Security inspection machine data dynamic correction method
By using the improved ResNet neural network model in the ray security machine to judge and correct the ray image, the problem that traditional security machine cannot be automatically calibrated is solved, and the automatic calibration and real-time recognition efficiency of the security machine are improved.
Patent Information
- Application Number
- CN202510042798.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-10
- Publication Date
- 2025-05-27
AI Technical Summary
Traditional ray security inspection machines cannot be automatically calibrated, resulting in artifacts and dirty images displayed when the belt is off or the edges are irregular, affecting the judgment of safety personnel.
Using a dynamic calibration method of security machine data, the ray image is acquired and preprocessed, and input it into the pre-trained improved ResNet neural network model to determine whether the image belongs to a dirty image, and the background data and fullness data of the ray image collected next time are determined based on the judgment results to correct it.
Automatic calibration of security inspection machines is realized, the efficiency of recognition of ray images is improved, the calculation amount and parameter amount are reduced, the computing speed is improved, and real-time calibration of security inspection machines is realized.
Smart Images
Figure CN120047362A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of security inspection detection, and particularly to a method for dynamically correcting security inspection machine data. Background Art
[0002] Due to its own unique advantages, ray security inspection technology has been highly regarded and widely used in the field of security inspections such as customs, aviation, transportation, and logistics. The irradiation position and angle of the ray can be adjusted and transformed according to the actual situation, mainly including bottom irradiation type and side irradiation type security inspection machines. With the development of technology, multi-view security inspection machines have gradually entered production and life.
[0003] The security inspection machine uses a linear detector array to collect images, and can obtain dual-energy signals through a dual-energy detector, so as to distinguish organic substances, inorganic substances, and mixtures. In traditional X-ray security inspection machines, due to the long-term operation of the conveyor belt, the thickness of the belt is uneven, burrs occur at the edges due to wear, and even the belt runs off. At the same time, due to the relationship between the environment and the purity and uniformity of the object substances, there will be some deviations. Therefore, when the ray penetrates the object and the belt, factors such as the drift of the background data and the drift of the saturated data caused by unstable operation will affect the background data and the saturated data, resulting in the gain parameter not reflecting the change of the system.
[0004] Traditional ray security inspection machines only perform a calibration operation on the ray image when starting up. When the belt runs off or the edge is irregular, the calibration parameter cannot automatically correct the current belt state, resulting in artifacts and dirty images in the displayed image, which in turn affects the judgment of security personnel and poses a hidden danger to the security prevention work.
[0005] Therefore, there is an urgent need for a technical solution to quickly calibrate the security inspection machine. Summary of the Invention
[0006] In view of the above analysis, the embodiments of the present invention aim to provide a method for dynamically correcting security inspection machine data to solve the problem that it is difficult to identify security inspection images due to the inability of existing security inspection machines to automatically calibrate.
[0007] The embodiments of the present invention provide a method for dynamically correcting security inspection machine data, and the correction method includes:
[0008] Obtain the ray image collected this time, preprocess the ray image to obtain a ray image to be judged;
[0009] Input the ray image to be judged into a pre-trained ray image judgment model to obtain a judgment result; wherein, the ray image judgment model is trained based on an improved ResNet neural network model, and two improved basic residual modules are used to form an improved residual module in the improved ResNet neural network model to replace the original residual module;
[0010] Determine the background data and full-scale data of the ray image to be collected next according to the judgment result, and correct the ray image to be collected next.
[0011] Based on a further improvement of the above correction method, the improved ResNet neural network model includes a first convolution module, a residual extraction module, and a fully connected module connected in sequence;
[0012] The first convolution module is used to receive the ray image to be judged, and sequentially perform convolution operation, batch normalization operation, and linear correction operation, and output the obtained first feature map to the residual extraction module;
[0013] The residual extraction module is used to extract features from the input first feature map, and output the obtained second feature map to the fully connected module; the residual extraction module includes a plurality of improved residual modules connected in sequence;
[0014] The fully connected module is used to make a prediction based on the input second feature map, obtain a judgment result and output it.
