A method for ultra-high-speed radiographic image recognition based on coding plate
By combining a coding plate and X-ray imaging technology with a BPNN neural network in laser processing, image recognition and feedback can be performed directly in the detector, solving the problem of difficult depth direction monitoring in laser processing and achieving efficient real-time monitoring and rapid feedback.
Patent Information
- Application Number
- CN202210035596.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-01-13
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2042-01-13
AI Technical Summary
Existing online monitoring technologies for laser processing are insufficient for real-time monitoring of the processing depth direction, and their low feedback frequency cannot meet the requirements of high-frequency pulsed laser processing. Conventional optical equipment also has difficulty observing the depth direction and is susceptible to external interference.
An ultra-high-speed X-ray image recognition method based on an encoding plate is adopted. This method combines X-ray imaging technology with artificial intelligence algorithms. By installing an encoding plate in front of the detector to attenuate the X-ray intensity, and combining it with a BPNN neural network, image recognition and feedback can be achieved directly in the detector, avoiding the complex analog-to-digital conversion process.
It enables real-time monitoring and high-frequency feedback during laser processing, improving monitoring accuracy and feedback rate, achieving nanosecond-level image recognition rate, possessing real-time monitoring capability in the depth direction, simplifying equipment structure and reducing power consumption.
Smart Images

Figure CN114596460B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of ultra-high-speed target state recognition and feedback using rays and online monitoring of laser processing with high time response requirements, and in particular, to an ultra-high-speed X-ray image recognition method based on a coding plate. Background Art
[0002] Laser processing refers to processing using the thermal effect generated by projecting a laser beam onto the surface of an object. The processing accuracy is affected by factors such as the processing material, laser power, focal position, nozzle diameter and height, and auxiliary gas pressure. Therefore, for different processing materials and processing requirements, it is necessary to adjust the process parameters in real time during the processing process to improve processing quality and reduce processing costs.
[0003] At the same time, laser processing is mostly faced with ceramics, metals and other materials with complex shapes and structures. The internal changes of materials during processing cannot be observed with conventional optical equipment, which brings difficulties to the real-time monitoring of processing and online processing parameter adjustment. At present, monitoring is mainly through the release of sound, light, heat, electricity and other signals generated during the processing, and then using the corresponding sensor to indirectly judge the processing conditions, but such a method depends on the specific processing materials in actual application, and is easily interfered by the external environment, signal judgment distortion and other problems. Even if multiple sensors are used for joint monitoring, it will also bring equipment space limitations, and it is generally only applicable to a certain specific processing method and material; with the emergence of higher frequency pulse processing technology, the online monitoring feedback frequency in some applications can no longer meet the application requirements; more importantly, the current laser processing online monitoring technology mainly monitors the surface phenomena in the processing area, and lacks effective online monitoring technology in the depth direction, and the monitoring in the depth direction of processing is crucial. Defects in the depth direction will lead to internal cracks, recasting layers, etc. In addition, the use of optical cameras and other equipment to directly observe the processing morphology, in addition to the difficulty in observing in the depth direction mentioned above, has a high time delay from camera imaging to feedback adjustment of processing parameters.
[0004] Therefore, it is necessary to develop a more efficient, faster-responding, and more direct observation method and equipment that can monitor and provide feedback on the internal working conditions of laser-processed materials in real time, so as to solve the monitoring difficulties in the processing depth, high-speed feedback, and high-frequency pulsed laser during laser processing, so as to improve the laser processing accuracy and simplify the processing supervision process. Summary of the invention
[0005] In view of the shortcomings of the prior art, the present invention provides an ultra-high-speed radiation image recognition method based on a coding plate, comprising: S1, obtaining a plurality of input images obtained by a detector after X-rays pass through an object; S2, building a neural network according to the probe pixels of a pixelated detector with multiple pixels and the number of categories of the input images, and using the plurality of input images to train and verify the neural network to obtain a trained and verified neural network, wherein the trained and verified neural network includes weight information and can output category information of the input image; S3, extracting the weight information, and constructing a plurality of coding plates corresponding to the input images, wherein the coding plates can implement a weight calculation process of the neural network by attenuating the intensity of the radiation to different degrees; S4, by installing the coding plate in front of the detector, the radiation is detected by the detector after the intensity of the radiation passes through the coding plate and is attenuated to different degrees at different positions of the detector, and the detector converts the radiation energy information after the attenuation to different degrees into a pulse photoelectric signal, and then the pulse photoelectric signal generated by each pixel of the detector is connected according to the corresponding relationship in the trained and verified neural network, so as to obtain the recognition result of the target object.
