CN114119556B - An automatic detection method for the quality of laser repair of surface defects of fused quartz components - Google Patents

Abstract

An automatic detection method for the quality of laser repair of surface defects of fused quartz components relates to the field of engineering optical technology and is used to detect the quality of repair of surface defects of fused quartz components. The technical points of the present invention include: changing the distance between the camera and the component to collect multiple images containing repair pits under different focusing states; performing depth of field fusion on multiple images under different focusing states to obtain a clear image containing the repair pits; inputting the clear image containing the repair pits into a pre-trained residual damage detection model to obtain the detection result. The present invention realizes the automatic acquisition of images of repair pits of different sizes by combining single-frame photography and scanning photography, obtains a complete full-depth image of the repair pits by using depth of field fusion and image stitching methods, and realizes the detection of residual damage of the repair pits by using a target detection method based on a convolutional neural network. The present invention does not require manual intervention and can be applied to the automatic detection of the repair quality after the surface defects of components are repaired.

Description

An automatic detection method for the quality of laser repair of surface defects of fused quartz components

Technical Field

The invention relates to the technical field of engineering optics, and in particular to an automatic detection method for the quality of laser repair of surface defects of a fused quartz component.

Background Art

Fused quartz components have the advantages of good light transmittance, stable chemical properties, and high temperature resistance. They are widely used in the terminal optical components of high-power solid-state laser devices, and undertake multiple functions such as beam focusing and debris shielding. However, fused quartz is a typical hard and brittle material, which will inevitably produce some surface defects such as pits and microcracks during processing, cleaning, and transportation. Surface defects will reduce the material properties of the component and make it more susceptible to damage. Studies have shown that once defects occur on the surface of the component, if they are not treated in time, the defect size will increase exponentially and the number of defects will also increase sharply. The generation and growth of defects will not only reduce the transmittance of the component itself, but also affect the downstream components of the optical path and cause damage to the downstream components, thereby threatening the stable operation of the laser device. Therefore, it is of great significance to use appropriate methods to repair the surface defects of the component in a timely manner.

In engineering, _CO2 laser repair methods are often used to perform local repairs on damaged components. This process is achieved by etching a cone in the defect area. This method can effectively increase the damage threshold of the material, inhibit the growth of surface defects, and does not affect the light transmission performance of the component. The repaired components can be loaded into the circuit for continued use, saving the maintenance cost of the device. During the repair, because individual defects have a deeper subsurface crack morphology than expected, in a few cases, they cannot be completely repaired using commonly used existing repair schemes, and there is residual damage in the repair pit. Once the incompletely repaired defects are reinstalled in the laser optical path, they will have an adverse effect on the service life of the optical component, so the quality of the surface defect laser repair must be tested.

Summary of the invention

In view of the above problems, the present invention proposes an automatic detection method for the laser repair quality of surface defects of fused quartz components, which is used to detect the repair quality of surface defects of fused quartz components.

An automatic detection method for the quality of laser repair of surface defects of a fused quartz component comprises the following steps:

Step 1: for each repair pit of the component, change the distance between the camera and the component, and collect multiple images containing the repair pit under different focusing states;

Step 2: Perform depth of field fusion on multiple images in different focus states to obtain a clear image containing the repaired pit;

Step 3: Input a clear image containing the repair pit into a pre-trained residual damage detection model to obtain a detection result; the detection result includes whether there is residual damage on the surface of the component.

Furthermore, in step one, when acquiring images corresponding to different repair pits, different acquisition methods are used for repair pits of different sizes: when the size of the repair pit is within the field of view of the camera, a single image containing the repair pit is acquired; when the size of the repair pit is not within the field of view of the camera, an array scan is performed on the partial area of the component containing the repair pit to obtain multiple sub-images containing the repair pit; wherein, two adjacent sub-images have an overlapping area.

Furthermore, when the size of the repair pit in step one is not within the camera's field of view, after obtaining multiple sub-images containing the repair pits, a template matching-based method is used to calculate the translation misalignment between the multiple sub-images, and then the overlapping areas of two adjacent sub-images are processed by weighted fusion, so as to perform image stitching on the multiple sub-images.

