Nuclear Radiation Image Enhancement System Based on Artificial Intelligence Analysis

Abstract

The present invention discloses a nuclear radiation image enhancement system based on artificial intelligence analysis, belonging to the field of nuclear radiation image processing, which includes a data acquisition module and an image enhancement module; the data acquisition module acquires multi-modal nuclear radiation image data, and the image enhancement module includes an image completion unit, an image compression unit, an image fusion unit, and an image correction unit; the image completion unit performs image completion, the image compression unit performs radiation intensity zoning and compression on the initially complete image output, the image fusion unit fuses the compressed image with two-channel feature maps, and the image correction unit performs non-linear correction and then inputs it into the image completion unit for final completion. Through the efficient cooperation of multiple modules, the present invention realizes high-quality nuclear radiation monitoring and analysis, improves the visual quality, enhances the image readability, repairs and enhances the image accuracy. Overall, it not only improves the image quality and analysis accuracy, but also provides strong support for nuclear radiation monitoring and safety protection.

Description

Technical Field

[0001] The invention belongs to the technical field of nuclear radiation image processing, and in particular relates to a nuclear radiation image enhancement system based on artificial intelligence analysis. Background Art

[0002] In recent years, with the development of nuclear radiation monitoring technology, the application of various detectors and imaging technologies in the fields of nuclear safety, environmental monitoring and medical imaging has received widespread attention. In particular, the synchronous application of gamma cameras and neutron imagers makes multimodal nuclear radiation imaging possible, which can simultaneously obtain image data of gamma radiation and neutron radiation, providing a more comprehensive analysis of nuclear radiation sources. However, the existing technology still faces some challenges, such as noise problems in the image acquisition process, difficulties in completing missing areas of the image, and problems in fusing characteristics of different radiation sources. These problems seriously affect the accuracy and real-time performance of radiation monitoring and hinder its effective application in practical applications. Therefore, the development of an efficient nuclear radiation image enhancement system, especially the combination of artificial intelligence technology for image processing to improve image quality and enhance the ability to identify radiation sources, has become an important issue that needs to be solved urgently.

[0003] In the existing technology, most image enhancement methods still rely on traditional image processing techniques, such as filtering, histogram equalization, etc. These methods often do not work well when processing nuclear radiation images with strong noise and missing information. In addition, the existing image enhancement system lacks intelligent dynamic partition management of radiation intensity areas, which leads to insufficient processing of high, medium and low radiation areas, which may cause misjudgment of nuclear radiation monitoring. By analyzing the shortcomings of the existing technology, it can be seen that the image enhancement method based on artificial intelligence can effectively overcome these limitations.

[0004] In view of the above-mentioned deficiencies in the prior art, our invention provides an innovative solution. Summary of the invention

[0005] In view of the above existing problems, the present invention solves the following technical problem: In nuclear radiation images, due to the limitations of detectors and environmental factors, blind spots and noise interference often occur, resulting in incomplete image information.

[0006] Specifically, it is: how to dynamically divide the radiation intensity areas in the image and adopt appropriate processing methods for different areas to ensure the accuracy and reliability of the image; how to effectively integrate data from different imaging devices to improve image quality, make the radionuclide information clearer and easier to analyze and judge later.

[0007] In order to solve the above technical problems, a nuclear radiation image enhancement system based on artificial intelligence analysis is proposed, including a data acquisition module and an image enhancement module;

[0008] A data acquisition module that acquires multi-modal nuclear radiation image data and synchronously records the physical parameter metadata of the gamma camera and the neutron imager;

[0009] An image enhancement module, including an image completion unit, an image compression unit, an image fusion unit, and an image correction unit;

[0010] The image completion unit inputs the acquired image into a two-stage generative adversarial network. The first-stage generator completes the missing area and outputs an initial completed image, and the second-stage generator performs texture repair;

[0011] The image compression unit partitions the radiation intensity of the output preliminary complete image, dynamically divides it into three high, medium, and low radiation intensity regions, and respectively uses a combination strategy of logarithmic transformation and local histogram equalization for compression;

[0012] The image fusion unit inputs the compressed image into a two-channel neural network. The two-channel neural network includes a physical feature channel and a visual feature channel. The physical feature channel analyzes the position of the gamma energy spectrum characteristic peak, and the visual feature channel extracts texture features. The feature maps of the two channels are fused through a spatial alignment module;

[0013] The image correction unit performs spatial feature analysis based on the fused two-channel feature map output by the image fusion unit, and matches the pre-stored geometric distortion parameter set for non-linear correction;

[0014] The image completion unit inputs the corrected image into the second-stage generator of the two-stage generative adversarial network, and outputs the final enhanced image according to the pre-trained optimized texture details.

[0015] As a preferred embodiment of the nuclear radiation image enhancement system based on artificial intelligence analysis of the present invention, wherein: the data acquisition module includes a synchronous acquisition device composed of a gamma camera and a neutron imager, which synchronously acquires gamma radiation images and neutron radiation images with time stamps aligned, and simultaneously obtains physical parameter metadata;

[0016] The gamma radiation image and the neutron radiation image are spliced to obtain spliced image data and input it into the image enhancement module.

