A method for intelligently identifying a repair site of a long-distance pipeline girth weld radiographic image

Abstract

The present application relates to the technical field of pipeline weld quality evaluation, in particular to a kind of long-distance pipeline girth weld radiographic image welding repair site intelligent identification method, the specific process is as follows: first, the girth weld radiographic image welding repair site of pipeline is masked and labeled to establish label database.Then a deep learning instance segmentation model is constructed based on data-driven.By building a main network and a neck network, the features are fully extracted.Then, a pixel-by-pixel segmentation loss is constructed to guide the training process.Finally, a probability graph model is constructed, and the final result is obtained based on the posterior probability distribution of the mechanism model.The method provided by the present application significantly improves the identification accuracy of the welding repair site.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of pipeline weld quality evaluation, in particular to a long-distance pipeline girth weld radiographic image welding repair site intelligent identification method. BACKGROUND

[0002] Ensuring the safe operation of long-distance pipelines is of great significance in the industrial field. Pipeline leaks caused by weld area quality problems occur from time to time, so increasing the evaluation of weld quality will help reduce accidents.

[0003] In the weld quality evaluation process, finding the welding repair site is a key step. Defects in the repair site usually have more serious hazards. Under normal circumstances, manual search for welding repair areas has strong subjective factors, resulting in high uncertainty in evaluation. Different technical personnel have different quality assessments, which makes the welding repair area search inefficient, and the recognition rate and accuracy are low.

[0004] Deep neural networks have strong feature extraction capabilities and are widely used in industrial fields. A large number of target detection and instance segmentation algorithms based on neural networks have emerged. However, there are still several problems with deep learning methods: first, deep learning requires a large amount of labeled data, and it is very laborious to obtain a large number of labels in fields with strong professional knowledge. Second, the model obtained based on local data training does not have strong generalization ability. There is no alignment between unlabeled images and labeled image features, which weakens the ability to identify unknown defects.

[0005] The mechanism model method has a wide range of applications in image segmentation. This paper builds a target area segmentation model based on deep learning and combining a probabilistic graph model, which can significantly improve the prediction accuracy of the model. SUMMARY

[0006] In view of the low efficiency of manual identification of welding repair sites at present, the present application proposes a long-distance pipeline girth weld radiographic image welding repair site intelligent identification method, which can combine neural network models and probabilistic image models well, and ultimately achieve high recognition rate.

[0007] The specific technical solutions are as follows:

[0008] A long-distance pipeline girth weld radiographic image welding repair site intelligent identification method includes the following steps:

[0009] Step 1: Based on a digital device to scan the industrial negative to obtain the original pipeline weld defect image and perform image denoising processing:

[0010] Step 1-1: After the industrial X-ray film is digitized, due to the influence of many factors such as photosensitive material, X-ray energy and radiation dose, the digitized image has the defects of low contrast and high noise; in order to make the target features in the image obvious, noise removal processing is needed, and mean filtering is used to realize image smoothing; first, a plurality of defect images of the same size are cut from the X-ray defect original image, here a square image is cut with the smaller side as the side length;

[0011] Step 1-2: In order to remove noise, median filtering which is effective for pulse noise removal is used to filter the image, and the formula is:

[0012] Y(i,j)=Med A [X(i,j)],i=1,2,L h;j=1,2,L w

[0013] Where X(i,j) represents the pixel value of the i-th row and j-th column of the original image, Y(i,j) is the image after median filtering of the i-th row and j-th column; h and w represent the height and width of the original image respectively, and A is the median filter, a 3*3 filter window is used in this section;

[0014] Step 2: The original pipe weld image has large changes in blackness value during the scanning process, so the image is subjected to adaptive brightness enhancement;

[0015] Step 3: Based on the enhanced X-ray image, labels are made and generalized:

[0016] Step 3-1: According to the repair area contour, the area is labeled by the nondestructive testing engineer to obtain area labeling information and form a training set Where X k is the k-th image, Y k is the corresponding label, and N is the number of training sets, which is 2163 label images here;

[0017] Step 3-2: The training set image is subjected to generalization operation, here image translation, image folding, image scaling, image splicing and other methods are used, and the corresponding processing of the label is obtained to obtain the expanded data set Here M=11789;

[0018] Step 4: Based on the Resnet50 backbone network, the weld repair part feature is extracted, and the made weld image and its corresponding label mask are input into the network to extract deep features:

[0019] Step 4-1: First, the labeled image is input to the Focus structure for image segmentation, which greatly reduces the image size while not losing information, increases the dimension, and reduces network overhead;

[0020] Step 4-2: Then, the labeled image is input to a resnet50 convolutional neural network with 5 stages; the network level gradually deepens, and the abstract level of the input image also improves in turn, and rich semantic information of the mask area can be extracted through feature extraction;

[0021] Step 5: Design a neck network to realize rich feature fusion, which is conducive to selecting and increasing reasonable features for subsequent classification regression operations;

[0022] Step 6: The prediction mechanism extracts candidate regions and generates loss to form a training process;

[0023] Step 7: Based on the posterior probability distribution, a Markov chain is used to construct a probability graph model.

[0024] The specific steps of the adaptive brightness enhancement of the image in step 2 are as follows:

[0025] Step 2-1: Data augmentation, adaptive brightness enhancement of the original ray image, first, use the improved histogram equalization to realize the cumulative pixel probability density calculation; count the pixel number n of each gray level of the original image i , where 0 < i < L, L is the gray level of all pixels in the image, L = 255;

[0026] Step 2-2: Determine the pixel occurrence probability:

[0027] p x (i) = p(x = i) = n i / n,

[0028] Where n is the number of all pixels in the image, p x (i) is the pixel histogram, and is normalized to the [0, 1] interval according to the formula p x (i) = p x (i) / max(p x (i));

[0029] Step 2-3: Calculate the cumulative probability distribution to determine the cumulative histogram:

[0030]

[0031] Where cdf x (i) represents the pixel distribution graph after the cumulative histogram calculation of the pixel;

[0032] Step 2-4: threshold selection, determine the pixel cutoff moving interval, determine the pixel center threshold p m , the pixel moving interval range is [0, m], p m <p tre , wherein p tre moving interval probability distribution threshold, here 0.8 is selected;

[0033] Step 2-5: reconstruct the image pixel, determine the enhanced image, the pixel moving interval range is [0, m], p m <p tre , wherein p tre moving interval probability distribution threshold, here 0.8 is selected, the enhanced image is obtained by reconstructing the image pixel; wherein the reconstruction formula of each pixel value is as follows:

[0034]

[0035] , wherein x i is the gray value of each pixel point, k is a proportional factor, in order to prevent the brightness change from being too large, here k = 1.1 is selected.

[0036] The design of the neck network in step 5 realizes rich feature fusion, which specifically includes the following steps:

[0037] Step 5-1: build a feature pyramid model after the backbone network to fuse features from top to bottom, and select the three stages of the top features extracted by the backbone network as: features fpn ={Z3, Z4, Z5}, the neck fusion mode is as follows: each stage is fused with the corresponding feature branch from the horizontal direction, and the specific formula is as follows:

[0038] Z5 = P5,

[0039] , wherein P3, P4, P5 are three stage feature maps when fused from top to bottom, Z3, Z4, Z5 are three stage feature maps when fused from bottom to top, Upsample is an up-sampling interpolation function; is a pixel superposition function;

[0040] Step 5-2: continue to build the PAN network fused from bottom to top, enhance the top feature position information, and select the feature fusion network layer features fpn ={Z3, Z4, Z5}, features pan{F3, F4, F5} from bottom to top fusion mode is as follows: each stage fusion and its from the corresponding feature branch is down-sampling fusion, the specific formula is as follows: Z3=F3,

[0041] Where, Conv 3,2,0 (·) represents the convolution operation of 3*3 convolution kernel step size 2, zero padding, and the final feature map is generated according to the above operation.

[0042] The prediction mechanism described in step 6 carries out candidate region extraction and generates loss to form a training process, which specifically includes the following steps:

[0043] Step 6-1: First, based on the feature map, the anchor box Anchor is constructed based on the K-means algorithm by clustering the pre-training set label, so as to achieve the goal that Anchor and the adjacent real value have greater coincidence; here, 9 cluster centers are given, which are set to 9 categories;

[0044] Step 6-2: Construct RPN loss to realize candidate region target selection, and design cascade threshold to determine positive and negative samples for training, where the positive and negative sample selection threshold is conf rpn_1 =0.5, conf rpn_1 =0.6, conf rpn_1 =0.7;

[0045] Step 6-3: According to the positive and negative sample selection method, the loss function is constructed and the training process is formed; the RPN loss process includes two parts of classification and regression: L rpn = L cls +L res , wherein the classification loss L cls is calculated according to the cross entropy loss

