Method for measuring laser-induced damage area of optoelectronic material

Abstract

The application discloses a kind of photoelectric material laser damage area measurement methods, this method is based on deep learning, by adopting U-Net network model to the segmentation of damage area, by traversing the number of damage area pixel points, and combining the calibration result of the size of single pixel of high-speed camera, it can realize the precise segmentation of laser damage picture damage area with dark background and low edge profile contrast, to achieve the purpose of accurate calculation of damage area.

Description

Technical Field

[0001] This invention relates to a method for measuring the laser damage area of ​​optoelectronic materials, belonging to the field of optoelectronic testing technology. Background Technology

[0002] With the development of high-power, high-energy laser technology, research on laser-induced damage to optoelectronic materials has received increasing attention, and various laser damage testing systems for optoelectronic materials have emerged. Among them, the laser damage area is an important indicator for measuring the damage performance of a material. The traditional calculation method for laser damage area usually includes the following steps: (1) Image preprocessing, which involves grayscale conversion, histogram equalization, filtering, and morphological processing of the image to eliminate interference from irrelevant information in the image, enhance useful information in the image, and improve the accuracy of image processing; (2) Damage target location extraction, which involves extracting the image mask or finding the image contour to extract the damage area from the image, which is a crucial step; (3) Obtaining the calculation results, which involves counting the number of pixels in the contour extracted in step (2) and then obtaining the results based on the calibrated single pixel size of the camera. In the process of observing the damage morphology of optoelectronic materials using a high-speed camera, in order to observe more damage details, the frame rate of the camera is usually set to be greater than 5000fps, and the captured video is slowed down. The higher frame rate darkens the background, making it more difficult to preprocess the damage image using traditional methods. Summary of the Invention

[0003] To address the shortcomings of traditional methods, such as the difficulty in segmenting damaged areas in dark-background damaged images and the low accuracy of damaged area calculation, the present invention aims to provide a method for measuring the laser damage area of ​​optoelectronic materials. This method uses deep learning to accurately segment the damaged area, thereby achieving precise calculation of the damaged area.

[0004] To achieve the above objectives, as shown in the appendix Figure 1 As shown, the technical solution adopted in this invention is: A method for measuring the laser damage area of ​​optoelectronic materials includes the following steps: Step 1: Calibrate the mapping relationship between pixel size and sample length to obtain the actual sample length p corresponding to a single pixel; Step 2: Select the U-net network model as the backbone network for damaged region segmentation; use the encoder to capture the context information of the image and use its decoder to recover the fine segmentation results; select the ReLU function and the sigmoid function as the activation functions for convolution and upsampling respectively. The ReLU function is represented by Equation (1) and the sigmoid function is represented by Equation (2). Step 3: Use a high-speed camera to capture multiple damage videos, and decompose the videos frame by frame to obtain n (n≥1000) damage images; Step 4: Determine the target region in the damaged images. Set the four coordinate points of the cropping rectangle. Use the programming language Python and the open-source image processing database OpenCV to perform batch cropping of n damaged images to obtain n damaged images with a resolution of 512*512 pixels. Step 5: Import the damaged image data obtained in Step 4 into the image annotation tool labelme, manually annotate the target areas, generate a JSON file for each image, and obtain the training dataset by processing the JSON file and performing mask operations, which includes n pairs of damaged images / images with labeled damage areas; Step 6: Train the U-Net model using the training dataset; during training, select the RMSpropz optimizer to accelerate the optimization process. Initially, the learning rate lr is set to 10^-4. When training the network, the learning rate lr is set to 10^-3, the weight decay coefficient is set to 10^-8, the machine's floating-point operation error limit eps is set to 10^-8, and betas is set to (0.9, 0.999). The network model is trained for a total of 20 epochs. The Dice loss function shown in Equation (3) is used to measure the similarity between the network prediction results and the labeled images of the sample set. In the formula, |X| and |Y| represent the accuracy of the algorithm segmentation (Ground True) and the corresponding Mask operation, respectively. The Dice loss function is shown in Equation (4): When the network model does not improve the Dice coefficient of the evaluation metric for three consecutive epochs on the validation set, the learning rate is reduced by half, and CrossEntropyLoss is used as the objective function. CrossEntropyLoss can be expressed by Equation (5). The final learning rate is guaranteed to be no less than 10^-6. The training parameters are output to obtain the trained U-Net model. In the formula, x is the input, y is the target, w is the weight, C is the number of classes, and dk represents the K-dimensional case; Step 7: Input the damaged image into the trained network, segment the damaged region according to the model and training parameters, traverse each pixel in the image, and count the number of pixels m in the damaged region; S = m × p 2 (6) Step 8: Calculate the damage area in the damage image according to formula (6).

[0005] Beneficial effects: The present invention provides a method for measuring the laser damage area of ​​optoelectronic materials, which incorporates deep learning methods to accurately segment the damaged area in laser damage images with dark backgrounds and low edge contour contrast, thereby achieving the purpose of accurate calculation of the damage area. Attached Figure Description

[0006] Figure 1 This is a flowchart illustrating a method for measuring the laser damage area of ​​optoelectronic materials. Detailed Implementation

[0007] Example 1: A method for measuring the laser damage area of ​​optoelectronic materials.

