X-ray detection method and detector for detecting various densities and materials

Abstract

The invention relates to the technical field of X-ray detection, and discloses an X-ray detection method and detector for detecting various densities and materials, and the method comprises the steps: carrying out the penetration imaging of an X-ray beam through a double-layer TFT detector structure; forming a first common energy X-ray image on the first TFT imaging layer for X-rays penetrating through the detected object, and forming a first high-energy X-ray image filtered by the copper filter layer on the second TFT imaging layer; performing preprocessing to obtain a second ordinary energy X-ray image and a second high-energy X-ray image; performing dual-energy-domain feature extraction, and constructing an effective atomic number map and a material density map; and executing feature space clustering analysis and decision tree classification, and outputting a final detection result. According to the invention, the image information of different energy domains is effectively integrated, the recognition capability of a mixed material and a complex structure is enhanced, the service life of a TFT detector is prolonged, and the system operation cost is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of X-ray detection, and in particular, to an X-ray detection method and a detector for detecting multiple densities and materials. Background Art

[0002] Traditional X-ray detection systems for multiple density materials mainly use single-layer TFT detectors, and multiple samplings are required by switching the energy of the X-ray source to obtain information on different density materials. This multiple sampling method not only increases the radiation dose received by the sample, but also causes difficulties in image registration due to the minute displacement of the object being detected during the shooting interval, thus affecting the accuracy of material identification.

[0003] Even more intractable is that when faced with dynamic scenarios such as a fast-moving suitcase on a conveyor belt or flowing products on a production line, traditional techniques are unable to achieve real-time and accurate material differentiation due to the need for multiple exposures. In addition, conventional energy spectrum detectors are costly and bulky, making it difficult to be widely applied in cost-sensitive scenarios. Existing dual-energy imaging techniques can provide a certain ability to distinguish materials, but the utilization of multi-level energy information at the algorithm level is limited, especially when quickly identifying mixed materials in complex objects. Summary of the Invention

[0004] The present invention provides an X-ray detection method and a detector for detecting multiple densities and materials. The present invention effectively integrates image information in different energy domains, enhances the ability to identify mixed materials and complex structures, and at the same time extends the service life of the TFT detector and reduces the system operation cost.

[0005] In a first aspect, the present invention provides an X-ray detection method for detecting multiple densities and materials, and the X-ray detection method for detecting multiple densities and materials includes: Performing penetration imaging on an X-ray beam through a double-layer TFT detector structure, where the double-layer TFT detector structure includes a first TFT imaging layer, a copper filtering layer, and a second TFT imaging layer; Forming a first ordinary energy X-ray image of the X-ray penetrating the object to be detected on the first TFT imaging layer, and at the same time forming a first high-energy X-ray image filtered by the copper filtering layer on the second TFT imaging layer; Preprocessing the first ordinary energy X-ray image and the first high-energy X-ray image to obtain a second ordinary energy X-ray image and a second high-energy X-ray image; Performing dual-energy domain feature extraction based on the second ordinary energy X-ray image and the second high-energy X-ray image to construct an effective atomic number map and a material density map; Based on the effective atomic number map and the material density map, perform feature space clustering analysis and decision tree classification, and output the final detection result.

[0006] In a second aspect, the present invention provides an X-ray detector for detecting multiple densities and materials, and the X-ray detector for detecting multiple densities and materials includes: A penetration imaging module for performing penetration imaging on an X-ray beam through a double-layer TFT detector structure, where the double-layer TFT detector structure includes a first TFT imaging layer, a copper filtering layer, and a second TFT imaging layer; An image formation module for forming a first ordinary energy X-ray image of the X-ray penetrating the object to be detected on the first TFT imaging layer, and simultaneously forming a first high-energy X-ray image filtered by the copper filtering layer on the second TFT imaging layer; A preprocessing module for preprocessing the first ordinary energy X-ray image and the first high-energy X-ray image to obtain a second ordinary energy X-ray image and a second high-energy X-ray image; A feature extraction module for performing dual-energy domain feature extraction based on the second ordinary energy X-ray image and the second high-energy X-ray image, and constructing an effective atomic number map and a material density map; An output module for performing feature space clustering analysis and decision tree classification based on the effective atomic number map and the material density map, and outputting the final detection result.

[0007] In the technical solution provided by the present invention, a double-layer TFT detector structure is adopted and a copper filtering layer is arranged between the two layers of TFTs, realizing the function of obtaining an ordinary energy image and a high-energy image simultaneously with a single X-ray exposure, without the need to switch the X-ray source energy for multiple exposures, fundamentally avoiding the image registration problem in the traditional technology. Special feature extraction paths are designed for low-density materials and high-density materials respectively, and the density difference features of different materials are highlighted through an energy difference enhancement module, improving the accuracy of material identification. The image information in different energy domains is effectively integrated, enhancing the recognition ability for mixed materials and complex structures, especially the discrimination ability for materials with similar densities but different atomic numbers. The direct material attenuation coefficient reconstruction algorithm of the present invention utilizes the spatial registration advantage of the double-layer TFT detector to directly establish an accurate mapping of the effective atomic number and the material density, without complex image registration preprocessing, greatly reducing the computational complexity. This method obtains dual-energy image information through a single X-ray exposure, significantly reducing the radiation dose, while prolonging the service life of the TFT detector and reducing the system operation cost.

[0008] Other features and advantages of the present invention will be set forth in the following description, and in part will be obvious from the description, or may be learned by practice of the present invention. The objectives and other advantages of the present invention are realized and attained by the structure particularly pointed out in the specification, claims as well as the drawings.

[0009] To make the above objectives, features and advantages of the present invention more obvious and understandable, the following specific preferred embodiments are given, in conjunction with the accompanying drawings, and are described in detail as follows. Brief Description of the Drawings

[0010] Figure 1 It is a schematic diagram of an embodiment of an X-ray detection method for detecting multiple densities and materials in an embodiment of the present invention; Figure 2 It is a schematic diagram of an embodiment of an X-ray detector for detecting multiple densities and materials in an embodiment of the present invention. Detailed Embodiments

[0011] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0012] The terms "including" and "having" and any variations thereof mentioned in the embodiments of the present invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device including a series of steps or units is not limited to the listed steps or units, but optionally further includes other unlisted steps or units, or optionally further includes other steps or units inherent to these processes, methods, products or devices.

[0013] For ease of understanding of this embodiment, first, a detailed introduction is given to an X-ray detection method for detecting multiple densities and materials disclosed in the embodiments of the present invention. As Figure 1 shown, the method includes the following steps: 101. Pass an X-ray beam through a double-layer TFT detector structure for penetration imaging, where the double-layer TFT detector structure includes a first TFT imaging layer, a copper filtering layer, and a second TFT imaging layer; It can be understood that the execution subject of the present invention can be an X-ray detector for detecting multiple densities and materials, or a terminal or a server. Specifically, it is not limited here. In the embodiments of the present invention, the server is taken as an example of the execution subject for illustration.

[0014] Specifically, a double-layer TFT detector structure with energy-layered response capabilities is constructed. The core lies in the synergistic effect of the first TFT imaging layer, the copper filtration layer, and the second TFT imaging layer. The first TFT imaging layer is prepared using amorphous silicon (a-Si) material. Since it has a high X-ray absorption efficiency and charge conversion sensitivity in the energy range of 15 - 30 keV, it is suitable for capturing low-energy X-ray signals. At the same time, this layer is combined with cesium iodide (CsI:Tl) scintillator prepared by deposition. Its columnar crystal structure has a good lateral light confinement effect, which can improve the spatial resolution. The thickness of the scintillator is controlled between 300 and 500 microns to ensure sufficient response to low-energy rays without excessive attenuation. The second TFT imaging layer selects polycrystalline silicon (poly-Si) as the substrate material, which has higher electron mobility and current-carrying capacity compared to amorphous silicon and is suitable for working under high-energy conditions. Its matching scintillator material is gadolinium oxysulfide (Gd2O2S:Tb), which has good quantum absorption efficiency for X-rays in the energy range of 50 - 120 keV, and the thickness is set between 500 and 800 microns to enhance the detection ability for high-energy rays after penetrating deep structures. A copper filtration layer made of high-purity oxygen-free copper is inserted between the two TFT layers. This copper layer needs to be precisely processed, and the thickness is maintained in the range of 0.2 to 0.5 mm. Its linear attenuation characteristics enable it to effectively block X-ray components below 50 keV, allowing only high-energy rays to penetrate to the lower-layer detector, thereby constructing the difference in energy spectrum response between the upper and lower detection layers. To ensure the pixel points of the two-layer detector array are accurately aligned in space, a high-precision Z-axis alignment device is used for overlapping mounting, so that the two-layer TFT arrays are completely coincident in the X-Y plane position, and the alignment error is strictly controlled within 5% of the single-pixel size. The entire double-layer structure is fixed within a rigid frame made of carbon fiber composite material. On the one hand, this frame provides sufficient structural strength to avoid micro-vibration interference, and on the other hand, it reduces the system load through the lightweight characteristics of the material, making it suitable for integrated deployment in dynamic imaging or mobile devices. After completing the above structure, the constructed double-layer TFT detector is placed under the X-ray source. The X-ray source uses high-frequency micro-focus emission technology to ensure imaging sharpness and reduce scattering errors. The energy range of the X-ray beam covers 20 to 140 keV to meet the penetration imaging requirements of various materials.