Furthermore, the process of calculating the translation misalignment between multiple sub-images using a template matching-based method during the image stitching process in step one is as follows: using the overlapping area of two adjacent sub-images as a template, using the template to traverse all pixel positions of the two adjacent sub-images, and determining the position of the template in the adjacent sub-image by comparing the similarity between the template and multiple areas of the adjacent sub-image, and then calculating the translation misalignment of the adjacent sub-images.

Furthermore, in the image stitching process in step 1, the similarity between the template and different regions of the adjacent sub-image is calculated according to the following formula:

Where R(x,y) represents the similarity between the template and the sub-image area centered at the position (x,y); T'(x',y') represents the value of the standardized template image at the position (x',y'); I'(x+x',y+y') represents the value of the standardized sub-image image at the position (x+x',y+y').

Furthermore, the depth of field fusion process described in step 2 is: each image containing the repair pit in the multiple images under different focusing states is divided into multiple image sub-regions, for the image sub-regions corresponding to the same position in the multiple images, the pixel gradient value in each image sub-region is calculated, and the image sub-region with the largest pixel gradient value is selected to extract; and the multiple image sub-regions extracted at different positions are combined into a new full-view depth image.

Furthermore, in step 2, the gradient value of the pixel point in the image sub-region is obtained by Laplace operator calculation.

Furthermore, the pre-trained residual damage detection model in step three is built based on the convolutional neural network YOLOV3.

Furthermore, the detection result in step three also includes the location and size of the residual damage.

The beneficial technical effects of the present invention are:

The present invention realizes the automatic acquisition of repair pit images of different sizes by combining single-frame photography and scanning photography; obtains a complete full-depth image of the repair pit by using depth of field fusion and image stitching methods; realizes the detection of residual damage of the repair pit by using a target detection method based on a convolutional neural network; the method realizes the automatic detection of the quality of laser repair of component surface defects without manual intervention.

BRIEF DESCRIPTION OF THE DRAWINGS

The present invention can be better understood by referring to the description given below in conjunction with the accompanying drawings, which together with the following detailed description are included in this specification and form a part of this specification, and are used to further illustrate the preferred embodiments of the present invention and explain the principles and advantages of the present invention.

FIG1 is a schematic structural diagram of an automatic detection device for surface defect repair quality according to an embodiment of the present invention;

FIG2 is a schematic diagram of a repair pit image acquisition process according to an embodiment of the present invention;

FIG3 is a schematic diagram of a scanned image stitching process according to an embodiment of the present invention;

Figure 4 is an example diagram of the residual damage detection results of the repair pit in an embodiment of the present invention; wherein, Figure (a) is the image before repair corresponding to the 1mm repair pit; Figure (b) is the microscopic image after repair using the 1mm repair pit; Figure (c) is the image before repair corresponding to the 2mm repair pit; and Figure (d) is the microscopic image after repair using the 2mm repair pit.

DETAILED DESCRIPTION

In order to enable those skilled in the art to better understand the scheme of the present invention, exemplary implementations or embodiments of the present invention will be described below in conjunction with the accompanying drawings. Obviously, the described implementations or embodiments are only implementations or embodiments of a part of the present invention, not all of them. Based on the implementations or embodiments of the present invention, all other implementations or embodiments obtained by ordinary technicians in the field without creative work should fall within the scope of protection of the present invention.

The present invention provides an automatic detection method for the quality of laser repair of surface defects of fused quartz components. By collecting and processing a microscopic image of a repair pit, it is determined whether the repair pit is completely repaired. The method provides technical support for the repair quality control of component surface defects.

An embodiment of the present invention provides an automatic detection method for the quality of laser repair of surface defects of a fused quartz component, the method comprising the following steps:

Step 1: for each repair pit of the component, change the distance between the camera and the component, and collect multiple images containing the repair pit under different focusing states;

Step 2: Perform depth of field fusion on multiple images in different focus states to obtain a clear image containing the repaired pit;

Step 3: Input a clear image containing the repair pit into a pre-trained residual damage detection model to obtain a detection result; wherein the detection result includes whether there is residual damage on the surface of the component.

In this embodiment, optionally, when acquiring images corresponding to different repair pits in step one, different acquisition methods are used for repair pits of different sizes: when the size of the repair pit is within the field of view of the camera, a single image containing the repair pit is acquired; when the size of the repair pit is not within the field of view of the camera, an array scan is performed on the partial area of the component containing the repair pit to acquire multiple sub-images containing the repair pit; wherein, two adjacent sub-images have an overlapping area.

