Crack identification model training method and device, electronic equipment and storage medium

Abstract

The embodiment provides a crack identification model training method and device, electronic equipment and a storage medium. The input data set of the model can be generated by using a random cosine function conforming to the formation crack characteristics, sample expansion in a small sample condition is realized, data acquisition work is simplified, cracks in a sample image can be automatically labeled, human and time costs are saved, a Cluster NMS is used to replace an NMS of a yolo v5s original model to perform a redundant frame deletion operation, a weight and a center point distance penalty term are added to complete non-maximum suppression, the model can more quickly and accurately obtain an optimal frame, and therefore, target detection effect is improved, which is of great significance for crack reservoir exploration.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of oil exploration and development, and particularly relates to a fracture identification model training method and device, an electronic device, and a storage medium. BACKGROUND

[0002] Fractured reservoirs have always been the focus and difficulty of exploration. In geological exploration work, it is very important to identify fractures and half fractures. Imaging logging images contain a lot of fracture information. At present, the main method to obtain fractures in imaging logging images is manual identification. However, this method has the disadvantages of large workload, large error, and slow efficiency. SUMMARY

[0003] The main purpose of the embodiments of the present application is to provide a fracture identification model training method and device, an electronic device, and a storage medium, so as to improve the efficiency of formation fracture identification.

[0004] In a first aspect, the embodiments of the present application provide a fracture identification model training method, comprising:

[0005] A plurality of sample images are generated by simulating a random cosine function of a formation fracture, and image enhancement is performed on each sample image. The generated sample images and the sample images obtained after enhancement are added to the sample image set;

[0006] Any sample image in the sample image set is input into a fracture identification model to obtain a plurality of prediction boxes corresponding to the sample image and a confidence of each prediction box. The fracture identification model is implemented based on a yolov5 network model, and the prediction box is used to indicate the position of the fracture in the sample image;

[0007] A final prediction box is determined from the plurality of prediction boxes according to the confidence corresponding to the plurality of prediction boxes, the area of the overlapping region of each pair of prediction boxes, and the center point of each prediction box;

[0008] A loss value is calculated according to the final prediction box, and the parameters of the fracture identification model are adjusted according to the loss value.

[0009] Optionally, the final prediction box is determined from the plurality of prediction boxes according to the confidence corresponding to the plurality of prediction boxes, the area of the overlapping region of each pair of prediction boxes, and the center point of each prediction box, comprising:

[0010] The plurality of prediction boxes corresponding to the sample image are arranged in descending order of confidence to obtain a prediction box sequence;

[0011] performing IOU calculation on each of the plurality of prediction boxes in the prediction box sequence to obtain an IOU matrix, wherein an element in the ith row and jth column of the IOU matrix is a result of performing IOU calculation on a prediction box ranked ith in confidence and a prediction box ranked jth in confidence;

[0012] determining at least one redundant box from the plurality of prediction boxes according to the IOU matrix, and performing a deletion operation on the at least one redundant box;

[0013] after deleting the redundant box, determining a prediction box with the highest confidence from the remaining prediction boxes, and calculating an area of an overlapping region of each of the remaining prediction boxes and the prediction box with the highest confidence;

[0014] determining a final prediction box according to the calculated area and a center point of each prediction box.

[0015] Optionally, before performing IOU calculation on each of the plurality of prediction boxes in the prediction box sequence to obtain an IOU matrix, the method further comprises:

[0016] performing IOU calculation on the prediction box with the highest confidence and each of the remaining prediction boxes to obtain a calculation result, and screening a prediction box with a calculation result greater than or equal to a non-maximum suppression (NMS) threshold from the remaining prediction boxes;

[0017] performing weighted summation on the prediction box with the highest confidence and the screened prediction box to obtain a new rectangular box, wherein a weight of each prediction box in the weighted summation is a corresponding confidence;

[0018] using the new rectangular box to replace the prediction box with the highest confidence to obtain an updated prediction box sequence.

[0019] Optionally, determining a final prediction box according to the calculated area and a center point of each prediction box comprises:

[0020] if there are at least two prediction boxes with the same area of the overlapping region with the prediction box with the highest confidence, calculating a distance between each of the at least two prediction boxes and a center point of the prediction box with the highest confidence, retaining a prediction box with a maximum distance, and deleting the remaining prediction boxes in the at least two prediction boxes;

[0021] if the areas of the overlapping regions corresponding to each prediction box are all different, recalculating a confidence of each prediction box in the prediction box sequence according to the following formula:

[0022]

[0023] wherein M represents the prediction box with the highest confidence, D represents any prediction box other than the prediction box with the highest confidence, and Score represents a confidence of the prediction box other than the prediction box with the highest confidence. j j ​D represents j a corresponding confidence, threshold is a preset threshold, d is the distance between the centers of two prediction boxes, c is the diagonal length of the minimum rectangular frame containing the two prediction boxes, and γ is a constant parameter;

[0024] According to the recalculated confidence, the prediction box with the highest confidence is selected as the final prediction box.

[0025] Optionally, according to the IOU matrix, at least one redundant box is determined from the plurality of prediction boxes for a deletion operation, comprising:

[0026] The following operations are repeatedly performed until the tensors obtained twice in succession are the same: the current IOU matrix is subjected to upper triangular operation to obtain an operated IOU matrix; the maximum value of each column is selected from the operated IOU matrix, and the maximum value is compared with an NMS threshold to obtain a tensor formed by 0 and / or 1, wherein 1 indicates that the maximum value is less than the NMS threshold, and 0 indicates that the maximum value is greater than the NMS threshold; the tensor is taken as the diagonal line of a matrix to generate a diagonal matrix E; the updated IOU matrix is obtained by multiplying the IOU matrix by the diagonal matrix E from the left;

[0027] After the same tensor is obtained twice in succession, it is judged according to the elements in the tensor whether the corresponding prediction box is a redundant box, and a deletion operation is performed on the determined redundant box, wherein an element value of 0 indicates that the corresponding prediction box is a redundant box, and an element value of 1 indicates that the corresponding prediction box is not a redundant box.