[0017] As a preferred embodiment of the nuclear radiation image enhancement system based on artificial intelligence analysis of the present invention, wherein: the image completion unit includes a two-stage generative adversarial network including a first-stage generator and a second-stage generator;

[0018] The first-stage generator receives the spliced image data of the gamma radiation image and the neutron radiation image. The operator selects the confidence threshold of the area to be repaired. The generator dynamically adjusts the structure generation intensity of the filling area according to the threshold to select the convolution kernel and selects the radiation intensity gradient feature of the adjacent area or the neutron radiation image for the filling method assignment. The first stage outputs the initial filled image;

[0019] The second-stage generator optimizes the texture filling of the initial filled image generated by the first-stage generator.

[0020] As a preferred solution of the nuclear radiation image enhancement system based on artificial intelligence analysis according to the present invention, wherein: the image compression unit includes partitioning the radiation intensity of the output initial filled image, performing linear normalization, mapping to the standard radiation intensity unit, generating a full-image radiation intensity distribution matrix of the same size as the input image, and calculating the global reference value and the spatial gradient according to the full-image radiation intensity distribution matrix;

[0021] When the full-image radiation intensity distribution matrix in the continuous pixel block is greater than the sum of the standard deviation of all image pixels and the global reference value, the current area is marked as a high-radiation area, and the high-radiation area is compressed by adaptive logarithmic transformation;

[0022] When the full-image radiation intensity distribution matrix is less than or equal to the sum of the standard deviation of all image pixels and the global reference value and greater than or equal to the difference between the global reference value and the standard deviation of all image pixels, the current area is marked as a medium-radiation area; when the full-image radiation intensity distribution matrix is less than the difference between the global reference value and the standard deviation of all image pixels, the current area is marked as a low-radiation area; the medium and low radiation areas are compressed by local histogram equalization with the window size adjusted according to the radiation intensity;

[0023] A transition zone with an adjustable width is generated at the boundary of the adjacent area for the medium-radiation area and the low-radiation area. The range of the transition zone is marked by a semi-transparent color band for the treatment of the transition zone, and linear interpolation is performed on the compression intensity of the pixels in the transition zone.

[0024] As a preferred solution of the nuclear radiation image enhancement system based on artificial intelligence analysis according to the present invention, wherein: the image fusion unit includes inputting the compressed image into a two-channel neural network, including physical feature and visual feature channels, and fusing the feature maps of the two channels through a spatial alignment module;

[0025] The physical feature channel analyzes the position of the gamma energy spectrum characteristic peak, selects the position of the characteristic peak according to the characteristic energy spectrum data of the radioactive nuclide contained in the gamma radiation image, calls the structure generation intensity parameter of the filling area of the initial filled image, and applies a compensation offset to the position of the energy spectrum characteristic peak of the filling area to generate a compensated physical feature map;

[0026] The visual feature channel extracts texture features, sets extraction weights for the texture features of the complemented region, and generates a weighted visual feature map;

[0027] A two-dimensional coordinate system is constructed in the compensated physical feature map and the weighted visual feature map. The energy spectrum peak position coordinates of the compensated physical feature map are matched with the texture centroid coordinates of the weighted visual feature map. Rigid registration is used for the non-complemented region, and elastic deformation registration is used for the complemented region. A registration error map is output for manual review. The region with excessive error is returned to the image completion unit stage for regeneration, and a fused feature map after matching is output.

[0028] As a preferred solution of the nuclear radiation image enhancement system based on artificial intelligence analysis according to the present invention, wherein: the image correction unit includes, based on the output fused feature map, performing spatial feature analysis, extracting distortion features, superimposing equally spaced orthogonal line arrays on the fused feature map, tracking the grid intersection offset of the feature peak, and identifying radial distortion and tangential distortion;

[0029] The radial distortion coefficient is calculated by fitting the offset - radius squared curve; the tangential distortion coefficient is determined by calculating the average value of the tangential offset vectors of the feature points;

[0030] A two-stream convolutional network including a radial distortion sensitive channel and a tangential distortion sensitive channel is constructed;

[0031] A geometric distortion parameter knowledge base is established to store the correction parameter combinations under typical working conditions. The cosine similarity between the radial distortion or tangential distortion feature vectors output by the feature analysis layer and the knowledge base is calculated, and the top 3 parameter combinations with the highest similarity are selected. The selected parameter combinations are weighted and averaged, and the weights are dynamically adjusted by the historical correction success rates of each combination. An automatic correction strategy is implemented according to the radial distortion coefficient and the tangential distortion coefficient, and the coordinates are adjusted according to the corrected coordinates, and the corrected image data is output.