In this embodiment, optionally, when the size of the repair pit in step one is not within the camera field of view, after obtaining multiple sub-images containing the repair pit, a template matching-based method is used to calculate the translation misalignment between the multiple sub-images, and then the overlapping areas of two adjacent sub-images are processed by weighted fusion, thereby performing image stitching on the multiple sub-images.

In this embodiment, optionally, the process of calculating the translation misalignment between multiple sub-images using a template matching-based method during the image stitching process in step one is: using the overlapping area of two adjacent sub-images as a template, using the template to traverse all pixel positions of the two adjacent sub-images, and determining the position of the template in the adjacent sub-image by comparing the similarity between the template and multiple areas of the adjacent sub-image, and then calculating the translation misalignment of the adjacent sub-images.

In this embodiment, optionally, in the image stitching process in step 1, the similarity between the template and different regions of the adjacent sub-image is calculated according to the following formula:

Where R(x,y) represents the similarity between the template and the sub-image area centered at the position (x,y); T'(x',y') represents the value of the standardized template image at the position (x',y'); I'(x+x',y+y') represents the value of the standardized sub-image image at the position (x+x',y+y').

In this embodiment, optionally, the process of depth of field fusion in step 2 is: divide each image containing the repair pit in multiple images under different focusing states into multiple image sub-regions, calculate the pixel gradient value in each image sub-region for image sub-regions corresponding to the same position in the multiple images, and select the image sub-region with the largest pixel gradient value to extract; combine the multiple image sub-regions extracted at different positions into a new full-view depth image.

In this embodiment, optionally, in step 2, the gradient value of the pixel point in the image sub-region is obtained by Laplacian operator calculation.

In this embodiment, optionally, the pre-trained residual damage detection model in step three is built based on the convolutional neural network YOLOV3.

In this embodiment, optionally, the detection result in step three also includes the location and size of the residual damage.

Another embodiment of the present invention provides an automatic detection method for the quality of laser repair of surface defects of fused quartz components, and the corresponding detection device is shown in Figure 1, including a motion platform and a microscopic detection system. The motion platform includes three motion axes, X, Y, and Z, and the motion directions of the X, Y, and Z motion axes are respectively consistent with the directions of the X, Y, and Z coordinate axes of the machine tool coordinate system; the motion platform X and Y motion axes can carry optical elements for two-dimensional high-precision movement, thereby achieving accurate positioning of the repair pit; the Z motion axis can carry a microscopic detection system for object distance adjustment, so that the repair pit image is clear. The microscopic detection system consists of an array CCD camera, a microscope lens, and a light source, wherein the array CCD camera has a resolution of 2456×2056 and a field of view of 1.5mm×1.3mm. The light source uses a backlight source. Since the part where the repair pit is completely repaired has better light transmittance than the part containing residual damage, the two have different imaging characteristics under backlight irradiation, which is conducive to the detection of residual damage. After the defective point of the component is repaired, the platform is first controlled to move, and the repair pits are moved one by one to the field of view of the microscope camera; then the microscope detection system is controlled to collect a clear image of the repair pit; finally, the repair pit image is processed and detected to determine whether there is residual damage in the repair pit and the specific location of the residual damage. The specific steps of this method are as follows:

Step 1: Develop a repair pit detection path.

According to an embodiment of the present invention, after the defect point is repaired, the repair pits need to be moved one by one to the microscopic field of view for residual damage detection. Since the laser repair file used in the repair already contains the coordinate information of the repair pits, these coordinates can be used to formulate the repair pit detection path. The laser repair file contains the size of the repair pit and the coordinates of the center point position, and the order of the repair pits is arranged according to the path planned by the greedy algorithm. By calibrating the position difference between the laser repair station and the microscopic detection station, the above coordinates can be converted into machine tool coordinates that position the center point of the repair pit to the center of the microscopic field of view, thereby realizing the positioning of the repair pit.

Step 2: Use a microscopic detection system to collect images of the repaired pits.