[0015] Based on a further improvement of the above correction method, each improved residual module includes a first improved basic residual module and a second improved basic residual module connected in sequence;
[0016] The first improved basic residual module and the second improved basic residual module have the same structure, and both include a second convolution module, a third convolution module, and a fourth convolution module connected in sequence.
[0017] Based on a further improvement of the above correction method, in the first improved basic residual module or the second improved basic residual module, the feature map input to the second convolution module is spliced with the feature map output by the fourth convolution module through a skip connection to obtain the output feature map of the first improved basic residual module or the second improved basic residual module for output.
[0018] Based on a further improvement of the above correction method, the first convolution module, the second convolution module, the third convolution module, and the fourth convolution module have the same structure, and all include a convolution layer, a batch normalization layer, and a linear correction layer connected in sequence;
[0019] The convolution layer is used to perform a convolution operation on the input feature map, and output the obtained feature map to the batch normalization layer;
[0020] The batch normalization layer is used to perform a batch normalization operation on the input feature map, and output the obtained feature map to the linear correction layer;
[0021] The linear correction layer is used to perform a linear correction operation on the input feature map, and output the obtained feature map.
[0022] Based on a further improvement of the above correction method, the convolution kernel size of the convolution layer of the first convolution module is 7*7;
[0023] The convolution kernel size of the convolution layer of the second convolution module is 1*1;
[0024] The convolution kernel size of the convolution layer of the third convolution module is 3*3;
[0025] The convolution kernel size of the convolution layer of the fourth convolution module is 1*1.
[0026] Based on a further improvement of the above correction method, the fully connected module includes a global average pooling layer and a linear layer connected in sequence;
[0027] The global average pooling layer is used to perform global average pooling on the input second feature map to obtain a one-dimensional vector;
[0028] The linear layer is used to determine the judgment result according to the one-dimensional vector.
[0029] Based on a further improvement of the above correction method, the preprocessing includes one or more of the following operations:
[0030] Normalization;
[0031] Edge blanking;
[0032] Edge enhancement;
[0033] Material filtering;
[0034] Distortion correction.
[0035] Based on a further improvement of the above correction method, the following formula is used to calculate the loss when training the improved ResNet neural network model:
[0036]
[0037] Among them, Loss represents the loss value, N represents the number of training samples, w represents the weight of the dirty image, w′ represents the weight of the non-dirty image, y i _true represents the label value that the i-th sample belongs to the dirty image, y i _pred represents the predicted value that the i-th sample belongs to the dirty image.
[0038] Based on a further improvement of the above correction method, the determining the background data and full-scale data of the ray image to be collected next according to the judgment result, and correcting the ray image to be collected next includes:
[0039] If the judgment result of the ray image to be judged is a dirty image, then turn on the ray to obtain the air value to get the full-scale data, and turn off the ray to obtain the background data;
[0040] The corrected radiographic image is calculated by the following formula:
[0041]
[0042] Where new_img represents the corrected radiographic image, yuantu represents the radiographic image before correction, bendi represents the updated background data, mandu represents the updated full-scale data, Upper and Lower represent the image upper limit threshold and the image lower limit threshold respectively.
[0043] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:
[0044] 1. The radiographic image judgment model is used to judge the radiographic images collected by the security inspection machine. After determining that the radiographic image belongs to a dirty image, the security inspection machine is calibrated, realizing the automatic calibration of the security inspection machine and improving the recognition efficiency of the radiographic images;
[0045] 2. In the radiographic image judgment model, two improved basic residual modules are used to form an improved residual module to replace the original residual module, reducing the dimension of the feature map of the radiographic image, reducing the amount of calculation and the number of parameters in the radiographic image judgment model, improving the operation speed, quickly obtaining the judgment result of the radiographic image, and realizing the real-time calibration of the security inspection machine.