[0028] Optionally, a plurality of sample images are generated by simulating a random cosine function of a stratum fracture, and each sample image is subjected to image enhancement, and the generated sample images and the sample images obtained after enhancement are added to the sample image set, comprising:

[0029] According to the random cosine function of the stratum fracture, a corresponding cosine curve is obtained;

[0030] The cosine curve is subjected to a plurality of adjustment operations to obtain a plurality of corresponding sample images; wherein the adjustment operation comprises at least one of the following: adding random noise, translation, adjusting thickness, adjusting inclination, adjusting the position of the curve in the image; the adjustment amount of each adjustment operation is different;

[0031] Each sample image is subjected to image enhancement to obtain a plurality of sample images after image enhancement;

[0032] The generated sample images and the sample images obtained after enhancement are added to the sample image set.

[0033] In a second aspect, an embodiment of the present application provides a fracture identification method, comprising:

[0034] obtain an imaging logging image to be identified;

[0035] input the imaging logging image to be identified into a fracture identification model to obtain a fracture position in the imaging logging image;

[0036] The fracture identification model is a model trained based on the method in any one of claims 1-6.

[0037] In a third aspect, an embodiment of the present application provides a training device of a fracture identification model, including:

[0038] The generating module is configured to generate a plurality of sample images by simulating a random cosine function of a formation fracture, perform image enhancement on each sample image, and add the generated sample images and the sample images obtained after enhancement to the sample image set.

[0039] The training module is configured to input any sample image in the sample image set into a fracture identification model to obtain a plurality of prediction boxes corresponding to the sample image and a confidence of each prediction box. The fracture identification model is implemented based on a yolov5 network model. The prediction box is used to indicate a position of a fracture in the sample image.

[0040] The determining module is configured to determine a final prediction box from the plurality of prediction boxes according to the confidence corresponding to each prediction box, an area of an overlapping region of each pair of prediction boxes, and a center point of each prediction box.

[0041] The adjusting module is configured to calculate a loss value according to the final prediction box, and adjust parameters of the fracture identification model according to the loss value.

[0042] In a fourth aspect, an embodiment of the present application provides an electronic device, including:

[0043] at least one processor; and a memory connected with the at least one processor in communication;

[0044] The memory stores instructions executable by the at least one processor. The instructions are executed by the at least one processor to enable the electronic device to perform the method in any one of the above aspects.

[0045] In a fifth aspect, an embodiment of the present application provides a computer readable storage medium, which stores computer execution instructions. When a processor executes the computer execution instructions, the method in any one of the above aspects is implemented.

[0046] The training method, apparatus, electronic device, and storage medium for the fracture recognition model provided in this embodiment include: generating multiple sample images by simulating a random cosine function of a formation fracture, performing image enhancement on each sample image, and adding the generated sample images and the enhanced sample images to the sample image set; inputting any sample image from the sample image set into the fracture recognition model to obtain multiple prediction boxes corresponding to the sample image and the confidence level of each prediction box; wherein, the fracture recognition model is implemented based on a YOLOv5 network model, and the prediction boxes are used to indicate the location of the fracture in the sample image; determining the final prediction box from the multiple prediction boxes based on the confidence level corresponding to the multiple prediction boxes, the area of ​​the overlapping region of each pair of prediction boxes, and the center point of each prediction box; calculating the loss value based on the final prediction box, and adjusting the parameters of the fracture recognition model based on the loss value. By using a stochastic cosine function that conforms to the characteristics of formation fractures to generate the model's input dataset, sample expansion is achieved under small sample conditions, simplifying data collection. Furthermore, fractures in sample images can be automatically labeled, saving manpower and time costs. At the same time, Cluster NMS is used instead of the original YOLO v5s model's NMS to perform redundant box removal operations, and weights and center point distance penalties are added to complete non-maximum suppression, enabling the model to obtain the optimal bounding box faster and more accurately, thereby improving the target detection effect. This is of great significance for the exploration of fractured reservoirs. Attached Figure Description

[0047] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0048] Figure 1 An application scenario diagram provided for an embodiment of this application;

[0049] Figure 2 A flowchart illustrating a training method for a crack recognition model provided in an embodiment of this application;

[0050] Figure 3 A sample image generated by a random cosine function simulating formation fractures, provided for an embodiment of this application;

[0051] Figure 4 A crack recognition image output by a trained crack recognition model, provided for an embodiment of this application;

[0052] Figure 5 A flowchart illustrating a crack identification method provided in an embodiment of this application;

[0053] Figure 6 A schematic diagram of the structure of a training device for a crack recognition model provided in an embodiment of this application;

[0054] Figure 7 A structural schematic diagram of an electronic device is provided for an embodiment of the present application.

[0055] The specific embodiments of the present application have been shown through the above-described drawings, and will be described in more detail hereinafter. These drawings and textual descriptions are not intended to limit the scope of the concept of the present application in any way, but to illustrate the concept of the present application to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0056] Exemplary embodiments will be described in detail herein with reference to the attached drawings. The same reference numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments are not meant to represent all implementations consistent with the present application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the present application as detailed in the appended claims.

[0057] Identification of cracks and half-cracks is of great significance for exploration of fractured reservoirs.

[0058] In some technologies, artificial identification methods are used to obtain cracks in imaging logging images, but such methods have the disadvantages of heavy workload, large errors, and slow efficiency.

[0059] In other technologies, model training methods can be used to identify cracks in imaging logging images. In order to improve the accuracy of model identification, a large number of cracks in imaging logging images need to be manually annotated before training the model, which still consumes huge human and time costs.

[0060] Therefore, the present application provides a training method of a crack identification model, which can automatically annotate cracks in sample images, save the tedious manual annotation, save a large amount of human and time costs, and introduce a center point distance penalty term to complete non-maximum suppression. According to the multiple prediction boxes obtained by the trained model, the final prediction box can be determined more quickly and accurately.