[0032] As a preferred solution of the nuclear radiation image enhancement system based on artificial intelligence analysis according to the present invention, wherein: the image completion unit further includes that the secondary stage generator in the two-stage generative adversarial network adopted by the image completion unit receives the corrected image data;

[0033] A shared texture feature library is established to store the standard texture templates under typical working conditions. According to the association mapping relationship between the corrected image data and the shared texture feature library, the standard texture template with the highest matching degree is automatically selected, and the matching process is based on the following priority order:

[0034] The first priority is the standard texture corresponding to the nuclide already identified in the currently processed image, the second priority is the texture of historical cases with similar radiation intensity distributions, and the third priority is the basic texture of common radioactive substances;

[0035] The sub - stage generator receives the corrected image data and the matching texture template, performs gradient detection on the edge connection band of the complemented area, and automatically starts the transition smoothing process when it detects that the gray - level jump exceeds the average gradient value of the adjacent area.

[0036] According to the divided radiation intensity regions, dynamically adjust the texture generation intensity. High - radiation regions use high - frequency texture enhancement, and low - radiation regions maintain the basic texture resolution. Call the corrected coordinate data to ensure that the directionality of the generated texture is consistent with the actual radiation distribution pattern.

[0037] Another object of the present invention is to provide a nuclear radiation image enhancement method based on artificial intelligence analysis.

[0038] As a preferred embodiment of the nuclear radiation image enhancement method based on artificial intelligence analysis described in the present invention, it includes:

[0039] The data acquisition module acquires multi - modal nuclear radiation image data and synchronously records the physical parameter metadata of the gamma camera and the neutron imager.

[0040] The image completion unit inputs the acquired image into a two - stage generative adversarial network. The first - stage generator completes the missing area and outputs an initial completed image, and the sub - stage generator performs texture repair.

[0041] The image compression unit performs radiation intensity partitioning on the output preliminary complete image, dynamically divides it into three radiation intensity regions: high, medium, and low, and respectively uses a combination strategy of logarithmic transformation and local histogram equalization for compression.

[0042] The image fusion unit inputs the compressed image into a two - channel neural network. The two - channel neural network includes a physical feature channel and a visual feature channel. The physical feature channel analyzes the position of the gamma energy spectrum characteristic peak, and the visual feature channel extracts texture features. The feature maps of the two channels are fused through a spatial alignment module.

[0043] The image correction unit performs spatial feature analysis based on the output fused feature map and matches the pre - stored geometric distortion parameter set for non - linear correction.

[0044] The image completion unit inputs the corrected image into the sub - stage generator of the two - stage generative adversarial network and outputs the final enhanced image according to the pre - trained optimized texture details.

[0045] Advantages of the present invention: Through the efficient cooperation of multiple modules, high-quality nuclear radiation monitoring and analysis are achieved. The data acquisition module uses the synchronous acquisition of a gamma camera and a neutron imager to ensure the accuracy and integrity of the data; the image completion unit in the image enhancement module dynamically optimizes image completion and texture repair through a two-stage generative adversarial network, improving the visual quality; the image compression unit differentiates compression strategies according to the radiation intensity to improve image readability; the image fusion unit integrates physical and visual features in a dual-channel neural network to ensure complete information; the image correction unit enhances image accuracy through geometric distortion repair; the sub-stage completion unit optimizes texture matching using a shared texture feature library to improve detail performance. Overall, the optimization of each link in the system not only improves the image quality and analysis accuracy, but also provides strong support for nuclear radiation monitoring and safety protection. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings, where:

[0047] Figure 1 It is a system schematic diagram of a nuclear radiation image enhancement system based on artificial intelligence analysis provided by an embodiment of the present invention.

[0048] Figure 2 It is a general flowchart of a nuclear radiation image enhancement method based on artificial intelligence analysis provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0049] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following will provide a detailed description of the specific embodiments of the present invention with reference to the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of them. 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 scope of protection of the present invention.

[0050] Example 1, referring to Figure 1 This is the first embodiment of the present invention. This embodiment provides a nuclear radiation image enhancement system based on artificial intelligence analysis, including a data acquisition module and an image enhancement module.

[0051] S1: The data acquisition module collects multi-modal nuclear radiation image data and synchronously records the physical parameter metadata of the gamma camera and the neutron imager.

[0052] It should be noted that a synchronous acquisition device composed of a gamma camera and a neutron imager is adopted. The detection plane of the acquisition device is fixedly installed at an adjustable angle within [30°, 60°], and gamma radiation images and neutron radiation images with time stamps aligned are synchronously acquired, while physical parameter metadata is obtained. By setting up the synchronous acquisition device of the gamma camera and the neutron imager, an efficient data acquisition mechanism is created, realizing the alignment of the time stamps of the gamma radiation images and the neutron radiation images. By recording the radiation parameters and blind area information in detail, the accuracy and integrity of the data are ensured.

[0053] The gamma radiation image contains radioactive nuclide characteristic energy spectrum data and spatial intensity characteristics.

[0054] The neutron radiation image contains light element attenuation coefficient distribution data and spatial density distribution.

[0055] The gamma radiation image and the neutron radiation image are spliced to obtain spliced image data, which is input into the image enhancement module.

[0056] The physical parameter metadata includes the radiation source activity value, detector distance parameter, and acquisition time length during acquisition, which are embedded in the image file header as metadata.