[0046] In the present invention, the above technical solutions can also be combined with each other to achieve more preferred combination schemes. Other features and advantages of the present invention will be described in the subsequent specification, and some advantages can be made obvious from the specification or understood by implementing the present invention. The objectives and other advantages of the present invention can be realized and obtained through the content specifically pointed out in the specification and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS
[0047] The drawings are only for the purpose of showing specific embodiments and are not considered as limiting the present invention. Throughout the drawings, the same reference numerals represent the same components.
[0048] Figure 1 It is a schematic flowchart of a method for dynamically correcting security inspection machine data provided by an embodiment of the present invention;
[0049] Figure 2 It is one of the schematic structural diagrams of the radiographic image judgment model provided by an embodiment of the present invention;
[0050] Figure 3 It is the second schematic structural diagram of the radiographic image judgment model provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0051] The preferred embodiments of the present invention will be specifically described below in conjunction with the accompanying drawings. The accompanying drawings form a part of this application and are used together with the embodiments of the present invention to explain the principles of the present invention, rather than to limit the scope of the present invention.
[0052] A specific embodiment of the present invention discloses a method for dynamically correcting security inspection machine data. As Figure 1 shown, the correction method includes:
[0053] Step S1: Obtain the ray image collected this time, preprocess the ray image to obtain the ray image to be judged;
[0054] Step S2: Input the ray image to be judged into the ray image judgment model that has been pre-trained, and obtain the judgment result. Among them, the ray image judgment model is trained based on an improved ResNet neural network model. In the improved ResNet neural network model, two improved basic residual modules are used to form an improved residual module to replace the original residual module;
[0055] Step S3: Determine the background data and full-scale data of the ray image to be collected next time according to the judgment result, and correct the ray image to be collected next time.
[0056] Specifically, as Figure 1 shown, in step S1, obtain the ray image collected by the security inspection machine this time, preprocess the ray image to obtain the ray image to be judged.
[0057] It can be understood that during the operation of the security inspection machine, the ray image of the package will be obtained in real time and output for the judgment of security personnel. If the belt runs off or the edge is irregular during the operation of the security inspection machine, the calibration parameters cannot automatically correct the current belt state, which will cause artifacts and dirty images in the ray image of the package. At this time, it is difficult for security personnel to make an accurate judgment on the dirty image, resulting in potential safety hazards in the work.
[0058] In step S1, obtain the ray image collected by the security inspection machine in real time, preprocess it to obtain the ray image to be judged, and input the ray image to be judged into the ray image judgment model to determine whether there are artifacts in the ray image and whether the ray image belongs to a dirty image.
[0059] Preferably, the preprocessing includes one or more of the following operations:
[0060] Normalization;
[0061] Edge whitening;
[0062] Edge enhancement;
[0063] Material filtering;
[0064] Distortion correction.
[0065] Specifically, normalization is to convert the original high and low energy data into values in the range of 0 to 65535 according to the background data and full scale data.
[0066] Specifically, edge blanking is to set the data of the blocked detector as empty according to the hardware design.
[0067] Specifically, edge enhancement is to perform low-pass filtering on the original image to obtain its blurred image, then subtract the blurred image from the original image to obtain an approximate high-frequency information image, and then magnify the approximate high-frequency information image by a certain ratio and superimpose it on the original image to obtain an edge-enhanced image.
[0068] Specifically, material filtering is to significantly reduce the data statistical fluctuations of the material image on the basis of basically not changing the edge detail material information.
[0069] Specifically, distortion correction is a method for automatically correcting deformed images based on the principle of consistent imaging ratio, which is designed for the deformation problems existing in the images collected by channel-type X-ray security inspection equipment.
[0070] Specifically, as Figure 1 shown, in step S2, the ray image judgment model needs to be pre-trained so that the ray image judgment model can quickly judge the ray image and determine whether the ray image belongs to a dirty image.
