Feature extraction method based on image processing

Abstract

This invention relates to the field of image processing technology and discloses a feature extraction method based on image processing, comprising: acquiring infrared grayscale data and a radiation intensity distribution image in an environment not lower than 15℃; calculating the gradient of the radiation intensity distribution image and normalizing it to generate a unit orientation field; calculating the gradient of the infrared grayscale data and constructing an original structure tensor based on the tensor product; constructing a projection operator based on the unit orientation field, and using the projection operator to project and transform the components of the original structure tensor to generate a reconstructed structure tensor field constrained by the radiation gradient direction; decomposing the reconstructed structure tensor field and extracting target contour features. This invention utilizes the radiation gradient directionality to generate modulation constraints on the infrared gradient space, eliminates frequency domain mismatch between high-frequency and low-frequency signals, effectively suppresses unrelated thermal interference, and improves the target contour positioning accuracy in complex environments based on physical property verification during the extraction process.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to a feature extraction method based on image processing. Background Technology

[0002] Currently, using multimodal feature extraction of infrared thermal imaging and radiation flux to identify key targets is a common technique. By using spatially registered thermal images and radiation spectrum images, the geometric contours and corresponding physical distribution attributes of the target are extracted. However, when dealing with complex environments with strong background interference, there is a significant mismatch in resolution dimensions between heterogeneous sensors. Infrared imaging acquires geometric edges with high-frequency step characteristics, while radiation imaging is limited by random particle scattering and the physical aperture of the collimator, and only acquires the potential energy distribution of low-frequency dispersion. Due to the impedance mismatch in the frequency domain characteristics of the two, the existing techniques have a fundamental flaw when performing feature fusion.

[0003] The commonly used pixel-level weighting or simple gradient superposition methods result in high-resolution thermal geometric features being smoothed by low-resolution radiative energy fields, causing target contour blurring and positioning errors. Even attempts to optimize the results by increasing the detector sampling frequency cannot eliminate the diffusion effect of radiated signals due to the signal-to-noise ratio constraints of the detector's physical flux. In this case, the infrared gradient features of non-radioactive heat sources in the background are extremely similar to those of the target, resulting in a lack of unique physical attribute-level discrimination criteria during the extraction process, making it impossible to accurately separate interference signals. Existing technologies mainly face the following shortcomings: 1. Mismatch in frequency domain characteristics of heterogeneous sensors leads to loss of geometric positioning accuracy during multimodal feature fusion; 2. Lack of an effective physical consistency coupling mechanism between low-resolution radiative potential energy fields and high-resolution thermal geometric features; 3. Complex thermal... The feature extraction process under background interference presents uncertainties in target attribute discrimination. In addition to differences in hardware resolution, software control algorithms also have shortcomings in handling heterogeneous feature fusion. For example, Chinese invention patent CN115731174B discloses an infrared small target detection method and device based on image information entropy and multi-scale local contrast. This technology is based on digital image statistical features and lacks coupling with the spatial evolution law of heterogeneous physical fields. Under dynamic conditions such as nuclear industrial facility inspection, the background non-radioactive heat source and the infrared gradient features of the target are similar. The image entropy value or contrast difference alone cannot eliminate the edge of the environmental heat source that is unrelated to the radiation distribution in principle. The core preset premise is out of sync with the actual physical boundary conditions, making the target attribute discrimination uncertain and the geometric positioning accuracy lost under complex thermal background interference.

[0004] Therefore, the technical problem to be solved by this invention is how to use the spatial attenuation trend of the radiation potential energy field to guide the vector screening of the infrared gradient field and construct a feature reconstruction mechanism based on anisotropic tensor projection so that the extraction results have deterministic constraints on physical properties while maintaining high-frequency positioning accuracy. Summary of the Invention

