conditions satisfy ls, = I, ; ri =p and La = Ju = Pt . A brief process of the BM3D-based self-adaptive TV denoising algorithm can be expressed as follows: Step 1, input real ESPI image as input sample for network training; Step 2, initialize parameters: fr dee [ nu n=0, At=04 1 =l, VI ‚ the parameters of the fidelity term À are constructed by | /.- | and|/,, | ; Step 3, rerun steps (1) and (2) if 7 is less than the maximum number of iterations ‘ = fr.
    The LA Step (1) N=N+1 , calculate Li by discrete iteration; VI” A ;( nor PVE) Step (2), perform an operation to obtain the diffusion term Vis, : Step 4, output the results as an output sample for network training after the iteration is complete.
    The detailed steps of the BM3D-based self-adaptive TV denoising algorithm can be expressed as follows: The first step is to write the operator of the BM3D denoising algorithm as BM3D(-), and the image of results obtained after denoising by a BM3D filter can be expressed as follows: BE2021/5108! yo = BM3D(1) where, { is an information fringed image with noise to be processed. According to the corresponding content of the total variation above, the self-adaptive TV algorithm can be expressed as follows: pley) 2 Jo) = ple x) {97° dev + (4/2) [1-14] ded where, P(X, Y) is a self-adaptive parameter and can be expressed as follows: p(x, y) =1+ li + VG, “1, (x, y | G, is a Gaussian filter, © >0, ©) is an area of the image, À is a Lagrange multiplier, ff is an object to be filtered, ie an image with noise, and /, is an image without noise.
    Therefore, the new fidelity term in the self-adaptive TV algorithm can be written as: À 2 Prev = [ZU Teas) dx By synthesizing the new fidelity elements Fa, A self-adaptive TV denoising model based on BM3D can be obtained, which is expressed as follows: . pley) 2 min Jot) = (/p(e”) {V7 dxdy + (3/2)[, [I To dxdy In addition, it is possible to obtain an Euler-Lagrange equation of the auto TV denoising model -adaptive, which can be expressed as: vpn vr} AU Tas) = 0 A gradient downflow in the self-adaptive TV denoising model can be expressed as: 61/01 = -{9I/ Nij "AT Toys) Figure 5 shows the structure of the network model of the present disclosure.
    Taking the actual ESPI images before and after processing in the corresponding step in Figure 4, the simulated fringed image without noise and the simulated fringed ESPI image with noise as a training sample of the two channels in the model network, the two channels are the actual image training sample channel and the simulated image training sample channel, or for one channel, the input is the ESPI image real and the output is the real processed ESPI image (self-adaptive BM3D-TV filtering and data reinforcement) inside the network model; and for the other channel, the input is the simulated ESPI image and the output is the simulated ESPI image without noise before adding noise, which are used to train the model.
    The depth of the convolutional nerve network in the present example is 17 layers, the layers of the convolution + activation functions (Conv + ReLU): as the first layer of the network, 64 filters with dimensions of 3*3*C are used to generate 64 characteristic images; where, © represents the number of image channels, the processed image is a grayscale image, so the number of channels is 1; the activation function in the network is a ReLU function; the convolution + batch normalization + activation (Conv + BN + ReLU) function layers: as layers 2 to 16, 64 filters with dimensions of 3x 3x 64 are used, and batch normalization is added between the functions convolution and activation; the convolution layer, as the last layer, consists of CE filters having dimensions of 3x 3x 64 to generate the output image. Then it is appropriate to combine the self-adaptive TV denoising algorithm based on BM3D and the convolutional nervous network above, the structure of the model after the combination being shown in Figure 5. We aim to predict the image after denoising on the original image with the convolutional nerve network, assuming the original image is Ji, the processed image is X, and the noise is % thus we get X AM and the end goal is obtain a mapping F(y)=x to predict the image after denoising; the convolutional nerve network in the present disclosure results in a residual mapping R(y;0)= V, with the residual, where, is a function adjustable with the training of the network, in order to predict the image X” YT R(y ,0) after denoising based on the residual mapping. Finally, taking the root mean square deviation between the actual residual and the predicted value as the loss function, a training adjustment on the updatable & parameter in the network structure will be performed. The loss function is expressed as 1 N (9) =)
    N WHERE, fx). represents N groups of images with and without noise, ie training groups of input and output images. During training, the loss function can be reduced by adjusting &, in order to get a proper residual mapping and further get a noise-free image by calculating the difference with the noise-free image. During back-propagation, trainable parameters in the network can be learned by minimizing the loss function by the self-adaptive moment estimation algorithm. ü Through backpropagation, the parameters in the CNN are updated again and again, and the value of the loss function is gradually reduced to complete the training of the model. For batch normalization during this process, the minimum training set size B is 128, and for each activation function x***, the output after normalization is vi The normalization operation of the training set is expressed as follows: BN 4 Xa 77 Vaag The EM process can be expressed as follows: The input is B=X%, ‚2; the parameters to learn is + and PB. The output is {x=BN, „(x)}, the average value of the minimum training set is 1 12 expressed as Ag “DEX, the variance of the minimum training set is expressed ia) TT ap ot has like 1285! , O,+E after normalization, the output after batch normalization is Y,=+X +B= BN, (X). Where, 7 and А can all be learned, and Ÿ is the input to the next network layer and will be passed forward, during training the input filter image block dimensions will be set to 50x50, 128x3000 image blocks will be cut to train the model, the initial learning rate is 10, and the number of iterations is 300. After the training is completed, the denoising network model will be tested with the sample images of the test set, the test set comprising 60 real ESPI images, the images being used as the input to the network and the output being the denoised image.
