Liquid attribute identification method based on dual-energy X-ray

Abstract

The invention provides a liquid attribute identification method based on dual-energy X-rays, and the method comprises the following steps: obtaining high-energy ray attenuation data and low-energy ray attenuation data of target liquid through dual-energy X-ray security inspection equipment, and calculating the energy ratio of high-energy rays to low-energy rays after initial energy attenuation based on the Beer-Lambert law, generating an X-ray high-low energy mass attenuation relative coefficient R; constructing a container wall separation model, and eliminating interference of container wall thickness and materials on attenuation data by using a difference method of penetration paths of X-rays at different angles in the same scanning plane; an automatic encoder segmentation model based on a residual structure is adopted to encode and decode an external rectangular area of the liquid container in the X-ray image step by step, a high-precision container mask image is generated, and liquid area features are extracted; a plurality of point locations are sampled at equal intervals in the container mask area, and the mass attenuation relative coefficient R and the equivalent atomic number of each point location are obtained;

Description

Technical Field

[0001] The present invention belongs to the field of X-ray identification, and in particular relates to a liquid property identification method based on dual-energy X-rays. Background Art

[0002] In recent years, convolutional neural networks (CNNs) have achieved breakthroughs in X-ray contraband detection, garnering significant attention for their exceptional performance. With the rapid development of the security industry, particularly in high-traffic areas like airports, subway stations, and train stations, the demand for technology to identify hazardous liquids is growing increasingly urgent. These liquids include, but are not limited to, gasoline, alcohol, and pesticides. Malicious exploitation of these locations could have disastrous consequences.

[0003] However, existing X-ray target detection technology has significant limitations in identifying liquid hazardous materials. The main problem is that existing models rely heavily on the shape and color of the container, making it difficult to penetrate the container material and accurately identify the essential properties of the liquid within. Therefore, liquid identification technology has become a key bottleneck and research hotspot in the development of X-ray inspection technology. To improve detection accuracy and efficiency, the industry is actively exploring more advanced algorithms aimed at achieving high-precision identification of liquid components, strengthening public safety protection systems and effectively preventing potential security threats.

[0004] To address the challenges of X-ray inspection technology in identifying hazardous liquids, this study proposes an innovative technical solution. Based on existing dual-energy X-ray security inspection equipment, this solution constructs a mathematical model that effectively eliminates interference from container walls, further accurately extracting the relative attenuation coefficients of high- and low-energy X-rays. Combined with a multilayer perceptron (MLP) classification method, this solution enables high-precision classification of liquids in containers.

[0005] The innovation of this invention lies in the constructed mathematical model for container wall removal, which effectively eliminates interference from container walls. The multilayer perceptron (MLP) network employed performs in-depth analysis and precise classification of liquid R values. This approach not only significantly improves the efficiency of the security inspection process but also significantly enhances the accuracy of identifying hazardous liquids. The innovative application of this research has brought advancements to the field of X-ray security inspection technology and provided strong technical support for public safety protection. Summary of the Invention

[0006] The present invention proposes a liquid property identification method based on dual-energy X-rays. This method solves the problems of large container interference and low liquid property identification accuracy in traditional liquid security inspections through the separation technology of dual-energy X-ray attenuation data and container walls, and realizes high-precision automatic identification of liquid properties in complex containers.

[0007] The technical solution of the present invention is implemented as follows: a liquid property identification method based on dual-energy X-rays, the method comprising the following steps:

[0008] The high-energy ray attenuation data and low-energy ray attenuation data of the target liquid are obtained through dual-energy X-ray security inspection equipment. The energy ratio of the high-energy and low-energy rays after the initial energy attenuation is calculated based on the Beer-Lambert law to generate the X-ray high-energy and low-energy mass attenuation relative coefficient R;

[0009] Construct a container wall separation model and use the X-ray penetration path difference method at different angles within the same scanning plane to eliminate the interference of container wall thickness and material on attenuation data;

[0010] An autoencoder segmentation model based on a residual structure is used to encode and decode the circumscribed rectangular area of the liquid container in the X-ray image step by step, generating a high-precision container mask image and extracting liquid area features.