[0005] This invention proposes a feature extraction method based on image processing, comprising the following steps: Step S101: Acquire spatially registered infrared grayscale data and radiation intensity distribution image. The infrared grayscale data and radiation intensity distribution image have the same pixel resolution. The ambient temperature when acquiring the infrared grayscale data is not lower than 15℃. Step S102: Calculate the gray-level gradient of each pixel position in the radiation intensity distribution image, generate a radiation gradient vector field, and perform normalization processing on the magnitude of each pixel in the radiation gradient vector field to generate a unit direction field characterizing the directionality of the radiation gradient. Step S103: Calculate the gradient vector of each pixel position in the infrared grayscale data, and construct the original structure tensor pixel by pixel based on the tensor product of the gradient vector and its transpose vector. The original structure tensor is used to define the gradient distribution characteristics of the infrared grayscale data in the local neighborhood. Step S104: Construct a second-order projection operator based on the direction vector of each pixel position in the unit direction field, and use the second-order projection operator to perform component projection transformation on the original structure tensor. The component projection transformation reconstructs the spatial components of the original structure tensor according to the directionality of the direction vector, generating a reconstructed structure tensor field constrained by the directionality of the radiation gradient. Step S105: Perform eigenvalue decomposition on the reconstructed structural tensor field to extract principal eigenvalues ​​and principal eigenvectors, and determine the target contour features characterizing the geometric boundary of the distribution of radioactive materials based on the principal eigenvalues ​​and principal eigenvectors.

[0006] Preferably, step S104 specifically includes: performing matrix multiplication on the second-order directional projection matrix corresponding to the second-order projection operator and the corresponding original structure tensor; defining an anisotropic projection space parallel to the gradient direction of the radiation distribution through the second-order directional projection matrix, so that the original structure tensor is reconstructed in the anisotropic projection space, suppressing the gradient components in the original structure tensor that are orthogonal to the gradient direction of the radiation distribution, thereby filtering out the background noise edges in the reconstructed structure tensor field.

[0007] Preferably, the reconstructed structural tensor field has an anisotropy metric value representing the local geometry. The method further includes: calculating the anisotropy metric value of each pixel position in the reconstructed structural tensor field, and determining the structural element attributes of the morphological operation based on the value of the anisotropy metric value; if the anisotropy metric value is lower than a preset shape threshold, the structural element is adjusted to an isotropic circular structure; if the anisotropy metric value is not lower than the shape threshold, the structural element is adjusted to a linear structural element along the extension direction of the principal feature vector, so as to suppress random edge signals while enhancing the target contour features.

[0008] Preferably, the construction of the projection operator P in step S104 follows the following rules: Where P is the projection operator. Let be the direction vector of the unit direction field at the corresponding pixel coordinates, and It is the transpose of the direction vector.

[0009] Preferably, the method further includes: performing spatial consistency smoothing on the reconstructed structural tensor field; and using the dominant flow direction information determined by neighboring pixels to perform weighted direction compensation at the feature break point, so that the reconstructed structural tensor field maintains the consistency of local topological extension at the singular point position, so as to repair the discontinuous pixel points in the target contour feature.

[0010] Preferably, the calculation of the gray-level gradient of each pixel position in the radiation intensity distribution image in step S102 specifically includes: performing Gaussian filtering and first derivative operation on the radiation intensity distribution image to generate a radiation gradient vector field; the normalization process includes mapping the vector magnitude of each pixel in the radiation gradient vector field to 1 to eliminate the fluctuation of radiation intensity values ​​in the radiation intensity distribution image and extracting spatial pointing features.

[0011] Preferably, determining the target contour features in step S105 specifically includes: extracting local maxima of gradient magnitude in the direction of the principal feature vector in the reconstructed structural tensor field, performing edge tracking in combination with a preset dual threshold criterion, and outputting the target structural contour after verification by the unit direction field.

[0012] Preferably, the structural element attributes include the ratio of the major axis to the minor axis and the deflection angle; the anisotropy metric is positively correlated with the ratio of the major axis to the minor axis, and the deflection angle is consistent with the direction of the principal eigenvector of the reconstructed structural tensor field.

[0013] Preferably, when performing component projection transformation in step S104, a gain coefficient is also introduced. The gain coefficient is adaptively updated based on the mutual information between the infrared grayscale data and the radiation intensity distribution image within the local window.

[0014] Preferably, the method further includes: superimposing the extracted target contour features onto the infrared grayscale data for fusion display, and using the target contour features to mark the boundary of the characteristic distribution area representing the radioactive material leakage area on the infrared grayscale data.

[0015] The beneficial effects of this invention are: 1. In the feature extraction method of image processing, the radiation flux potential energy field is used as an anisotropic projection operator to generate modulation constraints on the eigenspace of the infrared gradient, which solves the frequency domain mismatch problem between low-frequency diffuse radiation signals and high-frequency step thermal signals. This transforms low-resolution radiation data into a topological mask that guides high-resolution thermal feature extraction. While preserving the edge positioning accuracy of infrared images, the feature extraction process is verified with explicit physical properties, avoiding feature omissions or false detections caused by small spatial registration errors.