[0061] Figure 1 An application scenario diagram is provided for an embodiment of the present application. As shown in Figure 1As shown, first, a sample image is obtained according to a cosine function simulating a formation fracture, then the sample image is data enhanced to obtain an enhanced sample image, then the sample image and the enhanced sample image are combined together to form a sample image set, each sample image in the sample image set is labeled to obtain a minimum rectangular frame in which the fracture is located, then the sample images in the sample image set are sequentially input into a fracture identification model based on a yolov5 network model to obtain a plurality of prediction frames corresponding to the sample image and a confidence of each prediction frame, according to the plurality of prediction frames, an overlapping area of each pair of prediction frames and a center point of each prediction frame are determined, then according to the overlapping area, the center point and the confidence, a final prediction frame is determined, the final prediction frame is compared with the minimum rectangular frame corresponding to the fracture in the sample image to calculate a loss value, and the parameters of the fracture identification model are adjusted according to the loss value.

[0062] The training method of the fracture identification model provided in the application can directly generate a plurality of sample images through a random cosine function simulating a formation fracture, and a large number of sample images are obtained through image enhancement, thereby overcoming the shortcomings of high cost and insufficient quantity of imaging logging image acquisition, and the position of the fracture in the sample image is determined during the generation of the sample image, so that the position of the fracture can be automatically labeled, thereby saving the tediousness of manual labeling and saving time and labor cost. After obtaining the sample image, a plurality of prediction frames corresponding to the sample image and a confidence of each prediction frame can be obtained by inputting any sample image into the fracture identification model, and according to the confidence of the prediction frame, the area of the overlapping region of each pair of prediction frames and the position of the center point of the prediction frame, a final and unique prediction frame can be quickly and accurately obtained. The final prediction frame is compared with the rectangular frame in which the fracture is located to calculate a loss value, the parameters of the fracture identification model are adjusted using the loss value, and finally, the plurality of prediction frames output from the fracture identification model after multiple adjustments and the confidence corresponding to the prediction frame can be used to more quickly and accurately determine the final prediction frame.

[0063] The technical solutions of the application and how the technical solutions of the application solve the above technical problems will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes can not be described again in some embodiments. The embodiments of the application will be described below with reference to the accompanying drawings.

[0064] Figure 2 A flowchart of a training method of a fracture identification model provided in an embodiment of the application. As shown in 2, the training method of the fracture identification model provided in the embodiment of the application can include:

[0065] In step 201, a plurality of sample images are generated by simulating a random cosine function of a formation fracture, and image enhancement is performed on each sample image, and the generated sample images and the sample images obtained after enhancement are added to the sample image set.

[0066] The image enhancement refers to operations such as mirroring, rotating, scaling, cropping, changing illumination intensity, and changing contrast on the sample images.

[0067] Specifically, assuming that the filler in the fracture is a low-resistance medium, the surrounding rock is a high-resistance surrounding rock layer, and the well wall extends along the north, east, south, and west directions, the cosine function expression for simulating the formation fracture is as follows: wherein θ represents the fracture dip angle, and n represents random noise. According to the cosine function for simulating the formation fracture, a corresponding cosine curve is obtained, and then a plurality of sample images containing the cosine curve are generated. After that, data enhancement is performed on each sample image. Figure 3 A sample image generated by simulating a random cosine function of a formation fracture is provided in the embodiment of the present application. The generated sample images and the enhanced sample images can be randomly shuffled and then added to the sample image set together.

[0068] After obtaining the sample image set, each sample image in the sample image set can be labeled to obtain the position and size of the smallest rectangular frame in which the fracture in the sample image is located.

[0069] There are many methods for sample labeling. In an optional implementation, a data labeling tool Labelimg can be used to label the sample images to generate xml format labeling data. In another optional implementation, the fractures in the images can be automatically labeled during the generation of the sample images.

[0070] After labeling, the sample images in the sample image set are evenly divided into 10 parts, 9 parts are set as a training set, that is, a set of images used for training the model, and 1 part is set as a validation set, that is, a set of images used for verifying the effect of the model, and a fracture data set is produced using the VOC2007 standard data set format.

[0071] Optionally, a plurality of sample images are generated by simulating a random cosine function of a formation fracture, and image enhancement is performed on each sample image, and the generated sample images and the sample images obtained after enhancement are added to the sample image set, including:

[0072] According to the random cosine function for simulating the formation fracture, a corresponding cosine curve is obtained.

[0073] The cosine curve is adjusted multiple times to obtain a plurality of sample images corresponding to the adjustment operations; wherein the adjustment operations include at least one of the following: adding random noise, translation, adjusting thickness, adjusting inclination, adjusting the position of the curve in the image; the adjustment amount is different each time the adjustment operation is performed;

[0074] The sample images are subjected to image enhancement to obtain a plurality of sample images after image enhancement;

[0075] The generated sample images and the sample images obtained after enhancement are added to the sample image set.

[0076] Specifically, the expression of the cosine function simulating the formation fracture is For example, the expression of the cosine function simulating the formation fracture is

[0077] First, the cosine curve is drawn according to the expression corresponding to the cosine function, and then the cosine curve is subjected to adjustment operations, which can include: adding random noise, translation, adjusting thickness, adjusting inclination, adjusting the position of the curve in the image.

[0078] The n in the above cosine function represents random noise, and when the intensity of the added random noise is adjusted, only the value of n needs to be changed. When n=0, it indicates that no random noise is added, and when n≠0, it indicates that random noise has been added. The size of n represents the intensity of the added random noise, and the larger n is, the greater the intensity of the added random noise.

[0079] θ represents the inclination of the fracture, and when the inclination of the fracture is adjusted, only the value of θ needs to be changed. The larger the value of θ is, the greater the inclination of the fracture.

[0080] When the cosine curve is subjected to translation operation, only a fixed value m needs to be added or subtracted to x, and the fixed value m indicates that the cosine curve moves to the left by m distance, and the fixed value m indicates that the cosine curve moves to the right by m distance. The absolute value of m is changed to adjust the amplitude of the curve translation, and the larger the absolute value of m is, the greater the amplitude of the translation.

[0081] In addition, the thickness of the curve and the position of the curve in the image can also be adjusted. The curve can be at any position in the image, and at least one curve can be included in an image.

[0082] Optionally, when the cosine curve is subjected to adjustment operation, only one of the above operations can be selected for operation, or multiple operations can be selected for operation. For the same operation, the adjustment amount is different each time the operation is performed, so as to avoid the same sample image.