According to an embodiment of the present invention, since the repair pits vary in size and the field of view of the microscope camera is limited, the smaller repair pits can be directly captured, but for larger repair pits, a single shot cannot obtain the complete morphology of the repair pits. For this reason, the embodiment of the present invention uses scanning photography to obtain images of larger repair pits. Since the depth of field of the microscopic detection system is very small and the depth of the repair pit exceeds the depth of field of the camera, the camera can only focus locally and cannot obtain a clear global image. For this reason, the present invention captures images at different Z values according to the depth of the defect pit to achieve the acquisition of clear images of different areas of the repair pit. The specific process of repair pit image acquisition is:

Step 2-1: Use different collection methods according to the size of the repair pit.

The size range of the repair pit is 0.5mm to 3mm, and the camera field of view is 1.5mm×1.3mm. For repair pits smaller than the camera field of view, the repair pit center coordinates can be directly positioned to the center of the microscopic field of view for single-frame photography; for repair pits with larger sizes, the scanning photography method shown in Figure 2 is used to scan the 3×3 microscopic area in an array. To facilitate subsequent image stitching, the step value in the X-axis direction is less than 1.5mm, and the step value in the Y-axis direction is less than 1.3mm, so as to ensure that there is a 10% overlap between the two adjacent images in the X and Y axis directions. The overlapping part can be used as the basis for image stitching. This method can meet the detection needs of large-size repair pits. The solid arrow in Figure 2 represents the movement trajectory of the platform when the microscope camera takes pictures, and the red dot in the figure represents the shooting position of the bright field camera. To improve the scanning efficiency, an S-shaped path is used for scanning; to ensure the quality of the collected image, when the camera moves to the shooting position, the platform decelerates to zero, and controls the camera to automatically collect. After obtaining a clear image of the shooting position, the platform is controlled to continue moving to complete the scanning of the 3×3 area.

Step 2-2: Collect images at different focus states of the photo point.

The shape of the repair pit is conical, and its angle α is relatively fixed. Therefore, the depth h of the repair pit can be estimated by formula (1).

h＝Rtanα (1)

Among them, R is the repair pit size and α is the cone angle.

When taking single or scanning photos of the repair pit, the platform is controlled to move along the Z axis to collect images of different focus states at each photo point. The embodiment of the present invention collects microscopic images at three positions of 0, h/2, and h from the focal plane to the repair pit surface.

Step 3: Process the collected data to obtain a clear image of the area to be detected.

According to an embodiment of the present invention, a globally clear image of the area is obtained by performing depth of field fusion on images at different Z values of the same area; a complete image of a large-size repair pit is obtained by splicing the scanned images; and the spliced clear image of the damaged pit is processed to extract the area to be detected for residual damage detection.

Depth of field fusion refers to analyzing the repair pit image sequence collected when the microscope lens continuously changes the focal plane, selecting the area with the highest clarity in each image sequence to combine into a new full-depth image in which all areas of the repair pit are clear. In the embodiment of the present invention, the gradient value of each pixel point in the image sequence is obtained by the Laplace operator, and the value is used as the evaluation index of clarity. The pixel value of the image where the pixel point with the highest gradient value is located is used as the pixel value of the full-depth image. The process is shown in formula (2).

I(i,j)＝I _m (i,j),m＝index(max(L _k (i,j))),k＝1,2,3 (2)

Where _Lk represents the Laplace gradient map of the kth image in the image sequence, and the index function is used to obtain the sequence number of the image sequence.

For repair pits with larger sizes, the local images of the repair pits with overlapping parts are stitched into a seamless complete image of the repair pit. Since the array scanning process of the component is completed on an X-Y two-dimensional platform, there is only translational misalignment between the 9 collected images, but no rotational misalignment. Therefore, a template matching-based method is used to calculate the translational misalignment between images to achieve image stitching. The image stitching process is shown in Figure 3. Figures 3(a) and (b) are schematic diagrams of two adjacent images during the scanning process. In Figure 3(b), a partial image of the overlapping area of the two images is intercepted as a matching template. The template is used to traverse all pixel positions in Figure 3(a). The position of the template in Figure 3(a) is determined by comparing the similarity between the template and its covered area, thereby obtaining the translational misalignment of the two images. The similarity between the template and its covered area is evaluated by equations (3) to (5):

Where T and I represent the template image and target image respectively; w _T and h _T represent the width and height of the template image respectively; w _I and h _I represent the width and height of the template coverage area respectively.