[0015] 102. The X-rays penetrating the object to be detected form a first ordinary energy X-ray image on the first TFT imaging layer, and at the same time, a first high-energy X-ray image filtered by the copper filtration layer is formed on the second TFT imaging layer; Specifically, the working voltage and working current of the X-ray source are adjusted to control the energy spectrum and intensity distribution of the emitted X-ray beam. The working voltage of the X-ray source is set between 40 and 150 kV, and the current is controlled between 0.5 and 5 mA to ensure that the generated X-ray beam covers a broad spectrum range from low energy to high energy, meeting the requirements of penetration imaging for different materials. An aluminum filter is set in the X-ray source, with a thickness of 2 mm, to remove soft rays with energy lower than 15 keV, reduce noise components, improve signal purity, and protect the front-end components of the detector from unnecessary radiation loads of low-energy rays. After the adjustment is completed, the X-ray beam irradiates the object to be detected. During the penetration process, according to the different densities and atomic numbers of each region inside the object, the X-ray will undergo different degrees of attenuation. The higher the density and the larger the atomic number, the more significant the absorption of high-energy rays. At this time, the penetrated X-ray first reaches the first TFT imaging layer, which is most sensitive to rays with energy concentrated in the range of 15 - 50 keV. Therefore, it efficiently captures the attenuation characteristics of low-density and low-atomic-number materials (such as plastics, organic compounds, fibers, etc.) and converts them into the first ordinary energy X-ray image. The attenuation contours corresponding to these regions in the image are relatively soft, facilitating the subsequent material recognition model to perform refined feature extraction on lightweight materials. The X-ray that is not completely absorbed by the first TFT imaging layer continues to propagate downward. This part of the ray then passes through the copper filter layer set between the two layers. This filter layer utilizes the physical property that copper has a high attenuation coefficient for X-rays with energy lower than 50 keV to shield the remaining low-energy rays, so that the rays passing through this layer are mainly concentrated in the range of 50 - 140 keV, effectively increasing the average energy of the rays reaching the second TFT imaging layer and enabling this layer to focus on the detection of high-energy X-ray components. The second TFT imaging layer captures these high-energy transmitted rays, and the formed image is the first high-energy X-ray image. Since this image mainly reflects the attenuation response of high-density and high-atomic-number materials (such as metals, ceramics, composite claddings, etc.), it provides a key imaging basis for distinguishing high-density targets hidden inside low-density casings. To ensure that these two images are completely consistent in the time dimension, the first TFT imaging layer and the second TFT imaging layer are uniformly controlled by a high-speed synchronization control unit throughout the exposure process. The trigger logic must operate with a time resolution at the nanosecond level, and the two sets of detection data need to be synchronously read through a parallel data acquisition module. The sampling frequency is not less than 30 frames per second, and the data path bandwidth should be guaranteed to be above 3.2 Gbps to avoid frame loss or delay misalignment problems during high-speed data acquisition. Through the coordinated control of the structure and time mechanism, the first ordinary energy image and the first high-energy image have a strict one-to-one correspondence relationship.

[0016] 103. Preprocess the first ordinary energy X-ray image and the first high-energy X-ray image to obtain the second ordinary energy X-ray image and the second high-energy X-ray image; Specifically, adaptive median filtering is performed on the first ordinary energy X-ray image and the first high-energy X-ray image respectively to effectively remove the salt-and-pepper-like mutation noise introduced by detector instantaneous electrical noise, gamma interference, or unstable response of edge pixels. The window size of the adaptive median filter is dynamically adjusted according to the local noise variance of the image, and is set within the range of 3×3 to 7×7 to ensure that the image details are not overly blurred and the outlier noise points can be filtered out to the greatest extent, obtaining the first filtered image and the second filtered image respectively. The wavelet domain threshold denoising technology is used to perform a deeper Gaussian-like noise suppression process on the above filtered images. The five-layer DB4 wavelet decomposition structure is used to perform multi-scale transformation on the images, and soft threshold truncation is performed on the high-frequency noise components in the wavelet coefficient space through the BayesShrink threshold selection mechanism, and then the images are reconstructed to obtain the first denoised image and the second denoised image respectively. The non-linear response correction operation is performed on the denoised images to compensate for the gray-scale shift and quantization distortion caused by the non-linearity of the charge transfer function during the imaging process of the TFT detector. The correction process is based on a cubic polynomial fitting model, where the parameters a0 to a3 are obtained by performing least-squares fitting on the response values of the calibrated wedge-shaped aluminum block under different thickness conditions. Optimization coefficient tables are set for the ordinary energy image and the high-energy image respectively to ensure that the dynamic response curves of the two imaging channels are strictly consistent within the corresponding energy intervals, obtaining the first corrected image and the second corrected image. After completing the response correction, it enters the bad pixel identification and repair stage. The local variance analysis method is used to detect those abnormal pixel points that deviate from the mean by more than three standard deviations in the spatial neighborhood and mark them as bad point areas. Then, the bicubic spline interpolation algorithm is used to calculate the replacement value of each bad pixel by combining the surrounding 16 effective pixel values, constructing the first repaired image and the second repaired image with continuous structure and high gray-scale consistency. The first repaired image and the second repaired image are respectively subjected to geometric correction to eliminate the pixel-level geometric distortion caused by TFT manufacturing errors, packaging stress, or assembly micro-deviations. The geometric correction process depends on the pre-taken grid correction plate image, which is embedded with equally spaced metal marking points. By comparing the ideal coordinates of the marking points with the actual image coordinates, a two-dimensional transformation mapping function is established to remap the pixel positions of the entire image, generating the first geometric correction image and the second geometric correction image respectively. The first geometric correction image and the second geometric correction image are respectively subjected to spatial registration. Using the phase correlation method, Fourier transform is performed on the two images, the translation vector is extracted using the normalized cross-power spectrum, and sub-pixel resampling interpolation is performed on the original images to obtain the second ordinary energy X-ray image and the second high-energy X-ray image.

[0017] 104. Based on the second ordinary energy X-ray image and the second high-energy X-ray image, dual-energy domain feature extraction is performed to construct an effective atomic number map and a material density map; Specifically, structured feature extraction operations are performed on the images of the two energy channels respectively. In the second ordinary energy X-ray image, since it mainly reflects the attenuation behavior of low-energy X-rays in low-density and low-atomic-number materials, edge details and texture information of organic, plastic, and fiber materials are extracted through convolution operations and local contrast enhancement by the feature extraction network, forming a low-density material feature map. For the second high-energy X-ray image, because it focuses on the penetration characteristics in a higher energy range and is suitable for identifying high-density objects such as metals, ceramics, and composite claddings, a high-density material feature map is formed after deep convolution and activation function enhancement. These two feature maps together constitute the structural basis of the dual-energy information space. To achieve joint modeling of different energy responses, the second ordinary energy image and the second high-energy image are simultaneously input into an adaptive parameter forward neural network, which consists of a three-layer fully connected structure, with the hidden layer dimension uniformly set to 128, and ReLU is used as the non-linear activation function. The training objective of the network is to learn the optimal weighted parameters at each position in the energy spectrum based on multi-scale image statistical features, and output a parameter space distribution map. Each pixel position in this distribution map contains two adaptive weight coefficients α(x, y) and β(x, y), which are used to control the relative contribution degrees of the ordinary energy image and the high-energy image in weighted fusion respectively. Based on this parameter space distribution map, a pixel-by-pixel weighted difference operation is performed. The energy difference feature map formed by this difference processing can significantly enhance the regions with large response differences under the two energy imaging conditions, especially helping to distinguish composite structure regions with weak density or atomic number differences. After the above image enhancement, the low-density material feature map, the high-density material feature map, and the energy difference feature map are concatenated in the channel dimension, and a 1×1 convolution kernel is introduced for dimension compression and feature fusion to obtain a unified structure fusion feature map. Based on this fusion feature map, a regression model for estimating the effective atomic number Z and the material density ρ is constructed. The regression relationship is based on the attenuation model of X-rays for materials with different atomic numbers in different energy intervals, and a non-linear mapping function is constructed. This mapping relationship is realized by training an encoder-decoder neural network. The encoder extracts the spatial expression of the fusion features, and the decoder predicts the Z value and ρ value of the corresponding region pixel by pixel, thereby outputting an effective atomic number map and a material density map with the same spatial resolution as the input image, realizing a lossless estimation from image energy features to material physical parameters.

[0018] Perform channel number unification operations on feature maps from different sources to ensure dimensional consistency during subsequent fusion. Separately perform 1×1 convolutional operations on the low-density material feature map and the high-density material feature map to compress or expand their channel numbers to a preset unified channel dimension (e.g., 256 dimensions), obtaining the adjusted low-density material feature map and the adjusted high-density material feature map. Concatenate the adjusted low-density material feature map, the original energy difference feature map, and the adjusted high-density material feature map along the channel dimension to form a multi-source information joint feature map. This feature map simultaneously retains the texture structure of the low-density material, the energy spectrum profile of the high-density region, and the cross-energy response difference in the same tensor, effectively integrating the breadth and depth of material information. To extract key structural information from the joint feature map and dynamically enhance important regions, input the joint feature map into a feature enhancement module composed of a three-layer convolutional network. Each layer of convolution in this module uses receptive fields of 7×7, 5×5, and 3×3 respectively to cover information at different scales from global to local, and at the same time, insert spatial attention and channel attention mechanism units into the convolutional network. Among them, the spatial attention mechanism captures the pixel region in the image that contributes the most to the recognition result by calculating the activation distribution of the feature map in the spatial dimension, while the channel attention mechanism dynamically allocates convolutional kernel response resources by modeling the relative importance between channels of the feature map. After the two are fused, a highly fusion-aware attention map is formed. Perform an element-wise multiplication operation on the attention map and the adjusted low-density material feature map. The resulting weighted feature map emphasizes the key regions determined by the attention mechanism while retaining the structural features of the low-density material. Add this weighted feature map and the adjusted high-density material feature map element-wise to form a fused feature map. Input the fused feature map, the second ordinary energy X-ray image, and the second high-energy X-ray image into a deep neural network with an encoder-decoder architecture. The encoder part of this architecture consists of multiple residual convolutional blocks, which gradually extract high-level semantic features through downsampling and non-linear activation while retaining the spatial structure information of the image. After the encoding output, it is sent to two decoder branches respectively. The first decoder branch is responsible for generating an effective atomic number map. During the decoding process, upsampling and skip connection methods are used to gradually restore the spatial resolution and reconstruct the Z(x,y) distribution; the second decoder branch generates a material density map ρ(x,y). This branch introduces a density-specific normalization loss function to improve numerical stability and recognition sensitivity. Both branches use a per-pixel regression output method, and finally generate high-precision Z maps and ρ maps respectively.