[0071] Specifically, the ray image to be judged is input into the ray image judgment model to obtain the judgment result corresponding to the ray image to be judged. It should be noted that the ray image judgment model is trained based on an improved ResNet neural network model, and two improved basic residual modules are used in the improved ResNet neural network model to form an improved residual module to replace the original residual module.
[0072] It can be understood that in the ResNet neural network, the original residual module is used to extract image features. In the present invention, the improved residual module is used to replace the original residual module to extract image features, quickly extract the features of the ray image to be judged, and quickly obtain the judgment result of the ray image to be judged.
[0073] Specifically, in step S2, the judgment result of the ray image to be judged is obtained, and it is determined whether the ray image to be judged belongs to a dirty image or a non-dirty image.
[0074] Specifically, as Figure 1 shown, the ray image judgment model used in step S2 is trained based on an improved ResNet neural network model.
[0075] Preferably, the improved ResNet neural network model includes a first convolutional module, a residual extraction module, and a fully connected module connected in sequence;
[0076] The first convolutional module is configured to receive the ray image to be judged, and sequentially perform a convolution operation, a batch normalization operation, and a linear correction operation, and output the obtained first feature map to the residual extraction module;
[0077] The residual extraction module is configured to perform feature extraction on the input first feature map, and output the obtained second feature map to the fully connected module; the residual extraction module includes a plurality of improved residual modules connected in sequence;
[0078] The fully connected module is configured to make a prediction based on the input second feature map, obtain a judgment result and output it.
[0079] Specifically, as Figure 2 shown, the improved ResNet neural network model includes a first convolutional module, a residual extraction module, and a fully connected module connected in sequence from shallow to deep. The ray image to be judged is input into the first convolutional module, and a convolution operation, a batch normalization operation, and a linear correction operation are sequentially performed in the first convolutional module to obtain a first feature map P1. The first feature map P1 is then output to the residual extraction module; in the residual extraction module, the first feature map P1 sequentially passes through a plurality of improved residual modules from shallow to deep to obtain a second feature map P2, and the second feature map P2 is then output to the fully connected module; in the fully connected module, a prediction is made on the second feature map P2 to obtain a judgment result of the ray image to be judged.
[0080] Exemplarily, the residual extraction module includes 4 improved residual modules, and the first feature map P1 is subjected to feature extraction through 4 improved residual modules to obtain a second feature map P2.
[0081] Preferably, each improved residual module includes a first improved basic residual module and a second improved basic residual module connected in sequence;
[0082] The first improved basic residual module and the second improved basic residual module have the same structure, and both include a second convolutional module, a third convolutional module, and a fourth convolutional module connected in sequence.
[0083] Specifically, as Figure 2 shown, each improved residual module includes two improved basic residual modules connected in sequence, namely a first improved basic residual module and a second improved basic residual module.
[0084] Preferably, in the first improved basic residual module or the second improved basic residual module, the feature map input to the second convolutional module is concatenated with the feature map output by the fourth convolutional module through a skip connection to obtain the output feature map of the first improved basic residual module or the second improved basic residual module for output.
[0085] Specifically, as Figure 2 shown, the first improved basic residual module and the second improved basic residual module have the same structure, and both include a second convolutional module, a third convolutional module, and a fourth convolutional module connected in sequence.
[0086] It should be noted that in the first improved basic residual module or the second improved basic residual module, the feature map input to the second convolutional module is concatenated with the feature map output by the fourth convolutional module through a skip connection, ensuring that the gradient can propagate smoothly in the deep network and preventing the problem of gradient disappearance.