The template and target image are standardized by equations (3) and (4), which can eliminate the interference of uneven illumination on template matching. The correlation coefficient between the template and its coverage area can be calculated by equation (5). The closer the correlation coefficient is to 1, the higher the similarity between the template and the coverage area. Figure 3(c) is a schematic diagram of the spliced image obtained by this method. However, due to the influence of uneven illumination, there are obvious splicing marks at the splicing position during the image splicing process. Therefore, equation (6) is used to process the overlapping area of the two images, and the spliced image is slowly transitioned from the first image to the second image by weighted fusion.

Where I ₁ and I ₂ represent the overlapping area of two adjacent images; n and _nc represent the size and fusion position of the overlapping area respectively; α(·) is the fusion coefficient of the overlapping area, and the value is related to the fusion position.

Perform the above operations on all images to obtain a stitched image of the large-sized repaired pit.

Step 4: Detect the repair pit area using the residual damage detection model and save the detection results.

According to an embodiment of the present invention, in order to detect residual damage in repair pits, a residual damage detection model is built using a target detection method based on a convolutional neural network. Inputting the image of the area to be detected into the model can automatically determine whether there is residual damage in the image and give the specific location of the residual damage. Use the model to detect all repair pits, and save the detection results in a residual damage detection result file.

The residual damage detection model is built on the basis of YOLO-v3, which is a single-stage defect detection network. After the image is input into the network, the position, size and category of the regressed bounding box will be directly output in the output layer. Compared with other target detection methods, YOLO-v3 has the advantages of fast prediction speed, high accuracy, and strong recognition ability of small objects. Based on the original model, the present invention adjusts the model input size, convolution step size and other parameters to make the model more suitable for residual damage detection. The data set for model training is obtained by collecting the repair pit data in advance and annotating the data. The data set is enhanced by means of target area replication, image scale transformation, rotation transformation, etc. The data set is used for model training and evaluation, and finally a model for residual damage detection is obtained. The specific process of using this model for repair pit repair quality detection is as follows:

Step 4-1, obtain the repair pit detection area image.

Since the residual damage is small in size, directly adjusting the repaired pit image to the input size required by the model will be detrimental to the detection of small targets. Therefore, a sliding window method is used for detection, and the image is intercepted region by region using a sliding frame that meets the model input size requirements. To ensure the integrity of target detection, there is a 15% overlap between sliding frames.

Step 4-2: Input the intercepted areas one by one into the residual damage detection model to determine whether there is residual damage in the intercepted area. If there is residual damage, the center position and size of the residual damage prediction box are recorded.

Step 4-3: Integrate the test results and save them in a residual damage test result file.

If no residual damage is detected in the intercepted area, the defect point is completely repaired and the repair point is skipped. If residual damage is detected in the intercepted area, all prediction boxes are integrated according to the location of the residual damage prediction box to form the final prediction result. The result will select the residual damage area in the form of a rectangle. The repair pit ID, repair resistance size, and residual damage location and size of the residual damage are saved in the residual damage detection result file.

Another embodiment of the present invention provides an example analysis of an automatic detection method for the quality of laser repair of micro-defects on the surface of large-caliber components. The above method is used to detect a batch of components that have completed laser repair of defect points. The diameter of the component is 430mm×430mm, and it contains a total of 50 repair pits. The self-developed large-caliber fused quartz component surface defect automatic detection and repair control software can realize automatic detection of repair pits. The specific process of detection is described by taking 1mm and 2mm repair pits as examples:

(1) According to the coordinate information in the laser repair file, move the above repair pits to the center of the microscopic field of view;

(2) For repair pits with a size of 1 mm, a single photo is taken directly, and for repair pits with a size of 2 mm, a 3×3 scanning photo is taken. At each photo point, the Z axis is controlled to move, the focal plane position is continuously adjusted, and three images with different focus states are collected.

(3) Perform depth of field fusion on the image sequences at different depths of field at each photographing point to obtain a full depth of field image of the area. For a 2 mm repair pit, the fused image is used for image stitching to obtain a complete image of the repair pit. The repair pit image obtained after the above processing is shown in Figure 4. Figures 4 (a) and (b) are microscopic images before and after repair when a 1 mm repair pit is used for repair. The repair pit area is smaller than the microscopic field of view. A clear image of the repair pit shown in Figure 4 (b) can be obtained through single-frame photography and depth of field fusion; Figures 4 (c) and (d) are microscopic images before and after repair when a 2 mm repair pit is used for repair. The repair pit area is larger than the microscopic field of view. A clear image of the repair pit shown in Figure 4 (d) can be obtained through scanning photography and depth of field fusion.