[0046] Where, p i represents that the anchor point is determined as a defect, represents the real probability value, and the predicted position in the anchor box is 1 if it is a real defect, otherwise 0; N is a batch image processing parameter, in order to prevent memory overflow, here N=32 is selected;

[0047] The regression loss is calculated according to the IoU loss, L res =1-IoU, wherein S p is the predicted position, S t is the real position; the total loss function is the superposition loss of classification and regression:

[0048] Step 6-4: Construct mask loss L maskThe final loss is generated with the precision multi-classification and regression loss, wherein For the pixel-by-pixel classification loss, the final loss is as follows: Loss = L rpn + L cls + L res + L mask .

[0049] wherein C, H and W are respectively the channel number, height and width of the image; alpha is a class imbalance factor, and herein 0.5 is taken according to the class distribution of the data set; and beta is a hard example mining factor, and herein 1.5 is taken;

[0050] Step 6-5: Finally, the predicted position is corrected according to the real position of the label through continuous training, and when the loss function is stable and converges to a minimum value, a defect detection model is obtained.

[0051] Step 7: The posterior probability distribution is used to construct a probability graph model based on Markov chain, and the specific steps are as follows:

[0052] Step 7-1: Constructing a post-processing target function: the Markov post-processing target function is set as Adjacent pixels have high similarity, wherein gamma is a non-negative constant parameter for controlling the spatial smoothness level, and N i is the adjacent pixel of pixel i, and T ij is a spatial interaction term, and is defined as Here, the function delta (a, b) is defined as follows: Sigma is a scale parameter; the spatial interaction term T ij Actually contains information associated with the class boundary, and tends to make adjacent pixels have the same label, and by optimizing this term, the classification boundary alignment is helped, and specifically, for adjacent pixels i and j in the image plane region, Will be close to 1, and after model optimization, y i and y j may have the same class label; for adjacent pixels i and j across a strong edge, Is smaller, so after model optimization, y i and y j may use different labels;

[0053] Step 7-2: The energy minimization problem in step 7-1 is solved based on the graph cut method to obtain the final target label.

[0054] Compared with the prior art, the present application has the following beneficial technical effects:

[0055] First, the mask annotation of the pipe girth weld radiographic image welding repair site is performed to establish a signature database. Then, a deep learning instance segmentation model is constructed based on data driving. The feature is fully extracted by constructing a backbone network and a neck network. Then, a pixel-by-pixel segmentation loss is constructed to guide the training process. Finally, a probability graph model is constructed, and the final result is obtained based on the posterior probability distribution. The method provided by the application combines the neural network model and the probability image model well, and greatly improves the recognition accuracy of the welding repair site. BRIEF DESCRIPTION OF DRAWINGS

[0056] Figure 1 A pipe girth weld radiographic image welding repair site intelligent identification method structure diagram is provided for the embodiments of the application.

[0057] Fig. 2 is a schematic diagram of two groups of original crack images and self-adaptive brightness enhancement, wherein Fig. 2(a) is two groups of original crack images, and Fig. 2(b) is a schematic diagram of two groups of crack self-adaptive brightness enhancement. DETAILED DESCRIPTION

[0058] The application will be described in detail below in combination with the drawings and embodiments, but the protection scope of the application is not limited by the drawings and embodiments.

[0059] Embodiment 1:

[0060] A pipe girth weld radiographic image welding repair site intelligent identification method, which mainly includes the following parts: data preprocessing part, feature extraction and fusion part, prediction part, post-processing part, and the overall structure diagram is as shown in Figure 1 .

[0061] Step 1: Based on the digital device scanning the industrial film, the original pipe weld defect image is obtained and the image denoising processing is performed.

[0062] Step 1-1, after the industrial X-ray film is digitized, due to the influence of many factors such as photosensitive material, X-ray energy and radiation dose, the digitized image has the disadvantages of low contrast and high noise. Therefore, in order to make the target features in the image obvious, the image needs to be processed to remove noise, and the mean filtering method is used to realize image smoothing.

[0063] First, a plurality of defect images of the same size are cut from the X-ray defect original image, here the square image is cut with the smaller side of the length and width as the side length, in order to remove noise, the median filter which is effective for pulse noise removal is used for filtering operation on the image, and the formula can be expressed as (2.11):

[0064] Y(i,j)=Med A[X(i,j)], i = 1,2, L h; j = 1,2, L w (2.11)

[0065] Here X(i,j) represents the pixel value of the i-th row and j-th column of the original image, Y(i,j) is the image after median filtering of the i-th row and j-th column. h and w represent the height and width of the original image respectively, and A is the median filter, a 3*3 filter window is used in this section.