[0057] S2: The image enhancement module includes an image completion unit, an image compression unit, an image fusion unit, and an image correction unit.

[0058] S201: The image completion unit inputs the acquired images into a two-stage generative adversarial network. The first-stage generator completes the missing area and outputs an initial completed image, and the second-stage generator performs texture repair.

[0059] It should be noted that the image completion unit adopts a two-stage generative adversarial network including a first-stage generator and a second-stage generator. Image completion and texture repair are realized through the two-stage generative adversarial network, and the missing area can be quickly filled in a short time.

[0060] The first-stage generator receives the spliced image data of the gamma radiation image and the neutron radiation image. The operator selects the confidence threshold of the area to be repaired, and the generator dynamically adjusts the structure generation intensity of the completed area according to the threshold.

[0061] The first-stage generator receives the spliced image data and receives the confidence threshold parameter T set by the operator through the interactive interface c with the range set from 0.3 to 0.7. The first-stage generator calculates the structure generation intensity value according to the confidence threshold parameter and outputs the initial completed image at the first stage.

[0062] When the structure generation intensity value is greater than 0.6, the first-stage generator adopts a 3×3 small convolution kernel and preferentially uses the radiation intensity gradient characteristics of adjacent areas for completion.

[0063] When the structure generation intensity value is less than or equal to 0.6, the first-stage generator switches to a 5×5 large convolution kernel for regional feature speculation, and combines the light element attenuation coefficient distribution data in the neutron radiation image for completion, and calls the light element distribution data of the neutron image to fill large missing areas;

[0064] The structure generation intensity is defined as the structural continuity level between the completed area and the original area. The structure generation intensity determines the reuse priority of adjacent features during the completion process, and the calculation method is:

[0065] G s =α·(1 - T c ) + β

[0066] where G s is the structure generation intensity, T c is the confidence threshold set by the operator, dimensionless (the operator sets a pure number); α is the structure sensitivity coefficient (for example, it can be set to 0.8), dimensionless (the operator sets a pure number); β is the basic generation intensity (for example, it can be set to 0.2), dimensionless (the operator sets a pure number).

[0067] The first-stage generator dynamically optimizes the completion process according to the set confidence threshold, so that the completed area has a high structural similarity with the original image.

[0068] It should also be noted that the second-stage generator optimizes the texture completion of the image data generated in the first stage, specifically as in S205.

[0069] It should be noted that the first-stage generator and the second-stage generator can both be generators. For example, the first-stage generator can be, but is not limited to, a U-Net generator or a ResNet generator, and the second-stage generator can be, but is not limited to, an Autoencoder or a Transformer-based texture generation network;

[0070] In addition, the input layer of the first-stage generator accepts the spliced image data of the gamma radiation image and the neutron radiation image. Convolution layer: uses a 3×3 or 5×5 convolution kernel, and the specific selection is based on the structure generation intensity; Normalization layer: Batch Normalization is used to stabilize the training process; Activation function: uses ReLU as the non-linear transformation to improve the expression ability of the model; Output layer: generates the initial completed image; The loss function of the first stage includes structural loss and adversarial loss.

[0071] The secondary-stage generator, whose structural parameters include an input layer that receives the initial completed image output by the primary-stage generator, a texture feature extraction layer that performs local texture extraction using a convolutional window of 11×11 or 21×21, skip connections, adaptive normalization, and an output layer. The loss function of the secondary stage includes texture consistency loss and style loss.

[0072] The secondary-stage generator ensures that the texture of the generated image is natural, achieving the purpose of enhancing the visual effect. This not only improves the visual quality of the image but also provides a more reliable information basis for subsequent analysis.

[0073] S202: The image compression unit divides the output preliminary complete image into radiation intensity zones, dynamically divides it into three radiation intensity zones of high, medium, and low, and respectively uses a combined strategy of logarithmic transformation and local histogram equalization for compression. Dynamically dividing the radiation intensity zones and using different compression strategies effectively improves the readability and analysis efficiency of the image.

[0074] It should be noted that the output initial completed image is divided into radiation intensity zones, dynamically divided into three radiation intensity zones of high, medium, and low, and respectively uses a combined strategy of logarithmic transformation and local histogram equalization for compression.

[0075] The output initial completed image is divided into radiation intensity zones, linearly normalized, mapped to the standard radiation intensity unit, generates a full-image radiation intensity distribution matrix of the same size as the input image, and calculates the global reference value and spatial gradient based on the full-image radiation intensity distribution matrix.

[0076] The global reference value is to calculate the mean and standard deviation of all image pixels, and determine the final reference value according to the ratio of the completed area to the total area.

[0077] The spatial gradient calculates the comprehensive gradient of the eight neighborhoods for each pixel point through the Euclidean distance.

[0078] When the full-image radiation intensity distribution matrix of consecutive pixel blocks is greater than the sum of the standard deviation of all image pixels and the global reference value, the current area is marked as a high-radiation area, and the system automatically displays the contour of the high-radiation area, and the operator manually excludes misjudged areas.