[0016] 2. By constructing a direction screening criterion based on the geometric axioms of vector space, non-radioactive interference signals with high gradient modulus are orthogonalized during tensor reconstruction, thus achieving deterministic suppression of pseudo-edges such as steam pipes or power generation equipment in complex environments. This screening logic based on the physical field attenuation law does not depend on the setting of empirical thresholds, enhancing the stability of the feature extraction process under dynamic conditions.

[0017] 3. In the process of image feature extraction, the anisotropic factor decomposed from the structural tensor field is used to dynamically adjust the ratio of the major and minor axes and the deflection angle of the morphological operator. This causes the structural element to shrink into an isotropic circle in the textured and messy area, while automatically stretching into a directional probe in the feature extension direction. This enhances weak linear features while blocking the connection of background noise that is inconsistent with its flow direction, thus solving the edge blurring problem caused by traditional rigid operators when processing curved features. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the overall steps of the image processing feature extraction method of the present invention. Figure 2 This is a data processing architecture diagram for infrared and radiation data fusion and feature reconstruction of the present invention. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0021] A feature extraction method based on image processing, comprising the following steps: Step S101: Acquire spatially registered infrared grayscale data and radiation intensity distribution image. The infrared grayscale data and radiation intensity distribution image have the same pixel resolution. The ambient temperature when acquiring the infrared grayscale data is not lower than 15℃. Step S102: Calculate the gray-level gradient of each pixel position in the radiation intensity distribution image, generate a radiation gradient vector field, and perform normalization processing on the magnitude of each pixel in the radiation gradient vector field to generate a unit direction field characterizing the directionality of the radiation gradient. Step S103: Calculate the gradient vector of each pixel position in the infrared grayscale data, and construct the original structure tensor pixel by pixel based on the tensor product of the gradient vector and its transpose vector. The original structure tensor is used to define the gradient distribution characteristics of the infrared grayscale data in the local neighborhood. Step S104: Construct a second-order projection operator based on the direction vector of each pixel position in the unit direction field, and use the second-order projection operator to perform component projection transformation on the original structure tensor. The component projection transformation reconstructs the spatial components of the original structure tensor according to the directionality of the direction vector, generating a reconstructed structure tensor field constrained by the directionality of the radiation gradient. Step S105: Perform eigenvalue decomposition on the reconstructed structural tensor field to extract principal eigenvalues ​​and principal eigenvectors, and determine the target contour features characterizing the geometric boundary of the distribution of radioactive materials based on the principal eigenvalues ​​and principal eigenvectors.

[0022] Preferably, step S104 specifically includes: performing matrix multiplication on the second-order directional projection matrix corresponding to the second-order projection operator and the corresponding original structure tensor; defining an anisotropic projection space parallel to the gradient direction of the radiation distribution through the second-order directional projection matrix, so that the original structure tensor is reconstructed in the anisotropic projection space, suppressing the gradient components in the original structure tensor that are orthogonal to the gradient direction of the radiation distribution, thereby filtering out the background noise edges in the reconstructed structure tensor field.

[0023] Preferably, the reconstructed structural tensor field has an anisotropy metric value representing the local geometry. The method further includes: calculating the anisotropy metric value of each pixel position in the reconstructed structural tensor field, and determining the structural element attributes of the morphological operation based on the value of the anisotropy metric value; if the anisotropy metric value is lower than a preset shape threshold, the structural element is adjusted to an isotropic circular structure; if the anisotropy metric value is not lower than the shape threshold, the structural element is adjusted to a linear structural element along the extension direction of the principal feature vector, so as to suppress random edge signals while enhancing the target contour features.

[0024] Preferably, the construction of the projection operator P in step S104 follows the following rules: Where P is the projection operator. Let be the direction vector of the unit direction field at the corresponding pixel coordinates, and It is the transpose of the direction vector.

[0025] Preferably, the method further includes: performing spatial consistency smoothing on the reconstructed structural tensor field; and using the dominant flow direction information determined by neighboring pixels to perform weighted direction compensation at the feature break point, so that the reconstructed structural tensor field maintains the consistency of local topological extension at the singular point position, so as to repair the discontinuous pixel points in the target contour feature.