    Figure 6(a) to Figure 6(f) are respectively a view and a cross-sectional view of the real speckle interference fringed image after denoising with the DnCNN, SRCNN and DnCNN model; Figure 7(a) to Figure 7(b) are each a view and a cross-sectional view of the real speckle interference fringed image after denoising by the BM3D-TV based convolutional nerve network algorithm self-adaptive.
    Example II: This example aims to provide an ESPI image denoising system based on deep learning. For the above purpose, the present example provides a deep learning based ESPI image denoising system, which includes: a training data acquisition module, the acquired training data set including noise-free fringed images, simulated speckle interference fringed images
    N BE2021/5108 with noise and real speckle interference fringed images collected by ESPI technology; a model building module, for building the convolutional nerve network model based on the self-adaptive BM3D-TV algorithm and training said convolutional nerve network model with the training dataset; a real ESPI image acquisition module, to acquire ESPI fringed images of an object by electron speckle interferometry technology; a denoising processing module, for inputting the real ESPI fringed images to be processed into the trained model and outputting the denoised ESPI fringed images.
    Example HI: This example is intended to provide electrical equipment.
    The electrical equipment includes a memory, a processor, and an executable computer program stored in the memory, the processor performing the following steps in executing the program, including: constructing a training data set; build a convolutional nerve network model based on the self-adaptive BM3D-TV algorithm train the above convolutional nerve network model with the training dataset; inputting a speckle interference fringe image to be denoised into the trained model, and denoising the speckle interference fringe image.
    Example IV: This example is intended to provide a computer readable storage medium.
    A computer-readable storage medium is provided in which a computer program is stored, the processor performing the following steps in executing the program, including: constructing a training data set: constructing a convolutional nerve network model based on the self-adaptive BM3D-TV algorithm — train the above convolutional nerve network model with the training dataset; inputting a speckle interference fringe image to be denoised into the trained model, and denoising the speckle interference fringe image.
    The ESPI image denoising method and system based on deep learning provided in the examples below are quite feasible and have wide application prospects.
    The examples below are given only by way of example of preference, rather than limiting the present disclosure, and modifications and variations on the present disclosure by those skilled in the art are of course possible. Any modifications, equivalent replacements, and enhancements that adhere to the spirits and principles of this disclosure shall be included within the scope of the protection of this disclosure.
    Although the present disclosure has been described in detail by the embodiments with reference to the figures, they do not constitute a limitation of the scope of protection of the present disclosure. Those skilled in the art will understand that any modifications or variations without creative work should be included within the scope of the protection of this disclosure.
    权利要求:
    Claims (10)
    [1]
    1. Method for denoising ESPI images based on deep learning, characterized by comprising: constructing a training dataset: constructing a convolutional nerve network model based on the BM3D-TV algorithm auto- adaptively training said convolutional neural network model with the training dataset; inputting a speckle interference fringe image to be denoised into the trained model, and denoising the speckle interference fringe image.
    [2]
    A deep learning based ESPI image denoising method according to claim 1, characterized in that, said training data set comprises noiseless fringed images, simulated speckle interference fringed images with noise and real speckle interference fringed images collected by ESPI technology; said simulated speckle interference fringed images with noise are obtained by adding additive noise and multiplicative noise to the fringed images.
    [3]
    3. Method for denoising ESPI images based on deep learning according to claim 2, characterized in that, to guarantee the number of real images with speckle interference fringes in the training data set, the actual speckle interference fringed images collected will be clipped and rotated, in order to improve data and achieve an expansion of the number of actual images.
    [4]
    4. Method for denoising ESPI images based on deep learning according to claim 1, characterized in that the self-adaptive BM3D-TV algorithm is used to denoise the real images with speckle interference fringes, and the real denoised speckle interference fringed images are used as the output of convolutional nerve network training.
    [5]
    S. ESPI image denoising method based on deep learning according to claim 1, characterized in that, the target function of the self-adaptive BM3D-TV algorithm is expressed as follows: mind, ()=0/ plc 2) VI" ad + (A/2)[ 7 aus] ded where, £ is an information fringe image with noise to be processed, P(XY) is a self-adaptive parameter, Zap is a result image obtained after denoising by a BM3D filter, © is an image area and λ is a Lagrange multiplier.