[0079] When the full-image radiation intensity distribution matrix is less than or equal to the sum of the standard deviation of all image pixels and the global reference value and greater than or equal to the difference between the global reference value and the standard deviation of all image pixels, the current area is marked as a medium-radiation area.

[0080] When the full-image radiation intensity distribution matrix is less than the difference between the global reference value and the standard deviation of all image pixels, the current area is marked as a low-radiation area.

[0081] A transition zone with an adjustable width (e.g., 5 pixels) is generated at the boundary of the adjacent regions of the medium radiation area and the low radiation area.

[0082] Compression processing is performed on different radiation areas:

[0083] For the high radiation area, adaptive logarithmic transformation is used for dynamic range compression to avoid over - saturation while retaining details: L(x, y) = log(1 + λM(x, y)) / log(1 + λM max ), where L(x, y) is the dynamic compression result of M(x, y), λ is the compression factor that can be adjusted in real - time by the operator through a slider, dimensionless (the operator sets a pure number); M(x, y) is the radiation intensity distribution matrix of the whole image, dimensionless (the normalized pixel value, range 0 - 1); M max is the maximum value of the radiation intensity distribution matrix of the whole image. The system displays a highlight clipping warning in real - time and automatically reduces the α value when the proportion of over - saturated pixels detected exceeds the set threshold.

[0084] For the medium - low radiation area, local histogram equalization is used, and the window size is adjusted dynamically according to the radiation intensity. Among them, an 11×11 window is used for the medium radiation area, and a 21×21 window is used for the low radiation area.

[0085] For the transition zone processing, a semi - transparent color band is used to mark the range of the transition zone, and linear interpolation is performed on the compression intensity of the pixels in the transition zone.

[0086] Through efficient logarithmic transformation and local histogram equalization, the visualization effect of important information in the image is enhanced, ensuring that the high - radiation area is clearly distinguishable, which is helpful for further image analysis and monitoring, and improving the response ability to potential radiation risks.

[0087] S203: Image fusion unit. The compressed image is input into a two - channel neural network. The two - channel neural network includes a physical feature channel and a visual feature channel. The physical feature channel analyzes the position of the gamma energy spectrum characteristic peak, and the visual feature channel extracts texture features. The feature maps of the two channels are fused through a spatial alignment module, realizing more comprehensive information extraction.

[0088] It should be noted that the image fusion unit inputs the compressed image into a two - channel neural network, including physical features and visual feature channels, and fuses the feature maps of the two channels through a spatial alignment module;

[0089] The physical feature channel analyzes the position of the gamma energy spectrum characteristic peak, selects the position of the characteristic peak according to the characteristic energy spectrum data of the radionuclide contained in the gamma radiation image, calls the structure of the completion area of the initial completion image to generate an intensity parameter, and applies a compensation offset to the position of the energy spectrum characteristic peak in the completion area. The offset Δ is calculated to satisfy Δ = c×(1 - G s), where c is a preset basic compensation coefficient in pixel units (e.g., c = 2 means a maximum compensation of 2 pixels), to generate a compensated physical feature map.

[0090] The visual feature channel extracts texture features and extracts the weight w for the texture features of the completed region v Set as w v = γ × G s , where γ is an adjustable scale factor.

[0091] The operator adjusts the γ value and observes the feature map superposition effect in real time. When a feature misalignment region is detected, a semi-transparent marking frame is automatically generated in the misalignment region. The operator clicks on the marking frame to trigger a local realignment operation, and the system automatically optimizes the c value of this region to generate a weighted visual feature map.

[0092] A two-dimensional coordinate system is constructed in the compensated physical feature map and the weighted visual feature map. The energy spectrum peak position coordinates of the compensated physical feature map are matched with the texture centroid coordinates of the weighted visual feature map. Rigid registration is used for non-completed regions, and elastic deformation registration is used for completed regions. An output registration error map is generated for manual review, and regions with excessive errors need to be returned to the image completion unit stage for regeneration.

[0093] While ensuring information integrity, the fused feature map also improves the accuracy and efficiency of subsequent analysis, making the monitoring and management of nuclear radiation more efficient.

[0094] It should be noted that the physical feature channel mainly focuses on the analysis and compensation of gamma energy spectra to ensure the effective extraction of the characteristic energy spectrum data of radionuclides contained in gamma radiation images, and is optimized in combination with the structure completion results, based on a common architecture;

[0095] The physical feature channel adopts a convolutional neural network (CNN) architecture and combines spectral analysis;

[0096] The input layer receives the compressed image and generates intensity parameters in combination with the structure of the completed region of the initial completed image; the feature extraction layer (CNN layer) uses a 3×3 or 5×5 convolution kernel for feature extraction; during spectral analysis, the Fourier transform (FFT) or wavelet transform is used to extract the feature peak position; the output layer is to generate the physical feature map;

[0097] The visual feature channel adopts, but is not limited to, a deep residual network (ResNet) or a VGG-style feature extraction network;

[0098] The input layer receives the compressed image and the texture data of the initial completed image, calculates the texture extraction weights. The feature extraction layer (ResNet / VGG) can use, but is not limited to, 5×5 or 7×7 convolutional kernels to perform local texture feature extraction, generate a weighted visual feature map, and output the final fusion result.