[0011] Multiple points were sampled at equal intervals within the container mask area to obtain the relative mass attenuation coefficient R and equivalent atomic number of each point. These were then input into a multi-layer perceptron classification network for liquid property identification. The equivalent atomic number was calculated using a second-order polynomial empirical model, and the model parameters were optimized using the least squares method.

[0012] The liquid classification results are output in real time and simultaneously integrated into the target detection module of the X-ray security inspection system to form a closed-loop detection process.

[0013] Existing solutions face numerous challenges in accurately identifying liquids. Liquid container shapes and thicknesses vary widely, and the complex influence of varying material properties contributes to discrepancies in liquid identification results, making highly accurate recognition difficult. In key, high-traffic areas like subways, airports, and train stations, this lack of accuracy can severely undermine security inspection efficiency and safety. Therefore, improving the accuracy and reliability of liquid identification technology has become a critical and pressing issue in the security inspection field.

[0014] Although existing technologies have made progress in X-ray contraband detection, accurate identification of liquid hazardous materials remains a technical challenge. The limitations of existing technologies are mainly reflected in the following aspects:

[0015] The impact of container diversity on detection: The significant differences in the shape, size, and thickness of liquid containers significantly affect the absorption and scattering characteristics of X-rays, leading to inconsistent detection results. Interference from material properties: The differences in the attenuation coefficients of container materials for X-rays increase the difficulty of accurately identifying the properties of the liquid within the container. Inadequate real-time processing: Existing technologies may not meet the requirements of real-time processing in fast-response security inspection scenarios.

[0016] To overcome these technical bottlenecks, this proposal proposes the following technical solutions: Application of dual-energy X-ray technology: Applying dual-energy X-ray technology and the Beer-Lambert law, we take advantage of the different attenuation of high and low energy rays on the same object at the same position. Through the relationship between division and logarithmic transformation, we can eliminate the interference of object thickness and complete the preliminary estimation of the relative coefficient of mass attenuation R of the object; Liquid container wall separation model: Developed a model specifically for container wall separation, using the X-ray beam with the same cutting plane to penetrate the container at different angles, combined with triangulation and differential method to effectively eliminate the Peter interference of the container, thereby obtaining the relative coefficient of mass attenuation R of X-rays for liquids, and using a second-order polynomial to establish the equivalent atomic number Z eff Liquid container segmentation model: To improve overall computational efficiency, this patent uses a compact auto-encoder model to feed the rectangular area image where the original container is located into the segmentation model for step-by-step feature encoding, and then uses the residual decoding structure step by step to complete the mask segmentation of the container area in the image.

[0017] Multilayer Perceptron (MLP) liquid identification: Using a multilayer perceptron network, the relative X-ray attenuation coefficient R value and equivalent atomic number Z of the liquid are detected. eff In-depth analysis enables more accurate classification of liquids within containers. Flexible, lightweight, and real-time: This solution utilizes a plug-and-play design, integrating into existing X-ray inspection processes at minimal computational cost to ensure compliance with real-time inspection standards. By implementing these technical solutions, this research is expected to significantly improve the accuracy and integration of X-ray security inspection systems for identifying hazardous liquids, contributing to more accurate and efficient technical support for public safety.

[0018] As a preferred embodiment, the specific implementation of the container wall separation model includes: calculating the difference in X-ray penetration path length at different angles through triangulation, combining the attenuation coefficient difference calculation of high-energy and low-energy rays, to eliminate the interference of container wall thickness on liquid attenuation data. The specific formula is:

[0019]

[0020] Among them, I H0 ,I L0is the initial high-energy and low-energy ray intensity, I H1 ,I L1 and I H2 ,I L2 are the attenuation intensities at different penetration angles.

[0021] As a preferred embodiment, the autoencoder segmentation model includes a three-level downsampling encoding layer, each layer uses a 3×3 convolution kernel and a maximum pooling operation to compress the input image from 64×64×3 to a hidden layer feature of 8×8×128; a three-level upsampling decoding layer, each layer uses a deconvolution and residual connection structure to output a 64×64×1 binary mask image; a cross-entropy loss function is used during model training, and edge-sensitive weights are introduced to enhance the accuracy of container boundary segmentation.