[0006] Furthermore, random Gaussian noise is added to the images to expand the number of images so as to obtain a sufficient number of input images for algorithm training.
[0007] Furthermore, the method of extracting weight information in a neural network to construct a coding board includes: in a trained neural network, calculating and outputting the category information of the input image according to the following formula:
[0008]
[0009] Among them, C represents the target category, C i represents the i-th category, n represents the number of image pixels, t m represents the mth pixel, represents the weight of the mth pixel of the i-class image, b i represents bias, f is the activation function corresponding to the summing node to the final output value, the activation function does not change the relative size of the output value, but only makes the large value larger and the small value smaller, so as to obtain a clearer and more accurate classification result;
[0010] Extract each time obtained in the formula Weight values are determined and the weight values are trained to produce coding plates of different thicknesses.
[0011] Further, according to the weight parameter, the method for making the coding plate is as follows: the coding plate uses the same material and different material thicknesses to reflect different weight parameter processing of the ray energy; wherein the material thickness is selected according to the following formula:
[0012] M=M0Be -μR
[0013] Among them, M0 and M represent the ray intensity before and after passing through the material with a thickness of R, μ is the linear attenuation coefficient, the unit is 1 / cm, B is the accumulation factor, and μ and B are related to the material used.
[0014] By reflecting the relationship between material thickness and ray intensity, different weight parameters are mapped to process ray energy.
[0015] Furthermore, the activation function is softmax and its derivative function.
[0016] Furthermore, the method for training the neural network and setting the weight parameters is as follows: a cross entropy loss function is used in the neural network, and the neural network is trained using a stochastic gradient descent and a learning rate decay method.
[0017] Furthermore, the method for training the neural network also includes using an L2 regularization method to prevent overfitting.
[0018] Furthermore, the multiple images are divided into a training set and a validation set using programming software, and the ratio of the training set to the validation set is 8:2.
[0019] The present application also relates to an artificial intelligence rapid imaging device, comprising: a ray emitting device, emitting initial rays, the initial rays passing through an object to be detected to obtain detection rays with different energies; a detection device, comprising a detector and a coding plate processor, the pixelated array detector detects rays of different energies, the coding plate processor executes any of the above-mentioned recognition methods, and outputs the ray energy as target state information.
[0020] In view of the problems of low monitoring accuracy, difficulty in monitoring in the processing depth direction, and low feedback frequency in the current online monitoring of laser processing. This application proposes an artificial intelligence high-speed radiographic image recognition technology based on a non-uniform coding version combined with X-ray imaging technology. Relying on the relatively mature back propagation neural network (BPNN) in the current artificial intelligence algorithm, the weights in the trained neural network are combined with the coding plate that has a certain blocking effect on the rays, and the neural network weights are applied to the thickness of the coding plate according to the numerical size. Therefore, according to the rays with uneven intensity distribution after passing through the coding plate, the image after entering the pixelated probe is converted into data with image feature information processed by weights, and then the target is directly judged and feedback is given according to the above-mentioned difference signal received by the detector. There is no need to perform complex analog-to-digital conversion and subsequent cumbersome processes of information processing on the computer CPU. At present, according to the emission frequency of X-rays, the technology can reach image recognition and instant feedback rates of ns level at the highest. It can monitor the feedback efficiency of the laser processing process in real time, and has the real-time monitoring capability that cannot be observed by conventional optical means in the processing depth and other directions.
[0021] This method uses X-rays that have a certain degree of penetration into materials, combined with an artificial neural network algorithm, with the aim of transplanting the conventional image recognition algorithm that needs to be processed in digital chips such as computer CPUs into a hardware analog circuit for implementation, so that the analog signal output at the moment when the pixelated detector receives the X-rays that pass through the object to be measured can be used to judge the state and morphology of the target object, and immediately feed back to the processing equipment to adjust the processing parameters.
