Full-field displacement calculation method based on self-statistical noise reduction and Gaussian enhancement

Abstract

The invention relates to the technical field of structure test, and relates to a full-field displacement calculation method based on self-statistical noise reduction and Gaussian enhancement, comprising the following steps: scaling a first image and a second image to obtain a first sub-pixel image and a second sub-pixel image; moving the first sub-pixel image based on a displacement threshold, storing a plurality of obtained first intermediate images as a multi-channel first tensor image, and performing edge expansion on the second sub-pixel image based on the displacement threshold to obtain a second tensor image; determining second intermediate data according to the channel data difference value between the single channel of the second tensor diagram and each channel of the first tensor diagram, performing pixel cross-correlation calculation based on each channel data of the second intermediate data, and determining sub-pixel displacement data; according to the regional pixel statistical values of the sub-pixel displacement data under different convolution scales, carrying out self-noise reduction on the sub-pixel displacement data to obtain noise-reduced displacement data; and performing Gaussian enhancement on the noise reduction displacement data to obtain a target full-field displacement result.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of structural testing, and more particularly, to a full-field displacement calculation method based on self-statistical denoising and Gaussian enhancement. BACKGROUND

[0002] Surface displacement measurement of materials can quantify the slight position changes of the surface of materials under stress, temperature changes or other external conditions, and provide basic data support for scenarios such as mechanical property analysis of materials, structural health monitoring, dynamic behavior research, and process optimization. The measurement object can be the surface of any solid material such as metal, composite material, biological tissue, etc.

[0003] In current surface displacement testing technologies, contact measurement methods require the sample to be pasted or clamped, introducing additional mass / stiffness, changing the local stress state (especially for thin-walled / soft materials), and only obtaining local data at the pasting points (strain gauges) or gauge sections (extensometers), which cannot capture the full-field deformation distribution, and are prone to large deformation scenarios. Optical non-contact measurement, such as single-point scanning based on LDV (Laser Doppler Vibrometer), results in extremely long time consumption for full-field measurement, which cannot monitor dynamic deformation in real time. Depending on the surface optical properties, rough / solar surface generates noise due to speckle effect, resulting in low measurement accuracy. High cost and complex calibration restrict the popularization of engineering. The photoelastic method requires transparent birefringent materials (such as epoxy resin), which cannot be directly used for common engineering materials such as metal / concrete; and the displacement field needs to be calculated by secondary integration of the strain field, with significant error accumulation. The current measurement method based on DIC (Digital Image Correlation) technology still has the disadvantages of low efficiency, inability to meet real-time monitoring requirements, insufficient noise suppression, weak micro-displacement visualization, and low quality, and is limited by expensive licensing fees, with low deployment flexibility.

[0004] It should be noted that the information disclosed in the above background section is only used to strengthen the understanding of the background of the present disclosure, and therefore can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0005] The purpose of the present disclosure is to provide a full-field displacement calculation method based on self-statistical denoising and Gaussian enhancement, thereby improving the full-field displacement calculation speed, enhancing the denoising effect and realizing the enhancement of small displacement, realizing high real-time and high-quality full-field displacement calculation, damage area identification and detection, crack propagation path capture, etc.

[0006] Other characteristics and advantages of the present disclosure will become apparent from the following detailed description, or will be learned by practice of the present disclosure.

[0007] According to one aspect of the present disclosure, a full-field displacement calculation method based on self-statistical noise reduction and Gaussian enhancement is provided, comprising: performing scaling processing on a first image and a second image of a frame after the first image to obtain a first sub-pixel image and a second sub-pixel image, respectively, the first image and the second image being single-channel gray images; performing edge expansion and pixel movement on the first sub-pixel image based on a displacement threshold, storing a plurality of obtained first intermediate images as a multi-channel first tensor image, and performing edge expansion on the second sub-pixel image based on the displacement threshold to obtain a single-channel second tensor image, the second tensor image having the same size as the first tensor image; determining second intermediate data according to channel data difference values of the single channel of the second tensor image and each channel of the first tensor image, and performing pixel cross-correlation calculation based on each channel data of the second intermediate data to determine sub-pixel point displacement data or regional displacement data between the second image and the first image, the second intermediate data having the same number of channels as the first tensor image; performing self-noise reduction on the sub-pixel point displacement data or the regional displacement data according to regional pixel statistical values of the sub-pixel point displacement data or the regional displacement data under different convolution scales to obtain noise reduction displacement data; and performing Gaussian enhancement on the noise reduction displacement data to obtain a target full-field displacement result.

[0008] In an exemplary embodiment of the present disclosure, performing edge expansion and pixel movement on the first sub-pixel image based on a displacement threshold, and storing a plurality of obtained first intermediate images as a multi-channel first tensor image, and performing edge expansion on the second sub-pixel image based on the displacement threshold to obtain a single-channel second tensor image, comprises: determining the number of movable steps in the horizontal and vertical directions according to the displacement threshold; expanding the edges of the first sub-pixel image based on the displacement threshold, moving the first sub-pixel image in the horizontal and vertical directions within the limit range of the expanded edges according to the number of movable steps, and determining a plurality of first intermediate images according to the movement results to store the plurality of intermediate images as the first tensor image; and expanding the edges of the second sub-pixel image based on the displacement threshold to obtain the second tensor image, so that the second tensor image has the same size as the first tensor image.

[0009] In an exemplary embodiment of the present disclosure, determining second intermediate data according to channel data difference values of the single channel of the second tensor image and each channel of the first tensor image comprises: for each channel of the first tensor image, obtaining data difference values of the channel and the single channel of the second tensor image to obtain difference value data of each channel of the first tensor image; for the difference value data of each channel of the first tensor image, performing a squaring operation on the difference value data to obtain difference square data of each channel of the first tensor image; and according to the displacement threshold, removing the data of the peripheral edges from the difference square data of each channel, respectively, and determining the second intermediate data according to the difference square data after removing the data of the peripheral edges.

[0010] In an example embodiment of the present disclosure, the pixel cross-correlation calculation is performed based on the channel data in the second intermediate data to determine the sub-pixel point displacement data between the second image and the first image, including: determining, for each sub-pixel, a channel index of the minimum channel data according to the size of the channel data of the corresponding sub-pixel in the second intermediate data; determining the displacement data of the sub-pixel according to the channel index; and determining the sub-pixel point displacement data based on the displacement data of each sub-pixel.

[0011] In an example embodiment of the present disclosure, the pixel cross-correlation calculation is performed based on the channel data in the second intermediate data to determine the sub-pixel point displacement data between the second image and the first image, including: determining, for each sub-pixel, a channel index of the minimum channel data according to the size of the channel data of the corresponding sub-pixel in the second intermediate data; determining the displacement data of the sub-pixel according to the channel index; and determining the sub-pixel point displacement data based on the displacement data of each sub-pixel.

[0012] In an example embodiment of the present disclosure, the sub-pixel point displacement data or the region displacement data is self-de-noised according to the region pixel statistical values of the sub-pixel point displacement data or the region displacement data under different convolution scales to obtain de-noised displacement data, including: determining a first convolution scale and a second convolution scale, the step lengths of the first convolution scale and the second convolution scale are the same, and the kernel size of the second convolution scale is larger than the kernel size of the first convolution scale; for each convolution operation, performing convolution operation on the sub-pixel point displacement data or the region displacement data based on the first convolution scale to obtain a first operation result, and performing convolution operation on the sub-pixel point displacement data or the region displacement data based on the second convolution scale to obtain a second operation result; removing the data at the peripheral edge of the first operation result based on a displacement threshold, so that the size of the processed first operation result is the same as that of the second operation result; comparing the second operation result with the processed first operation result, and self-de-noising the sub-pixel point displacement data or the region displacement data according to the obtained comparison result to obtain the de-noised displacement data.