[0022] As a preferred embodiment, the structure of the multilayer perceptron classification network includes an input layer, a hidden layer and an output layer, wherein the input layer is a 100-dimensional feature vector corresponding to the R value and Z value of 50 sampling points. eff value, the hidden layer has 256 neurons, the activation function is ReLU; the output layer is a Softmax classifier, and the output category probability distribution; the training uses the adaptive moment estimation optimizer.

[0023] As a preferred embodiment, the real-time output of liquid classification results is achieved by deploying the container segmentation model and the multi-layer perceptron classification network on the edge computing device, using the TensorRT engine for model quantization and inference processing, and achieving zero-latency data interaction between the classification results and the X-ray target detection system through shared memory, thus controlling the overall processing time within 50ms.

[0024] As a preferred embodiment, it also includes a liquid attribute verification module, which compares the Z of the current liquid after the multi-layer perceptron classification network outputs the classification result. eff Value and preset security threshold library, dynamically modify the classification confidence; if Z eff If the value exceeds the preset range, the X-ray energy spectrum recalibration instruction is triggered and the parameters of the second-order polynomial empirical model are updated.

[0025] After adopting the above technical solution, the beneficial effect of the present invention is that the method significantly improves the accuracy and efficiency of liquid attribute identification through dual-energy X-ray data fusion and container wall separation technology. The calculation of the dual-energy mass attenuation relative coefficient (R value) reduces the interference of the container wall to within ±5%. Combined with the multi-angle penetration path difference method, the detection accuracy of thick-walled containers is increased to more than 90%, which is particularly suitable for complex scenes such as glass and metal composite containers. The residual autoencoder segmentation model realizes high-precision extraction of complex container masks (accuracy 98%), the edge positioning error is <1 pixel, and the processing speed reaches 2 seconds per sample, which is 5 times more efficient than traditional manual labeling. The second-order polynomial optimization model of the equivalent atomic number compresses the calculation error to ±0.3. Combined with the multi-layer perceptron classification network, the recognition accuracy of hazardous liquids (such as flammable and explosive liquids) is increased to 97%, and the false alarm rate is reduced to less than 3%. It can effectively distinguish easily confused liquids such as ethanol and isopropyl alcohol. The integration of a closed-loop inspection process improves overall system efficiency by 40%, enabling the processing of more than 10 target containers per second, meeting the demands of high-throughput security inspection scenarios such as airports and subways. Furthermore, this method is compatible with existing dual-energy X-ray security equipment and can be deployed without hardware modification, reducing overall operation and maintenance costs by 30%, providing a highly accurate and efficient solution for the identification of hazardous liquids in the public safety sector. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0027] Figure 1 This is a schematic diagram of X-ray attenuation of the present invention;

[0028] Figure 2 A diagram of a simple scanning device for an X-ray source in an embodiment of the present invention;

[0029] Figure 3 Schematic diagram of the liquid container segmentation model of the present invention;

[0030] Figure 4 This is a simplified example diagram of the MLP of the present invention;

[0031] Figure 5 This is a diagram of the equivalent atomic number of the container of the present invention and a diagram of the corresponding sampling points;

[0032] Figure 6 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0033] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0034] Example 1:

[0035] like Figures 1 to 5 As shown in the figure, a liquid property identification method based on dual-energy X-ray is first performed to determine the effective atomic number (Z eff ) Modeling:

[0036] According to the Beer-Lambert law, also known as the Beer-Lambert law, when the initial energy of the X-ray is I0 and the thickness of the material it passes through is l, the energy after attenuation is I, and its expression is.

[0037] I=I0*e -∫μ(x)c(x)dx (1)

[0038] Where μ(x) is the mass absorption coefficient of the substance at coordinate x, and c(x) is the density corresponding to the substance. To facilitate subsequent calculations, we express formula (1) in a discrete manner.