[0066] Step 2: The original pipe weld image has large changes in blackness value during the scanning process, so the image is subjected to adaptive brightness enhancement. The brightness enhancement effect is shown in Figure 2, which shows two groups of crack original images and adaptive brightness enhanced images.

[0067] Step 2-1: Data enhancement is performed, and the original ray image is subjected to adaptive brightness enhancement. First, the cumulative pixel probability density is calculated using the improved histogram equalization. The pixel number n of each gray level of the original image is counted i , where 0 < i < L, L is the total number of gray levels in the image, L = 255.

[0068] Step 2-2: Determine the probability of pixel occurrence, p x (i) = p(x = i) = n i / n, where n is the total number of pixels in the image, p x (i) is the pixel histogram, and is normalized to the interval [0, 1] according to the formula p x (i) = p x (i) / max(p x (i)).

[0069] Step 2-3: Calculate the cumulative probability distribution to determine the cumulative histogram,

[0070] Step 2-4: Threshold selection, determine the pixel cutoff moving interval, determine the pixel center threshold p m , the pixel moving interval range is [0, m], p m < p tre , where p tre is the moving interval probability distribution threshold, which is selected as 0.8 here.

[0071] Step 2-5: Reconstruct the image pixel to determine the enhanced image, the pixel moving interval range is [0, m], p m < p tre , where p tre is the moving interval probability distribution threshold, which is selected as 0.8 here.

[0072]

[0073] reconstructing the image pixels to obtain an enhanced image X;

[0074] Step 3: label making based on the enhanced ray image and generalization.

[0075] Step 3-1: region annotation by nondestructive testing engineers according to the repair area contour to obtain region annotation information and form a training set where X k is the kth image, Y k is the label corresponding thereto, and N is the number of training sets, which is 2163 label images.

[0076] Step 3-2: generalization operation on the training set images, which adopts image translation, image folding, image scaling, image splicing and the like, and corresponding processing of the labels is performed to obtain an expanded data set Here, M = 11789.

[0077] Step 4: feature extraction of the weld repair site based on a Resnet50 backbone network. The prepared weld image and the corresponding label mask are input into the network to extract deep features.

[0078] Step 4-1: first, the annotated image is input into the Focus structure for image segmentation, which greatly reduces the image size while not losing information, increases the dimension, and reduces the network overhead.

[0079] Step 4-2: then, the annotated image is input into a resnet50 convolutional neural network with 5 stages. The network level gradually deepens, and the abstract level of the input image also improves in turn, and rich semantic information of the mask region can be extracted through feature extraction.

[0080] Step 5: design of a neck network to realize rich feature fusion, which is conducive to selecting and increasing reasonable features for subsequent classification and regression operations.

[0081] Step 5-1: a feature pyramid model is established after the backbone network to fuse features from top to bottom. The three-stage top layer extracted by the backbone network is denoted as: features fpn ={Z3, Z4, Z5}. The neck fusion mode is as follows: each stage is fused with the corresponding feature branch from the horizontal direction, and the specific formula is as follows:

[0082] Z5 = P5,

[0083] Step 5-2: Continue to build the bottom-up fusion PAN network, enhance the top feature position information. Select the feature fusion network layer features fpn = {Z3, Z4, Z5}, features pan = {F3, F4, F5} from bottom to top fusion method is as follows: each stage fusion and its from the corresponding feature branch is down-sampling fusion, the specific formula is as follows: Z3= F3.

[0084] Where, Conv 3,2,0 (·) represents a 3*3 convolution kernel step size of 2, zero padding under the convolution operation, according to the above operation to generate the final feature map.

[0085] Step 6: The prediction mechanism extracts candidate regions and generates loss to form the training process.

[0086] Step 6-1: First, based on the feature map, the anchor box is constructed based on the K-means algorithm by clustering the pre-training set label, so that the anchor and the adjacent true value have greater coincidence degree target, here 9 clustering center points are given, which are set to 9 categories.

[0087] Step 6-2: Construct RPN loss to realize candidate region target selection, design cascade threshold to determine positive and negative samples for training, here the positive and negative sample selection threshold is conf rpn_1 = 0.5, conf rpn_1 = 0.6, conf rpn_1 = 0.7.