[0087] Preferably, the first convolutional module, the second convolutional module, the third convolutional module, and the fourth convolutional module have the same structure, and each includes a convolutional layer, a batch normalization layer, and a rectified linear unit layer connected in sequence;
[0088] The convolutional layer is used to perform a convolution operation on the input feature map and output the obtained feature map to the batch normalization layer;
[0089] The batch normalization layer is used to perform a batch normalization operation on the input feature map and output the obtained feature map to the rectified linear unit layer;
[0090] The rectified linear unit layer is used to perform a rectified linear unit operation on the input feature map and output the obtained feature map.
[0091] Specifically, as Figure 2 shown, the first convolutional module is used to receive the preprocessed ray image to be judged, and perform a convolution operation, a batch normalization operation, and a rectified linear unit operation in sequence in the convolutional layer, the batch normalization layer, and the rectified linear unit layer.
[0092] Preferably, as Figure 3 shown, the convolution kernel size of the convolutional layer of the first convolutional module is 7*7;
[0093] The convolution kernel size of the convolutional layer of the second convolutional module is 1*1;
[0094] The convolution kernel size of the convolutional layer of the third convolutional module is 3*3;
[0095] The convolution kernel size of the convolutional layer of the fourth convolutional module is 1*1.
[0096] Specifically, as Figure 3As shown, the ray image to be judged is input into the first convolutional module Conv1 through the input end Input. The first convolutional module includes a convolutional layer Conv2d, a batch normalization layer BatchNorm2d, and a rectified linear unit layer Relu. The convolutional kernel size of the convolutional layer Conv2d is 7*7.
[0097] Specifically, ResBlk1, …, ResBlk4 are improved residual modules included in the residual extraction module, as Figure 3 shown. The first improved basic residual module BasicBlk1 and the second improved basic residual module BasicBlk2 are included in the first improved residual module ResBlk1. The first improved basic residual module BasicBlk1 and the second improved basic residual module BasicBlk2 are included in the fourth improved residual module ResBlk4.
[0098] Specifically, as Figure 3 shown, the first improved basic residual module BasicBlk1 and the second improved basic residual module BasicBlk2 have the same structure and both include a second convolutional module, a third convolutional module, and a fourth convolutional module connected in sequence. Each convolutional module includes a convolutional layer, a batch normalization layer, and a rectified linear unit layer connected in sequence. That is, as Figure 3 shown, in the first improved basic residual module BasicBlk1 or the second improved basic residual module BasicBlk2, Conv2d(1*1), BN+Relu, Conv2d(3*3), BN+Relu, Conv2d(1*1), and BN+Relu are included and connected in sequence.
[0099] Specifically, the convolutional kernel size of the convolutional layer Conv2d in the second convolutional module is 1*1, the convolutional kernel size of the convolutional layer Conv2d in the third convolutional module is 3*3, and the convolutional kernel size of the convolutional layer Conv2d in the fourth convolutional module is 1*1. It should be noted that by replacing the large convolutional kernels 5*5 or 7*7 in the original residual module with multiple small convolutional kernels 1*1, 3*3, and 1*1, the expression ability of the residual extraction module is improved, and at the same time, the number of parameters is reduced and the network depth is increased.
[0100] Specifically, the batch normalization layer BN in the second convolutional module, the third convolutional module, and the fourth convolutional module is used for batch normalization operations, and the rectified linear unit layer Relu is used for rectified linear operations.
[0101] Preferably, the fully connected module includes a global average pooling layer and a linear layer connected in sequence;
[0102] The global average pooling layer is used to perform global average pooling on the input second feature map to obtain a one-dimensional vector;
[0103] A linear layer for determining a judgment result based on a one-dimensional vector.
[0104] Specifically, as Figure 2 and Figure 3 shown, the fully connected module FC Linear includes a global average pooling layer Pooling and a linear layer Linear; the global average pooling layer Pooling is used to perform global average pooling on the input second feature map P2 to convert the second feature map into a one-dimensional vector; the linear layer Linear is used for classification output, mapping the one-dimensional vector to the classification space, and outputting the final classification result, that is, outputting the judgment result.