[0022] At the same time, this method uses a non-uniform thickness coding plate combined with a pixelated scintillator to implement it. By individually setting the thickness of the coding plate at each position, it can play a role in programming the incident X-rays, thereby realizing the neural network algorithm in the array detector. This method only needs to add a low-power bias power supply at the detector end. After the pixel density of the detector is increased, there is no need to worry about complex power supply and circuit design issues. The circuit scheme is simple, which brings lower power consumption and structural stability. At the same time, it has the characteristics of a wide field of view, which has unique advantages for some scenes that require wide-angle X-ray image recognition.
[0023] In summary, this application transfers the neural network algorithm that originally needs to be implemented in the computer to the encoding board and pixelated probe based on mathematical principles and methods, so that when the algorithm is used for image recognition, there is no need for subsequent signal analog-to-digital conversion and processing by computer chips, etc., and the image recognition rate can be increased from ms level to ns level.
[0024] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0025] Figure 1 This is a schematic diagram of the process of the rapid ray imaging method of this application.
[0026] Figure 2 Schematic diagram of the initial target image and the image after adding Gaussian noise selected in the embodiment of the present application.
[0027] Figure 3 Shows the one-dimensionalization process of the input image.
[0028] Figure 4 The process of calculating and outputting category information for the neural network structure used in this method.
[0029] Figure 5 This is a simulation diagram of the application of the coding plate in actual detection.
[0030] Figure 6 A model diagram of monitoring using the rapid imaging device of the present application during laser processing. DETAILED DESCRIPTION
[0031] The following is a more detailed description of the specific embodiments of the present invention in conjunction with the accompanying drawings and examples, so that the scheme of the present invention and its advantages in various aspects can be better understood. However, the specific embodiments and examples described below are only for the purpose of illustration, rather than for limiting the present invention.
[0032] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the accompanying drawings, which are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first" and "second" are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, "multiple" means two or more, unless otherwise clearly and specifically defined.
[0033] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection, an electrical connection, or can communicate with each other; it can be directly connected, or indirectly connected through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.
[0034] The disclosure below provides many different embodiments or examples to realize different structures of the present invention. In order to simplify the disclosure of the present invention, the parts and settings of specific examples are described below. Of course, they are only examples, and the purpose is not to limit the present invention. In addition, the present invention can repeat reference numbers and / or reference letters in different examples, and this repetition is for the purpose of simplicity and clarity, which itself does not indicate the relationship between the various embodiments and / or settings discussed. In addition, the present invention provides various specific examples of processes and materials, but those of ordinary skill in the art can be aware of the application of other processes and / or the use of other materials.
[0035] Figure 1 This is a schematic diagram of the process of the rapid ray imaging method of this application.
[0036] like Figure 1 As shown, the method for rapid ray imaging in this application mainly includes the following steps:
[0037] S1, obtaining training data, which is a plurality of input images obtained by a detector after X-rays pass through an object, such as images detected in non-destructive testing. Gaussian noise is added to the target image to obtain a sufficient number of training image data.
[0038] S2, using the above-mentioned number of image data classifications and the probe pixels in the pixelated detector with multiple pixels in the actual state, a neural network is built, and the image data is used to train and verify the neural network to obtain a trained and verified neural network, and the trained and verified neural network includes weight information. The trained and verified neural network can output the category of the input image, based on which the information of the original image can be known. Among them, the weight information is equivalent to the processing process of the original input image, which will be described in detail later.
[0039] S3, constructing a coding plate corresponding to the plurality of input images according to the weight information, wherein the coding plate can cause different degrees of attenuation on the intensity of incident rays.