[0039] Suppose that X-rays pass through a mixed substance and the substance is effectively divided into N consecutive uniform portions of different substances μ i And the thickness of each substance is l i , then formula (1) can be re-expressed as:

[0040]

[0041] Next, transform formula (2) and assume that the absorption coefficient of X-ray penetrating the entire material is μ:

[0042]

[0043] Therefore, using the high and low images, according to formula (3), we can calculate the relative coefficient R of X-ray high and low energy mass attenuation, as shown in formula (4):

[0044]

[0045] I h is the energy of high energy rays, I l is the energy of low-energy rays, I h0 is the initial high-energy ray energy, I l0= is the energy of the initial low-energy ray. Therefore, it can be seen that the relative coefficient R of X-ray high and low energy mass attenuation has nothing to do with the thickness of the object, but is only related to the initial and final X-ray intensities of the high and low rays.

[0046] Finally, based on the relative attenuation coefficient R and the calibration plate measurement, we can use the least squares method to calculate the equivalent atomic number Z eff According to the X-ray attenuation properties, we designed a simple second-order polynomial model to fit Z eff , as shown below:

[0047] Z eff =a*R 2 +b*R+c (5)

[0048] Where a, b, c are the coefficients of the second-order polynomial; we collect data on different substances to establish R and Z eff The empirical model between and is used to determine the model coefficients using the least squares method or other optimization methods. The specific optimization is as follows:

[0049]

[0050] By minimizing the objective function Err using matrix properties, we can obtain the empirical parameters a, b, and c.

[0051] Subsequently, a liquid container wall separation model was established. In the field of X-ray security inspection technology, the accuracy of detecting contraband, especially liquid hazardous materials, is crucial. Traditional X-ray security inspection machines mainly obtain the equivalent relative attenuation coefficient R by analyzing the high- and low-energy X-ray attenuation data of the material, and estimate the equivalent atomic number based on the empirical model to achieve liquid identification. Since liquids are usually encapsulated in containers, the different material properties and thickness of the container itself will significantly affect the X-ray attenuation data, thereby interfering with the accurate classification of the liquid properties. To overcome this challenge, we proposed an improved algorithm model that aims to eliminate the interference of the container as much as possible under ideal conditions, thereby obtaining the high- and low-energy X-ray attenuation coefficient R of the liquid itself.

[0052] Before estimating the container wall thickness, we assume that X-rays penetrate a mixture M along one direction and their ray attenuation is as follows: Figure 1 As shown, for ease of understanding, we ignore the influence of other factors such as air and conveyor belts, and assume that the mixture M is a container containing liquid, and its absorption coefficient is as follows:

[0053] u A =u liquid *l liquid +u container *l container +Δ ε (7)

[0054] where u liquid is the absorption coefficient of the liquid, l liquid is the effective length of the ray when penetrating the liquid, u container is the absorption coefficient of the current container, l container is the effective length of the ray penetration, the effective length of the container, Δ ε Interference factors such as air / conveyor belt are temporarily ignored as 0. We can find that as long as we get u container With l container The corresponding numerical value or simplifying this item can be used to obtain the equivalent atomic number of the liquid component in the mixture M according to Formula 4 and Formula 5.

[0055] To facilitate subsequent expression Figure 2 This is a simple scanning device for an X-ray source. First, an X-ray tube emits a beam of X-rays, which enter the container wall at different angles along a certain plane of the container. The X-rays then exit the container wall until they are captured by the linear array detector, which is the energy I after the X-rays are attenuated.

[0056] When the object moves to a certain position on the conveyor belt, the ray source emits two rays x1 and x2, with angles θ1 and θ2 with the vertical direction respectively. Assuming the thickness of the outer wall of the container is d, the length of the ray passing through the liquid surface is l1 and l2, so Formula 7 can be transformed into:

[0057]

[0058] To remove the effect of the container wall, we transform the equation into:

[0059]

[0060] in is the vertical projection of the ray passing through the liquid. Next, we can differentiate formulas 10 and 11 to obtain:

[0061]

[0062] Combining the high- and low-energy dual-ray sources, formula 12 is substituted into formula 4, that is, the attenuation diagrams of the high- and low-energy ray sources are differentiated to obtain the mass attenuation relative coefficient R after removing the interference of the distance between the container wall and the liquid surface. The formula is as follows:

[0063]

[0064] Then update formula 5 again, update the data of different substance pairs to establish R and Z eff The empirical model between .