[0026] Preferably, the calculation of the gray-level gradient of each pixel position in the radiation intensity distribution image in step S102 specifically includes: performing Gaussian filtering and first derivative operation on the radiation intensity distribution image to generate a radiation gradient vector field; the normalization process includes mapping the vector magnitude of each pixel in the radiation gradient vector field to 1 to eliminate the fluctuation of radiation intensity values ​​in the radiation intensity distribution image and extracting spatial pointing features.

[0027] Preferably, determining the target contour features in step S105 specifically includes: extracting local maxima of gradient magnitude in the direction of the principal feature vector in the reconstructed structural tensor field, performing edge tracking in combination with a preset dual threshold criterion, and outputting the target structural contour after verification by the unit direction field.

[0028] Preferably, the structural element attributes include the ratio of the major axis to the minor axis and the deflection angle; the anisotropy metric is positively correlated with the ratio of the major axis to the minor axis, and the deflection angle is consistent with the direction of the principal eigenvector of the reconstructed structural tensor field.

[0029] Preferably, when performing component projection transformation in step S104, a gain coefficient is also introduced. The gain coefficient is adaptively updated based on the mutual information between the infrared grayscale data and the radiation intensity distribution image within the local window.

[0030] Preferably, the method further includes: superimposing the extracted target contour features onto the infrared grayscale data for fusion display, and using the target contour features to mark the boundary of the characteristic distribution area representing the radioactive material leakage area on the infrared grayscale data.

[0031] Example 1: When the system processes the combined thermal radiation field conditions at a nuclear industry decommissioning facility, the radioactive waste leakage pipeline to be identified is located in a non-radioactive high-temperature steam background. The infrared grayscale data collected by the infrared imaging sensor contains high-frequency step geometric features and non-target heat source edges. The radiation intensity distribution image collected by the gamma camera is limited by particle scattering and presents a low-frequency diffuse distribution. The heterogeneous sensors have frequency domain characteristic mismatch. The system acquires spatially registered infrared grayscale data and radiation intensity distribution images within the same field of view in an environment of not less than 15°C to construct an initial heterogeneous data field. An ambient temperature of not less than 15°C is selected. Utilizing the response characteristics of the uncooled infrared detector in this temperature range, the original temperature difference between the target boundary and the background is made higher than the detector noise equivalent temperature difference. By measuring the signal-to-noise ratio distribution of the detector at different ambient temperatures, it is determined that when the ambient temperature is higher than 15°C, the gradient magnitude of the infrared grayscale data is stably above three standard deviations of the detector's background noise, providing a physically significant initial gradient vector field for constructing the original structure tensor.

[0032] The system's central processing unit calculates the gray-level gradient at each pixel location in the radiation intensity distribution image, generating a radiation gradient vector field. The processor normalizes the magnitude of each pixel in this radiation gradient vector field and outputs a unit-direction field characterizing the spatial attenuation trend of ionizing radiation. The processor synchronously calculates the gradient vector at each pixel position in the infrared grayscale data, calculates the tensor product of this gradient vector and its transpose, and constructs a pixel-by-pixel original structure tensor. The constructed original structure tensor records the local geometric gradient features within the field of view, including the target pipe and interfering steam. The system constructs a second-order projection operator P based on the unit direction field. The second-order projection operator P is based on the formula... Confirmed, among which Represents the unit direction field vector The transpose vector is used by the system to perform component projection transformation on the original structure tensor using a second-order projection operator. This second-order directional projection matrix defines an anisotropic projection space parallel to the radiation distribution gradient direction, driving the original structure tensor to be reconstructed in this projection space by weighted spatial components. This projection transformation preserves the gradient components in the infrared grayscale data that are collinear with the radiation attenuation direction, so that the thermal gradient vector from non-radioactive high-temperature steam is attenuated in the matrix product because it is orthogonal to the radiation attenuation direction, generating a reconstructed structure tensor field constrained by the radiation gradient direction.