(4) The 1mm and 2mm repair pit images are input into the residual damage detection model. The model automatically determines whether there is residual damage in the repair pit and frames the residual damage area in the form of a rectangular frame. The detection results are shown in Figure 4. For the 1mm repair pit, no residual damage is found, and the repair pit is completely repaired; for the 2mm repair pit, there are two residual damages, and the damaged area is framed by a red rectangular frame, as shown in Figure 4 (b) and (d). The number of the 2mm repair pit and the residual damage position are saved in the residual damage detection result file.

Although the present invention has been described according to a limited number of embodiments, it will be apparent to those skilled in the art, with the benefit of the above description, that other embodiments are contemplated within the scope of the invention thus described. The disclosure of the present invention is intended to be illustrative rather than restrictive of the scope of the invention, which is defined by the appended claims.

Claims (4)

1. The automatic detection method for the laser repair quality of the surface defects of the fused quartz element is characterized by comprising the following steps of:

Step one, for each repair pit of an element, changing the distance between a camera and the element, and collecting a plurality of images containing the repair pit under different focusing states; wherein, adopt different collection modes to the repair hole of equidimension:

when the size of the repair pit is in the visual field of the camera, acquiring a single image containing the repair pit;

When the size of the repair pit is not in the visual field of the camera, carrying out array scanning acquisition on a part of the area of the element containing the repair pit to obtain a plurality of sub-images containing the repair pit, wherein two adjacent sub-images have overlapping areas; after a plurality of subgraphs containing the repair pit are obtained, calculating the translation dislocation quantity among the subgraphs by adopting a template matching-based method, and specifically comprising the following steps: firstly, using an overlapping area of two adjacent subgraphs as a template, and traversing all pixel positions of the two adjacent subgraphs by using the template; then, determining the position of the template in the adjacent subgraph by comparing the similarity degree of the template and the plurality of regions of the adjacent subgraph, wherein the similarity degree of the template and the different regions of the adjacent subgraph is calculated according to the following formula:

Wherein R (x, y) represents the similarity of the template to the sub-region centered at the (x, y) position; t ' (x ', y ') represents the value of the normalized template image at the (x ', y ') position; i ' (x+x ', y+y ') represents the value of the normalized sub-image at the (x+x ', y+y ') position; then, calculating the translation dislocation quantity of the adjacent subgraphs; and then processing the overlapping area of the two adjacent subgraphs in a weighted fusion mode, so that the plurality of subgraphs are subjected to image stitching, wherein the overlapping area of the two adjacent subgraphs is processed according to the following formula:

Wherein I ₁、I₂ represents an overlapping region of two adjacent images; n and n _c respectively represent the size and fusion position of the overlapping region; α (·) is the overlap region fusion coefficient;

Performing depth of field fusion on a plurality of images in different focusing states to obtain a clear image containing a repair pit; the depth of field fusion process comprises the following steps: dividing each image containing a repair pit in a plurality of images in different focusing states into a plurality of image subareas, calculating pixel point gradient values in each image subarea for the same image subareas corresponding to the positions of the plurality of images, and selecting the image subarea with the maximum extracted pixel point gradient value; taking the pixel value of the image with the largest gradient value as the pixel value of the panoramic deep image, so that a plurality of extracted image subregions at different positions are combined into a new panoramic deep image;

Inputting a clear image containing a repair pit into a pre-trained residual damage detection model to obtain a detection result; the detection result includes the presence or absence of residual damage to the surface of the component.

2. The automatic detection method for laser repair quality of surface defects of fused quartz elements according to claim 1, wherein in the second step, gradient values of pixel points in image subregions are obtained through calculation of Laplacian.

3. The method for automatically detecting the quality of laser repair of surface defects of a fused silica component according to claim 2, wherein the pre-trained residual damage detection model in the third step is built based on a convolutional neural network YOLOV.

4. The method for automatically detecting the quality of laser repair of surface defects of a fused silica component according to claim 3, wherein the detection result in the third step further comprises the position and the size of residual damage.