[0013] In an example embodiment of the present disclosure, the kernel size of the second convolution scale is twice the kernel size of the first convolution scale.

[0014] In an example embodiment of the present disclosure, the second operation result is compared with the processed first operation result, and sub-pixel point displacement data or region displacement data is self-de-noised according to a comparison result to obtain de-noised displacement data, including: for each convolution operation, if the size of the processed first operation result and the second operation result satisfies a preset size relationship, it is determined that the data before convolution corresponding to the second operation result is a noise point, and de-noising is performed based on the noise point; if the size of the processed first operation result and the second operation result does not satisfy the preset size relationship, the data before convolution corresponding to the second operation result is retained.

[0015] In an example embodiment of the present disclosure, the preset size relationship indicates that the second operation result is smaller than the processed first operation result by a preset multiple; if the size of the processed first operation result and the second operation result satisfies the preset size relationship, the data before convolution corresponding to the second operation result is assigned a value of 0, thereby achieving the purpose of de-noising.

[0016] In an example embodiment of the present disclosure, the de-noised displacement data is Gaussian enhanced to obtain a target full-field displacement result, including: establishing a Gaussian probability density function model based on displacement data; determining a displacement value interval to be enhanced in the de-noised displacement data, and dividing the displacement value interval into a plurality of sub-intervals; for each sub-interval, the displacement data of the sub-interval is enhanced according to a Gaussian function value corresponding to the Gaussian probability density function model of the sub-region, to obtain enhanced displacement data of the sub-interval; and the calculation result of the target full-field displacement is determined according to the enhanced displacement data of each sub-region.

[0017] The full-field displacement calculation method based on self-statistical noise reduction and Gaussian enhancement in the example embodiment of the present disclosure, on the one hand, performs scaling processing on the first image and the second image of the next frame after the first image, respectively obtaining a first sub-pixel image and a second sub-pixel image, performing edge expansion and pixel movement on the first sub-pixel image based on a displacement threshold, storing a plurality of obtained first intermediate images as a multi-channel first tensor image, performing edge expansion on the second sub-pixel image based on the displacement threshold, obtaining a single-channel second tensor image, and determining second intermediate data according to the channel data difference value of the single channel of the second tensor image and each channel of the first tensor image, and performing pixel cross-correlation calculation based on each channel data of the second intermediate data to determine sub-pixel point displacement data or regional displacement data between the second image and the first image. The process converts the image translation operation into tensor channel storage, and calculates the sum of squares of all translation positions by storing the multi-channel first tensor image and the single-channel second tensor image, compared with the image cross-correlation calculation using CPU, the present disclosure greatly reduces the time complexity of calculation by parallel design based on GPU operation, which is beneficial to reduce the calculation delay and realize real-time full-field displacement feedback. Moreover, by pre-scaling the image to generate sub-pixel level data, the boundary displacement calculation precision can reach 0.1 pixel unit, which helps to improve the displacement detection precision.

[0018] On the other hand, according to the regional pixel statistical value of the sub-pixel point displacement data or the regional displacement data under different convolution scales, the sub-pixel point displacement data or the regional displacement data is self-de-noised to obtain de-noised displacement data. Without introducing other judgment criteria, this process can benefit from the statistical characteristics of the displacement field itself, decouple the de-noising decision and the displacement calculation, identify noise points in any displacement data value region, and achieve effective self-de-noising. On the other hand, the de-noised displacement data is subjected to Gaussian enhancement to obtain the target full-field displacement result. This process can enhance the small displacement values after the pre-processing process, improve the visualization of the micro-displacement segment, and improve the visualization of the micro-displacement segment.

[0019] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present disclosure. BRIEF DESCRIPTION OF DRAWINGS

[0020] The above and other objects, features and advantages of the present disclosure exemplary embodiments will be more apparent from the following detailed description read in conjunction with the accompanying drawings, in which several embodiments of the present disclosure are shown by way of example, and not limitation. In the drawings:

[0021] Figure 1 A flowchart of a full-field displacement calculation method based on self-statistical noise reduction and Gaussian enhancement according to an example embodiment of the present disclosure is shown;

[0022] Figure 2 A schematic diagram of a pre-processing procedure of image data is shown according to an example embodiment of the present disclosure;

[0023] Figure 3 A schematic diagram of extending the edge of a second sub-pixel image by one circle is shown according to an example embodiment of the present disclosure;

[0024] Figure 4 A flowchart of obtaining a first tensor map and a second tensor map is shown according to an example embodiment of the present disclosure;

[0025] Figure 5 A schematic diagram of moving a first sub-pixel image horizontally and vertically once within the limit range of the extended edge according to the number of movable steps is shown according to an example embodiment of the present disclosure;

[0026] Figure 6 A schematic diagram of a first tensor map and a second tensor map is shown according to an example embodiment of the present disclosure;

[0027] Figure 7 A schematic diagram of determining second intermediate data is shown according to an example embodiment of the present disclosure;

[0028] Figure 8 A flowchart of determining sub-pixel point displacement data or region displacement data is shown according to an example embodiment of the present disclosure;

[0029] Figure 9 A schematic diagram of obtaining the channel index of the minimum channel data is shown according to an example embodiment of the present disclosure;

[0030] Figure 10 A flowchart of another determination of sub-pixel point displacement data or region displacement data is shown according to an example embodiment of the present disclosure;

[0031] Figure 11 A schematic diagram of performing a convolution operation once on the second intermediate data channel by channel is shown according to an example embodiment of the present disclosure;

[0032] Figure 12 A schematic diagram of determining the channel index of the minimum convolution result (the minimum value index in the channel dimension of the convolution result) is shown according to an example embodiment of the present disclosure.

[0033] Figure 13 A schematic diagram of a convolution operation is shown according to an example embodiment of the present disclosure.

[0034] Figure 14 A schematic diagram of denoising the noise points existing in one-dimensional data is shown according to an example embodiment of the present disclosure.

[0035] Figure 15 A schematic diagram of a first image is shown according to an example embodiment of the present disclosure.

[0036] Figure 16 A schematic diagram of a second image is shown according to an example embodiment of the present disclosure.

[0037] Figure 17 A schematic diagram of a first processing result (sub-pixel displacement result) is shown according to an example embodiment of the present disclosure.

[0038] Figure 18 A schematic diagram of a second processing result is shown according to an example embodiment of the present disclosure.

[0039] Figure 19 A schematic diagram of a third processing result is shown according to an example embodiment of the present disclosure.

[0040] Figure 20 A schematic diagram of a fourth processing result is shown according to an example embodiment of the present disclosure.

[0041] Figure 21 A schematic diagram of a fifth processing result is shown according to an example embodiment of the present disclosure.

[0042] Figure 22 A schematic diagram of a sixth processing result is shown according to an example embodiment of the present disclosure.

[0043] Figure 23 A schematic diagram of a seventh processing result is shown according to an example embodiment of the present disclosure.

[0044] In the drawings, the same or corresponding parts are denoted by the same or corresponding reference numerals. DETAILED DESCRIPTION

[0045] Example embodiments now will be described more fully hereinafter with reference to the accompanying drawings. Example embodiments, however, can be implemented in many different forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of example embodiments to those skilled in the art. Like reference numerals refer to like elements throughout the figures, and thus description of the same will be omitted.