[0019] Perform a preliminary transformation in the physical sense on the original image data, specifically including performing a logarithmic operation on the second ordinary energy X-ray image. According to the physical model of X-ray attenuation, perform logarithmic processing on the original image I to obtain an ordinary energy logarithmic attenuation map. Similarly, perform the same logarithmic operation on the second high-energy X-ray image to obtain a high-energy logarithmic attenuation map. Concatenate the above two logarithmic attenuation maps and the fusion feature map in the channel dimension to construct a multi-modal feature map, which simultaneously contains physical channel data processed based on the energy attenuation model and high-dimensional material structure expressions obtained through the forward extraction path. This multi-modal input is uniformly fed into a neural network with an encoder-decoder architecture. The encoder part consists of four stacked residual blocks. Each residual block contains two 3×3 convolutional layers with skip connections and realizes non-linear mapping through the ReLU activation function. This structure can effectively suppress the problem of gradient disappearance and improve the depth of feature extraction. In the encoder, a spatial downsampling operation is performed after each layer of the residual block. By continuously compressing the channel space and enhancing the semantic expression, a deep feature map is output, which condenses the structure boundary information, density distribution trend, and local material discrimination characteristics in the image. The deep feature map enters two independent decoder branches respectively for regression prediction. Among them, the first decoder branch is used to generate an effective atomic number map. This branch consists of four transposed convolutional modules, and the feature map is upsampled layer by layer to restore the original spatial resolution. After each transposed convolution, the ReLU activation function is connected to retain the positive numerical distribution characteristics, while the linear activation function is used in the last layer, so that the network output can be directly mapped to the true atomic number value range, such as continuous values between 1 and 92, to ensure that the effective atomic number map has accurate mathematical representation ability and meets the actual physical requirements. The second decoder branch is used to generate a material density map. Its network structure also consists of four transposed convolutional layers. The ReLU activation function is still used in the first three transposed convolutions to enhance the non-linear expression ability, while the Sigmoid activation function is used in the last layer to limit the output value between [0,1] for density normalization. Then, through a density value range mapping function, the normalized density is restored to the actual physical range to ensure that the output density value has both accurate proportional relationships and does not violate the physical upper and lower limit constraints, and a material density map is obtained.

[0020] 105. Based on the effective atomic number map and the material density map, perform feature space clustering analysis and decision tree classification, and output the final detection result.

[0021] Specifically, the effective atomic number value Z(x, y) corresponding to each pixel is combined with the density value ρ(x, y) to construct a two-dimensional Z-ρ joint feature space. In this space, the atomic number and density serve as two orthogonal axes. Due to the differences in atomic structure and physical density of different materials, distinct and highly discriminable cluster structures are formed in this joint space. After construction, adaptive threshold segmentation and clustering analysis operations are performed on this joint feature space. The adaptive threshold segmentation determines the boundary point between Z and ρ dynamically based on the overall statistical characteristics of the image and guides the distribution of initial clustering points. The clustering analysis preferably uses an improved density-based clustering algorithm (such as DBSCAN), which dynamically adapts to different distribution patterns by setting the neighborhood radius ε and the minimum number of core samples, to identify multiple preliminary classification clusters such as organic matter, light metals, heavy metals, and mixed materials, obtaining a preliminary clustering result. Each clustering region corresponds to a candidate category of substances with similar physical properties. The preliminary clustering result is jointly modeled with the energy difference feature map E(x, y) to construct a composite feature vector containing three-dimensional attributes of Z, ρ, and E. This three-dimensional feature vector reflects the static physical distribution of substances in terms of structural attributes (Z and ρ), and at the same time introduces the dynamic energy spectrum difference E extracted from the dual-energy response, enabling it to more sensitively capture the boundaries of composite materials and the nested structural relationships. On this basis, the constructed three-dimensional feature vector is input into a trained decision tree classifier for per-pixel classification. The classifier adopts a multi-way tree structure with a maximum depth of 5 and uses the Gini coefficient as the node splitting criterion to generate a material label map containing the category information of each pixel. This label map annotates the specific classification numbers of all material regions in the image, with high recognition rate and boundary retention ability, and can effectively identify fine-grained material changes and complex wrapping structures. After obtaining the material label map, combined with the second ordinary energy X-ray image and the second high-energy X-ray image, the final result generation and visualization construction of the detection area are carried out. The pseudo-color fusion display technology is used to perform color channel mapping on the dual-energy image: the energy difference feature map is mapped to the hue channel H of the HSV color space, the high-energy image is mapped to the saturation channel S, and the ordinary energy image is mapped to the value channel V to generate a color fusion image with both structural clarity and energy sensitivity. A color marker map is generated based on the material label map, with different colors assigned to different material categories to intuitively display the spatial distribution state of various materials in the measured area. Combining the pixel depth information and label distribution of the dual-energy image, three-dimensional volume reconstruction is performed through an improved algebraic reconstruction algorithm to obtain a three-dimensional voxel structure with the same resolution as the original image, and a three-dimensional model image that can be rotated and scaled is generated for internal observation of complex structures. All independent object regions are identified based on the connected component analysis algorithm, and an identification result table containing each object ID is generated.The spatial position of each object (by calculating the centroid of the region), the boundary dimensions (based on the minimum bounding rectangle), the main axis direction (the rotation angle obtained by principal component analysis), the material composition (statistically calculating the pixel proportion of each category in the label map), and the hazard level calculated based on the cosine similarity between the material combination vector and the hazardous material feature library (marked as high risk if the similarity exceeds 0.85) are listed in the table.

[0022] Extract structured feature information at the pixel level. Based on the preliminary clustering results obtained from the Z-ρ joint feature space clustering analysis in the early stage, in the effective pixel region identified by this result, the effective atomic number Z(x,y) and the material density ρ(x,y) corresponding to each pixel position are extracted respectively, and a two-dimensional physical property vector is formed. This vector accurately expresses the static physical composition of the material at this position at each pixel. To enhance the connection between this physical expression and the material response, a dynamic energy spectrum difference index is introduced, that is, the energy difference feature map E(x,y) is normalized. The linear normalization strategy is used to uniformly map the E value to the [0,1] interval to eliminate the scale inconsistency problem caused by the fluctuation of the E value range in different regions and retain its relative distribution trend. After the normalization process, the two-dimensional feature vector (Z,ρ) at each pixel position is merged with the normalized energy difference value E at its corresponding position to form a three-dimensional feature vector (Z,ρ,E). This three-element combination can fully reflect the material intrinsic characteristics of this position in the dimension of static attributes (such as atomic number and density), and incorporates the dynamic attenuation behavior of the response difference between the high and low energy channels at this point, thus possessing the ability to identify the same material across energy spectra and multiple levels. Organize the three-dimensional feature vectors of all effective pixels in the image into a structured input sample set and uniformly input it into the trained decision tree classifier. The decision tree classifier adopts a multi-layer splitting mechanism and the minimum Gini coefficient criterion in its design. According to the feature distribution law in the three-dimensional feature space, the samples are divided into predefined material category regions. Each decision node of the classifier corresponds to a set feature threshold in a certain dimension to determine whether the sample meets the condition to enter this path. Finally, the sample is mapped to a leaf node under the tree structure, and this node represents a specific material category identification number. After the three-dimensional feature vector of each pixel is input into the classification tree, it is judged step by step according to its feature splitting path in the tree structure, and finally a material label is assigned. After completing the entire pixel-level feature classification, remap it back to the original image space according to the pixel position to construct a material label map. This label map has the same resolution as the original X-ray image in space and assigns an accurate material identification number to each pixel in terms of semantics.

[0023] In the embodiments of the present invention, a double-layer TFT detector structure is adopted and a copper filtering layer is arranged between the two layers of TFTs, realizing the function of simultaneously obtaining a normal energy image and a high-energy image by a single X-ray exposure, without switching the energy of the X-ray source for multiple exposures, fundamentally avoiding the image registration problem in the traditional technology. Special feature extraction paths are respectively designed for low-density materials and high-density materials, and the density difference features of different materials are highlighted through an energy difference enhancement module, improving the accuracy of material identification. The image information in different energy domains is effectively integrated, enhancing the identification ability for mixed materials and complex structures, especially the discrimination ability for materials with similar densities but different atomic numbers. The direct material attenuation coefficient reconstruction algorithm of the present invention utilizes the spatial registration advantage of the double-layer TFT detector to directly establish an accurate mapping of the effective atomic number and material density, without complex image registration preprocessing, greatly reducing the computational complexity. This method simultaneously obtains dual-energy image information by a single X-ray exposure, significantly reducing the radiation dose, while extending the service life of the TFT detector and reducing the system operation cost.

[0024] In a specific embodiment, the process of executing step 101 may specifically include the following steps: The first TFT imaging layer is prepared by using amorphous silicon material, and the sensitivity range of the first TFT imaging layer is 15 - 30 keV. The second TFT imaging layer is prepared by using polycrystalline silicon material, and the sensitivity range of the second TFT imaging layer is 50 - 120 keV; The cesium iodide material is deposited to obtain the scintillator of the first TFT imaging layer, and the gadolinium oxysulfide material is deposited to obtain the scintillator of the second TFT imaging layer; The copper filtering layer is processed by using oxygen-free copper material, and the first TFT imaging layer and the second TFT imaging layer are overlapped and mounted in the Z-axis direction through an alignment device, so that the pixel points of the first TFT imaging layer and the second TFT imaging layer correspond one by one in the spatial position. The copper filtering layer is placed between the first TFT imaging layer and the second TFT imaging layer to obtain a double-layer TFT detector structure; Through the double-layer TFT detector structure, an X-ray beam is emitted from the X-ray source to irradiate the object to be detected for penetration imaging.