[0088] Step 6-3: According to the positive and negative sample selection method, the loss function is constructed and the training process is formed. The RPN loss process includes two parts of classification and regression. L rpn = L cls + L res Where the classification loss L cls is calculated according to the cross entropy loss Where p i represents the anchor point in the probability is determined as a defect, the real probability value, if the predicted position in the anchor box is a real defect, it is 1, otherwise it is 0.

[0089] The regression loss is calculated according to the IoU loss, L res = 1-IoU, where S p is the predicted position, S t is the real position. The total loss function is the superposition loss of classification and regression:

[0090] Step 6-4: Constructing the mask loss L mask The final loss is generated with the precision multi-classification and regression loss, where is the pixel-wise classification loss. The final loss is as follows: Loss = L rpn + L cls + L res + L mask :

[0091] Step 6-5: Finally, through continuous training, the predicted position is corrected according to the true position of the label. When the loss function is stable and converges to the minimum value, the defect detection model is obtained.

[0092] Step 7: Obtain the probability map model:

[0093] Step 7-1: Construct the post-processing target function: Set the Markov post-processing target function as The adjacent pixels have high similarity, where γ is a non-negative constant parameter that controls the spatial smoothness level, N i is the adjacent pixel of pixel i, T ij is the spatial interaction term, defined as Here, the function δ(a, b) is defined as follows:

[0094] σ is a scale parameter.

[0095] The spatial interaction term T ij actually contains information associated with the class boundary and tends to make adjacent pixels have the same label. By optimizing this term, it helps to align the classification boundary. Specifically, for adjacent pixels i and j in the image plane region, will be close to 1, and after model optimization, y i and y j may have the same class label. For adjacent pixels i and j that cross a strong edge, is small, so after model optimization, y i and y j may use different labels.

[0096] Step 7-2: Based on the graph cut method, solve the energy minimization problem in step 7-1 to obtain the final target label.

Claims

1. A method for intelligent identification of welding repair areas in radiographic images of circumferential welds in long-distance pipelines, characterized in that, Includes the following steps: Step 1: Obtain the original image of the pipe weld defect by scanning industrial film using digital equipment and perform image denoising processing: Step 1-1: Industrial X-ray film is digitized and smoothed using mean filtering. First, multiple defect images of the same size are cropped from the original X-ray defect image. Here, a square image is cropped with the side with the smaller length and width as the side length. Step 1-2: To remove noise, a median filter, effective for impulse noise removal, is used to filter the image. The formula is: ； in, This represents the pixel values ​​in the i-th row and j-th column of the original image. This is the image after median filtering of the i-th row and j-th column; h and w represent the height and width of the original image, respectively, and A is the median filter with a 3*3 filtering window. Step 2: Adaptive brightness enhancement is performed on the original pipe weld image; Step 3: Create and generalize labels based on the enhanced ray image: Step 3-1: The non-destructive testing engineer marks the area according to the outline of the rework area to obtain the area marking information and form a training set. ,in Let be the k-th image. For each corresponding label, N is the number of training images, N = 2163 labeled images; Step 3-2: Perform generalization operations on the training set images, using image translation, image flipping, image scaling, and image stitching, and perform corresponding processing on their labels to obtain the expanded dataset. M=11789; Step 4: Extract features from weld repair areas based on the ResNet50 backbone network, extracting deep features: Step 4-1: First, input the labeled image into the Focus structure for image segmentation; Step 4-2: Then, the labeled image is fed into a ResNet50 convolutional neural network with 5 stages; Step 5: Design the neck network to achieve feature fusion; Step 5, designing the neck network to achieve feature fusion, specifically includes the following steps: Step 5-1: After the backbone network, build a feature pyramid model to fuse features from top to bottom. Select the top-level features of the three stages extracted by the backbone network and denote them as: , The neck fusion method is as follows: each stage of fusion is fused with the corresponding feature branch from the horizontal direction, and the specific formula is shown below: ， ； Among them, P5, P4, and P3 are the feature maps of the three stages during top-down fusion. The feature maps represent the three stages of bottom-up fusion, and Upsample is the upsampling interpolation function. This is a pixel overlay function; Step 5-2: Continue building the bottom-up fused PAN network, enhance the top-level feature location information, and select the feature fusion network layer from the previous stage. , The bottom-up fusion method is as follows: Each stage of fusion is downsampled and fused with the feature branch corresponding to the horizontal branch from which it comes, as shown in the following formula: , , ; in, The convolution operation with a stride of 2 represents a 3*3 convolution kernel, and the final feature map is generated according to the above calculation. Step 6: The prediction mechanism extracts candidate regions and generates loss to form the training process; Step 7: Construct a probabilistic graphical model using Markov chains based on the posterior probability distribution.