[0008] This invention provides a method for measuring the laser damage area of ​​optoelectronic materials, combined with the attached... Figure 1 Follow these steps: Step 1: Calibrate the mapping relationship between pixel size and sample length to obtain the actual sample length p corresponding to a single pixel; Step 2: Select the U-net network model as the backbone network for damaged region segmentation, use the encoder in it to capture the context information of the image, and use its decoder to restore the fine segmentation result; select the ReLU function and the sigmoid function as the activation functions of convolution and upsampling respectively. The ReLU function is represented by Equation (1), and the sigmoid function is represented by Equation (2). Step 3: Use a high-speed camera to capture multiple damage videos, and decompose the videos frame by frame to obtain n (n≥1000) damage images; Step 4: Determine the target region in the damaged images. Set the four coordinate points of the cropping rectangle. Use the programming language Python and the open-source image processing database OpenCV to perform batch cropping of n damaged images to obtain n damaged images with a resolution of 512*512 pixels. Step 5: Import the damaged image data obtained in Step 4 into the image annotation tool labelme, manually annotate the target areas, generate a JSON file for each image, and obtain the training dataset by processing the JSON file and performing mask operations, which includes n pairs of damaged images / images with labeled damage areas; Step 6: Train the U-Net model using the training dataset; during training, select the RMSpropz optimizer to accelerate the optimization process. Initially, the learning rate lr is set to 10^-4. When training the network, the learning rate lr is set to 10^-3, the weight decay coefficient is set to 10^-8, the machine's floating-point operation error limit eps is set to 10^-8, and betas is set to (0.9, 0.999). The network model is trained for a total of 20 epochs. The Dice loss function shown in Equation (3) is used to measure the similarity between the network prediction results and the labeled images of the sample set. In the formula, |X| and |Y| represent the accuracy of the algorithm segmentation (Ground True) and the corresponding Mask operation, respectively. The Dice loss function is shown in Equation (4): When the network model does not improve the Dice coefficient of the evaluation metric for three consecutive epochs on the validation set, the learning rate is reduced by half, and CrossEntropyLoss is used as the objective function. CrossEntropyLoss can be expressed by Equation (5). The final learning rate is guaranteed to be no less than 10^-6. The training parameters are output to obtain the trained U-Net model. In the formula, x is the input, y is the target, w is the weight, C is the number of classes, and dk represents the K-dimensional case; Step 7: Input the damaged image into the trained network, segment the damaged region according to the model and training parameters, traverse each pixel in the image, and count the number of pixels m in the damaged region; S = m × p 2 (6) Step 8: Calculate the damage area in the damage image according to formula (6).

[0009] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for measuring the laser damage area of ​​optoelectronic materials, characterized in that... This method consists of the following steps: Step 1: Calibrate the mapping relationship between pixel size and sample length to obtain the actual sample length p corresponding to a single pixel; Step 2: Select the U-net network model as the backbone network for damaged region segmentation; use the encoder to capture the context information of the image and use its decoder to recover the fine segmentation results; select the ReLU function and the sigmoid function as the activation functions for convolution and upsampling respectively. The ReLU function is represented by Equation (1) and the sigmoid function is represented by Equation (2). (1) (2) Step 3: Use a high-speed camera to capture multiple damage videos, and decompose the videos frame by frame to obtain n (n≥1000) damage images; Step 4: Determine the target region in the damaged images. Set the four coordinate points of the cropping rectangle. Use the programming language Python and the open-source image processing database OpenCV to batch crop the n damaged images, obtaining n images with a resolution of 512. A damaged image with 512 pixels; Step 5: Import the damaged image data obtained in Step 4 into the image annotation tool labelme, manually annotate the target areas, generate a JSON file for each image, and obtain the training dataset by processing the JSON file and performing mask operations, which includes n pairs of damaged images / images with labeled damage areas; Step 6: Train the U-Net model using the training dataset; During training, the RMSpropz optimizer was selected to accelerate the optimization process. The initial learning rate lr was set to 10^-4. When training the network, the learning rate lr was set to 10^-3, the weight decay coefficient was set to 10^-8, the machine's floating-point operation error limit eps was set to 10^-8, and betas was set to (0.9, 0.999). The network model was trained for a total of 20 epochs. The Dice loss function shown in Equation (3) was used to measure the similarity between the network prediction results and the labeled images of the sample set. (3) In the formula, |X| and |Y| represent the accuracy of the algorithm segmentation (Ground True) and the corresponding Mask operation, respectively. The Dice loss function is shown in Equation (4): (4) When the network model does not improve the Dice coefficient of the evaluation metric for three consecutive epochs on the validation set, the learning rate is reduced by half, and CrossEntropyLoss is used as the objective function. CrossEntropyLoss can be expressed by Equation (5). The final learning rate is guaranteed to be no less than 10^-6. The training parameters are output to obtain the trained U-Net model. (5) In the formula, x is the input, y is the target, w is the weight, and C is the number of classes; Step 7: Input the damaged image into the trained network, segment the damaged region according to the model and training parameters, traverse each pixel in the image, and count the number of pixels m in the damaged region; (6) Step 8: Calculate the damage area in the damage image according to formula (6).

Images