[0040] S4, since the detector recognizes the object through the uneven distribution of the intensity of the rays after passing through it, the above-mentioned coding plate can be installed in the actual detector so that the rays pass through the object to be identified and then pass through the coding plate. Therefore, the processing of the intensity of the rays is equivalent to the weight processing process of the neural network mentioned above. After that, the ray information entering the detector has already completed the weight processing process, and the detector can convert the energy deposited in the rays after different degrees of attenuation into a pulse photoelectric signal. Subsequently, it is only necessary to connect the photoelectric signal according to the corresponding relationship in the neural network after training and verification, and then output it after processing to obtain the recognition result of the actual target to be measured.
[0041] Each step will be described in detail below.
[0042] Step S1, obtaining sufficient training data. Figure 2 Schematic diagram of the initial target image and the image after adding Gaussian noise selected in the embodiment of the present application. Figure 2 As shown in , the number of initial target images is too small, such as Figure 2 There are 4 images on the left, which are A, F, I, and J. Therefore, random Gaussian noise is added to expand the data volume. As shown in the figure, the input image is 4*4 pixels, and noise with signal-to-noise ratios of 3, 1, and 0.5 is added to it. The image obtained is as follows Figure 2 As shown on the right. It can be seen that by adding noise, the clarity of the image is reduced to a certain extent. The lower the signal-to-noise ratio, the worse the clarity. Through the above method and combined with the actual number of target classifications, graphic complexity, etc., the input data can be expanded to the order of tens of thousands or even millions of images. Training in this way can greatly improve the fault tolerance of the detection equipment when the uniformity of the input graphic is not ideal or there are other interference factors.
[0043] Step S2, building a neural network and training the neural network. Building a neural network based on the pixels of the probe in actual situations and the number of classifications of the images of the training data. And using the training data images to train the neural network, a trained neural network is obtained.
[0044] Among them, the number of image categories can be set according to actual conditions, such as the image category in this embodiment is 4, that is, the number of categories is 4. In this embodiment, one layer of encoding board is used, so the number of hidden layers of the corresponding neural network is one layer. The method of building a neural network based on the input image and the number of categories is a well-known method in the industry and will not be repeated here.
[0045] When training a neural network, the initial training data needs to be classified and divided into a training set and a validation set. It is common practice to divide the initial training data through programming software. This application divides the input data into a training set and a validation set. As the name suggests, the training set is responsible for training the neural network to obtain the optimal output data set, and the validation set needs to verify the accuracy of the trained neural network. The training set: validation set ratio set in this article is 0.8:0.2 (it can be divided into other ratios according to the actual application effect, that is, 80% of all data is used for training and the rest is used for verification).
[0046] Step S3, extracting weight information in the neural network for making a coding board.
[0047] When training a neural network, the input image is converted into a one-dimensional image and then processed to output the category information of the input image. During this process, the neural network parameters, i.e., weight information, are extracted, which is required to build the coding board.
[0048] Figure 3 It shows the one-dimensionalization process of the input image. Figure 3 As shown, the leftmost image is the original image, the middle image is the image after adding Gaussian noise to simulate the image detected by rays. The rightmost image is the pixel matrix parameters generated after the image is one-dimensionalized.
[0049] It should be noted that in this embodiment, all inputs and neurons in the neural network are fully connected. The activation function of the neural network uses softmax, and the loss function uses the cross entropy loss function. These two functions have good performance in multi-classification problems in previous machine learning. At the same time, the random gradient descent and learning rate decay methods are used for network training. In order to prevent overfitting, the L2 regularization method is also applied. The loss value after 1000 iterations is stabilized at about 0.75, and the final verification algorithm accuracy is 97.8%. The network output selects the one-hot encoding method, that is, only one value is "lit up" in each output result. This method can maximize the accuracy of the output value.
[0050] Figure 4 The process of calculating and outputting category information for the neural network structure used in this method. Figure 4 The calculation process represented can be explained by the following formula.
[0051] According to the following formula:
[0052]
[0053] Among them, C represents the target category, C i represents the i-th category, n represents the number of image pixels, tm represents the mth pixel, represents the weight of the mth pixel of the i-class image, b i represents bias, f is the activation function corresponding to the summing node to the final output value, the activation function does not change the relative size of the output value, but only makes the large value larger and the small value smaller, so as to obtain a clearer and more accurate classification result;
[0054] The trained neural network weights are then extracted and a coding plate of corresponding thickness is made in proportion.