[0099] S204: The image correction unit performs spatial feature analysis based on the fused two-channel feature map output by the image fusion unit, and matches the pre-stored geometric distortion parameter set for non-linear correction.

[0100] Furthermore, based on the fused two-channel feature map output by the image fusion unit, perform spatial feature analysis, extract distortion features, superimpose equally spaced orthogonal line arrays on the feature map, track the grid intersection offset of the feature peak, and identify radial distortion and tangential distortion; realize the repair of image geometric distortion and ensure the geometric accuracy of the output image.

[0101] Among them, the radial distortion is defined as: calculating the growth rate of the offset from the central region to the edge region, and marking it as radial distortion when the growth rate exceeds the linear threshold;

[0102] The tangential distortion is defined as: detecting the asymmetric offset of the feature points along the diagonal direction of the image, and the offset direction is associated with the installation inclination angle of the detector;

[0103] Calculate the radial distortion coefficients k1 and k2 by fitting the offset - radius squared curve; determine the tangential distortion coefficients p1 and p2 by calculating the average value of the tangential offset vectors of the feature points. Among them, k1 represents the quadratic radial distortion amount from the image center to the edge (for example, correcting the "barrel distortion" or "pincushion distortion" caused by the curvature of the gamma camera lens), k2 represents the quartic radial distortion compensation (for example, correcting the residual error of k1 in the high radiation intensity region), p1 represents the rotational asymmetry distortion in the X - Y plane, and p2 represents the diagonal direction distortion coupling amount;

[0104] Construct a two-stream convolutional network including a radial distortion sensitive channel and a tangential distortion sensitive channel:

[0105] Among them, the radial sensitive channel uses a ring convolutional kernel group to detect the curvature anomaly features in the edge region of the image; the tangential sensitive channel uses an inclined direction convolutional kernel to capture the pixel offset features in the diagonal direction;

[0106] Establish a geometric distortion parameter knowledge base, store the correction parameter combinations under typical working conditions, calculate the cosine similarity between the radial distortion or tangential distortion feature vectors output by the feature analysis layer and the knowledge base, select the top 3 parameter combinations with the highest similarity, and perform weighted averaging on the selected parameter combinations. The weights are dynamically adjusted by the historical correction success rates of each combination;

[0107] Implement an automatic correction strategy. According to the radial distortion coefficient and tangential distortion coefficient, execute the following correction equations to eliminate the radial distortion and tangential distortion caused by the gamma camera lens:

[0108] x corrected = x(1 + k1 * r 2 + k2 * r 4 ) + 2p1 * xy + p2(r 2 + 2x 2 )

[0109] y corrected = y(1 + k1 * r 2 + k2 * r 4 ) + p1(r 2 + 2y 2 ) + 2p2 * xy

[0110] where (x, y) is the original coordinate, in pixel units; (x corrected , y corrected ) is the corrected coordinate, r 2 = x 2 + y 2 , in pixel square units;

[0111] Adjust according to the corrected coordinates and output the corrected image data. The present invention combines the detection of radial and tangential distortions, can accurately identify feature offsets, greatly reduce misjudgments caused by distortions, and improve the credibility of images and their analysis.

[0112] S205: The image completion unit further includes that the secondary stage generator in the two-stage generative adversarial network adopted by the image completion unit receives the corrected image data, effectively improving the detail performance of the image.

[0113] Establish a shared texture feature library to store standard texture templates under typical working conditions. According to the association mapping relationship between the corrected image data and the shared texture feature library, automatically select the standard texture template with the highest matching degree. The matching process is based on the following priority order:

[0114] The first priority is the standard texture corresponding to the radionuclide already identified in the currently processed image, the second priority is the historical case texture with a similar radiation intensity distribution, and the third priority is the basic texture of common radioactive substances;

[0115] The secondary stage generator receives the corrected image data and the matched texture template, and performs gradient detection on the edge connection band of the completion area. When it detects that the gray level jump exceeds the average gradient value of the adjacent area, it automatically starts the transition smoothing process;

[0116] According to the divided radiation intensity regions, dynamically adjust the texture generation intensity. For high-radiation regions, use high-frequency texture enhancement, and for low-radiation regions, maintain the basic texture resolution. Call the corrected coordinate data to ensure that the directionality of the generated texture is consistent with the actual radiation distribution pattern; The present invention can automatically optimize texture matching according to historical cases, ensure the propagation effect and information integrity of the image in the real scene, and help to better reflect the actual radiation distribution.