[0083] After the above adjustment operation is completed, a plurality of sample images are obtained, and the sample images are image-enhanced, and then the sample images before and after image enhancement are placed together, and after being randomly shuffled, an image set of the sample is obtained.

[0084] In practical applications, according to the expression of the cosine function of the simulated formation fracture Draw the cosine curve, and then modify the cosine curve by changing the intensity of random noise, adjusting the dip angle, etc. to obtain 1000 sample images. According to the principle of subtracting the right from the left, the cosine curve is shifted left and right, and 1000 sample images are generated again. Place the 1000 sample images obtained above together and randomly shuffle them to obtain 2000 sample images after shuffling. Then, randomly perform mirror image, rotation, scaling, cropping, changing the intensity of light, changing the contrast, and other image enhancement operations on the 2000 cosine curve images to obtain 10000 sample images after image enhancement. Place the 10000 sample images after image enhancement together with the 2000 sample images before image enhancement and randomly shuffle them to obtain a sample data set containing 12000 sample images.

[0085] In this way, by performing multiple adjustment operations on the cosine curve and image enhancement on the sample images, more diverse sample images can be obtained on the basis of the existing sample images, making the types of sample images in the sample image set more diverse.

[0086] Step 202, input any sample image in the sample image set into the crack identification model to obtain a plurality of prediction boxes corresponding to the sample image and the confidence of each prediction box; wherein the crack identification model is realized based on a yolov5 network model, and the prediction box is used to indicate the position of the crack in the sample image;

[0087] The confidence of the prediction box is the probability that the position of the prediction box is accurate.

[0088] Specifically, any one sample image in the sample image set is input into the crack identification model realized based on the yolov5 network model, and a plurality of prediction boxes corresponding to the sample image and the confidence of each prediction box can be obtained.

[0089] Step 203, determining a final prediction box from the plurality of prediction boxes according to the confidence corresponding to the plurality of prediction boxes, the area of the overlapping region of each pair of prediction boxes, and the center point of each prediction box;

[0090] The area of the overlapping region of each pair of prediction boxes refers to the area of the overlapping region of any two prediction boxes in the plurality of prediction boxes corresponding to the same sample image, and the center point of each prediction box refers to the geometric center point of each prediction box.

[0091] After the sample image is input into the crack identification model, a plurality of prediction boxes can be obtained, and the plurality of prediction boxes can be positions of cracks. According to the confidence of the plurality of prediction boxes, the area of the overlapping region of each pair of prediction boxes, and the center point of each prediction box, redundant prediction boxes can be found and deleted. Finally, the only prediction box left is the final prediction box.

[0092] Optionally, the final prediction box is determined from the plurality of prediction boxes according to the confidence of the plurality of prediction boxes, the area of the overlapping region of each pair of prediction boxes, and the center point of each prediction box, and includes:

[0093] The plurality of prediction boxes corresponding to the sample image are arranged in descending order of confidence to obtain a prediction box sequence.

[0094] IOU (Intersection over Union) calculation is performed on each pair of prediction boxes in the prediction box sequence to obtain an IOU matrix. An element in the ith row and jth column of the IOU matrix is a result of IOU calculation of a prediction box with a confidence rank of i and a prediction box with a confidence rank of j.

[0095] At least one redundant box is determined from the plurality of prediction boxes according to the IOU matrix, and a deletion operation is performed.

[0096] After the redundant box is deleted, a prediction box with the highest confidence is determined from the remaining prediction boxes, and the area of the overlapping region of each prediction box and the prediction box with the highest confidence is calculated.

[0097] The final prediction box is determined according to the calculated area and the center point of each prediction box.

[0098] The IOU is a ratio of the intersection to the union of the areas of two prediction boxes.

[0099] Specifically, the plurality of prediction boxes corresponding to a sample image are arranged in descending order of confidence to obtain a prediction box sequence: B1, B2, B3, B4, …, BN. N where N is the total number of the plurality of prediction boxes, the confidence of the prediction box B1 is s1, the confidence of the prediction box B2 is s2, …, and the confidence of the prediction box BN is SN. N The confidence of the prediction box B1 is s1, the confidence of the prediction box B2 is s2, …, and the confidence of the prediction box BN is SN. N Therefore, s1≥s2≥…≥SN. N .

[0100] IOU calculation is performed on each pair of prediction boxes in the prediction box sequence to obtain an IOU matrix. An element in the ith row and jth column of the IOU matrix x ij is a result of IOU calculation of B i and B jThe result obtained by performing IOU calculation, i.e., x ij = IOU(B i , B j ). When B i is completely overlapped with B j , x ij = 1, when B i is completely not overlapped with B j , x ij = 0, when B i is overlapped with B j but not completely overlapped, 0 < x ij < 1, and x ij is always equal to x ji .

[0101] Optionally, according to the IOU matrix, at least one redundant box is determined from the plurality of prediction boxes to perform a deletion operation, comprising:

[0102] The following operations are repeatedly performed until the tensors obtained in two consecutive times are the same: performing upper triangular operation on the current IOU matrix to obtain an operated IOU matrix; selecting the maximum value of each column on the operated IOU matrix, and comparing the maximum value with an NMS threshold to obtain a tensor formed by 0 and / or 1, wherein 1 represents that the maximum value is less than the NMS threshold, and 0 represents that the maximum value is greater than the NMS threshold; taking the tensor as the diagonal line of the matrix to generate a diagonal matrix E; multiplying the IOU matrix by the diagonal matrix E on the left to obtain an updated IOU matrix;

[0103] After the same tensor is obtained in two consecutive times, whether the corresponding prediction box is a redundant box is determined according to the elements in the tensor, and a deletion operation is performed on the determined redundant box, wherein the element value of 0 represents that the corresponding prediction box is a redundant box, and the element value of 1 represents that the corresponding prediction box is not a redundant box.

[0104] Specifically, assuming that the IOU matrix X1 obtained when N = 8 is as follows:

[0105]

[0106] Comparing any element x ij in the matrix X1 with an NMS (non maximum suppression) threshold, the NMS threshold can be set to 0.8, when x ij ≥ 0.8, x ij = 0, when x ij < 0.8, x ij = 1, then the matrix X1 becomes:

[0107]

[0108] The matrix X1' is operated by using the triu function of Pytorch to obtain the matrix X1" after operation as follows:

[0109]

[0110] The maximum value of each column of the matrix X1" is selected to form a tensor b1.