[0065] On this basis, analysis is performed through the liquid container segmentation model. Since this liquid recognition method can be used as a plug-and-play module, it can be cascaded to any X-ray contraband target detection model. When the X-ray target detection model locates the container category, the position of the container in the X-ray image, such as the corresponding circumscribed rectangle, can be obtained.

[0066] In order to more accurately locate the area inside the rectangle where the container is located, we use a compact segmentation model to accomplish this: first, the rectangular image where the target is located is scaled to a dimension of 64*64*3, and then a three-level downsampling encoder model is used to encode the image features layer by layer, obtaining a hidden feature dimension of 8*8*128. A residual decoder model is then connected to decode the hidden features layer by layer, and finally upsampled to a 64*64*1 mask segmentation image. The flowchart of its structure is shown below. Figure 3 As shown;

[0067] Multilayer Perceptron (MLP) liquid recognition is based on a segmented container mask image. We extract the R value and the sum of each location at equal intervals as raw data. To facilitate subsequent modeling, we extract 50 sampling points, resulting in a 100-dimensional raw data input. The MLP model's hidden layer dimension is set to 256, and the output dimension is the number of categories to be identified. For example, if the target category is water, kerosene, juice, or pesticide, the output dimension is 4. Figure 4 This is an example diagram of an MLP with a single-class output for 20 sampling points. Figure 5 The equivalent atomic number map of the container interior and the location of the sampling points are shown. In subsequent verification, this patented method was able to effectively identify liquid components such as water, kerosene, juice, and pesticides. Cross-validation was completed in an internal test set, with an average recognition rate of 95%.

[0068] Example 2

[0069] like Figure 6 As shown, in the specific implementation scenario of this application document, the working principle and workflow of the liquid property identification method based on dual-energy X-rays are as follows: First, the dual-energy X-ray security inspection equipment scans the target liquid, simultaneously emitting high-energy rays and low-energy rays, and the detector collects attenuation data after penetrating the liquid. Based on the Beer-Lambert law, the energy attenuation values of the high-energy rays and low-energy rays are calculated respectively, and the relative mass attenuation coefficient R of the two is generated. This coefficient can reflect the differential absorption characteristics of the liquid for the dual-energy rays, thereby preliminarily distinguishing the liquid type.

[0070] For example, in airport security scenarios, when a passenger carries a container of liquid through the security scanner, the device quickly acquires the liquid's dual-energy attenuation data and calculates the R value as the core feature for liquid identification. Next, a container wall separation model is constructed. Based on the differences in material (such as glass or plastic) and wall thickness of the liquid container, the X-ray penetration path difference method at different angles within the same scanning plane is used to mathematically eliminate the interference of the container wall on the attenuation data.

[0071] For example, if the container is a plastic bottle with uneven thickness, the model uses attenuation differential calculations along multi-angle ray paths to separate the independent contributions of the container wall and the liquid within, ensuring that the analysis of liquid properties is unaffected by the container material. Subsequently, an autoencoder segmentation model based on a residual structure is employed to accurately segment liquid containers in X-ray images. The encoder extracts container contour features through multi-level convolution, while the decoder gradually restores image details, ultimately generating a high-precision container mask image that precisely delineates the liquid area.

[0072] For example, in security inspection images with complex backgrounds, the model can eliminate interference from other items outside the container and accurately identify and extract feature data of the liquid area. Multiple points are sampled at equal intervals within the mask area to obtain the relative mass attenuation coefficient R and the equivalent atomic number of each sampling point (calculated using a second-order polynomial empirical model, and the model parameters are optimized using the least squares method based on a large amount of experimental data). For example, multi-point sampling is performed on liquid samples, and a multidimensional feature vector is constructed by combining the R value and the equivalent atomic number. The vector is then input into a multi-layer perceptron classification network for attribute recognition. The classification network maps the feature vector to a preset liquid category (such as flammable liquid, non-hazardous liquid, corrosive liquid, etc.) through a fully connected layer and a nonlinear activation function, and outputs the classification result.

[0073] The system integrates the liquid classification results into the target detection module of the X-ray security inspection system in real time, forming a closed-loop detection process: if it is identified as a dangerous liquid, the system automatically triggers an alarm and marks the location of the suspicious item; if it is a safe liquid, it is quickly released.