[0025] Specifically, the first TFT imaging layer is prepared using amorphous silicon material. Its structure is based on a-Si:H technology, featuring high uniformity, mature processes, and low costs. It is suitable for X-ray imaging scenarios where the response energy range is between 15 and 30 keV. In this energy range, the X-ray penetration ability is limited, making it suitable for detecting low-density and low-atomic-number materials such as organic substances, plastics, or carbon-based composite structures. When designing the first TFT imaging layer, the charge collection efficiency for low-energy rays needs to be enhanced. The capacitance per unit area and the photoelectric conversion efficiency are optimized to ensure that the dynamic response range of the image gray scale covers the signal differences caused by subtle energy changes. To match the photoelectric response ability of the first TFT imaging layer, a cesium iodide (CsI:Tl) scintillator material with a thickness controlled between 300 and 500 microns is deposited on its surface. This type of material is selected because it has a highly columnar crystal arrangement structure, which can achieve a highly directional fluorescence channel, efficiently convert X-ray energy into visible light signals that can be read by TFT pixels, and its emission wavelength is near 550 nm, highly overlapping with the response spectrum of amorphous silicon photodiodes, with high conversion efficiency and strong energy utilization. The deposition process uses a combination of low-temperature vacuum evaporation and auxiliary heat treatment to control the crystal orientation and density uniformity, and optimize the thallium doping concentration during deposition to increase the emission brightness and response speed, ensuring stable performance output under high frame rate imaging conditions. On this basis, to establish dual-energy separation imaging ability, a second TFT imaging layer is prepared. This layer uses polycrystalline silicon (poly-Si) technology as the base material, which has a higher electron mobility and a smaller switching response time compared to amorphous silicon, and is suitable for responding to the 50 to 120 keV X-ray range with higher energy and stronger penetration. The polycrystalline silicon structure can withstand higher-frequency data sampling and signal processing, which is particularly crucial for image acquisition in high-speed penetration scenarios. At the same time, gadolinium oxysulfide (Gd2O2S:Tb) is deposited on the surface of this imaging layer as its dedicated scintillator material. This material has a high absorption rate for high-energy X-rays and the characteristic that its emission spectrum band highly matches the sensitive area of polycrystalline silicon. The deposition thickness is controlled between 500 and 800 microns to enhance the absorption depth in the high-energy segment, and the thermal plasma-assisted deposition technology is used to improve its crystallization density and interlayer adhesion strength, thereby ensuring stable imaging performance under high irradiation loads. A copper filter layer is inserted between the two TFT imagers to construct an energy shielding interface. This copper layer is processed using high-purity oxygen-free copper material, with its thickness controlled between 0.2 and 0.5 millimeters, and the surface flatness error needs to be strictly controlled within ±5 microns. The linear attenuation coefficient of copper material for X-rays in the energy region below 50 keV is significantly greater than that in the high-energy region. Therefore, it can effectively absorb the remaining low-energy rays after passing through the first TFT layer, preventing them from entering the second TFT layer and generating secondary responses, thereby strengthening the selectivity of the second TFT imaging layer for high-energy X-rays and increasing the energy difference and decoupling ability between the images of the two layers of imaging.The production process of the copper filter layer requires the use of high-precision CNC machining and electrolytic polishing processes to ensure its overall smoothness and perpendicularity, so as to avoid the light spot scattering effect caused by uneven edges of the filter layer during the penetration imaging process, which affects the clarity of the image boundary. After the preparation of the three-layer structure is completed, the first TFT imaging layer, the copper filter layer, and the second TFT imaging layer are precisely laminated and mounted vertically through a nano-level Z-axis alignment device. This mounting process is achieved by the cooperation of a bilateral vacuum adsorption and a vision recognition guidance system. By matching and calibrating the pixel array marking points of the two-layer TFT arrays, the pixel position error between the upper and lower layers is controlled within 5% of the single pixel size, ensuring a corresponding relationship during the final image synthesis and feature fusion processes, and avoiding misalignment, ghosting, or geometric distortion. The entire three-layer structure is installed in a rigid frame made of carbon fiber reinforced composite materials. This frame also has the structural properties of being lightweight, high-strength, radiation-resistant, and having good thermal stability, and can maintain the stable adhesion between layers during long-term operation, preventing structural displacement caused by thermal expansion and contraction or mechanical vibration. After the construction of the above detector structure is completed, it is integrated into the X-ray penetration imaging system, and the detector module is geometrically aligned with the high-frequency micro-focus X-ray source. The X-ray source is set with an energy range of 20 to 140 keV, and an X-ray beam with a specific energy spectrum form is generated through an adjustable voltage and current and a pulse controller. The soft rays are pre-filtered through a 2 mm aluminum filter to ensure that the incident rays cover the sensitive energy regions of the first and second TFT imaging layers. During the actual imaging process, the object to be detected is placed between the X-ray beam and the detector. Under the irradiation of the X-ray beam, its internal structure selectively attenuates the X-rays according to different densities and atomic numbers. The penetrated rays form the first image and the second image respectively in the double-layer TFT structure, reflecting the attenuation characteristics of different materials under low-energy and high-energy conditions.

[0026] In a specific embodiment, the process of performing step 102 may specifically include the following steps: Adjust the working voltage and working current of the X-ray source to generate an X-ray beam; Capture the X-ray beam passing through the object to be detected on the first TFT imaging layer to obtain a first ordinary energy X-ray image, and the first ordinary energy X-ray image corresponds to the characteristic information of low-density and low-atomic-number materials; Filter the energy of the X-ray beam passing through the first TFT imaging layer through the copper filter layer to obtain a transmitted ray; Capture the transmitted ray through the second TFT imaging layer to obtain a first high-energy X-ray image. At the same time, the first TFT imaging layer and the second TFT imaging layer are synchronously triggered and data is collected through the synchronous control unit to ensure that the first ordinary energy X-ray image and the first high-energy X-ray image are completely synchronous in the time dimension.

[0027] Specifically, the working voltage and working current of the X-ray source are adjusted to generate a continuous energy spectrum X-ray beam that meets the requirements of multi-energy imaging. The X-ray source has a high-frequency excitation and micro-focus structure, with its working voltage adjustment range controlled between 40 and 150 kV, and the current adjustment range set at 0.5 to 5 mA, enabling it to flexibly match the penetration imaging requirements for high- and low-density materials and taking into account the radiation response characteristics of objects with different sizes and structural complexities. In actual imaging tasks, according to the structural complexity, material type, and thickness distribution of the object to be detected, the input parameters of the X-ray source are dynamically adjusted so that its output power density can penetrate the high-density core area without over-saturating the low-density edge area, thus forming a good contrast range in the image gray space. At the same time, to eliminate soft-ray interference and improve the image signal-to-noise ratio, a 2-mm aluminum filter is set at the front end of the X-ray source to pre-filter soft X-rays with energies below 15 keV to avoid unnecessary responses of the front-layer detector and image blurring caused by these low-energy components that are easily absorbed and do not participate in imaging. When the X-ray beam with a high energy spectrum width is successfully generated, it irradiates the object to be detected placed at the center of the imaging system. Due to the large differences in atomic number and density among different materials, their X-ray attenuation capabilities also show significant differences, resulting in changes in the energy spectrum of the X-ray beam during penetration. The rays passing through the object first reach the first TFT imaging layer located above the system. This layer is constructed of amorphous silicon material and combined with a cesium iodide (CsI:Tl) scintillator with a thickness of 300 to 500 microns, which has a high-sensitivity response to X-rays in the energy range of 15 to 50 keV. In this energy region, low-density and low-atomic-number materials such as organic substances, plastics, and fiber fabrics exhibit relatively small absorption coefficients, so that the low-energy rays still have a certain amount of energy for the imaging system to capture after transmission. The image recorded by the first TFT imaging layer under these conditions is the first ordinary energy X-ray image, which retains a large amount of information about the low-density structure contours, edge details, and texture levels, especially showing a higher signal contrast when detecting thin-walled claddings, flexible material coverings, or internal cavity structures. At the same time, the rays not absorbed by the first TFT imaging layer will continue to penetrate downward and pass through the copper filter layer sandwiched between the two imaging layers. This filter layer is made of high-purity oxygen-free copper material with a thickness between 0.2 and 0.5 mm and has an extremely high linear attenuation coefficient in the low-energy X-ray segment. The addition of the copper layer can effectively shield the remaining soft-ray components below 50 keV and only allow X-rays with energies higher than 50 keV to pass through, thus achieving energy-selective filtering of the penetration spectrum. The transmitted rays passing through this copper filter layer have an energy spectrum composition that is relatively more concentrated in the range of 50 to 120 keV. The X-rays in this energy region have stronger penetration capabilities and can deeply detect high-density and high-atomic-number materials such as metals, ceramics, minerals, and lead-containing components, which show more intense attenuation contrast characteristics in imaging.These high-energy filtered transmitted rays are efficiently received after entering the second TFT imaging layer. The second TFT layer adopts a polysilicon structure and is combined with a gadolinium oxysulfide (Gd2O2S:Tb) scintillator, with a thickness of 500 to 800 microns. It exhibits extremely high quantum efficiency and light output ability within the high-energy response range, and can convert high-energy rays into optical signals that can be read by the TFT, forming the first high-energy X-ray image. The information recorded in this image is mainly related to the position, thickness, and edge contour of high-density materials. The entire double-layer TFT detector structure is connected to a high-speed synchronous control unit. Based on a nanosecond-level high-precision clock control mechanism, this unit sends synchronous trigger signals to the two imaging layers respectively before the start of each frame exposure, and conducts unified timing coordination for their sampling processes, gain responses, analog-to-digital conversion, and data output links. The synchronous control module uses a combination of a phase-locked loop and a digital clock manager to ensure precise synchronization of the sampling timing, with a maximum error not exceeding 1 nanosecond; at the same time, a 16-bit high-precision ADC module is used to parallel-sample the two-channel image signals, with a sampling frequency of up to 60 MHz, and the data is transmitted to the backend image processing module through a high-speed data bus. Due to the complete synchronization of the dual-channel data acquisition, it avoids the problems of time inconsistency and spatial registration errors caused by multiple exposures or X-ray source energy switching in traditional dual-energy image acquisition methods, and is suitable for real-time material identification in high-speed dynamic scenarios.

[0028] In a specific embodiment, the process of executing step 103 may specifically include the following steps: Perform adaptive median filtering on the first ordinary-energy X-ray image and the first high-energy X-ray image respectively to obtain a first filtered image and a second filtered image; Perform wavelet-domain threshold denoising processing on the first filtered image and the second filtered image respectively to obtain a first denoised image and a second denoised image; Perform TFT non-linear response correction on the first denoised image and the second denoised image respectively to obtain a first corrected image and a second corrected image; Perform bad pixel detection and repair on the first corrected image and the second corrected image respectively to obtain a first repaired image and a second repaired image; Perform geometric correction on the first repaired image and the second repaired image respectively to obtain a first geometrically corrected image and a second geometrically corrected image; Perform spatial registration on the first geometrically corrected image and the second geometrically corrected image respectively to obtain a second ordinary-energy X-ray image and a second high-energy X-ray image.