[0046] Moreover, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided to give a thorough understanding of embodiments of the disclosure. One skilled in the relevant art will recognize, however, that the techniques described can be practiced without one or more of the specific details, or with other methods, components, materials, and so forth. In other instances, well-known structures, materials, or operations are not shown or described in detail in order to avoid obscuring aspects of the disclosure.

[0047] The block diagrams in the drawings show only the functionality of the features and can not imply a necessity of any particular physical or architectural arrangement. That is, the functionality of the features can be implemented in software, hardware, or a combination thereof, and can be implemented in one or more modules or components. The functionality of the features can be implemented in one or more software modules or components, or in a hardware module or component, or in a hardware module or component that is software-configured.

[0048] In the current surface displacement test technology, the contact measurement method (such as strain gauge / extensometer method) needs to paste or clamp the sample, introduces additional mass / stiffness, changes the local stress state (especially for thin-walled / soft materials), can only obtain local data at the pasting point (strain gauge) or the gauge length (extensometer), cannot capture the full-field deformation distribution, and is prone to large deformation scenarios. Optical non-contact measurement, such as single-point scanning based on LDV, results in extremely long time consumption for full-field measurement, and cannot monitor dynamic deformation in real time. Depending on the surface optical properties, rough / slightly absorbing surfaces produce noise due to speckle effect, and the measurement accuracy is low. High cost and complex calibration restrict the popularization of engineering. The photoelastic method requires transparent birefringent materials (such as epoxy resin), and cannot be directly used for common engineering materials such as metals and concrete; and the displacement field needs to be calculated by secondary integration of the strain field, and the error accumulation is significant. The current measurement method based on DIC technology still has the disadvantages of low efficiency, inability to meet real-time monitoring requirements, insufficient noise suppression, and weak micro-displacement visualization, and is limited by expensive licensing fees and low flexibility.

[0049] Specifically, the current DIC technology has low calculation efficiency, high-resolution image processing delay > 10 seconds / frame, and displacement calculation processing after all photos are taken, which cannot meet the real-time monitoring demand. The root cause is the CPU serial traversal of pixel cross-correlation calculation architecture. The noise suppression of this method is insufficient, and isolated noise points (caused by speckle interference) frequently occur in the displacement field. The traditional maximum and minimum filtering method deletes real micro-deformation data. The root cause is the lack of regional statistical correlation analysis mechanism. Moreover, the micro-displacement visualization of this method is weak, and the contrast in the cloud chart tends to zero in the micro-strain area (<0.1 pixel, such as the crack initiation section), which is difficult to identify with the naked eye. The root cause is that linear enhancement cannot specifically improve the small displacement section, and the post-processing commercial software of the DIC system (such as Vic-3D) has no enhancement algorithm interface, and users cannot independently optimize the micro-strain display. In addition, the commercial DIC software is closed source and has high authorization fees, and the hardware relies on high-end workstations, which has poor deployment flexibility.

[0050] Based on one or more of the above problems, the full-field displacement calculation method based on self-statistical denoising and Gaussian enhancement is provided in the example embodiments of the present disclosure. Through the pixel cross-correlation calculation of the deformable convolution of the sub-pixel level adjustable displacement threshold, the regional deformable self-denoising operation based on statistics, and the Gaussian self-enhancement processing based on the relative size of the displacement data, the variable displacement threshold, low delay, real-time full-field displacement calculation and small displacement segment enhancement are realized, and the accuracy and real-time performance of the full-field displacement calculation are improved.

[0051] It should be noted that the full-field displacement calculation method of the example embodiments of the present disclosure can be applied to material mechanical property analysis, structural health monitoring, dynamic behavior research, preparation process optimization, etc. For example, material fatigue test, high-precision scene such as rock mass instability warning, field scene (such as highway detection, dam monitoring) for full-field displacement detection of common engineering materials such as metal, concrete, rock mass, etc. The example embodiments of the present disclosure do not limit the specific application scenarios.

[0052] As Figure 1 A flowchart of a full-field displacement calculation method based on self-statistical denoising and Gaussian enhancement according to the example embodiments of the present disclosure is shown, as shown in Figure 1 The method can include steps S110 to S150:

[0053] Step S110: scaling the first image and the second image of the next frame after the first image to obtain the first sub-pixel image and the second sub-pixel image, respectively. The first image and the second image are single-channel grayscale images.

[0054] Step S120: edge expansion and pixel movement are performed on the first sub-pixel image based on the displacement threshold value, a plurality of first intermediate images obtained are stored as a multi-channel first tensor image, and edge expansion is performed on the second sub-pixel image based on the displacement threshold value to obtain a single-channel second tensor image, the second tensor image has the same size as the first tensor image.

[0055] Step S130: second intermediate data is determined according to a difference value of channel data of a single channel of the second tensor image and each channel of the first tensor image, and pixel cross-correlation calculation is performed based on each channel data of the second intermediate data to determine sub-pixel point displacement data or region displacement data between the second image and the first image, the second intermediate data has the same number of channels as the first tensor image.

[0056] Step S140: self-de-noising is performed on the sub-pixel point displacement data or the region displacement data according to region pixel statistical values of the sub-pixel point displacement data or the region displacement data under different convolution scales to obtain de-noised displacement data.

[0057] Step S150: Gaussian enhancement is performed on the de-noised displacement data to obtain a target full-field displacement result.

[0058] The full-field displacement calculation method based on self-statistical de-noising and Gaussian enhancement according to the example embodiments of the present disclosure improves the full-field displacement recognition speed, enhances the de-noising effect and realizes enhancement on small displacement, and realizes high real-time and high-quality full-field displacement calculation.

[0059] The steps S110 to S150 are described in more detail below.

[0060] In step S110, the first image and the second image after the first image are scaled to obtain the first sub-pixel image and the second sub-pixel image, respectively.

[0061] In the example embodiments of the present disclosure, the first image and the second image are single-channel grayscale images, and the first image and the second image can be adjacent two frames of images. The image can be scaled by interpolation processing to obtain sub-pixel level image data scaled to a required multiple size.

[0062] For example, as shown in FIG. 1, the first image and the second image are two frames of images, and the size of each frame of image is WxH. The two frames of images are scaled to sub-pixel level images by interpolation processing, and the size of each frame of sub-pixel image is nW*nH, where n is the scaling multiple. Figure 2 As shown in FIG. 2, it is a schematic diagram of a pre-processing process of image data. The size of the input front and rear two frames of images is WxH. The two frames of images are scaled to sub-pixel level images by interpolation processing, and the size of each frame of sub-pixel image is nW*nH, where n is the scaling multiple.

[0063] In an exemplary embodiment, the first sub-pixel image and the second sub-pixel image are transmitted from a main memory or other storage location to a GPU (Graphics Processing Unit) for parallel computation and / or graphics processing using the GPU.

[0064] In step S120, the first sub-pixel image is edge-extended and pixel-shifted based on the displacement threshold value, and a plurality of first intermediate images obtained are stored as a multi-channel first tensor image. The second sub-pixel image is edge-extended based on the displacement threshold value to obtain a single-channel second tensor image, which has the same size as the first tensor image.