[0055] Furthermore, for example, Figure 4 When a specific pattern is input to the coding board on the left, each pixel value in the pattern represents the corresponding pixel n in the computer. i ,Right now Figure 4 The white circle on the left is the n in computer language i *P ni In the detector, it is equivalent to the intensity change of the incident image after the encoding plate. After summing up with different weights, we get Figure 4 The four values of 0.1, 0.6, 0, and 0.1 are processed by the f function to obtain the
[0100] matrix result. This result corresponds to the C2 graph in the setting. Then, we can use Figure 2 There are four image categories in the inference. The original image of the image is F.
[0056] Furthermore, if n=16 and the number of categories is 4, all 16 image pixels need to perform corresponding n*P operations, so for each category to output category information, 16 sets of n*P operations need to be performed. That is, for a specific image pixel, there are several output image categories (several Cs), which need to be multiplied by several weights P. The intensity change of the image by the weight P in the actual physical environment is analogous to the intensity change of the ray by the code plate. Therefore, the weight information P is extracted and used to construct the code plate.
[0057] Moreover, the number of coding plates is the same as the total number of weights, which is the product of the number of elements and the number of categories. For example, if the number of categories is 4 and the number of input pixels is n=16, then the number of coding plates is arranged in an 8*8 matrix.
[0058] In fact, the calculation process of the neural network consists of two parts. The first part is to re-encode each pixel with weights, and the second part is to process all the processed pixel matrices with the sum and f functions to generate matrix categories to determine the category information of the image.
[0059] Figure 5 This is a simulation diagram of the application of the coding plate in actual detection.
[0060] like Figure 5As shown, the coding plate can block the rays to different degrees of attenuation, which is equivalent to the weight processing of the image. This is the first part of the above-mentioned weight processing image process. When manufacturing the coding plate, it is known that the rays attenuate differently when passing through materials of different thicknesses. Therefore, by using the same material of different thicknesses, the intensity of the rays passing through it can be attenuated to the original 0-100%. The arrangement and combination of different thicknesses will form a mask matrix, which is equivalent to encoding the rays, so it is called a coding plate. At the same time, the neural network algorithm in artificial intelligence can be used to construct any algorithm structure, so the weights in the trained algorithm can be extracted and converted into the required coding plate thickness corresponding to each weight, so the neural network structure in the physical sense can be realized.
[0061] Material thickness selection can be based on the following formula:
[0062] M=M0Be -μR …………(2)
[0063] Among them, M and M0 represent the intensity of the ray passing through the material with a thickness of R, μ is the linear attenuation coefficient, the unit is 1 / cm, B is the accumulation factor, and μ and B are related to the material used.
[0064] The formula maps the relationship between material thickness and ray intensity to process ray energy with different weight parameters.
[0065] Step S4, installing the above-mentioned known encoding plate structure into the detector. When the detector recognizes the attenuation change of the incident ray energy, different images are formed in the detector. According to the different images at this time, the corresponding image category can be obtained.
[0066] Specifically, if Figure 5 As shown in the figure, the ray beam passes through the observed object and then passes through the encoding plate to attenuate the ray energy to different degrees. After the rays have been attenuated to different degrees, they are imaged on the detector, and the imaged electrical signals are connected in parallel, which simulates the separate summation process in the neural network. The output signal is then as follows: Figure 5 As shown, it is displayed as the second image category.
[0067] The present application also relates to an artificial intelligence rapid image recognition device, comprising: a ray emitting device, emitting an initial ray, the initial ray passing through an object to be detected, and obtaining a detection ray with different intensities. A detection device, comprising a detector and a coding plate processor, the pixel array detector detecting rays of different intensities and then generating electrical signals with different amplitudes, the coding plate processor executing any imaging method described in 1-7, and outputting the electrical signal as target state information.
[0068] The process of online monitoring of laser processing using the fast image recognition device in this application is as follows: Figure 6 During the laser processing process, the image recognition device in this application (the ray in this embodiment is X-ray) is placed near the workpiece, and X-rays are used to perform real-time monitoring while the laser beam is processing the workpiece.