[0105] It should be noted that in the embodiments of the present invention, the judgment results are divided into two categories: dirty pictures and non-dirty pictures, and the classification space includes two categories.
[0106] Preferably, the following formula is used to calculate the loss when training the improved ResNet neural network model:
[0107]
[0108] where Loss represents the loss value, N represents the number of training samples, w represents the weight of the dirty picture, w' represents the weight of the non-dirty picture, and y i _true represents the label value that the i-th sample belongs to the dirty picture, and y i _pred represents the predicted value that the i-th sample belongs to the dirty picture.
[0109] Specifically, when training the improved ResNet neural network model, the loss value of the training samples is calculated by the above formula, and the parameters of the improved ResNet neural network model are adjusted according to the loss value until the training is completed.
[0110] Specifically, as Figure 1 shown, in step S3, according to the judgment result of the ray image to be judged, the background data and full-scale data corresponding to the ray image to be collected next time are determined.
[0111] Specifically, when the judgment result of the ray image to be judged is a non-dirty picture, it means that there is no abnormal situation in the security inspection machine at this time or the abnormal situation does not affect the generation of the real ray image. For example, there is no belt deviation or irregular edge situation. At this time, the generated ray image output can be used for the security personnel to judge.
[0112] Preferably, determining the background data and full-scale data of the ray image to be collected next time according to the judgment result and correcting the ray image to be collected next time includes:
[0113] If the judgment result of the ray image to be judged is a dirty image, turn on the ray to obtain the air value to get the full-scale data, and turn off the ray to obtain the air value to get the background data;
[0114] Calculate the corrected ray image through the following formula:
[0115]
[0116] Among them, new_img represents the corrected ray image, yuantu represents the ray image before correction, bendi represents the updated background data, mandu represents the updated full-scale data, Upper and Lower represent the image upper limit threshold and the image lower limit threshold respectively.
[0117] Specifically, the image lower limit threshold refers to the data value of the detector when the ray is not turned on, and the image upper limit threshold refers to the data value of the detector passing through the air when the ray is turned on.
[0118] Specifically, when the judgment result of the ray image to be judged is a dirty image, it means that there is an abnormal situation in the security inspection machine at this time or the abnormal situation has affected the generation of the real ray image. If the existing background data and full-scale data are continued to be used, the subsequent generated ray images will continue to appear as dirty images, seriously affecting the judgment of security personnel.
[0119] It can be understood that by turning on the ray to obtain the air value to get the full-scale data, and turning off the ray to obtain the air value to get the background data, the ray image is corrected by using the newly obtained full-scale data and background data, the artifacts in the subsequent ray images are eliminated, the appearance of dirty images is avoided, and the recognition and judgment efficiency of security personnel is improved.
[0120] Compared with the prior art, a security inspection machine data dynamic correction method provided by an embodiment of the present invention judges the ray image collected by the security inspection machine through a ray image judgment model. After determining that the ray image belongs to a dirty image, the security inspection machine is calibrated, realizing the automatic calibration of the security inspection machine and improving the recognition efficiency of the ray image; at the same time, in the ray image judgment model, an improved residual module composed of two improved basic residual modules is used to replace the original residual module, reducing the dimension of the feature map of the ray image, reducing the calculation amount and the number of parameters in the ray image judgment model, improving the operation speed, quickly obtaining the judgment result of the ray image, and realizing the real-time calibration of the security inspection machine.
[0121] Those skilled in the art can understand that to implement all or part of the processes of the above embodiment methods, it can be completed by instructing relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium. Among them, the computer-readable storage medium is a disk, an optical disc, a read-only memory or a random access memory, etc.
[0122] As described above, it is only the preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.