[0065] In the exemplary embodiments of the present disclosure, the displacement threshold value is an adjustable displacement threshold value, which is a maximum displacement range allowed in the cross-correlation search, and can be used to control the displacement calculation region to control the number of channels in the generated first tensor image. It can be understood that the displacement threshold value can be used to control the maximum number of moving steps of the first sub-pixel image in the horizontal and vertical directions each time, and to control the number of times of edge extension of the second sub-pixel image. The displacement threshold value can be determined according to actual detection requirements, and can be flexibly adjusted in actual detection practice, which is not limited. The number of moving steps can refer to the number of sub-pixel units of movement.

[0066] In the exemplary embodiments of the present disclosure, the displacement threshold value is an adjustable displacement threshold value, which is a maximum displacement range allowed in the cross-correlation search, and can be used to control the displacement calculation region to control the number of channels in the generated first tensor image. It can be understood that the displacement threshold value can be used to control the maximum number of moving steps of the first sub-pixel image in the horizontal and vertical directions each time, and to control the number of times of edge extension of the second sub-pixel image. The displacement threshold value can be determined according to actual detection requirements, and can be flexibly adjusted in actual detection practice, which is not limited. The number of moving steps can refer to the number of sub-pixel units of movement.

[0067] The edge extension of the second sub-pixel image means adding zero values or copying edge pixels at the edge of the second sub-pixel image, so that the size of the extended image reaches the required size, such as consistent with the size of the first tensor image or meeting other specific requirements. It should be understood that the edge extension in the exemplary embodiments of the present disclosure is to extend the edge of the image by a number of times equal to the displacement threshold value, for example, the displacement threshold value = 1, the edge of the second sub-pixel image is extended by one time, the displacement threshold value = 2, the edge of the second sub-pixel image is extended by two times, and so on. Exemplarily, Figure 3 An exemplary diagram showing the edge of the second sub-pixel image being extended by one time is shown. The size of the second sub-pixel image with a size of nW*nH is changed to (nW+2)*(nH+2) after being edge-extended by one time.

[0068] For example, if the displacement threshold is 1, the first sub-pixel image is moved in the horizontal direction and the vertical direction by -1, 0, +1 respectively (where a negative number indicates moving left / down, and a positive number indicates moving right / up), and according to the combination, 9 possible moving modes are obtained, thereby corresponding to 9 channels. Correspondingly, the edge of the second sub-pixel image is expanded by one circle to obtain a single-channel second tensor graph. If the displacement threshold is 2, the first sub-pixel image is moved in the horizontal direction and the vertical direction by -2, -1, 0, +1, +2 respectively, and according to the combination, 25 possible moving modes are obtained, thereby corresponding to 25 channels. Correspondingly, the edge of the second sub-pixel image is expanded by two circles to obtain a single-channel second tensor graph. That is, if the displacement threshold is R, the channel dimension (channel number) of the first tensor graph obtained by moving the first sub-pixel image is:

[0069] L = (2R + 1) 2 Formula 1

[0070] It should be noted that the exemplary embodiments of the present disclosure are described with a displacement threshold of 1, and subsequent descriptions will not be repeated.

[0071] Figure 4 A flowchart for obtaining the first tensor graph and the second tensor graph is shown as Figure 4 The first sub-pixel image is moved based on the displacement threshold, the obtained plurality of first intermediate images are stored as a multi-channel first tensor graph, and the edge of the second sub-pixel image is expanded based on the displacement threshold to obtain a single-channel second tensor graph, which can include:

[0072] Step S410: determining the number of movable steps in the horizontal and vertical directions according to the displacement threshold.

[0073] Step S420: expanding the edge of the first sub-pixel image based on the displacement threshold, moving the first sub-pixel image in the horizontal and vertical directions within the limit range of the expanded edge according to the number of movable steps, and determining a plurality of first intermediate images according to the moving results to store the plurality of intermediate images as a first tensor graph.

[0074] Step S430: expanding the edge of the second sub-pixel image based on the displacement threshold to obtain a second tensor graph, so that the second tensor graph has the same size as the first tensor graph.

[0075] As described above, if the displacement threshold is 1, it is determined that the horizontal direction can move 1 step left, the horizontal direction does not move, and the horizontal direction moves 1 step right, and it is determined that the vertical direction can move 1 step up, the vertical direction does not move, and the vertical direction moves 1 step down.

[0076] Further, the edge of the first sub-pixel image can be expanded by one round according to the displacement threshold, and the number of rounds is also determined according to the displacement threshold. After expansion, the limit range of moving the first sub-pixel image is determined. The edge of the first sub-pixel image is expanded by one round Figure 3 The expansion mode is shown.

[0077] Exemplarily, Figure 5 A schematic diagram showing that the first sub-pixel image is moved horizontally and vertically once in the limit range of the expanded edge according to the number of movable steps is shown, Figure 5 is that the first sub-pixel image is moved horizontally left by 1 step (such as 1 sub-pixel unit), and is not moved in the vertical direction. According to the number of movable steps, 9 moving results of the first sub-pixel image in the limit range of the expanded edge are obtained, that is, 9 first intermediate images are obtained.

[0078] Figure 6 A schematic diagram of the first tensor graph and the second tensor graph is shown. The first intermediate images of 9 channels are stored as the first tensor graph, the edge of the second sub-pixel image is expanded by one round to obtain the second tensor graph, and the size of the first tensor graph and the second tensor graph is (nW+2)*(nH+2). The first intermediate image numbered (3) is obtained by moving the first sub-pixel image right by one sub-pixel unit and up by one sub-pixel unit, and the data of the first intermediate image numbered (3) is recorded as the 3rd channel data of the first tensor graph. The other numbered first intermediate images are similar, and will not be described in detail.

[0079] In step S130, the second intermediate data is determined according to the channel data difference value of the single channel of the second tensor graph and each channel of the first tensor graph, and the pixel cross-correlation calculation is performed based on the channel data of the second intermediate data to determine the sub-pixel point displacement data or the area displacement data between the second image and the first image. The second intermediate data has the same number of channels as the first tensor graph.

[0080] In the exemplary embodiments of the present disclosure, the second intermediate data is determined according to the channel data difference value of the single channel of the second tensor graph and each channel of the first tensor graph, and the moving condition of the first image and the second image is calculated based on the second intermediate data. The cross-correlation is an index for measuring the similarity between two images or two signals, and the pixel cross-correlation calculation is performed based on the channel data of the second intermediate data in the exemplary embodiments of the present disclosure.

[0081] Specifically, the second intermediate data is determined according to the channel data difference value of the single channel of the second tensor graph and each channel of the first tensor graph, including:

[0082] First, for each channel of the first tensor graph, the data difference value of the channel and the single channel of the second tensor graph is obtained to obtain the difference value data of each channel of the first tensor graph.

[0083] Secondly, the difference value data of each channel of the first tensor graph is squared to obtain difference value square data of each channel of the first tensor graph.

[0084] Finally, according to the displacement threshold, the difference value square data of each channel is removed from the peripheral edge data, and the second intermediate data is determined according to the difference value square data after removing the peripheral edge data.

[0085] The following will be described in combination with Figure 7 The process of determining the second intermediate data will be described. The channel data of the first tensor graph numbered (1) to numbered (9) and the single-channel data difference of the second tensor graph are obtained respectively, and the difference value data corresponding to each channel of the first tensor graph is obtained, that is, 9 difference value data. Then, the square sum operation is performed on the difference value data of each channel respectively, and the difference value square data corresponding to each channel is obtained, that is, 9 difference value square data. Further, according to the displacement threshold (taking 1 as an example), the outermost peripheral data of the 9 difference value square data is removed, and the difference value square data of 9 channels with a size of nW*nH is obtained, that is, the second intermediate data. Alternatively, the CUDA (Compute Unified Device Architecture, computing unified device architecture) core can be used to synchronize the calculation of the difference value square data of all moving modes (channels).