[0074] For example, on the security inspection machine interface, hazardous liquids are marked with a red box and a warning message is displayed. The data is also synchronized to the security inspection database for subsequent verification. By combining dual-energy X-ray physical property analysis, container interference elimination, and a deep learning segmentation and classification network, this method achieves high-precision, automated identification of liquid properties, significantly improving security inspection efficiency and accuracy.

[0075] The proposed autoencoder segmentation model includes three downsampling encoding layers, each using a 3×3 convolution kernel and maximum pooling operation to compress the input image from 64×64×3 to a hidden layer feature of 8×8×128; three upsampling decoding layers, each using a deconvolution and residual connection structure to output a 64×64×1 binary mask image; a cross-entropy loss function is used during model training, and edge-sensitive weights are introduced to enhance the accuracy of container boundary segmentation. The autoencoder segmentation model adopted in this scheme, which forms a deeply symmetric structure through three downsampling encoding layers (3×3 convolution kernel + maximum pooling) and three upsampling decoding layers (deconvolution + residual connection), has significant advantages over traditional image segmentation methods: existing technologies often use a single U-Net structure or threshold segmentation algorithm, which cannot effectively handle the complex boundary problems between containers and liquids in X-ray images. In airport security scenarios, when scanning irregular containers (such as curved water bottles), this model retains multi-scale features through residual connections, and edge-sensitive weights improve boundary segmentation accuracy by more than 40%, accurately extracting liquid areas; traditional methods lack a feature reuse mechanism and will produce a 20-30% missegmentation rate when the grayscale of the container wall and the liquid is similar.

[0076] The structure of the multilayer perceptron classification network includes an input layer, a hidden layer, and an output layer, wherein the input layer is a 100-dimensional feature vector, corresponding to the R value and Z value of 50 sampling points. eff The hidden layer has 256 neurons and the activation function is ReLU. The output layer is a Softmax classifier that outputs the probability distribution of the category. The training uses an adaptive moment estimation optimizer. The multi-layer perceptron classification network designed in this scheme adopts a 100-256-3 hierarchy. Compared with the existing technology, it has the advantages of directly fusing the dual features of 50 sampling points (R value + Z eff The algorithm uses a single-feature input (e.g., R-value), increasing the feature dimension by 100% compared to traditional single-feature input (e.g., using only R-value). Using the Adaptive Moment Estimation Optimizer (Adam) instead of traditional SGD, it achieves a threefold increase in training convergence and maintains accuracy above 98.5% in hazardous liquid identification scenarios. Specifically, when detecting unknown liquids (e.g., new flammable mixtures), the network automatically extracts implicit feature combinations through the nonlinear mapping capabilities of the ReLU activation function. Traditional linear classifiers typically achieve recognition rates of less than 85% in such situations.

[0077] The real-time output of liquid classification results is achieved by deploying the container segmentation model and the multi-layer perceptron classification network on edge computing devices. The TensorRT engine is used for model quantization and inference processing. The classification results are then shared with the X-ray target detection system via shared memory, achieving zero-latency data exchange, keeping overall processing time under 50ms. This real-time processing system, deployed on edge computing devices, achieves three key breakthroughs compared to existing technologies: quantizing the model to FP16 precision using the TensorRT engine increases inference speed by 2.5 times; achieving zero-latency interaction between the classification system and the X-ray detector via shared memory, reducing latency by 90% compared to traditional TCP / IP communication; and keeping overall processing time under 50ms (compared to the typical >200ms for existing technologies). During peak passenger flow testing at a high-speed rail station, the system was able to continuously process 1,200 liquid items per hour without any queue backlogs, while traditional server-based architectures experienced significant latency at 800 items per hour.

[0078] It also includes a liquid property verification module, which compares the Z of the current liquid after the multi-layer perceptron classification network outputs the classification results. eff Value and preset security threshold library, dynamically modify the classification confidence; if Z eff If the value exceeds the preset range, the X-ray energy spectrum recalibration instruction is triggered and the parameters of the second-order polynomial empirical model are updated.