[0117] Example 2, referring to Figure 2 , which is the second embodiment of the present invention. This embodiment provides a nuclear radiation image enhancement method based on artificial intelligence analysis, including:

[0118] The data acquisition module acquires multi-modal nuclear radiation image data and synchronously records the physical parameter metadata of the gamma camera and the neutron imager;

[0119] The image completion unit inputs the acquired image into a two-stage generative adversarial network. The first-stage generator completes the missing regions and outputs an initial completed image, and the second-stage generator performs texture repair;

[0120] The image compression unit performs radiation intensity zoning on the output initial completed image, dynamically divides it into three radiation intensity regions of high, medium, and low, and uses a combination strategy of logarithmic transformation and local histogram equalization for compression;

[0121] The image fusion unit inputs the compressed image into a two-channel neural network. The two-channel neural network includes a physical feature channel and a visual feature channel. The physical feature channel analyzes the position of the gamma energy spectrum characteristic peak, and the visual feature channel extracts texture features, and fuses the feature maps of the two channels through a spatial alignment module;

[0122] The image correction unit performs spatial feature analysis based on the fused two-channel feature map output by the image fusion unit, and matches the pre-stored geometric distortion parameter set for non-linear correction;

[0123] The image completion unit inputs the corrected image into the second-stage generator of the two-stage generative adversarial network, and outputs the final enhanced image according to the pre-trained optimized texture details.

[0124] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not restrictive. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. A nuclear radiation image enhancement system based on artificial intelligence analysis, characterized by: Including data acquisition module and image enhancement module; Data acquisition module, which collects multi-modal nuclear radiation image data and synchronously records the physical parameter metadata of the gamma camera and neutron imager; An image enhancement module, including an image completion unit, an image compression unit, an image fusion unit and an image correction unit; The image completion unit inputs the collected image into a two-stage generative adversarial network. The first-stage generator completes the missing area and outputs the initial completed image, and the second-stage generator performs texture restoration. The first-stage generator is based on the architecture of the U-Net generator, and the second-stage generator is based on the architecture of the texture generation network. The image compression unit performs radiation intensity partitioning on the output preliminary complete image, dynamically divides it into three radiation intensity areas of high, medium and low, and compresses it using a combination strategy of logarithmic transformation and local histogram equalization; The image fusion unit inputs the compressed image into a dual-channel neural network, wherein the dual-channel neural network includes a physical feature channel and a visual feature channel, wherein the physical feature channel analyzes the characteristic peak position of the gamma energy spectrum, and the visual feature channel extracts the texture feature, and fuses the two-channel feature maps through a spatial alignment module; wherein the physical feature channel adopts a convolutional neural network combined with a spectrum analysis architecture, and the visual feature channel adopts a deep residual network or a VGG style feature extraction network architecture; The image correction unit performs spatial feature analysis based on the fused two-channel feature map output by the image fusion unit, and matches the pre-stored geometric distortion parameter set to perform nonlinear correction; The image completion unit inputs the corrected image into the second-stage generator of the two-stage generative adversarial network and outputs the final enhanced image based on the pre-trained optimized texture details; The image compression unit includes: partitioning the output initial complement image by radiation intensity, performing linear normalization, mapping to a standard radiation intensity unit, generating a full-image radiation intensity distribution matrix of the same size as the input image, and calculating a global reference value and a spatial gradient according to the full-image radiation intensity distribution matrix; When the radiation intensity distribution matrix of the whole image in the continuous pixel block is greater than the sum of the pixel standard deviation of the whole image and the global reference value, the current area is marked as a high radiation area, and the high radiation area is compressed by adaptive logarithmic transformation; When the full-image radiation intensity distribution matrix is ​​less than or equal to the sum of the full-image pixel standard deviation and the global benchmark value and greater than or equal to the difference between the global benchmark value and the full-image pixel standard deviation, the current area is marked as a medium radiation area; when the full-image radiation intensity distribution matrix is ​​less than the difference between the global benchmark value and the full-image pixel standard deviation, the current area is marked as a low radiation area; for medium and low radiation areas, the local histogram equalization window size is used to compress according to the radiation intensity; For medium radiation area and low radiation area, a transition zone with adjustable width is generated at the boundary of adjacent areas. The transition zone range is marked with a semi-transparent color band, and the compression intensity of the pixels in the transition zone is linearly interpolated.

2. The nuclear radiation image enhancement system based on artificial intelligence analysis as claimed in claim 1, characterized in that: The data acquisition module includes a synchronous acquisition device composed of a gamma camera and a neutron imager, which synchronously acquires and generates gamma radiation images and neutron radiation images with aligned timestamps, and simultaneously acquires physical parameter metadata; The gamma radiation image and the neutron radiation image are spliced ​​to obtain spliced ​​image data which is input into an image enhancement module.

3. The nuclear radiation image enhancement system based on artificial intelligence analysis as claimed in claim 2, characterized in that: The image completion unit includes, using a two-stage generative adversarial network including a first-stage generator and a second-stage generator; In the first stage, the generator receives the spliced ​​image data of the gamma radiation image and the neutron radiation image. The operator selects the confidence threshold of the area to be repaired. The generator dynamically adjusts the structural generation intensity of the complement area according to the threshold to select the convolution kernel and select the radiation intensity gradient characteristics of the adjacent area or the neutron radiation image to allocate the complement method. The first stage outputs the initial complement image. The secondary stage generator optimizes texture completion of the initial completion image generated by the primary stage generator.