[0086] The example embodiments of the present disclosure can greatly reduce the time complexity of traversal calculation by introducing the concept of image data translation in sub-pixel units, storing the translated data in the concept of channels in the calculation variable, controlling the sub-core region in the pixel cross-correlation calculation process, and completing the entire operation process in the GPU. The image cross-correlation calculation process with adjustable displacement threshold is realized at a very low time cost.

[0087] In an alternative example embodiment, as Figure 8 As shown, the pixel cross-correlation calculation based on the channel data in the second intermediate data to determine the sub-pixel point displacement data between the second image and the first image can include:

[0088] Step S810: For each sub-pixel, according to the size of the channel data of each sub-pixel in the second intermediate data, the channel index of the minimum channel data is determined.

[0089] Step S820: According to the channel index, the displacement data of the sub-pixel is determined.

[0090] Step S830: Based on the displacement data of each sub-pixel, the sub-pixel point displacement data is determined.

[0091] For each sub-pixel, the channel data corresponding to the position of each channel in the second intermediate data can be determined according to the position of the sub-pixel, and the minimum channel data can be obtained. For example, for a single sub-pixel located in the i-th row and the j-th column, the channel data of the 9 channels in the i-th row and the j-th column of the second intermediate data can be obtained and directly compared to obtain the channel index of the minimum channel data in the 9 channels in the i-th row and the j-th column. For example, by calculation, it is determined that the minimum channel data of the 9 channels of the second intermediate data of a certain sub-pixel point appears in the (3) channel, and the channel index is (3).

[0092] Further, the displacement data of the sub-pixel is determined by decoding according to the channel index. Taking the channel index (3) as an example, it is determined by decoding that the displacement direction of the sub-pixel is horizontally left by one sub-pixel unit and vertically down by one sub-pixel unit. It can be understood that the displacement data is obtained by inversely deducing the obtaining mode of the first intermediate image corresponding to the channel index. For example, the channel index (3) is obtained by horizontally right moving one sub-pixel unit and vertically up moving one sub-pixel unit of the first sub-pixel image, and then the displacement data is determined by decoding the channel index (3) as “horizontally left moving one sub-pixel unit and vertically down moving one sub-pixel unit”.

[0093] Exemplarily, Figure 9 A schematic diagram of obtaining the channel index of the minimum channel data is shown as Figure 9 On the GPU, the movement of a single sub-pixel in the 9 channels of the second intermediate data can be directly compared, and then the channel index obtained after the comparison is decoded into displacement data, that is, the data is the displacement data of the sub-pixel point, that is, the movement of a single sub-pixel.

[0094] In an alternative exemplary embodiment, as Figure 10 As shown, the pixel cross-correlation calculation based on the channel data in the second intermediate data to determine the regional displacement data between the second image and the first image can include:

[0095] Step S1010: determining a target convolution kernel according to a set target pixel region size.

[0096] Step S1020: performing convolution operation on each channel data of the second intermediate data based on the target convolution kernel to obtain the convolution result corresponding to each channel.

[0097] Step S1030: determining the minimum value index in the channel dimension of the convolution result according to the size of the convolution result corresponding to each channel.

[0098] Step S1040: decoding according to the minimum value index to determine the regional displacement data of the sub-pixel region.

[0099] Step S1050: determining the region displacement data between the second image and the first image based on the displacement data of each sub-pixel region.

[0100] In this example embodiment, the sub-pixel displacement data is the movement of the pixel region. The target pixel region can be set according to the detection requirements, or it can be adjusted flexibly during the detection practice. The target pixel region can be a 2*2 pixel region, a 4*4 pixel region, etc. This example embodiment is described by taking a 2*2 target pixel region as an example.

[0101] Therefore, the target convolution kernel is determined to be a 2*2 convolution kernel with a weight of 1. Based on the target convolution kernel, the channel data of the 9 channels of the second intermediate data is convolved to obtain the difference sum of squares data of the 9 channels of the pixel region with a size of (nW-1)*(nH-1), i.e., the convolution result. As shown in Figure 11 As shown in FIG. 6, the second intermediate data is convolved once per channel.

[0102] Further, as shown in Figure 12 As shown in FIG. 6, the second intermediate data is convolved once per channel.

[0103] It should be noted that the way of decoding the displacement data according to the channel index in this example embodiment is similar to the above-mentioned way of obtaining the movement of a single sub-pixel.

[0104] As described above, the example embodiments of the present disclosure introduce the concept of shifting the image data by a sub-pixel unit, and store the shifted data in the calculation variable in the form of channels, so as to control the sub-kernel region in the pixel cross-correlation calculation process. At the same time, the variable-size convolution kernel is introduced to realize the variable-size pixel region cross-correlation identification process. By completing the entire operation process in the GPU, the time complexity of the traversal calculation can be greatly reduced, and the image cross-correlation calculation process with adjustable displacement threshold is realized at a very low time cost. On the contrary, taking a 3*3 sub-kernel region as an example, when the image cross-correlation calculation is performed based on the CPU, the scheme needs to be calculated 9*(nW-2)*(nH-2) times to calculate the displacement of the entire image, that is, when the image cross-correlation calculation is performed based on the CPU, the sub-kernel region needs to be calculated successively until the entire image is completely traversed. In this example embodiment, taking a unit pixel displacement as an example, the GPU channel storage converts the 9-direction displacement search into a single tensor operation (compared with the traditional 9×(W×H) traversal). The convolution kernel replaces the pixel traversal, and the calculation complexity is reduced from O(n 2) to O(1) (only related to the number of channels), which is conducive to achieving real-time level full-field feedback.

[0105] In step S140, the sub-pixel point displacement data or the regional displacement data is self-de-noised according to the regional pixel statistical values of the sub-pixel point displacement data or the regional displacement data under different convolution scales, to obtain de-noised displacement data.

[0106] In the example embodiments of the present disclosure, different convolution scales are used to determine the size and step length of the mean calculation region, so as to generate a specific convolution kernel for convolution and mean calculation, the weights of the convolution kernel are equal, and the sum is 1. The first convolution scale and the second convolution scale can be determined, the step lengths of the first convolution scale and the second convolution scale are the same, and the kernel size of the second convolution scale is greater than the kernel size of the first convolution scale. Alternatively, the kernel size of the second convolution scale is twice the kernel size of the first convolution scale.

[0107] Figure 13 A schematic diagram of a convolution operation is shown, s represents the step length of the convolution operation, and k represents the kernel size. The processing object of the example embodiments is the sub-pixel point displacement data or the regional displacement data obtained in step S130. Here, the size of the sub-pixel point displacement data is W*H, which is used as an example for description. In fact, the size is determined according to the sub-pixel point displacement data or the regional displacement data obtained in step S130.

[0108] Specifically, the self-de-noising of the sub-pixel point displacement data or the regional displacement data according to the regional pixel statistical values of the sub-pixel point displacement data or the regional displacement data under different convolution scales can include:

[0109] First, for each convolution operation, the sub-pixel point displacement data or the regional displacement data is subjected to a convolution operation based on the first convolution scale to obtain a first operation result, and the sub-pixel point displacement data or the regional displacement data is subjected to a convolution operation based on the second convolution scale to obtain a second operation result.

[0110] Second, the first operation result is processed to remove peripheral edge data based on the displacement threshold, so that the size of the processed first operation result is the same as that of the second operation result.