[0029] Specifically, the first ordinary energy X-ray image and the first high energy X-ray image are subjected to adaptive median filtering respectively to effectively eliminate the salt-and-pepper isolated abnormal points caused by scintillator lattice defects, TFT switching noise or transient electrical interference during the imaging process. The median filter dynamically adjusts its window size according to the noise variance of the local area of the image, and the window size varies between 3×3 and 7×7. A larger window is used in the low texture area to improve the noise reduction efficiency, while a smaller window is retained in the edge and high frequency area to reduce the boundary fuzziness. This processing obtains the first filtered image and the second filtered image respectively. The first filtered image and the second filtered image are subjected to wavelet domain threshold denoising. This method is based on the multi-scale wavelet decomposition theory, expands the image signal in the space-frequency joint domain, and decomposes the image into low-frequency approximation coefficients and multiple groups of high-frequency detail coefficients through a 5-level discrete wavelet transform (Daubechies4 wavelet basis is selected). For these high-frequency coefficients, the BayesShrink adaptive soft threshold strategy is used to adjust their amplitudes, suppressing the random Gaussian white noise components while retaining the contour edges. Then, the first and second denoised images are generated by wavelet reconstruction, which are smooth, continuous and clearly structured, so that the images achieve a better balance between noise suppression effect and edge protection. The two denoised images are corrected for TFT nonlinear response. Since the TFT detector has nonlinear charge conversion characteristics in the process of photoelectric signal transmission, the image grayscale value cannot linearly correspond to the incident X-ray energy. Therefore, the image pixels are corrected using a cubic polynomial model, in which the order coefficients a0 to a3 are obtained by fitting the response curve of the stepped wedge-shaped standard aluminum block under multiple thickness conditions. The first ordinary energy image and the first high energy image are different in the sensitive energy region, so two nonlinear response models are established respectively to obtain the first corrected image and the second corrected image. During the operation of the detector, some pixels fail or continuously output abnormal values due to scintillator ablation, pixel node penetration or reading circuit failure, so bad pixel detection and repair are performed on the corrected images respectively. The anomaly detection algorithm based on local statistical variance is used to calculate the mean and standard deviation of each pixel and its 8 neighbors. If a pixel value deviates from the mean by more than 3 times the standard deviation, it is marked as a bad pixel. Then, the bicubic spline interpolation algorithm is used to select 16 valid pixels around each bad pixel, and the local polynomial function is established through spatial interpolation to restore it, and the first repaired image and the second repaired image are obtained. Since the TFT detector has a small arrangement error or array warping problem in the actual manufacturing process, the image shows slight geometric distortion, so the repaired image is accurately geometrically corrected. The correction method uses a grid calibration plate, which provides high-contrast metal markers with a fixed geometric arrangement in the imaging field of view. By comparing the actual pixel coordinates of the markers in the image with the theoretical reference coordinates, a two-dimensional spatial mapping function is fitted, and the entire image is subjected to pixel-level coordinate remapping processing to obtain the first geometrically corrected image and the second geometrically corrected image, respectively.Perform spatial registration on the first geometrically corrected image and the second geometrically corrected image respectively. This operation adopts a Fourier domain registration method based on phase correlation, performs Fourier transforms on the first geometrically corrected image and the second geometrically corrected image respectively, constructs a normalized cross-power spectrum, calculates the translational offset in the spatial domain of the image by finding the frequency domain offset vector corresponding to the maximum peak of the cross-spectrum, and completes image alignment through sub-pixel resampling interpolation. Finally, the fully registered second ordinary energy X-ray image and the second high-energy X-ray image are obtained respectively.

[0030] In a specific embodiment, the process of performing step 104 may specifically include the following steps: Extract features from the second ordinary energy X-ray image to obtain a low-density material feature map; Extract features from the second high-energy X-ray image to obtain a high-density material feature map; Input the second ordinary energy X-ray image and the second high-energy X-ray image into an adaptive parameter forward neural network for parameter space distribution fusion, and output a parameter space distribution map. The adaptive parameter forward neural network includes 3 fully connected layers, and the hidden layer dimension is 128; Perform weighted difference calculation on the second ordinary energy X-ray image and the second high-energy X-ray image based on the parameter space distribution map to obtain an energy difference feature map; Perform feature fusion on the low-density material feature map, the high-density material feature map, and the energy difference feature map to obtain a fused feature map, and construct an effective atomic number map and a material density map based on the fused feature map.

[0031] Specifically, for the feature extraction process of the second ordinary energy X-ray image in the low-density direction, since this image mainly reflects the X-ray transmission response with the energy range concentrated between 15 and 50 keV, it has higher signal sensitivity to low atomic number and low-density materials such as organic matter, polymers, lightweight packaging materials, and carbon-based fibers. In order to capture the fine-grained features of these structures, a multi-scale convolutional neural network is used to perform hierarchical feature extraction on this image, emphasizing the preservation of edge details in the shallow structure and enhancing the expression of local texture and continuity information in the deep structure, and outputting a low-density material feature map. This map retains the structural boundaries, uniformity, and spatial distribution characteristics of lightweight materials and is an important data representation reflecting lightweight foreign objects, thin-layer materials, and cavity contours. At the same time, for the second high-energy X-ray image, the feature extraction work in the high-density direction is carried out. The image records the response formed by high-energy X-rays with strong penetrability after passing through the copper filter layer in the object under test. Its main energy range is 50 to 120 keV, so it has significant absorption difference characteristics for high-density or high atomic number objects such as steel, aluminum, ceramics, metal cores, conductive traces, and composite laminates. A residual connection mechanism is adopted in the feature extraction network to construct a deep feature extraction structure, which can prevent gradient disappearance while enhancing the network's expression ability, enabling it to have the ability to identify complex material morphologies. Especially when identifying metal foreign objects, multi-layer shields, and embedded devices, it shows high-resolution performance. The extraction result is a high-density material feature map, which is used to characterize the strong absorption regions of heavy objects, local high atomic number aggregation regions, and the core features of structural entities. To effectively fuse these two energy spectrum information and enhance their resolution for mixed material regions, the second ordinary energy image and the second high-energy image are input into an adaptive parameter forward neural network. This neural network contains a three-layer fully connected structure, and the hidden layer dimension is uniformly set to 128, and a non-linear activation function (such as ReLU) is used to enhance the expression ability of parameter learning. The input of the network is the gray values and their local statistical features at each pixel position from the two images, and the output of the network is a pair of position-corresponding adaptive fusion parameters α(x, y) and β(x, y), which respectively represent the weight distributions of the ordinary energy image and the high-energy image in the fusion process at this position. The training objective of the network is to learn the optimal fitting intervals of Z values and ρ values under different energy spectrum inputs in a preset material response database, so as to output a parameter space distribution map α(x, y), β(x, y) covering the entire image space. This map effectively reflects the differential weights of local material energy spectrum features and helps to enhance the structural difference signals in complex object regions. Based on the aforementioned obtained parameter space distribution map, a weighted difference operation of the dual-channel image is performed, that is, the second ordinary energy image and the second high-energy image are fused at each pixel point position.The weighted difference strategy can dynamically capture the dominant information sources of each energy spectrum image in different spatial regions, form an enhanced response to mixed materials or intermediate density junction regions, and generate an energy difference feature map. This map visually shows characteristic textures with significantly enhanced local contrast, which can highlight the variation rules caused by energy spectrum attenuation between materials and is an important basis for material demarcation, abnormal point recognition, and structural discontinuity detection. The low-density material feature map, high-density material feature map, and energy difference feature map are concatenated along the channel dimension to form a multi-channel fusion input map, and then a structured deep fusion module is used to perform feature encoding on it. This fusion module includes a group of parallel convolutional paths to process inputs of different scales respectively, adds a spatial attention mechanism in the middle to guide the network to focus on important structural regions, and adds a channel attention mechanism to weight-adjust the multi-source feature dimensions, and outputs a fusion feature map. The fusion feature map is input into a deep prediction network composed of an encoder-decoder architecture. The encoder consists of multiple residual blocks to extract compressed semantic features. The decoder consists of two branches. The first branch gradually restores the spatial size through transposed convolution to output the effective atomic number value Z(x, y) of each pixel, and the second branch restores the material density value ρ(x, y) under a similar structure. The output results respectively form an effective atomic number map and a material density map, which show the atomic composition information and mass density structure of all regions in the detected object in a high-resolution and physically meaningful way.

[0032] In a specific embodiment, the process of performing steps to perform feature fusion on the low-density material feature map, high-density material feature map, and energy difference feature map to obtain a fusion feature map, and constructing an effective atomic number map and a material density map based on the fusion feature map may specifically include the following steps: Adjust the number of channels of the low-density material feature map and the high-density material feature map to obtain an adjusted low-density material feature map and an adjusted high-density material feature map; Concatenate the adjusted low-density material feature map, energy difference feature map, and adjusted high-density material feature map along the channel dimension to obtain a joint feature map; Process the joint feature map through a three-layer convolutional network, and at the same time use spatial attention and channel attention units to output an attention map; Multiply the attention map by the adjusted low-density material feature map to obtain a weighted feature map, and add the weighted feature map to the adjusted high-density material feature map to obtain a fusion feature map; Input the fusion feature map, the second ordinary energy X-ray image, and the second high-energy X-ray image into a deep neural network with an encoder-decoder architecture. After extracting features through the encoder part, use the first decoder branch to generate an effective atomic number map, and at the same time use the second decoder branch to generate a material density map.