4. The nuclear radiation image enhancement system based on artificial intelligence analysis as claimed in claim 3, characterized in that: The image fusion unit includes inputting the compressed image into a dual-channel neural network, including physical feature and visual feature channels, and fusing the two-channel feature maps through a spatial alignment module; The physical feature channel analyzes the characteristic peak position of the gamma energy spectrum, selects the characteristic peak position according to the characteristic energy spectrum data of the radionuclide contained in the gamma radiation image, calls the structure generation intensity parameter of the complement area of ​​the initial complement image, applies a compensation offset to the characteristic peak position of the energy spectrum of the complement area, and generates a compensated physical feature map; The visual feature channel extracts texture features, sets extraction weights for the texture features of the complemented area, and generates a weighted visual feature map; A two-dimensional coordinate system is constructed in the compensated physical feature map and the weighted visual feature map, and the energy spectrum peak position coordinates of the compensated physical feature map are matched with the texture centroid coordinates of the weighted visual feature map. Rigid registration is used for the non-completed area, and elastic deformation registration is used for the completed area. The registration error map is output for manual review, and the area with excessive errors is returned to the image completion unit stage for regeneration, and the matched fusion feature map is output.

5. The nuclear radiation image enhancement system based on artificial intelligence analysis as claimed in claim 4, characterized in that: The image correction unit includes, based on the output fusion feature map, performing spatial feature analysis, extracting distortion features, superimposing an equally spaced orthogonal line array on the fusion feature map, tracking the grid intersection offset of the feature peak, and identifying radial distortion and tangential distortion; The radial distortion coefficient is calculated by fitting the offset-radius square curve; the tangential distortion coefficient is determined by calculating the average value of the tangential offset vector of the feature point; Construct a two-stream convolutional network containing radial distortion sensitive channels and tangential distortion sensitive channels; A geometric distortion parameter knowledge base is established to store correction parameter combinations under typical working conditions. The radial distortion or tangential distortion feature vector output by the feature analysis layer is used to calculate the cosine similarity with the knowledge base. The top three parameter combinations with the highest similarity are selected, and the selected parameter combinations are weighted averaged. The weights are dynamically adjusted according to the historical correction success rate of each combination. An automatic correction strategy is implemented according to the radial distortion coefficient and the tangential distortion coefficient. Adjustments are made according to the corrected coordinates, and the corrected image data is output.

6. The nuclear radiation image enhancement system based on artificial intelligence analysis as claimed in claim 5, characterized in that: The image completion unit further includes: a second-stage generator in a two-stage generative adversarial network adopted by the image completion unit receives the corrected image data; A shared texture feature library is established to store standard texture templates under typical working conditions. According to the correlation mapping relationship between the corrected image data and the shared texture feature library, the standard texture template with the highest matching degree is automatically selected. The matching process is based on the following priority order: The first priority is the standard texture corresponding to the identified nuclides in the current processed image, the second priority is the historical case texture with similar radiation intensity distribution, and the third priority is the general radioactive material basic texture; The second-stage generator receives the corrected image data and the matching texture template, performs gradient detection on the edge connection zone of the completed area, and automatically starts transition smoothing when it detects that the grayscale jump exceeds the average gradient value of the adjacent area; According to the divided radiation intensity areas, the texture generation intensity is dynamically adjusted. High-frequency texture enhancement is used in high-radiation areas, and the basic texture resolution is maintained in low-radiation areas. The corrected coordinate data is called to ensure that the directionality of the generated texture is consistent with the actual radiation distribution pattern.

7. A method for enhancing nuclear radiation images based on artificial intelligence analysis, applied to a nuclear radiation image enhancement system based on artificial intelligence analysis as claimed in any one of claims 1 to 6, characterized in that: include, The data acquisition module collects multi-modal nuclear radiation image data and synchronously records the physical parameter metadata of the gamma camera and neutron imager; The image completion unit inputs the collected image into a two-stage generative adversarial network. The first-stage generator completes the missing area and outputs the initial completed image, and the second-stage generator performs texture restoration. The image compression unit partitions the output preliminary complete image into radiation intensity zones, dynamically divides the zone into three radiation intensity zones: high, medium and low, and compresses the zone using a combination of logarithmic transformation and local histogram equalization strategies. The image fusion unit inputs the compressed image into a dual-channel neural network, wherein the dual-channel neural network includes a physical feature channel and a visual feature channel, wherein the physical feature channel analyzes the characteristic peak position of the gamma energy spectrum, and the visual feature channel extracts the texture feature, and the feature maps of the two channels are fused through a spatial alignment module; The image correction unit performs spatial feature analysis based on the fused two-channel feature map output by the image fusion unit, and matches the pre-stored geometric distortion parameter set to perform nonlinear correction; The image completion unit inputs the corrected image into the second-stage generator of the two-stage generative adversarial network, and outputs the final enhanced image based on the pre-trained optimized texture details.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the nuclear radiation image enhancement system based on artificial intelligence analysis described in any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the nuclear radiation image enhancement system based on artificial intelligence analysis described in any one of claims 1 to 6 are implemented.

Images