[0111] Finally, the second operation result is compared with the processed first operation result, and the sub-pixel point displacement data or the regional displacement data is self-de-noised according to the obtained comparison result to obtain de-noised displacement data.

[0112] It should be noted that when the first operation result is removed from the peripheral edge, the number of edge data rings removed is also determined according to the displacement threshold. For example, if the displacement threshold is 1, one ring of peripheral edge data is removed from the first operation result.

[0113] The following is an example of a first convolution scale with a step size s = 5, a convolution kernel size k = 10, and a second convolution scale with a step size s = 5, a convolution kernel size k = 20.

[0114] In each convolution operation, the sub-pixel point displacement data or the regional displacement data is convolved based on the first convolution scale to obtain a first operation result, and the size of the first operation result is (W / 5-1)*(H / 5-1). The sub-pixel point displacement data or the regional displacement data is convolved based on the second convolution scale to obtain a second operation result, and the size of the second operation result is (W / 5-3)*(H / 5-3). Then, the first operation result is removed from the mean value data of the outer periphery of the region, so that the size becomes (W / 5-3)*(H / 5-3), that is, the mean value data of the small region corresponding to the first convolution scale and the mean value data of the large region corresponding to the second convolution scale have the same size.

[0115] Further, the processed first operation result and the second operation result are compared to perform self-de-noising on the sub-pixel point displacement data or the regional displacement data according to the obtained comparison result to obtain de-noised displacement data. For each convolution operation, if the size of the processed first operation result and the second operation result satisfies a preset size relationship, the data before convolution corresponding to the second operation result is determined as a noise point, and de-noising processing is performed based on the noise point; if the size of the processed first operation result and the second operation result does not satisfy the preset size relationship, the data before convolution corresponding to the second operation result is retained.

[0116] Specifically, the preset size relationship indicates that the second operation result is smaller than the processed first operation result by a preset multiple, which can be obtained according to multiple tests or adjusted flexibly during detection. For example, the preset multiple = 0.6.

[0117] Taking the preset multiple as 0.6 as an example, if the second operation result is smaller than the processed first operation result by 0.6 times, it is determined that the size of the processed first operation result and the second operation result satisfies the preset size relationship, otherwise it does not satisfy.

[0118] If the size of the processed first operation result and the second operation result satisfies a preset size relationship, the data before convolution corresponding to the second operation result is assigned a value of 0 to remove the noise point. Otherwise, the data before convolution corresponding to the second operation result is reserved. For example, if the second operation result (large region mean value data) of the i-th row and the j-th column is less than 0.6 times the processed first operation result (small region mean value data) of the i-th row and the j-th column, the data before convolution corresponding to the second operation result (large region mean value data) of the i-th row and the j-th column is considered to be a noise point, and the displacement value of the region is directly assigned a value of 0, thereby realizing the self-noise reduction process. If the second operation result (large region mean value data) of the i-th row and the j-th column is greater than or equal to 0.6 times the processed first operation result (small region mean value data) of the i-th row and the j-th column, the data before convolution corresponding to the second operation result (large region mean value data) of the i-th row and the j-th column is considered to be a normal point, and the displacement data of the region is reserved.

[0119] Exemplary, Figure 14 Taking a noise point existing in one-dimensional data as an example, a self-noise reduction process based on statistics is represented. Taking a step size s = 5, a first convolution scale (small region calculation size) k = 10, and a second convolution scale (large region calculation size) k = 20 as examples, in this example, the first operation result (small region mean value data) is 0.65, the second operation result (large region mean value data) is 0.325, and the large region mean value data is less than 0.6 times the small region mean value data, i.e., 0.325 < 0.6 * 0.65 = 0.39, so it is considered that the large region has a noise point, and all data values of the large region are directly removed, i.e., assigned a value of 0. (a) is a noise point data type of one-dimensional data, and (b) is a self-noise reduction result based on statistics.

[0120] In the self-noise reduction process based on statistics, the size of the convolution kernel is adjusted (different convolution scales) to realize the elimination of noise points of different scales. The smaller the size of the convolution kernel, the stronger the ability to identify small region noise points; the larger the size of the convolution kernel, the stronger the ability to identify large region noise points.

[0121] The whole process is to first perform convolution mean value operation on all data based on GPU convolution operation process, and then execute the original data retention according to the mean value judgment standard, so in the process of noise reduction and mean value operation, it is not to calculate the mean value of a region and then execute the original data retention of the region, thereby avoiding the influence of the noise reduction data on the noise reduction mean value operation of the adjacent region. It can be understood that, in the example embodiment, the statistical mean value judgment method is introduced as the basis for judging whether the region is a noise point by considering the surrounding region of the large displacement data value region, and according to the displacement continuity, the isolated large displacement point can be judged as a noise point. Without pre-setting the threshold value, the noise criterion is derived from the statistical characteristics of the displacement field itself. In this process, the noise reduction decision and displacement calculation are decoupled, and the global statistics are performed first and then the local zero is cleared, which maintains the strain coordination.

[0122] In step S150, the denoised displacement data is enhanced by Gauss to obtain the target full-field displacement result.

[0123] In the example embodiment of the present disclosure, the small displacement data can also have strong display capability by performing nonlinear enhancement on the denoised displacement data. Specifically, the denoised displacement data is enhanced by Gauss to obtain the target full-field displacement result, which can include:

[0124] First, a Gaussian probability density function model based on displacement data is established.

[0125] Secondly, the displacement value interval to be enhanced in the denoised displacement data is determined, and the displacement value section is divided into a plurality of subintervals.

[0126] Next, for each subinterval, the displacement data of the subinterval is enhanced according to the Gaussian function value corresponding to the Gaussian probability density function model of the subregion, to obtain the enhanced displacement data of the subregion.

[0127] Finally, the target full-field displacement result is determined according to the enhanced displacement data of each subregion.

[0128] The Gaussian probability density function model based on displacement data is a Gaussian distribution function with variable mean and variance, that is, the mean and variance can be determined by multiple tests, or can be flexibly adjusted during detection. For example, the Gaussian distribution density function constructed by the example embodiment with a mean of 0 and a variance of 2 is:

[0129] G(x, μ = 0, σ 2 = 2) Equation 2

[0130] Wherein, x is displacement data, as a random variable, used to obtain a Gaussian function value.

[0131] Specifically, a nonlinear displacement sensitivity enhancement operator is constructed, the core of which is to construct a differential homeomorphism mapping from the physical displacement space to the enhanced visualization space by using the monotonic decay property of the Gaussian probability density function. The normalized displacement observation value (normalized denoising displacement data) d ∈ [0, 1] of the input is defined as a random variable, and a parameterized Gaussian probability density function is established:

[0132]

[0133] wherein the established parameterized Gaussian probability density function presents a super-linear growth characteristic when d→0 + .

[0134] Further, the displacement value interval to be enhanced in the denoising displacement data can be determined, which can be flexibly set according to actual needs or manually set in the detection process. Accordingly, the displacement value section can be divided into multiple subintervals.

[0135] wherein the displacement value interval to be enhanced can be decomposed by Lebesgue measure:

[0136]

[0137] Δk is each subinterval, k is the number of subintervals, that is, 50 subintervals are taken as an example for illustration here.

[0138] For each subinterval, data enhancement can be performed by the following formula 5:

[0139]

[0140] is the enhanced displacement data of the subinterval, d k is the displacement data before enhancement, and G() is the Gaussian probability density function model.