[0111] b1 = [0 1 1 1 0 1 1 1]

[0112] The tensor b1 is used as the diagonal line of the matrix to generate a diagonal matrix E1

[0113]

[0114] The IOU matrix X1 is multiplied by the diagonal matrix E1 on the left to obtain the updated IOU matrix X2 as follows:

[0115]

[0116] Similarly, comparing any element NMS threshold value 0.8 in the matrix X2, the matrix X2 becomes:

[0117]

[0118] The matrix X2' is operated by using the triu function of Pytorch to obtain the matrix X2" after operation as follows:

[0119]

[0120] The maximum value of each column of the matrix X2" is selected to form a tensor b2.

[0121] b2 = [1 1 1 1 1 1 1 1]

[0122] Comparing the tensors b1 and b2, it is found that b1≠b2, then the tensor b2 is used as the diagonal line of the matrix to generate a diagonal matrix E2, the IOU matrix X2 is multiplied by the diagonal matrix E2 on the left to obtain the updated IOU matrix X3, and the above operation is repeated to obtain the tensor b3, and the tensors b2 and b3 are continuously compared.

[0123] If b2≠b3, the above operation is repeated until the tensors obtained continuously are the same;

[0124] If b2=b3, it is judged according to the elements in the tensor b2 or b3 whether the corresponding prediction box is a redundant box. The element value of 0 indicates that the corresponding prediction box is a redundant box, and the element value of 1 indicates that the corresponding prediction box is not a redundant box. After the judgment is finished, the redundant box is deleted.

[0125] For example, if the element of the i-th column of the tensor is 1, the prediction box B i is not a redundant box and does not need to be deleted; if the element of the j-th column of the tensor is 0, the prediction box B j is a redundant box and needs to be deleted.

[0126] If the tensors obtained by two consecutive operations are both [0 1 0 0 1 0 0 0], the prediction boxes B1, B3, B4, B6, B7 and B8 are redundant boxes and need to be deleted, and the final remaining prediction boxes are B2 and B5.

[0127] The NMS model in the yolov5 network model is time-consuming to calculate because it uses the idea of sequential iteration. In the present application, the Cluster NMS is used to replace the NMS model in the yolov5 network model, and the matrix left multiplication method is adopted, so that only a cluster with the largest number of prediction boxes needs to be selected for iterative calculation, and calculation on all clusters is not needed, which can improve the operation efficiency without losing accuracy.

[0128] After deleting the redundant boxes, the prediction box with the highest confidence is found in the remaining prediction boxes, and the areas of the overlapping regions of all prediction boxes in the remaining prediction boxes except the prediction box with the highest confidence and the prediction box with the highest confidence are calculated.

[0129] For example, if the remaining prediction boxes after deleting the redundant boxes are B2, B4, B5 and B7, the prediction box with the highest confidence is B2, and the areas of the overlapping regions of the prediction boxes B4, B5 and B7 and the prediction box B2 are calculated respectively.

[0130] Optionally, according to the calculated areas and the center points of the prediction boxes, the final prediction box is determined, comprising:

[0131] If there are at least two prediction boxes with the same area of the overlapping region with the prediction box with the highest confidence, the distance between each prediction box in the at least two prediction boxes and the center point of the prediction box with the highest confidence is calculated, and the prediction box with the largest distance is retained, and the remaining prediction boxes in the at least two prediction boxes are deleted;

[0132] If the areas of the overlapping regions corresponding to the prediction boxes are all different, the confidence of each prediction box in the prediction box sequence is recalculated according to the following formula:

[0133]

[0134] wherein M represents the prediction box with the highest confidence, D j represents any prediction box other than the prediction box with the highest confidence, Score j represents D jThe corresponding confidence, threshold is a preset threshold, d is the distance between the centers of the two prediction boxes, c is the diagonal length of the minimum rectangular frame that completely contains the two prediction boxes, and γ is a constant parameter.

[0135] According to the recalculated confidence, the prediction box with the highest confidence is selected as the final prediction box.

[0136] Specifically, if there are two or more prediction boxes with the same area of the overlapping region with the prediction box with the highest confidence, all prediction boxes with the same area of the overlapping region with the prediction box with the highest confidence are regarded as a set, only the prediction box with the largest distance from the center of the prediction box with the highest confidence in the set is retained, and the other prediction boxes in the set are deleted. It should be noted that the prediction box with the highest confidence here refers to the prediction box with the highest confidence among the remaining prediction boxes after deleting the redundant boxes.

[0137] For example, if the remaining prediction boxes after deleting the redundant boxes are B2, B4, B5, and B7, the prediction box with the highest confidence is B2, and the areas of the overlapping regions of prediction boxes B4, B5, and B7 with prediction box B2 are calculated respectively.

[0138] If the areas of the overlapping regions of prediction boxes B4 and B5 with prediction box B2 are equal, the distances between the centers of prediction boxes B4 and B5 and prediction box B2 are calculated respectively, if the distance between the centers of prediction boxes B4 and B2 is greater, prediction box B4 is retained and prediction box B5 is deleted, otherwise, prediction box B5 is retained and prediction box B4 is deleted.

[0139] If the areas of the overlapping regions of prediction boxes B4, B5, and B7 with prediction box B2 are equal, the distances between the centers of prediction boxes B4, B5, and B7 and prediction box B2 are calculated respectively, if the distance between the centers of prediction boxes B4 and B2 is greater, prediction box B4 is retained and prediction boxes B5 and B7 are deleted, if the distance between the centers of prediction boxes B5 and B2 is greater, prediction box B5 is retained and prediction boxes B4 and B7 are deleted, and if the distance between the centers of prediction boxes B7 and B2 is greater, prediction box B7 is retained and prediction boxes B4 and B5 are deleted.