[0033] Specifically, the number of channels of the low-density material feature map and the high-density material feature map is adjusted. This adjustment process uses a 1×1 convolution operation to perform a linear transformation on the original feature map, unifying its feature channels while keeping the spatial resolution of the image unchanged. For example, the two feature maps are respectively compressed or expanded to a unified dimension (such as 256 dimensions), so as to eliminate the structural asymmetry problem during the channel dimension splicing process, and at the same time achieve the effects of feature compression, redundancy removal, and important pattern enhancement. The adjusted low-density material feature map, the energy difference feature map obtained through the aforementioned weighted difference calculation, and the adjusted high-density material feature map are spliced in the channel dimension to form a joint feature map, which contains three energy dimension features obtained from three paths in the tensor structure. The low-density material feature map reflects the contour and texture of the lightweight material structure, the high-density material feature map strengthens the response patterns of high-attenuation materials such as metals and heavy elements, and the energy difference feature map forms a discriminative material boundary and mixed response signal through energy spectrum cross-enhancement. In order to extract the key regions in the joint feature map and perform dynamic weighted adjustment on the information in different spatial and channel dimensions, an attention-guided feature extraction module containing three layers of convolution is constructed. This module embeds a spatial attention unit and a channel attention unit in the backbone structure, modeling the importance of image information from the two-dimensional plane dimension and the channel distribution dimension respectively. The spatial attention part is spliced after average pooling and max pooling of the joint feature map, and a two-dimensional attention heat map is generated through convolution operations to enhance local structure contrast, boundary region response, and texture distribution perception; the channel attention mechanism calculates the response importance of each channel through global pooling and generates a channel weighting vector through a fully connected network, enabling the model to allocate different degrees of learning weights between different channels, thereby highlighting the dimensions with high semantic importance. The output of the attention mechanism is an attention map with the same size as the low-density material feature map. This attention map is multiplied pixel by pixel with the low-density material feature map after channel adjustment to form a weighted feature map. In this process, the model assigns more weights to those image regions that are highly sensitive to material discrimination in terms of spatial position and feature dimension, and applies an attenuation factor to the redundant parts to achieve the synergistic goal of structure highlighting and noise suppression. The weighted low-density feature map and the adjusted high-density feature map are subjected to an element-wise addition operation in the channel to complete a cross-energy-level feature fusion and generate a fused feature map. In order to recover the material parameter map with physical interpretation significance from the fused feature map, the fused feature map, the second ordinary energy X-ray image, and the second high-energy X-ray image are input into a deep neural network with an encoder-decoder architecture. This network uses a symmetric structure of U-Net or ResNet-UNet hybrid form as the basic framework, where the encoder part is composed of multiple residual blocks stacked together. Each residual block contains two layers of convolution and a ReLU activation function, and is combined with a skip connection mechanism to retain the shallow detail information.During the encoding stage, the fused feature map and two energy maps undergo multiple downsamplings and semantic abstractions, gradually compressing the spatial dimension layer by layer and extracting deep structural features, finally outputting a deep feature map with rich semantic information and compressed representation. This deep feature map is respectively input into two independent branches during the decoding stage. The first decoder branch consists of four transposed convolutional modules, and each layer is combined with the ReLU activation function to maintain the continuity of the numerical forward propagation. The linear activation function is used in the last layer to adapt to the output range of the continuous atomic number Z, and the output result is the effective atomic number map Z(x, y). This map represents the equivalent atomic number of each pixel's corresponding area, indirectly mapped through the X-ray attenuation coefficient, reflecting the basic atomic composition characteristics of the material. The structure of the second decoder branch is the same as that of the first branch, but its output target is the material density map ρ(x, y). The Sigmoid activation function is used in the last layer of this branch to limit the output within the range of [0, 1], and then it is multiplied by a density range scaling factor to map the network output value to the actual physical density interval, thus meeting the real constraints of material characterization.

[0034] In a specific embodiment, the process of performing the steps of inputting the fused feature map, the second ordinary energy X-ray image, and the second high-energy X-ray image into the deep neural network of the encoder-decoder architecture, extracting features through the encoder part, and generating the effective atomic number map using the first decoder branch and simultaneously generating the material density map using the second decoder branch may specifically include the following steps: Perform logarithmic operation processing on the second ordinary energy X-ray image to obtain the ordinary energy logarithmic attenuation map, and perform logarithmic operation processing on the second high-energy X-ray image to obtain the high-energy logarithmic attenuation map; Perform channel dimension concatenation on the ordinary energy logarithmic attenuation map, the high-energy logarithmic attenuation map, and the fused feature map to obtain a multi-modal feature map, and input the multi-modal feature map into the deep neural network of the encoder-decoder architecture, and perform feature extraction through the encoder containing 4 residual blocks to obtain a deep feature map; Process the deep feature map through the first decoder branch. The first decoder branch consists of 4 transposed convolutional layers and uses the ReLU activation function, and the linear activation function is used in the last layer to adapt to the numerical range of the atomic number to obtain the effective atomic number map; Process the deep feature map through the second decoder branch. The second decoder branch consists of 4 transposed convolutional layers and uses the Sigmoid activation function, and the density value range mapping is performed in the last layer to conform to the physical constraints to obtain the material density map.

[0035] Specifically, a mathematical transformation for physical model consistency is performed on the input image, that is, a logarithmic operation is carried out, so as to be transformed into a linear expression form based on the absorption model. Considering that during the process of X-ray penetrating an object, the relationship between its intensity I and the initial intensity I0 follows the Beer-Lambert law, that is , after taking the logarithm, we have , where is the attenuation coefficient, This is the material thickness or density path. To conform to the calculation form of this model, a per-pixel logarithmic operation is performed on the second ordinary energy X-ray image to obtain an ordinary energy logarithmic attenuation map, which presents a numerical distribution linearly related to the material attenuation ability at the gray level; similarly, the same logarithmic transformation is also performed on the second high-energy X-ray image to obtain a high-energy logarithmic attenuation map. The above two logarithmic attenuation maps are concatenated in the channel dimension with the fusion feature map constructed previously to obtain a multi-modal feature map containing three information sources. In this three-dimensional tensor structure, the ordinary energy logarithmic map provides the response characteristics of low-density and low-atomic number regions, the high-energy logarithmic map provides the deep response characteristics of high-density and high-Z value regions, and the fusion feature map supplements the non-linear difference relationship and structural fusion representation between low and high energy spectra. The combination of the three constitutes a high-dimensional information spectrum that is complementary at three levels: physical model, energy spectrum structure, and spatial perception. The multi-modal feature map is input into a deep neural network with an encoder-decoder architecture. The encoder part is composed of 4 residual blocks stacked together. Each residual block contains two 3×3 convolutional layers, a batch normalization module, and a ReLU activation unit inside, and a skip connection is established between the input and output to alleviate gradient disappearance and improve feature stability. After passing through each residual block, the spatial resolution of the feature map is downsampled once (e.g., through convolution or pooling with a stride of 2), and at the same time, the number of channels doubles to enhance the network's perception ability. Through four consecutive residual structures, the original input feature map is gradually compressed into a deep feature map with stronger expression ability and richer semantic information. This deep feature map fuses energy spectrum information, logarithmic attenuation model constraints, and fusion expression levels in the tensor structure and has the comprehensive modeling ability from edges, textures to material responses in the semantic dimension. The deep feature map is respectively input into two independent decoder branches, which are used to reconstruct the effective atomic number map and the material density map. In the first decoder branch, a four-layer transposed convolution module is used to gradually restore the image size. Each layer of transposed convolution upsamples the output of the previous layer to a higher spatial dimension and ensures the continuity of the forward value range and the stable propagation of the gradient through the ReLU activation function; the output of the last layer uses a linear activation function to ensure that the predicted value of each pixel in the generated image is a continuous real number, which can accurately correspond to the effective atomic number Z(x,y) of the material, for example, the value range is between 1 and 92, covering all actually existing elements. This predicted map is consistent with the input image in terms of spatial structure and can perform pixel-level reconstruction of the Z value changes on the material surface, edges, interlayer regions, and embedded foreign objects, with high physical interpretability and map accuracy. In the second decoder branch, the structure is basically the same as that of the first branch, but the activation function of the last layer uses the Sigmoid function to map the output to the [0,1] interval. This normalized output is then multiplied by a preset maximum density constant to obtain the material density map ρ(x,y) that conforms to the physical dimension constraints.This density map reflects the volumetric density change characteristics of materials in the detected area at the pixel level, and can accurately distinguish the spatial distribution of materials with different density levels such as air cavities, foam structures, plastic materials, and metal skeletons. During the decoding process, by introducing a skip connection mechanism (concatenating the feature maps at each level in the encoder with the corresponding decoding layer), the structural details and edge information of the shallow-layer images are effectively retained, avoiding problems such as blurring, edge jitter, or disappearance of small structures during the decoding process.

[0036] In a specific embodiment, the process of performing step 105 may specifically include the following steps: Combine the effective atomic number map and the material density map to construct a Z-ρ joint feature space, and perform adaptive threshold segmentation and clustering analysis on different types of materials in the Z-ρ joint feature space to obtain a preliminary clustering result; Construct a three-dimensional feature vector jointly with the preliminary clustering result and the energy difference feature map, and classify the three-dimensional feature vector through a decision tree classifier to obtain a material label map; Generate the final detection result of the object to be detected based on the material label map, the second ordinary energy X-ray image, and the second high-energy X-ray image. The final detection result includes the pseudo-color fusion display of the dual-energy image, the color-coded map of the material classification result, the three-dimensional volume reconstruction map, and the recognition result table containing the object attribute information. The recognition result table contains the ID, location, size, material composition, and hazard level of each object.