[0141] For example, if the displacement value interval to be enhanced in the denoising displacement data is [0, 0.50] displacement value interval, the displacement value interval is divided into 50 subintervals with a step size of Δx = 0.01, and each subinterval is denoted as [x i , x i + Δx], then each subinterval is enhanced one by one, and the specific enhancement method is:

[0142]

[0143] wherein [y i ] represents the enhanced displacement data of the displacement value interval [x i , x i + Δx] to be enhanced, which contains the same number of data as [x i , x iThe number of data contained in the enhanced data [x+Δx] is consistent and is n. According to different processing data, the number of data in each cell is not the same, but the number of enhanced data and the number of data to be enhanced remains the same.

[0144] As can be seen from formula 6, the smaller the displacement value to be enhanced, the larger the Gaussian function value obtained, and the greater the degree of enhancement. Nonlinear enhancement of linear data is achieved, so that small displacement data also has strong display capability in visual display. That is, not only does the small displacement (nonlinearly magnified to a recognizable range from a displacement (<0.1 pixel) that cannot be distinguished by the naked eye), but also maintains the physical meaning (enhancement only changes the visualization contrast, without affecting the accuracy of the original displacement value), that is, enhancement only acts on the visualization mapping, without changing the accuracy of the original displacement data, significantly improving the visualization recognition capability of micro-crack initiation and propagation path.

[0145] The exemplary embodiments of the present disclosure convert sub-visual threshold displacement (<0.5 pixel) into a strong distinguishable signal in visual display through mathematical mapping reconstruction of visual perception. Different forms of enhancement of input data are achieved by adjusting the variance and mean of the Gaussian probability density function, the center point of the displacement to be enhanced is achieved by adjusting the mean of the Gaussian probability density function, and the contrast of the displacement to be enhanced is achieved by adjusting the variance of the Gaussian probability density function. Different degrees of enhancement mode of adjustable center point concentrated position and contrast are achieved.

[0146] In addition, in the results obtained in each of the above steps, in order to compare and process the results by visualization, the obtained data can be converted into three-channel uint8 image data, and then the size is scaled to the size of the original image (the first image and the second image), that is, the vertical and horizontal displacements are mapped to the original image, so that the displacement condition can be observed by visualization. For example, sub-pixel displacement data or regional displacement data can be converted into three-channel image data and scaled in size to realize the visualization of the results. For the processing results of each step, if another processing is required, the processing results of the previous step can be normalized to ensure the unity and accuracy of the overall processing flow, and the specific normalization method is not limited.

[0147] The following will fully describe the process of full-field displacement calculation based on the method of the exemplary embodiments of the present disclosure with specific examples.

[0148] Firstly, Figure 15 For the first image, Figure 16The first image and the second image are processed according to the method as in the above example to obtain a first tensor map and a second tensor map, and the second intermediate data is determined according to the channel data difference value between a single channel of the second tensor map and each channel of the first tensor map, and pixel cross-correlation calculation is performed based on the channel data of each channel of the second intermediate data to determine the sub-pixel displacement data or regional displacement data between the second image and the first image, and obtain Figure 17 The sub-pixel displacement results are shown. Figure 17 , the surface displacement of the material during loading can be identified, and the location and propagation direction of the crack can be determined by the relative size of the displacement value, but Figure 17 In the sub-pixel displacement result, in addition to useful information, the result is also affected by many noise points.

[0149] Secondly, Figure 17 The sub-pixel displacement results shown in the figure can be subjected to a maximum noise reduction process (conventional method) first, with the thresholds set to 0.1 and 0.95. That is, for the normalized data input in this step, the displacement data less than 0.1 and greater than 0.95 are removed, and then the remaining results are normalized to enhance the contrast of the data with a displacement of 0.1 to 0.95. The processing results are shown in FIG. Figure 18 shown.

[0150] Then, Figure 18 The corresponding processing results (sub-pixel displacement data or regional displacement data) are subjected to self-denoising according to their regional pixel statistics at different convolution scales to obtain denoised displacement data. The regional convolution operation process with s = 5 and convolution kernel size k = 10 is still set, and the large regional mean data (second operation result) is set to be less than 0.6 times the small regional mean data (the first operation result after processing). After processing, the denoised displacement data is obtained and normalized. The results are as follows: Figure 19 shown.

[0151] Further, yes Figure 19 The processing results shown in the figure are subjected to Gaussian enhancement processing, with the mean set to 0, the variance set to 2, the enhanced displacement value interval selected to be [0, 0.50], the step to be Δx = 0.01, and the self-enhancement processing is performed. The processed results are then normalized. Figure 20 shown.

[0152] Furthermore, Figure 20 The processing result shown in the figure still has some noise points that appear alone. Figure 20The processing results shown in the figure are self-denoised according to the regional pixel statistics at different convolution scales. For example, the 10*10 region can be denoised first to eliminate small noise points. The processing results are shown in the figure. Figure 21 Then, it is subjected to 16*16 area noise reduction processing to eliminate relatively large noise points. The processing results are as follows Figure 22 shown.

[0153] Finally, you can also Figure 22 The processing result shown is smoothed, for example, 5 times of convolution with a 7*7 convolution kernel to obtain the mean value to smooth the cloud map, and the result is as follows Figure 23 shown.

[0154] It should be noted that the detailed processing methods of each processing step in this example have been recorded in the above exemplary embodiments and will not be repeated here.

[0155] In an exemplary embodiment of the present disclosure, a full-field displacement calculation method based on self-statistical noise reduction and Gaussian enhancement is implemented. First, a first image and a second image one frame after the first image are scaled to obtain a first sub-pixel image and a second sub-pixel image, respectively. The first sub-pixel image is pixel-edge-extended and shifted based on a displacement threshold. The resulting multiple first intermediate images are stored as a multi-channel first tensor map. The second sub-pixel image is edge-extended based on the displacement threshold to obtain a single-channel second tensor map. Second intermediate data is then determined based on the channel data difference between a single channel of the second tensor map and each channel of the first tensor map. Pixel cross-correlation calculation is then performed based on the channel data of each channel of the second intermediate data to determine sub-pixel point displacement data or regional displacement data between the second image and the first image. This process converts image translation operations into tensor channel storage and calculates the sum of squared differences of all translation positions between the stored multi-channel first tensor map and the single-channel second tensor map. Compared to image cross-correlation calculations performed on a CPU, the present disclosure significantly reduces computational complexity through a parallel design based on GPU operations, which helps reduce computational latency and achieve real-time full-field displacement feedback. Moreover, by pre-scaling the image to generate sub-pixel data, the boundary displacement calculation accuracy can reach 0.1 pixel units, which helps to improve the accuracy of displacement detection.

[0156] On the other hand, the sub-pixel displacement data or the regional displacement data is self-de-noised according to the regional pixel statistical values of the sub-pixel displacement data or the regional displacement data under different convolution scales, to obtain de-noised displacement data. Without introducing other judgment criteria, this process can be beneficial to the noise source from the statistical characteristics of the displacement field itself, decouple the de-noising decision and the displacement calculation, identify the noise points in the region of any displacement data value, and realize effective self-de-noising. On the other hand, the de-noised displacement data is enhanced by Gaussian to obtain the target full-field displacement result. This process can enhance the small displacement values after the pre-processing de-noising, and improve the visualization display degree of the micro-displacement segment.