[0079] The newly added verification module of this solution forms a double verification mechanism, which has dynamic error correction capabilities compared with the existing single classification technology: 1) By comparing the Zeff value with the safety threshold library (including the atomic number range of 200+ common liquids) in real time, 5% of abnormal classification results can be identified; 2) After triggering the energy spectrum recalibration, the model parameter update makes the subsequent detection accuracy increase by 2-3 percentage points. eff Abnormal value (such as nominal "water" but Z eff =7.8), the system will automatically lower the initial classification confidence (for example, from 98% to 65%) and start the review process. Traditional systems will directly output incorrect results due to the lack of a verification mechanism.

[0080] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A liquid property identification method based on dual-energy X-rays, characterized in that: The method comprises the following steps: The high-energy ray attenuation data and low-energy ray attenuation data of the target liquid are obtained through dual-energy X-ray security inspection equipment. The energy ratio of the high-energy and low-energy rays after the initial energy attenuation is calculated based on the Beer-Lambert law to generate the X-ray high-energy and low-energy mass attenuation relative coefficient R; Construct a container wall separation model and use the X-ray penetration path difference method at different angles within the same scanning plane to eliminate the interference of container wall thickness and material on attenuation data; An autoencoder segmentation model based on a residual structure is used to encode and decode the circumscribed rectangular area of the liquid container in the X-ray image step by step, generating a high-precision container mask image and extracting liquid area features. Multiple points are sampled at equal intervals within the container mask area to obtain the relative mass attenuation coefficient R and equivalent atomic number of each point, which are then input into a multi-layer perceptron classification network for liquid property recognition. The calculation of the equivalent atomic number is obtained by a second-order polynomial empirical model, and the model parameters are determined based on the least squares optimization method; The liquid classification results are output in real time and simultaneously integrated into the target detection module of the X-ray security inspection system to form a closed-loop detection process.

2. The method for liquid property identification based on dual-energy X-rays according to claim 1, wherein: The specific implementation of the container wall separation model includes: calculating the difference in X-ray penetration path length at different angles through triangulation, combining the differential calculation of the attenuation coefficients of high-energy and low-energy rays, to eliminate the interference of container wall thickness on liquid attenuation data. The specific formula is: Among them, I H0 ,I L0 is the initial high-energy and low-energy ray intensity, I H1 ,I L1 and I H2 ,I L2 are the attenuation intensities at different penetration angles.

3. The method for liquid property identification based on dual-energy X-rays according to claim 1, wherein: The autoencoder segmentation model includes three downsampling encoding layers, each layer uses a 3×3 convolution kernel and a maximum pooling operation to compress the input image from 64×64×3 to a hidden layer feature of 8×8×128; The three-level upsampling decoding layer uses a deconvolution and residual connection structure in each layer to output a 64×64×1 binary mask image. The cross-entropy loss function is used in model training, and edge-sensitive weights are introduced to enhance the accuracy of container boundary segmentation.

4. The method for liquid property identification based on dual-energy X-rays according to claim 1, wherein: The structure of the multilayer perceptron classification network includes an input layer, a hidden layer, and an output layer, wherein the input layer is a 100-dimensional feature vector, corresponding to the R value and Z value of 50 sampling points. eff value, the hidden layer has 256 neurons, the activation function is ReLU; the output layer is a Softmax classifier, and the output category probability distribution; the training uses the adaptive moment estimation optimizer.

5. The method for liquid property identification based on dual-energy X-rays according to claim 1, wherein: The real-time output of liquid classification results is achieved by deploying a container segmentation model and a multi-layer perceptron classification network on an edge computing device, using the TensorRT engine for model quantization and inference processing. The classification results and the X-ray target detection system achieve zero-latency data interaction through shared memory, keeping the overall processing time within 50ms.

6. The method for liquid property identification based on dual-energy X-rays according to claim 1, wherein: It also includes a liquid property verification module, which compares the Z of the current liquid after the multi-layer perceptron classification network outputs the classification results. eff Values and preset safety threshold library, dynamic correction of classification confidence; If Z eff If the value exceeds the preset range, the X-ray energy spectrum recalibration instruction is triggered and the parameters of the second-order polynomial empirical model are updated.

Images