    [6]
    6. ESPI image denoising method based on deep learning according to claim 1, characterized in that, the convolutional nervous network comprises a plurality of layers, each of the layers adopts the ReLU activation function, except of the first and last layers of the
    © BE2021/5108 convolutional nerve network, the middle layers will undergo batch normalization after the convolution operation.
    [7]
    7. Method for denoising ESPI images based on deep learning according to claim 1, characterized in that, the convolutional nervous network model based on the self-adaptive BM3D-TV algorithm makes it possible to carry out a training of the model in taking the real ESPI image, the simulated fringed image with no noise, and the simulated fringed ESPI image with noise as the training sample of the two channels in the network model, and taking the real ESPI image denoised by the self-adaptive BM3D-TV algorithm as model output.
    [8]
    8. A deep learning based ESPI image denoising system characterized by comprising: a training data acquisition module, the acquired training data set comprising fringed images without noise, simulated speckle interference fringed images with noise and real speckle interference fringed images collected by ESPI technology: a model building module, to build the convolutional nerve network model based on the self-adaptive BM3D-TV algorithm and train said convolutional nerve network model with the training data set: a real ESPI image acquisition module, to acquire the ESPI fringed images of an object by the technology electron speckle interferometry; a denoising processing module, for inputting the real ESPI fringed images to be processed into the trained model and outputting the denoised ESPI fringed images.
    [9]
    9. Electrical equipment, characterized in that it comprises a memory, a processor and an executable computer program stored in the memory, the processor implementing said ESPI image denoising method based on deep learning according to one any of claims 1 to 7 by running the program.
    [10]
    10. Computer-readable storage medium, characterized in that in which a computer program is stored, the processor implements said ESPI image denoising method based on deep learning according to any one of claims 1 to 7 by running the program.
    类似技术:
    公开号 | 公开日 | 专利标题
    EP3138079B1|2019-09-18|Method of tracking shape in a scene observed by an asynchronous light sensor
    EP3221841B1|2018-08-29|Method and device for the real-time adaptive filtering of noisy depth or disparity images
    EP2584533A1|2013-04-24|Method for 3-D reconstruction of a scene object
    CN108447113B|2021-06-08|Impulse type intensity correlation three-dimensional imaging method based on deep learning model
    BE1027609B1|2022-01-17|ESPI image denoising method and system based on deep learning
    EP3572976A1|2019-11-27|Method for processing a video image stream
    Zhou et al.2021|Photoacoustic microscopy with sparse data by convolutional neural networks
    EP2294552B1|2011-11-09|Improved device for processing images
    EP0410826B1|1994-05-04|Iterative motion estimation process, between a reference image and a current image, and device for canying out the process
    Yang et al.2010|Remote sensing image restoration using estimated point spread function
    FR2854247A1|2004-10-29|Treatment of cubed seismic data with differing delays and angles of incidence, comprising estimating intersection corresponding with differing delays and angles of incidence
    EP2943935B1|2017-03-08|Estimation of the movement of an image
    EP3918576A1|2021-12-08|Dynamic three-dimensional imaging method
    FR2933520A1|2010-01-08|METHOD AND DEVICE FOR RESTORING A VIDEO SEQUENCE
    Zhang et al.2021|Hyperspectral Image Denoising With Realistic Data
    BE1026937B1|2020-07-28|Image segmentation method
    Krishnan et al.2020|A Novel Underwater Image Enhancement Technique using ResNet
    FR3075437A1|2019-06-21|IMAGE SYNTHESIS METHOD
    Zhang et al.2011|Super-resolution reconstruction Algorithm for PMD range image based on regularization
    EP3707676A1|2020-09-16|Method for estimating the installation of a camera in the reference frame of a three-dimensional scene, device, augmented reality system and associated computer program
    WO2018185104A1|2018-10-11|Method for estimating pose, associated device, system and computer program
    Prajapati et al.2021|Direct Unsupervised Super-Resolution Using Generative Adversarial Network | for Real-World Data
    Huang et al.2020|Deblurring approach for motion camera combining FFT with α-confidence goal optimization
    FR3066633A1|2018-11-23|METHOD FOR DEFLOWING AN IMAGE
    Su et al.2017|A motion deblurring method of a light field camera
    同族专利:
    公开号 | 公开日
    CN111768349A|2020-10-13|
    BE1027609A1|2021-04-26|
    引用文献:
    公开号 | 申请日 | 公开日 | 申请人 | 专利标题
    法律状态:
    优先权:
    申请号 | 申请日 | 专利标题
    CN202010517755.4A|CN111768349A|2020-06-09|2020-06-09|ESPI image noise reduction method and system based on deep learning|
    [返回顶部]