[0157] In addition, the detection method of the example embodiment of the present disclosure only needs an ordinary gray camera + a consumer-level GPU, which can replace high-end workstations and commercial closed-source software (such as traditional DIC systems). Moreover, it has open source compatibility: the algorithm is based on a general tensor operation library, supports embedded device transplantation, and has low hardware deployment cost, which is beneficial to flexible deployment. At the same time, as can be known from the above detection process, the detection method of the example embodiment of the present disclosure can overcome the limitation of the photoelastic method on transparent materials, and is suitable for common engineering materials such as metals, concretes, and rock masses. Moreover, it has environmental robustness, i.e., the self-de-noising process effectively resists speckle noise caused by surface roughness and uneven illumination, which is beneficial to improve the reliability of detection in extreme scenes (such as roads and dam bodies) in the field.

[0158] In addition, the above-described drawings are only schematic illustrations of the processes included in the method according to the example embodiment of the present disclosure, and are not for limitation purposes. It is easy to understand that the processes shown in the above-described drawings do not indicate or limit the time sequence of these processes. In addition, it is also easy to understand that these processes can be executed synchronously or asynchronously, for example, in multiple modules.

[0159] Other embodiments of the present disclosure will be apparent to those skilled in the art upon consideration of the specification and practice of the present disclosure disclosed. The present disclosure is intended to cover any variations, uses, or adaptations of the present disclosure following the general principles thereof and including the general principles disclosed herein and the art known to those skilled in the art. The specification and examples are to be regarded as illustrative only, and the true scope and spirit of the present disclosure are indicated by the claims.

Claims

1. A full-field displacement calculation method based on self-statistical noise reduction and Gaussian enhancement, characterized in that: include: Scaling a first image and a second image that is one frame subsequent to the first image to obtain a first sub-pixel image and a second sub-pixel image, respectively, where the first image and the second image are single-channel grayscale images; Performing edge expansion and pixel shifting on the first sub-pixel image based on a displacement threshold, storing the obtained multiple first intermediate images as a multi-channel first tensor map, and performing edge expansion on the second sub-pixel image based on the displacement threshold to obtain a single-channel second tensor map, where the second tensor map has the same size as the first tensor map; determining second intermediate data based on a difference value between channel data of a single channel of the second tensor image and each channel of the first tensor image, and performing pixel cross-correlation calculation based on the channel data of each of the second intermediate data to determine sub-pixel displacement data or regional displacement data between the second image and the first image, wherein the second intermediate data has the same number of channels as the first tensor image; performing self-denoising on the sub-pixel displacement data or the regional displacement data according to regional pixel statistics at different convolution scales to obtain denoised displacement data; Gaussian enhancement is performed on the noise-reduced displacement data to obtain a calculation result of the target full-field displacement.

2. The method according to claim 1, characterized in that The method further comprises: performing edge expansion and pixel shifting on the first sub-pixel image based on a displacement threshold, storing the obtained multiple first intermediate images as a multi-channel first tensor map, and performing edge expansion on the second sub-pixel image based on the displacement threshold to obtain a single-channel second tensor map. Determining the number of movable steps in the horizontal and vertical directions according to the displacement threshold; extending an edge of the first sub-pixel image based on the displacement threshold, horizontally and vertically moving the first sub-pixel image within a limited range of the extended edge each time according to the movable step number, and determining the plurality of first intermediate images based on the results of each movement, so as to store the plurality of intermediate images as the first tensor map; The edge of the second sub-pixel image is expanded based on the displacement threshold to obtain a second tensor map, so that the second tensor map has the same size as the first tensor map.

3. The method according to claim 1, characterized in that The determining of the second intermediate data according to the channel data difference value between a single channel of the second tensor graph and each channel of the first tensor graph includes: For each channel of the first tensor map, obtaining a data difference between the channel and a single channel of the second tensor map to obtain difference data for each channel of the first tensor map; performing a square operation on the difference data of each channel of the first tensor map to obtain squared difference data of each channel of the first tensor map; According to the displacement threshold, peripheral edge data are removed from the square difference data of each channel, and the second intermediate data is determined according to the square difference data after the peripheral edge data is removed.

4. The method according to claim 1, wherein Performing pixel cross-correlation calculation based on each channel data in the second intermediate data to determine sub-pixel displacement data between the second image and the first image includes: For each sub-pixel, determining a channel index of minimum channel data according to the size of each channel data corresponding to the sub-pixel in the second intermediate data; Decoding is performed according to the channel index to determine displacement data of the sub-pixel; The sub-pixel point displacement data is determined based on the displacement data of each sub-pixel.

5. The method according to claim 1, wherein Performing pixel cross-correlation calculation based on each channel data in the second intermediate data to determine regional displacement data between the second image and the first image includes: Determine the target convolution kernel according to the set target pixel area size; Performing a convolution operation on each channel data of the second intermediate data based on the target convolution kernel to obtain a convolution result corresponding to each channel; According to the size of the convolution result corresponding to each channel, determine the minimum index of the convolution result in the channel dimension; Decoding is performed according to the minimum value index to determine regional displacement data of the sub-pixel region; Based on the regional displacement data of each of the sub-pixel regions, regional displacement data between the second image and the first image is determined.

6. The method according to claim 1, characterized in that The step of performing self-noising on the sub-pixel displacement data or the regional displacement data according to regional pixel statistics at different convolution scales to obtain denoised displacement data includes: Determining a first convolution scale and a second convolution scale, where the first convolution scale and the second convolution scale have the same step size, and a convolution kernel size of the second convolution scale is larger than the convolution kernel size of the first convolution scale; For each convolution operation, performing a convolution operation on the sub-pixel displacement data or the regional displacement data based on the first convolution scale to obtain a first operation result, and performing a convolution operation on the sub-pixel displacement data or the regional displacement data based on the second convolution scale to obtain a second operation result; performing data processing on the first operation result by removing the outer edge based on the displacement threshold, so that the processed first operation result and the second operation result have the same size; The second operation result is compared with the processed first operation result, and self-noise reduction is performed on the sub-pixel displacement data or the regional displacement data according to the obtained comparison result to obtain noise-reduced displacement data.

7. The method according to claim 6, characterized in that The convolution kernel size of the second convolution scale is twice the convolution kernel size of the first convolution scale.

8. The method according to claim 6, characterized in that The step of comparing the second operation result with the processed first operation result, and performing self-noise reduction on the sub-pixel displacement data or the regional displacement data according to the comparison result to obtain the noise-reduced displacement data, includes: For each convolution operation, if the sizes of the processed first operation result and the second operation result satisfy a preset size relationship, determining that the data before the convolution corresponding to the second operation result is a noise point, and performing noise reduction processing based on the noise point; If the sizes of the processed first operation result and the second operation result do not satisfy a preset size relationship, the data before convolution corresponding to the second operation result is retained.

9. The method according to claim 8, characterized in that The preset size relationship indicates that the second operation result is smaller than the processed first operation result by a preset multiple; If the sizes of the processed first operation result and the second operation result satisfy the preset size relationship, the data before convolution corresponding to the second operation result is assigned a value of 0.

10. The method according to claim 1, characterized in that The Gaussian enhancement is performed on the noise reduction displacement data to obtain a calculation result of the target full-field displacement, including: Establish a Gaussian probability density function model based on displacement data; Determining a displacement value interval to be enhanced in the noise-reduced displacement data, and dividing the displacement value interval into a plurality of subintervals; For each sub-interval, enhancing the displacement data of the sub-interval according to the Gaussian function value of the sub-region corresponding to the Gaussian probability density function model to obtain enhanced displacement data of the sub-interval; The calculation result of the target full-field displacement is determined according to the enhanced displacement data of each sub-area.

Images