[0037] Specifically, the effective atomic number map Z(x,y) and the material density map ρ(x,y) are combined at the pixel level, and the atomic number and density of each pixel are used to form a two-dimensional feature vector, thereby establishing a Z-ρ joint feature space in the entire image space. In this space, due to the high separability of different materials in the two dimensions of atomic number and density, for example, organic substances usually have a distribution characteristic of Z < 10 and ρ < 1.5 g / cm 3 while light metals mostly fall in the region where Z is 10 - 30 and ρ is 1.5 - 5 g / cm 3 and heavy metals are concentrated in the region where Z > 30 and ρ > 5 g / cm 3Region. Therefore, different category boundaries are set by adaptive threshold division in this two-dimensional space. Adaptive threshold segmentation analyzes the global statistical features of the Z-ρ distribution and the local density change trend, dynamically setting the position of the segmentation line instead of using a fixed threshold division to adapt to the distribution offsets brought by different imaging conditions and different object structures. At the same time, it combines density-aware clustering algorithms, such as improved density-based spatial clustering of applications with noise (DBSCAN) or density peak clustering (DPC), to automatically identify material clusters according to the clustering trend formed by each type of material in the Z-ρ space, and outputs a preliminary clustering result map. Although the preliminary clustering result has a high separation degree in the two-dimensional space, there are still problems such as some blurred boundaries and unstable recognition of material overlapping regions, especially in mixed materials, bonding layers, and composite wrapping structures. Therefore, an energy difference feature map E(x,y) is introduced, and it is jointly expressed with the preliminary clustering result to construct a three-dimensional feature vector V(x,y)=[Z(x,y),ρ(x,y),E(x,y)]. This three-dimensional vector not only has the physical property characteristics of static materials but also integrates the response difference ability from the energy spectrum dimension, and is applicable to the recognition enhancement under complex material mixing areas, bonding boundary areas, and heterogeneous material distributions. After constructing the three-dimensional feature vectors of all pixels, they are input into a pre-trained decision tree classifier for multi-level splitting classification. The decision tree uses the Gini coefficient as the splitting criterion, and the depth is set to 5 to 8 layers to ensure that the model has strong generalization ability while ensuring the classification accuracy. The decision tree classifier learns the distribution pattern of each material in the feature space, realizes the assignment of the material category number for each pixel position, and outputs a material label map. This image clearly identifies different material regions in the image in the form of spatial structure and semantic expression. The material label map is integrated with the second ordinary energy X-ray image I1(x,y) and the second high-energy X-ray image I2(x,y) to construct a visualization output module for the final detection result. By means of pseudo-color fusion, the dual-energy image and the energy difference map are mapped to the HSV color space, where the H channel maps E(x,y), the S channel maps I2(x,y), and the V channel maps I1(x,y), forming a visual enhancement map combining hue and lightness, which is used to assist in observing the difference features of the responses of different materials in the dual-energy image. At the same time, a color marking map is generated according to the material label map, and different category materials are mapped to distinguishable color-coded regions, enabling the operator to quickly identify the spatial distribution state of various materials in the object. When the image is output, to support the volume recognition and in-depth understanding of the measured object at the spatial structure level, a three-dimensional volume reconstruction process is further carried out on the object based on the dual-energy image. This process uses an improved algebraic reconstruction technique algorithm. Based on the known X-ray penetration grayscale of each pixel, combined with the physical properties in the Z and ρ maps, the absorption ratio of each voxel on the X-ray projection path is inversely deduced, and then the three-dimensional voxel intensity distribution image is gradually inversely deduced to construct a three-dimensional reconstruction volume with the same resolution as the original image.To improve the intelligence level of the system, a structured recognition result table is finally generated based on all connected regions in the material label map. All independent object regions are recognized by the 8-neighborhood connectivity rule, and a unique ID is assigned to each object; then the material composition feature vector composed of its centroid position, minimum circumscribed rectangle size, major axis direction, and pixel ratio is calculated. The material vector is matched with the built-in dangerous goods material database by cosine similarity. When the matching similarity is higher than the preset threshold (such as 0.85), the object is marked as a high-risk category and displayed in red highlight in the result table. The final table includes the number, two-dimensional spatial position, geometric size, material composition ratio, and danger level label of each object.

[0038] In a specific embodiment, the process of performing the steps of jointly constructing a three-dimensional feature vector from the preliminary clustering result and the energy difference feature map and classifying the three-dimensional feature vector by a decision tree classifier to obtain the material label map may specifically include the following steps: Based on the preliminary clustering result, the effective atomic number value and the material density value at each pixel position are respectively extracted to form a two-dimensional feature vector, and the energy difference feature map is normalized to obtain a normalized energy difference feature map; The two-dimensional feature vector is merged with the value of the normalized energy difference feature map at the corresponding position to obtain a three-dimensional feature vector (Z, ρ, E) at each pixel position, where Z is the effective atomic number value, ρ is the material density value, and E is the normalized energy difference value; All the three-dimensional feature vectors of the pixels are input into the decision tree classifier, and the corresponding material category number is assigned to each pixel position through the category identifier output by the decision tree in the decision tree classifier to generate the material label map.

[0039] Specifically, taking the preliminary clustering result as a reference framework, within the effective region calibrated by this result, the effective atomic number value Z(x, y) and the material density value ρ(x, y) corresponding to the current pixel are extracted pixel by pixel, and these two physical property values are combined to form a two-dimensional feature vector. The effective atomic number map Z(x, y) is derived from the map previously inverted by the decoder branch of the deep neural network. The value of each pixel represents the equivalent atomic number of the material at that position. It not only reflects the elemental composition of the material but is also highly correlated with the attenuation ability of X-rays in this energy range. The material density map ρ(x, y), on the other hand, provides the macroscopic density information of the material, which is used to distinguish materials with the same atomic number but different densities, such as hollow metal structures and solid metal cores. The combined features of Z and ρ can comprehensively characterize the physical composition and structural distribution of the material, constituting an important basic feature for identifying different types of materials. The third dimension reflecting the energy spectrum attenuation difference is introduced, that is, the energy difference feature map E(x, y). This map is obtained by performing a weighted difference calculation on the second ordinary energy X-ray image and the second high-energy X-ray image. Normalization processing is performed on E(x, y). The normalization uses a linear mapping method to compress the E value into the interval [0, 1]. The normalized energy difference feature map E′(x, y) has good scale stability and structural recognizability and can be used as the third dimension to supplement the Z-ρ space to construct a higher-dimensional material characterization space. The Z(x, y), ρ(x, y) of each pixel above and the normalized energy difference value E′(x, y) are combined at the same spatial position to form a three-dimensional feature vector V(x, y) = [Z, ρ, E]. The construction process of the three-dimensional feature vector is carried out in units of pixels, and finally a set of three-dimensional feature vector sets are formed within the image size range, constituting the input sample set for driving the classifier to complete fine-grained recognition. This sample set is input into a pre-trained decision tree classifier. This classifier uses the supervised learning method and is constructed based on a large number of known material category samples. Its core consists of a series of decision nodes and leaf nodes. Each decision node performs a splitting operation on the input sample according to the threshold of a certain feature dimension. Using the Gini coefficient or information gain as the splitting criterion, the samples are gradually divided into subspaces with higher purity in the three-dimensional feature space until all samples are assigned to the optimal leaf nodes or reach the preset maximum depth. During the training process, the classifier learns the typical distribution regions occupied by different materials in the three-dimensional space of Z, ρ, and E and maps them to specific category numbers. This number corresponds to the identity labels of various actual materials (such as organic plastics, aluminum, steel, copper, ceramics, etc.) and has semantic meanings that can be directly used for visual segmentation and structural analysis. When the three-dimensional feature vector of each pixel is input into the decision tree, the system automatically assigns it to the corresponding category according to the decision path where the feature value is located, and fills the category number as the material label value of this pixel into the material label map.The label map maintains the same spatial resolution as the input image. All pixels are labeled one by one with the material type they belong to, forming a classification result map with a clear spatial distribution, distinct boundaries, and consistent semantics. The label map can not only reflect the actual distribution of various materials inside the object but also reveal the material composition and boundary relationships in complex structures such as mixtures, coatings, and foreign object embeddings. The finally obtained material label map serves as the core input basis for subsequent three-dimensional reconstruction, material statistical analysis, and risk assessment modules. By further fusing the label map with the three-dimensional reconstruction data, the system automatically extracts each independent object region and calculates its structural parameters, such as volume, shape, centroid position, boundary contour, etc., generating a complete recognition result table containing object ID, position, size, material composition ratio, and risk level.

[0040] The X-ray detection method for detecting multiple densities and materials in the embodiments of the present invention has been described above. Next, the X-ray detector for detecting multiple densities and materials in the embodiments of the present invention will be described. Please refer to Figure 2 , an embodiment of the X-ray detector for detecting multiple densities and materials in the embodiments of the present invention includes: A penetration imaging module 201 for performing penetration imaging on the X-ray beam through a double-layer TFT detector structure, where the double-layer TFT detector structure includes a first TFT imaging layer, a copper filter layer, and a second TFT imaging layer; An image formation module 202 for forming a first ordinary energy X-ray image of the X-ray penetrating the object to be detected on the first TFT imaging layer, and simultaneously forming a first high-energy X-ray image filtered by the copper filter layer on the second TFT imaging layer; A preprocessing module 203 for preprocessing the first ordinary energy X-ray image and the first high-energy X-ray image to obtain a second ordinary energy X-ray image and a second high-energy X-ray image; A feature extraction module 204 for performing dual-energy domain feature extraction based on the second ordinary energy X-ray image and the second high-energy X-ray image to construct an effective atomic number map and a material density map; An output module 205 for performing feature space clustering analysis and decision tree classification based on the effective atomic number map and the material density map and outputting the final detection result.

[0041] Through the collaborative cooperation of the above-mentioned various components, a double-layer TFT detector structure is adopted and a copper filtering layer is arranged between the two layers of TFTs, realizing the function of obtaining ordinary energy images and high-energy images simultaneously with a single X-ray exposure, without the need to switch the X-ray source energy for multiple exposures, fundamentally avoiding the image registration problem in traditional technologies. Special feature extraction paths are designed for low-density materials and high-density materials respectively. The density difference features of different materials are highlighted through the energy difference enhancement module, improving the accuracy of material identification. The image information in different energy domains is effectively integrated, enhancing the ability to identify mixed materials and complex structures, especially the ability to distinguish materials with similar densities but different atomic numbers. The direct material attenuation coefficient reconstruction algorithm of the present invention utilizes the spatial registration advantage of the double-layer TFT detector to directly establish an accurate mapping of the effective atomic number and material density, without complex image registration preprocessing, greatly reducing the computational complexity. This method obtains dual-energy image information through a single X-ray exposure, significantly reducing the radiation dose, while extending the service life of the TFT detector and reducing the system operation cost.

[0042] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the above-described systems, systems, and units can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated herein.

[0043] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs that can store program codes.

[0044] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. An X-ray detection method for detecting multiple densities and materials, characterized in that, Including: Performing penetration imaging on an X-ray beam through a double-layer TFT detector structure, where the double-layer TFT detector structure includes a first TFT imaging layer, a copper filter layer, and a second TFT imaging layer; Forming a first ordinary energy X-ray image of the X-rays penetrating the object to be detected on the first TFT imaging layer, and simultaneously forming a first high-energy X-ray image filtered by the copper filter layer on the second TFT imaging layer; Preprocessing the first ordinary energy X-ray image and the first high-energy X-ray image to obtain a second ordinary energy X-ray image and a second high-energy X-ray image; Performing dual-energy domain feature extraction based on the second ordinary energy X-ray image and the second high-energy X-ray image to construct an effective atomic number map and a material density map; Based on the effective atomic number map and the material density map, performing feature space clustering analysis and decision tree classification to output a final detection result.