2. The intelligent identification method for welding repair parts of circumferential welds in long-distance pipelines using radiographic images according to claim 1, characterized in that: The specific steps involved in adaptive brightness enhancement of the original pipe weld image described in step 2 are as follows: Step 2-1: Perform data augmentation by adaptively enhancing the brightness of the original pipe weld image. First, an improved histogram equalization is used to calculate the cumulative pixel probability density; then, the number of pixels at each gray level of the original image is counted. ,in L represents all gray levels in the image, L=255; Step 2-2: Determine the probability of pixel occurrence: ； Where n is the total number of pixels in the image. For pixel histograms, and according to the formula Normalize the interval [0, 1]; Step 2-3: Calculate the cumulative probability distribution and determine the cumulative histogram: ； in, This represents the pixel distribution map after calculating the cumulative histogram of pixels. Steps 2-4: Threshold selection, determining the pixel cutoff movement range, and determining the pixel center threshold. Let the range of pixel movement be . ,in The threshold for the probability distribution of the movement interval is set to 0.8 here; Steps 2-5: Reconstruct image pixels to determine the enhanced image; the reconstruction formula for each pixel value is as follows: ； in, The grayscale value of each pixel, k is a scaling factor, k=1.

1.

3. The intelligent identification method for welding repair parts of circumferential welds in long-distance pipelines using radiographic images according to claim 1, characterized in that: Step 6 describes a prediction mechanism that extracts candidate regions and generates a loss to form a training process, specifically including the following steps: Step 6-1: First, the anchor box is constructed based on the feature map. This is based on the K-means algorithm, which uses clustering to pre-train the label set, so as to achieve a greater overlap between the anchor and the nearby ground truth values. Here, 9 cluster center points are given, and 9 classes are set. Step 6-2: Construct the RPN loss to achieve candidate region target selection, and design cascaded thresholds to determine positive and negative samples for training. Here, the thresholds for positive and negative samples are 0.7 and 0.3, respectively. Step 6-3: Construct the loss function and form the training process based on the positive and negative sample selection method; the RPN loss process consists of two parts: classification and regression. Classification loss Calculated based on cross-entropy loss , in, This indicates that the anchor point is determined to be a defect based on probability. This represents the true probability value; the predicted position within the anchor box is 1 if it is a real defect, otherwise it is 0; N is the batch image processing parameter, N=32; Regression loss is calculated based on IoU loss. ,in ,in To predict the location, The true location is the sum of the classification loss and the regression loss. Step 6-4: Constructing the mask loss ,in The final loss is as follows, calculated using pixel-by-pixel classification loss: ; Where C, H, and W are the number of image channels, height, and width, respectively; α is the class imbalance factor, which is set to 0.5 based on the class distribution of the dataset, and β is set to 1.5; Step 6-5: Finally, through continuous training, the predicted position is corrected according to the actual position of the label. When the loss function is stable and converges to the minimum value, the defect detection model will be obtained.

4. The intelligent identification method for welding repair parts of radiographic images of circumferential welds in long-distance pipelines according to claim 1, characterized in that: Step 7 describes constructing a probabilistic graphical model using Markov chains based on the posterior probability distribution. The specific steps are as follows: Step 7-1: Construct the post-processing objective function: Set the Markov post-processing objective function as follows: Adjacent pixels have a high degree of similarity, among which It is a non-negative constant parameter that controls the level of spatial smoothness. Let i be the neighboring pixels of pixel i. For spatial interaction items, defined as ,here The function is defined as follows: , is a scale parameter; yi and yj are the predicted categories of neighboring pixels i and j, respectively, and ci and cj are the pixel values ​​at the positions of neighboring pixels i and j in the image; for neighboring pixels i and j within a planar region of the image, It will approach 1. and They have the same class tag; For adjacent pixels i and j that cross a strong edge, Smaller, therefore and Use different labels; Step 7-2: Solve the energy minimization problem in Step 7-1 based on the graph cut method to obtain the final target label.

Images