[0140] If the calculated areas of the overlapping regions are not the same, first calculate the DIOU (Distance-IOU, IOU with center point distance penalty term) value according to the IOU value, and the calculation formula is as follows:

[0141]

[0142] wherein d is the distance between the centers of the two prediction boxes, c is the diagonal length of the smallest rectangular box that can completely contain the two prediction boxes, and γ is a constant parameter, usually set between 2 and 10. When γ gradually increases to infinity, DIOU is converted into IOU, and then DIOU_NMS is consistent with the operation of the standard NMS; and when γ gradually decreases to 0, all the prediction boxes with a distance of 0 from the center of the prediction box with the highest confidence are retained.

[0143] Then, according to the calculated DIOU value, the confidence of each prediction box in the prediction box sequence composed of the remaining prediction boxes after the redundant boxes are deleted is recalculated:

[0144]

[0145] wherein M represents the prediction box with the highest confidence in the prediction box sequence, D j represents any prediction box in the prediction box sequence except the prediction box with the highest confidence, and threshold is a preset threshold, which is set to 0.8 in the present application.

[0146] That is, the DIOU of any prediction box in the prediction box sequence except the prediction box with the highest confidence is calculated with the prediction box with the highest confidence, and if the calculated DIOU value is less than the preset threshold, the previous confidence is retained, and if the calculated DIOU value is greater than the preset threshold, the confidence becomes zero.

[0147] Finally, according to the recalculated confidence, the prediction box with the highest confidence in the prediction box sequence is selected as the final prediction box.

[0148] In this way, the addition of the center point distance penalty term can make the model select the optimal prediction box faster and more accurately, and improve the effect of target detection.

[0149] Optionally, before the IOU matrix is obtained by performing IOU calculation on each two prediction boxes in the prediction box sequence, the method further comprises:

[0150] performing IOU calculation on the prediction box with the highest confidence and the remaining prediction boxes in the prediction box sequence respectively to obtain calculation results, and screening the prediction boxes with a calculation result greater than or equal to the NMS threshold from the remaining prediction boxes;

[0151] performing weighted summation on the prediction box with the highest confidence and the screened prediction boxes to obtain a new rectangular box; wherein the weight of each prediction box in the weighted summation is the corresponding confidence;

[0152] using the new rectangular box to replace the prediction box with the highest confidence to obtain an updated prediction box sequence.

[0153] wherein the rest of the prediction boxes refer to the prediction boxes in the initial prediction box sequence except the prediction box with the highest confidence, and the NMS threshold in the present application is 0.8.

[0154] Specifically, the prediction box with the highest confidence and the screened prediction box can be weighted and summed by using the following formula to obtain a new rectangular box:

[0155]

[0156] wherein w j is the confidence of the prediction box C j , j = 1, 2, 3…n, n is the total number of prediction boxes in the prediction box sequence, N’ is the new rectangular box, N is the prediction box with the highest confidence in the initial prediction box sequence, ThresholdNMS is the set NMS threshold, D i is any prediction box with an IOU calculation result greater than or equal to the NMS threshold from the rest of the prediction boxes, i = 1, 2, 3…(n-1).

[0157] The prediction box with the highest confidence is deleted from the prediction box sequence, and the new rectangular box is added to obtain an updated prediction box sequence. The IOU matrix can be calculated according to the updated prediction box sequence, and subsequent operations can be performed.

[0158] Since the prediction box with the highest confidence is not necessarily accurately positioned, the screened prediction box with an IOU calculation result greater than or equal to the threshold also has reference value. Therefore, the prediction box with the highest confidence and the screened prediction box are weighted and summed, the positions and sizes of multiple prediction boxes are comprehensively considered, and a new rectangular box is obtained for subsequent operations, which can further improve the effect of target detection.

[0159] Step 204, calculating a loss value according to the final prediction box, and adjusting the parameters of the crack identification model according to the loss value.

[0160] Specifically, the final prediction box and the automatically or manually labeled prediction box in step 201 are compared to calculate a loss value, and then the parameters of the crack identification model are adjusted according to the loss value.

[0161] Finally, other sample images can be replaced to repeat steps 202 to 204 until the number of executions reaches a preset maximum number or the calculated loss value is less than a preset threshold.

[0162] Optionally, when testing the trained model, FPS (Frame Per Second, frame rate per second) is used for speed evaluation.

[0163] Figure 4A crack identification image output by the trained crack identification model is provided for the embodiments of the present application. As shown in Figure 4 crack means crack, and the number after crack represents the probability of correct crack identification. A total of 3 cracks are identified, and the probabilities of correct crack identification from left to right are 0.86, 0.89 and 0.79.

[0164] In summary, the training method of the crack identification model provided in the embodiments includes: generating a plurality of sample images by simulating a random cosine function of a formation crack, and performing image enhancement on each sample image; adding the generated sample images and the sample images obtained after enhancement to the sample image set; inputting any sample image in the sample image set into a crack identification model to obtain a plurality of prediction boxes corresponding to the sample image and the confidence of each prediction box; wherein the crack identification model is realized based on a yolov5 network model, and the prediction box is used to indicate the position of the crack in the sample image; determining a final prediction box from the plurality of prediction boxes according to the confidence corresponding to the plurality of prediction boxes, the area of the overlapping region of each pair of prediction boxes and the center point of each prediction box; calculating a loss value according to the final prediction box, and adjusting the parameters of the crack identification model according to the loss value. The input data set of the model is generated by using the random cosine function conforming to the characteristics of the formation crack, the sample expansion under the condition of small sample is realized, the data acquisition work is simplified, the cracks in the sample images can be automatically labeled, the manpower and time cost are saved, the Cluster NMS is used to replace the NMS of the original yolo v5s model to perform the redundant box deletion operation, the weight and the center point distance penalty term are added to complete the non-maximum suppression, the model can select the optimal box faster and more accurately, and thus the target detection effect is improved, which is of great significance for the exploration of the crack reservoir.

[0165] Figure 5 A flowchart of a crack identification method provided in the embodiments of the present application is shown in FIG. 5. As shown in FIG. 5, the crack identification method provided in the embodiments of the present application can include:

[0166] Step 501: acquiring an imaging logging image to be identified;

[0167] Step 502: inputting the imaging logging image to be identified into a crack identification model to obtain the position of the crack in the imaging logging image;

[0168] The crack identification model is a model trained based on any of the above methods.