2. The X-ray detection method for detecting multiple densities and materials according to claim 1, characterized in that, The performing penetration imaging on an X-ray through a double-layer TFT detector structure, where the double-layer TFT detector structure includes a first TFT imaging layer, a copper filter layer, and a second TFT imaging layer, includes: Preparing the first TFT imaging layer using amorphous silicon material, with the sensitivity range of the first TFT imaging layer being 15 - 30 keV, and preparing the second TFT imaging layer using polycrystalline silicon material, with the sensitivity range of the second TFT imaging layer being 50 - 120 keV; Performing deposition treatment on cesium iodide material to obtain a scintillator of the first TFT imaging layer, and performing deposition treatment on gadolinium oxysulfide material to obtain a scintillator of the second TFT imaging layer; Processing an oxygen-free copper material to obtain a copper filter layer, and overlapping and mounting the first TFT imaging layer and the second TFT imaging layer in the Z-axis direction through an alignment device so that the pixel points of the first TFT imaging layer and the second TFT imaging layer correspond one-to-one in spatial position, and placing the copper filter layer between the first TFT imaging layer and the second TFT imaging layer to obtain a double-layer TFT detector structure; Through the double-layer TFT detector structure, emitting an X-ray beam from an X-ray source to irradiate the object to be detected for penetration imaging.

3. The X-ray detection method for detecting multiple densities and materials according to claim 1, characterized in that The forming a first ordinary energy X-ray image of the X-rays penetrating the object to be detected on the first TFT imaging layer, and simultaneously forming a first high-energy X-ray image filtered by the copper filter layer on the second TFT imaging layer, includes: Adjusting the working voltage and working current of the X-ray source to generate an X-ray beam; Capturing the X-ray beam passing through the object to be detected on the first TFT imaging layer to obtain a first ordinary energy X-ray image, and the first ordinary energy X-ray image corresponds to the characteristic information of low-density and low-atomic number materials; Performing energy filtering on the X-ray beam passing through the first TFT imaging layer through the copper filter layer to obtain a transmitted ray; The transmitted rays are captured by the second TFT imaging layer to obtain a first high-energy X-ray image. At the same time, the first TFT imaging layer and the second TFT imaging layer are synchronously triggered and data is collected by the synchronous control unit to ensure that the first ordinary energy X-ray image and the first high-energy X-ray image are completely synchronous in the time dimension.

4. The X-ray detection method for detecting multiple densities and materials according to claim 1, characterized in that The preprocessing of the first ordinary energy X-ray image and the first high-energy X-ray image to obtain a second ordinary energy X-ray image and a second high-energy X-ray image includes: Adaptive median filtering is respectively performed on the first ordinary energy X-ray image and the first high-energy X-ray image to obtain a first filtered image and a second filtered image; Wavelet domain threshold denoising processing is respectively performed on the first filtered image and the second filtered image to obtain a first denoised image and a second denoised image; TFT non-linear response correction is respectively performed on the first denoised image and the second denoised image to obtain a first corrected image and a second corrected image; Bad pixel detection and repair are respectively performed on the first corrected image and the second corrected image to obtain a first repaired image and a second repaired image; Geometric correction is respectively performed on the first repaired image and the second repaired image to obtain a first geometrically corrected image and a second geometrically corrected image; Spatial registration is respectively performed on the first geometrically corrected image and the second geometrically corrected image to obtain a second ordinary energy X-ray image and a second high-energy X-ray image.

5. The X-ray detection method for detecting multiple densities and materials according to claim 1, characterized in that, The dual-energy domain feature extraction is performed based on the second ordinary energy X-ray image and the second high-energy X-ray image to construct an effective atomic number map and a material density map, including: Feature extraction is performed on the second ordinary energy X-ray image to obtain a low-density material feature map; Feature extraction is performed on the second high-energy X-ray image to obtain a high-density material feature map; The second ordinary energy X-ray image and the second high-energy X-ray image are input into an adaptive parameter forward neural network for parameter space distribution fusion, and a parameter space distribution map is output. The adaptive parameter forward neural network includes 3 fully connected layers, and the hidden layer dimension is 128; Based on the parameter space distribution map, weighted difference calculation is performed on the second ordinary energy X-ray image and the second high-energy X-ray image to obtain an energy difference feature map; Feature fusion is performed on the low-density material feature map, the high-density material feature map and the energy difference feature map to obtain a fusion feature map, and an effective atomic number map and a material density map are constructed based on the fusion feature map.

6. The X-ray detection method for detecting multiple densities and materials according to claim 5, characterized in that, The feature fusion of the low-density material feature map, the high-density material feature map and the energy difference feature map to obtain a fusion feature map, and the construction of an effective atomic number map and a material density map based on the fusion feature map includes: Channel number adjustment is performed on the low-density material feature map and the high-density material feature map to obtain an adjusted low-density material feature map and an adjusted high-density material feature map; Perform channel - dimension concatenation on the adjusted low - density material feature map, the energy difference feature map, and the adjusted high - density material feature map to obtain a combined feature map; Perform three - layer convolutional network processing on the combined feature map, and at the same time use spatial attention and channel attention units to output an attention map; Multiply the attention map with the adjusted low - density material feature map to obtain a weighted feature map, and add the weighted feature map to the adjusted high - density material feature map to obtain a fused feature map; Input the fused feature map, the second ordinary energy X - ray image, and the second high - energy X - ray image into a deep neural network with an encoder - decoder architecture. After extracting features through the encoder part, use the first decoder branch to generate an effective atomic number map, and at the same time use the second decoder branch to generate a material density map.

7. The X-ray detection method for detecting multiple densities and materials according to claim 6, wherein The step of inputting the fused feature map, the second ordinary energy X - ray image, and the second high - energy X - ray image into a deep neural network with an encoder - decoder architecture, extracting features through the encoder part, using the first decoder branch to generate an effective atomic number map, and at the same time using the second decoder branch to generate a material density map includes: Perform logarithmic operation processing on the second ordinary energy X - ray image to obtain an ordinary energy logarithmic attenuation map, and perform logarithmic operation processing on the second high - energy X - ray image to obtain a high - energy logarithmic attenuation map; Perform channel - dimension concatenation on the ordinary energy logarithmic attenuation map, the high - energy logarithmic attenuation map, and the fused feature map to obtain a multi - modal feature map, and input the multi - modal feature map into a deep neural network with an encoder - decoder architecture. Perform feature extraction through an encoder containing 4 residual blocks to obtain a deep feature map; Process the deep feature map through the first decoder branch. The first decoder branch consists of 4 transposed convolutional layers and uses the ReLU activation function. The last layer uses a linear activation function to adapt to the numerical range of the atomic number to obtain an effective atomic number map; Process the deep feature map through the second decoder branch. The second decoder branch consists of 4 transposed convolutional layers and uses the Sigmoid activation function. The last layer performs density value range mapping to conform to physical constraints to obtain a material density map.

8. The X-ray detection method for detecting multiple densities and materials according to claim 7, characterized in that, The step of performing feature - space clustering analysis and decision - tree classification based on the effective atomic number map and the material density map and outputting the final detection result includes: Combine the effective atomic number map and the material density map to construct a Z - ρ joint feature space, and perform adaptive threshold segmentation and clustering analysis on different types of materials in the Z - ρ joint feature space to obtain a preliminary clustering result; Construct a three - dimensional feature vector jointly with the preliminary clustering result and the energy difference feature map, and classify the three - dimensional feature vector through a decision - tree classifier to obtain a material label map; Generate the final detection result of the object to be detected based on the material label map, the second ordinary energy X-ray image, and the second high-energy X-ray image. The final detection result includes the pseudo-color fusion display of the dual-energy image, the color marking map of the material classification result, the three-dimensional volume reconstruction map, and the recognition result table containing object attribute information. The recognition result table contains the ID, position, size, material composition, and danger level of each object.

9. The X-ray detection method for detecting multiple densities and materials according to claim 8, characterized in that, The method of jointly constructing a three-dimensional feature vector from the preliminary clustering result and the energy difference feature map, and classifying the three-dimensional feature vector by a decision tree classifier to obtain a material label map includes: Based on the preliminary clustering result, extract the effective atomic number value and the material density value at each pixel position respectively to form a two-dimensional feature vector, and perform normalization processing on the energy difference feature map to obtain a normalized energy difference feature map; Merge the two-dimensional feature vector with the value of the normalized energy difference feature map at the corresponding position to obtain a three-dimensional feature vector (Z, ρ, E) at each pixel position, where Z is the effective atomic number value, ρ is the material density value, and E is the normalized energy difference value; Input the three-dimensional feature vectors of all pixels into the decision tree classifier, and assign the corresponding material category number to each pixel position according to the category label output by the decision tree in the decision tree classifier to generate a material label map.

10. An X-ray detector for detecting multiple densities and materials, characterized in that, An X-ray detector for detecting multiple densities and materials, which is used to execute the X-ray detection method according to any one of claims 1-9, includes: A penetration imaging module for performing penetration imaging on an X-ray beam through a double-layer TFT detector structure, where the double-layer TFT detector structure includes a first TFT imaging layer, a copper filtering layer, and a second TFT imaging layer; An image formation module for forming a first ordinary energy X-ray image on the first TFT imaging layer for the X-ray penetrating the object to be detected, and simultaneously forming a first high-energy X-ray image filtered by the copper filtering layer on the second TFT imaging layer; A preprocessing module for preprocessing the first ordinary energy X-ray image and the first high-energy X-ray image to obtain a second ordinary energy X-ray image and a second high-energy X-ray image; A feature extraction module for performing dual-energy domain feature extraction based on the second ordinary energy X-ray image and the second high-energy X-ray image, and constructing an effective atomic number map and a material density map; An output module for performing feature space clustering analysis and decision tree classification based on the effective atomic number map and the material density map, and outputting a final detection result.

Images