[0069] By using this application and method, not only the image recognition rate is increased from the ms level to the ns level, but also the neural network algorithm that originally needs to be implemented in the computer is transferred to the coding plate and pixelated probe through mathematical principles and methods, so that when the algorithm is used for image recognition, there is no need to perform subsequent signal analog-to-digital conversion and computer chip processing, which is more convenient. In addition, the hardware construction of the neural network is realized by designing coding plates of different thicknesses, which minimizes the complexity of the equipment. The whole set of equipment has a simple structure and strong stability, and can achieve wide-angle imaging.
[0070] Obviously, the above embodiments are merely examples for clearly explaining the present invention, and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the protection scope of the present invention.
Claims
1. An ultra-high-speed radiographic image recognition method based on a coding plate, characterized in that: include: S1, obtaining a plurality of input images obtained by the detector after the X-ray passes through the object; S2, building a neural network according to the probe pixels of the pixelated detector with multiple pixels and the number of classifications of the input image, and training and verifying the neural network using the multiple input images to obtain a trained and verified neural network, wherein the trained and verified neural network includes weight information and can output category information of the input image; In the trained neural network, the category information of the input image is calculated and output according to the following formula: Among them, C represents the target category, Ci represents the i-th category, n represents the number of image pixels, tm represents the m-th pixel, represents the weight of the mth pixel of the i-class image, bi represents the bias, and f is the activation function corresponding to the summation node to the final output value. The activation function does not change the relative size of the output value, but only makes the large value larger and the small value smaller, so as to obtain a clearer and more accurate classification result; extract the value obtained each time in the formula weight values, and training the weight values; S3, extracting the weight information, constructing a plurality of coding plates corresponding to the input images, wherein the coding plates can implement a weight calculation process of a neural network by attenuating the ray intensity to different degrees; The method for extracting weight information from a neural network to construct a coding plate includes: the coding plate uses the same material with different material thicknesses to reflect different weight information processing of the ray energy; wherein the material thickness is selected according to the following formula: M=M0Be -μR M0 and M represent the ray intensity before and after passing through the material with a thickness of R, respectively. μ is the linear attenuation coefficient, with the unit of 1 / cm. B is the accumulation factor. μ and B are related to the material used. By reflecting the relationship between the material thickness and the ray intensity, different weight information processing is performed on the ray energy in a mapped manner. S4, by installing the coding plate in front of the detector, the intensity of the rays after passing through the coding plate is attenuated to different degrees at different positions of the detector and then detected by the detector. The detector converts the energy information of the rays after attenuation to different degrees into pulse photoelectric signals, and then connects the pulse photoelectric signals generated by each pixel of the detector according to the corresponding relationship in the neural network after training and verification, so as to obtain the recognition result of the target object.
2. The ultra-high-speed X-ray image recognition method based on a coding plate according to claim 1, characterized in that: By adding random Gaussian noise to the image, the number of images is expanded to obtain a sufficient number of input images for algorithm training.
3. The ultra-high-speed radiographic image recognition method based on a coding plate according to claim 2, characterized in that: The activation function is softmax and its derivative function.
4. The ultra-high-speed radiographic image recognition method based on a coding plate according to claim 3, characterized in that: The method for training the neural network and setting the weight information is as follows: a cross entropy loss function is used in the neural network, and the neural network is trained using a stochastic gradient descent and a learning rate decay method.
5. The ultra-high-speed radiographic image recognition method based on a coding plate according to claim 4, characterized in that: The method of training the neural network also includes using an L2 regularization method to prevent overfitting.
6. The ultra-high-speed radiographic image recognition method based on a coding plate according to claim 5, characterized in that: The multiple images are divided into a training set and a validation set using programming software, and the ratio of the training set to the validation set is 8:
2.
7. An artificial intelligence rapid imaging device, characterized in that: include: A ray emitting device emits an initial ray, wherein the initial ray passes through the object to be detected to obtain detection rays with different energies; The detection device includes a pixelated array detector and a coding plate processor, wherein the pixelated array detector detects rays of different energies, and the coding plate processor executes any one of the identification methods described in 1-6 to output the ray energy as target state information.