[0169] Specifically, the imaging logging image to be identified is acquired first, and then the imaging logging image is input into the crack identification model trained based on any of the above methods to obtain the position of the crack in the imaging logging image.

[0170] Corresponding to the training method of the crack identification model, the embodiment of the application further provides a crack identification model training device. Figure 6 A structural schematic diagram of a crack identification model training device provided by the embodiment of the application is shown in FIG. 7. Figure 6 As shown in the figure, the device comprises:

[0171] The generation module 601 is configured to generate a plurality of sample images by simulating a random cosine function of a formation fracture, and perform image enhancement on each sample image, and add the generated sample images and the sample images obtained after enhancement to the sample image set.

[0172] The training module 602 is configured to input any sample image in the sample image set into a crack identification model to obtain a plurality of prediction boxes corresponding to the sample image and a confidence of each prediction box, wherein the crack identification model is realized based on a yolov5 network model, and the prediction box is used to indicate a position of a crack in the sample image.

[0173] The determination module 603 is configured to determine a final prediction box from the plurality of prediction boxes according to the confidence corresponding to the plurality of prediction boxes, an area of an overlapping region of two prediction boxes, and a center point of each prediction box.

[0174] The adjustment module 604 is configured to calculate a loss value according to the final prediction box, and adjust parameters of the crack identification model according to the loss value.

[0175] The specific implementation principle and effect of the crack identification model training device provided by the embodiment of the application can be referred to the foregoing embodiments, which will not be described here.

[0176] Figure 7 A structural schematic diagram of an electronic device provided by the embodiment of the application is shown in FIG. 8. Figure 7 As shown in the figure, the electronic device of the embodiment can comprise:

[0177] at least one processor 701; and

[0178] a memory 702 in communication connection with the at least one processor;

[0179] The memory 702 stores instructions executable by the at least one processor 701, and the instructions are executed by the at least one processor 701 to enable the electronic device to perform the method of any of the foregoing embodiments.

[0180] Optionally, the memory 702 can be independent or integrated with the processor 701.

[0181] The implementation principles and technical effects of the electronic device provided in this embodiment can be referred to the foregoing embodiments, and will not be described here again.

[0182] The embodiment of the present application further provides a computer readable storage medium, wherein computer execution instructions are stored in the computer readable storage medium, and when a processor executes the computer execution instructions, the method in any of the foregoing embodiments is implemented.

[0183] The embodiment of the present application further provides a computer program product, comprising a computer program, and when the computer program is executed by a processor, the method in any of the foregoing embodiments is implemented.

[0184] In several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other manners. For example, the device embodiments described above are only schematic; the division of the modules is only a logical function division, and there can be another division manner in actual implementation; for example, a plurality of modules can be combined or integrated into another system, or some features can be ignored or not executed.

[0185] The integrated module implemented in the form of a software functional module can be stored in a computer readable storage medium. The software functional module stored in the storage medium includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to execute some steps of the method described in each embodiment of the present application.

[0186] It should be understood that the processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor and the like. The steps of the method disclosed in the application can be directly embodied as the execution of the hardware processor, or the execution of the combination of hardware and software modules in the processor. The memory can include a high-speed RAM memory, and can also include a non-volatile storage NVM, for example, at least one disk memory, and can also be a U disk, a mobile hard disk, a read-only memory, a magnetic disk or an optical disk, etc.

[0187] The storage medium can be implemented by any type of volatile or non-volatile storage devices or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk. The storage medium can be any available medium that can be accessed by a general or special purpose computer.

[0188] An exemplary storage medium is coupled to the processor such that the processor can read information from, and write information to, the storage medium. Of course, the storage medium can be a part of the processor. The processor and the storage medium can be located in an Application Specific Integrated Circuits (ASIC). Of course, the processor and the storage medium can exist as separate components in the electronic device or the host device.

[0189] It should be noted that, in the present document, the terms "comprising", "containing" or any other similar term are intended to encompass non-exclusive inclusion, such that a process, method, article or apparatus that comprises a list of elements does not only include those elements, but can also include other elements not expressly listed or inherent to such process, method, article or apparatus. Without more limitations, the element defined by the phrase "comprising a" does not exclude the presence of additional identical elements in the process, method, article or apparatus comprising the element.

[0190] The above-mentioned sequence numbers of the embodiments of the present application are only for description, and do not represent the advantages or disadvantages of the embodiments.

[0191] From the above description of the embodiments, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be realized by means of software and necessary general hardware platforms, of course, they can also be realized by hardware, but in many cases the former is a better embodiment. Based on such understanding, the technical solutions of the present application can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes a plurality of instructions for making a terminal device (which can be a mobile phone, computer, server, air conditioner or network device, etc.) execute the methods described in the various embodiments of the present application.

[0192] The above merely preferred embodiments of the present application and are not intended to limit the patent scope of the present application, any equivalent structure or equivalent process transformation using the content of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for training a crack identification model, characterized in that, The method comprises the following steps: A plurality of sample images are generated by simulating a random cosine function of a formation fracture, and each sample image is image-enhanced, and the generated sample images and the sample images obtained after enhancement are added to a sample image set; Any sample image in the sample image set is input into a fracture identification model to obtain a plurality of prediction boxes corresponding to the sample image and a confidence degree of each prediction box; the fracture identification model is realized based on a yolov5 network model, and the prediction box is used to indicate a position of a fracture in the sample image; A final prediction box is determined from the plurality of prediction boxes according to the confidence degrees corresponding to the plurality of prediction boxes, areas of overlapping regions of two prediction boxes and center points of the prediction boxes; A loss value is calculated according to the final prediction box, and parameters of the fracture identification model are adjusted according to the loss value; A plurality of sample images are generated by simulating a random cosine function of a formation fracture, and each sample image is image-enhanced, and the generated sample images and the sample images obtained after enhancement are added to a sample image set, which comprises the following steps: A corresponding cosine curve is obtained according to the random cosine function simulating the formation fracture; The cosine curve is subjected to a plurality of adjustment operations to obtain a plurality of corresponding sample images; the adjustment operation comprises at least one of the following: adding random noise, translation, adjusting thickness, adjusting inclination and adjusting a position of the curve in the image; an adjustment amount of each adjustment operation is different; Each sample image is image-enhanced to obtain a plurality of sample images after image enhancement; The generated sample images and the sample images obtained after enhancement are added to the sample image set.