Claims
1. A method for dynamic correction of security inspection machine data, characterized in that: The correction method comprises: Acquire the radiographic image collected this time, pre-process the radiographic image, and obtain the radiographic image to be judged; Inputting the radiographic image to be judged into a pre-trained radiographic image judgment model to obtain a judgment result; wherein the radiographic image judgment model is trained based on an improved ResNet neural network model, and two improved basic residual modules are used in the improved ResNet neural network model to form an improved residual module for replacing the original residual module; The background data and full-scale data of the next collected radiographic image are determined according to the judgment result, and the next collected radiographic image is corrected.
2. The calibration method according to claim 1, characterized in that: The improved ResNet neural network model includes a first convolution module, a residual extraction module and a fully connected module connected in sequence; A first convolution module is used to receive the radiographic image to be judged, and sequentially perform convolution operations, batch normalization operations, and linear correction operations, and output the obtained first feature map to the residual extraction module; A residual extraction module, used for performing feature extraction on the input first feature map, and outputting the obtained second feature map to the fully connected module; the residual extraction module includes a plurality of improved residual modules connected in sequence; The fully connected module is used to make predictions based on the input second feature map, obtain the judgment result and output it.
3. The calibration method according to claim 2, characterized in that: Each improved residual module includes a first improved basic residual module and a second improved basic residual module connected in sequence; The first improved basic residual module and the second improved basic residual module have the same structure, and both include a second convolution module, a third convolution module and a fourth convolution module which are connected in sequence.
4. The calibration method according to claim 3, characterized in that: In the first improved basic residual module or the second improved basic residual module, the feature map input to the second convolution module is spliced with the feature map output by the fourth convolution module through a jump connection to obtain the output feature map of the first improved basic residual module or the second improved basic residual module for output.
5. The calibration method according to claim 3, characterized in that: The first convolution module, the second convolution module, the third convolution module and the fourth convolution module have the same structure, and all include a convolution layer, a batch normalization layer and a linear correction layer connected in sequence; The convolution layer is used to perform convolution operations on the input feature map and output the obtained feature map to the batch normalization layer; The batch normalization layer is used to perform batch normalization on the input feature map and output the obtained feature map to the linear correction layer; The linear correction layer is used to perform a linear correction operation on the input feature map and output the obtained feature map.
6. The calibration method according to claim 5, characterized in that: The convolution kernel size of the convolution layer of the first convolution module is 7*7; The convolution kernel size of the convolution layer of the second convolution module is 1*1; The convolution kernel size of the convolution layer of the third convolution module is 3*3; The convolution kernel size of the convolution layer of the fourth convolution module is 1*1.
7. The calibration method according to claim 2, characterized in that: The fully connected module includes a global average pooling layer and a linear layer connected in sequence; The global average pooling layer is used to perform global average pooling on the input second feature map to obtain a one-dimensional vector; The linear layer is used to determine the judgment result based on the one-dimensional vector.
8. The calibration method according to claim 1, characterized in that: The preprocessing includes one or more of the following operations: Normalization; Whiten the edges; Edge enhancement; Texture filtering; Distortion correction.
9. The calibration method according to claim 1, characterized in that: The following formula is used to calculate the loss when training the improved ResNet neural network model: Among them, Loss represents the loss value, N represents the number of training samples, w represents the dirty image weight, w′ represents the non-dirty image weight, y i _truee indicates that the i-th sample belongs to the label value of the dirty image, y i _pred indicates the predicted value that the i-th sample belongs to the dirty image.
10. The calibration method according to claim 1, characterized in that: The step of determining the background data and full-scale data of the radiographic image to be collected next time according to the judgment result, and correcting the radiographic image to be collected next time, comprises: If the judgment result of the radiographic image to be judged is a dirty image, the radiographic image is turned on to obtain the air value to obtain the full-scale data, and the radiographic image is turned off to obtain the air value to obtain the background data; The corrected radiographic image is calculated using the following formula: Among them, new_img represents the corrected radiographic image, yuantu represents the radiographic image before correction, bendi represents the updated background data, mandu represents the updated full-scale data, and Upper and Lower represent the upper and lower thresholds of the image, respectively.