2. The method of claim 1, wherein, A final prediction box is determined from the plurality of prediction boxes according to the confidence degrees corresponding to the plurality of prediction boxes, areas of overlapping regions of two prediction boxes and center points of the prediction boxes, which comprises the following steps: The plurality of prediction boxes corresponding to the sample image are arranged in descending order of confidence degree to obtain a prediction box sequence; IOU calculation is performed on the plurality of prediction boxes in the prediction box sequence two by two to obtain an IOU matrix, and an element in the i-th row and the j-th column of the IOU matrix is a result of IOU calculation of a prediction box with a confidence degree ranking of i and a prediction box with a confidence degree ranking of j; At least one redundant box is determined from the plurality of prediction boxes according to the IOU matrix, and a deletion operation is performed on the at least one redundant box; After the deletion of the redundant box, a prediction box with the highest confidence degree is determined from the remaining prediction boxes, and an area of an overlapping region of each prediction box in the remaining prediction boxes and the prediction box with the highest confidence degree is calculated; The final prediction box is determined according to the calculated area and the center points of the prediction boxes.

3. The method of claim 2, wherein, Before the IOU calculation is performed on the plurality of prediction boxes in the prediction box sequence two by two to obtain the IOU matrix, the following step is further included: IOU calculation is performed on the prediction box with the highest confidence degree in the prediction box sequence and the remaining prediction boxes respectively to obtain calculation results, and a prediction box with a calculation result greater than or equal to a non-maximum suppression NMS threshold is screened out from the remaining prediction boxes; The prediction box with the highest confidence degree and the screened prediction box are subjected to weighted summation to obtain a new rectangular box; when the weighted summation is performed, a weight of each prediction box is a corresponding confidence degree. Using the new rectangular frame, instead of the highest confidence prediction frame, an updated prediction frame sequence is obtained.

4. The method of claim 2, wherein, According to the calculated area and the center point of each prediction frame, a final prediction frame is determined, including: If there are at least two prediction frames with the same area of the overlap region with the highest confidence prediction frame, the distance between each prediction frame in the at least two prediction frames and the center point of the highest confidence prediction frame is calculated, and the prediction frame with the largest distance is retained, and the remaining prediction frames in the at least two prediction frames are deleted; If the areas of the corresponding overlap regions of each prediction frame are not the same, the confidence of each prediction frame in the prediction frame sequence is recalculated according to the following formula: wherein M represents the prediction box with the highest confidence, D j represents any prediction box other than the prediction box with the highest confidence, Score j represents D j corresponding confidence, threshold is a preset threshold, d is the distance between the center points of the two prediction boxes, c is the diagonal length of the minimum rectangular frame that completely contains the two prediction boxes, and γ is a constant parameter. According to the recalculated confidence, the prediction frame with the highest confidence is selected as the final prediction frame.

5. The method of claim 2, wherein, According to the IOU matrix, at least one redundant frame is determined from the plurality of prediction frames for deletion, including: Repeat the following operations until the tensors obtained in two consecutive times are the same: perform upper triangular operation on the current IOU matrix to obtain the operated IOU matrix; select the maximum value of each column on the operated IOU matrix, compare the maximum value with the NMS threshold to obtain a tensor formed by 0 and / or 1, wherein 1 represents that the maximum value is less than the NMS threshold, and 0 represents that the maximum value is greater than the NMS threshold; take the tensor as the diagonal line of the matrix to generate a diagonal matrix E; multiply the diagonal matrix E with the IOU matrix to obtain an updated IOU matrix; After the same tensor is obtained in two consecutive times, it is judged according to the elements in the tensor whether the corresponding prediction frame is a redundant frame, and the determined redundant frame is deleted, wherein the element value of 0 represents that the corresponding prediction frame is a redundant frame, and the element value of 1 represents that the corresponding prediction frame is not a redundant frame.

6. A crack identification method characterized by, It includes: Obtaining an imaging logging image to be identified; Inputting the imaging logging image to be identified into a fracture identification model to obtain a fracture position in the imaging logging image. The fracture identification model is a model trained based on the method of any one of claims 1-5. 7.A device for training a crack identification model, characterized by, It includes: A generation module is configured to generate a plurality of sample images by simulating a random cosine function of a formation fracture, and perform image enhancement on each sample image, and add the generated sample images and the sample images obtained after enhancement to a sample image set; A training module is configured to input any sample image in the sample image set into a fracture identification model to obtain a plurality of prediction frames corresponding to the sample image and a confidence of each prediction frame; wherein the fracture identification model is implemented based on a yolov5 network model, and the prediction frame is used to indicate a position of a fracture in the sample image; A determination module is configured to determine a final prediction frame from the plurality of prediction frames according to the confidence corresponding to each prediction frame, an area of an overlap region between each two prediction frames, and a center point of each prediction frame; An adjustment module is configured to calculate a loss value according to the final prediction frame, and adjust parameters of the fracture identification model according to the loss value; The generation module is specifically configured to: According to the random cosine function of the simulated formation fracture, a corresponding cosine curve is obtained. The cosine curve is adjusted multiple times to obtain a plurality of corresponding sample images; wherein the adjustment operation includes at least one of the following: adding random noise, translation, adjusting thickness, adjusting the angle, adjusting the position of the curve in the image; the adjustment amount of each adjustment operation is different; The sample images are subjected to image enhancement to obtain a plurality of sample images after image enhancement; The generated sample images and the obtained sample images after enhancement are added to the sample image set.

8. An electronic device, comprising: Comprise: At least one processor; And the memory connected with the at least one processor in communication; Wherein, the memory has instructions executable by the at least one processor, the instructions are executed by the at least one processor to enable the electronic device to perform the method of any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer execution instructions, and when the processor executes the computer execution instructions, the method of any one of claims 1-6 is realized.

Images