[0033] The system decomposes and reconstructs the structural tensor field, extracting the principal eigenvalues ​​and principal eigenvectors at each pixel location. The processor extracts local maxima of the gradient magnitude along the principal eigenvector direction based on the image matrix, and tracks edges using a dual-threshold criterion. When calculating the anisotropy metric of pixel locations in the reconstructed structural tensor field, the principal eigenvalues ​​are obtained by performing eigenvalue decomposition on the tensor matrix. and secondary eigenvalues and according to the formula Determine the anisotropy metric value ;according to The ratio of the major and minor axes of the morphological structural elements is determined by comparing the results with the preset shape threshold of 0.7. When the value is below 0.7, a circular structural element with a major-to-minor axis ratio of 1:1 is selected for noise reduction. When the ratio of major axis to minor axis is not less than 0.7, the ratio of major axis to minor axis is adjusted to 5:1 and the direction of major axis coincides with the direction of principal feature vector. This enhances the continuity of target contour and blocks the topological connection of random noise points. The target contour features that characterize the geometric boundary of radioactive material distribution are determined. The output target contour features are the distribution area boundary marked on infrared grayscale data. This boundary is composed of connected pixels with thermal gradients and whose geometric normal direction is consistent with the ionization radiation attenuation path. This filters out the edges of environmental heat sources that are unrelated to radiation distribution.

[0034] Example 2: The system processes the image feature extraction task of the nuclear industry thermal radiation composite field, and obtains a spatially registered image dataset jointly generated by the computational fluid dynamics simulation model based on the Navier-Stokes equations and the Monte Carlo particle transport model. The spatial resolution of the infrared grayscale data and the radiation intensity distribution image in this dataset is uniformly limited to 640×480 pixels, and the physical environment temperature for simulating data acquisition is kept constant at 25℃. Gaussian white noise with a signal-to-noise ratio of 20dB is injected into the experimental data source to simulate the sensor's background thermal noise, and Poisson-distributed scattering scintillation noise is simultaneously superimposed on the radiation intensity distribution image. The system configures the local structure tensor calculation window size, calculates the spatial variance inverse of the local infrared grayscale data gradient to construct a dynamic constraint model, and shrinks the calculation window size when the local variance increases. Based on this judgment logic, the experiment determines the optimal calculation window size to be 5×5 pixels under noisy conditions, and establishes the initial physical environment and underlying computing conditions of the experiment.

[0035] Based on the above conditions, the system constructs the present invention sample group, the first comparison sample group, and the second comparison sample group. The first comparison sample group eliminates the second-order projection operator mechanism, and the system calculates the original structure tensor of the infrared grayscale data and extracts edges. The second comparison sample group uses a 15×15 pixel over-limit calculation window. At the input reference end, the original infrared grayscale data containing noise exhibits a structure with 85.2% connectivity, 64.3% vapor interference edges, and a measured potential energy dispersion radius of 12.6 pixels for the original radiation intensity distribution image. The processor calculates and normalizes the grayscale gradient of the radiation intensity distribution image, generating a unit direction field. According to the formula Construct a second-order projection operator P, where The system calculates the matrix product of the second-order projection operator P and the original structure tensor of the infrared grayscale data, using the transpose of the unit direction field vector. A gain coefficient is introduced by performing component projection transformation, determined based on the mutual information between the infrared grayscale data and the radiation intensity distribution image within a 5×5 pixel local window. The mutual information value is obtained by calculating the joint probability distribution of the two heterogeneous data streams within the local window and the Koolbek-Leibler divergence of each edge probability distribution, and compared with a preset registration confidence threshold. If the value is higher than the threshold, the gain coefficient is set to 1.2 to enhance the weighting of the radiation field on the infrared gradient modulation. The calibration process for the preset registration confidence threshold is as follows: Under standard calibration conditions, the pixel offset between the geometric center of the infrared target and the radiation extremum point is calculated. If the offset is less than 2.0 pixels, the confidence level is mapped to 0.95; if the offset is between 2.0 and 5.0 pixels... Between these values, the confidence level decreases downwards at a fixed slope of 0.15 per pixel. When the confidence level is below 0.50, the system determines that the registration has failed and locks the current gain coefficient. If it is not higher than the threshold, the gain coefficient is reduced to 0.8 to suppress projection distortion caused by residual spatial registration errors, thus ensuring that the reconstructed structure tensor field is aligned. Quantitative tests show that in the reconstructed structure tensor field of the sample group of this invention, the gradient magnitude of the non-radioactive steam region decreases from 145.6 before projection to 12.3, while the gradient magnitude of the target pipe edge, which is collinear with the radiation attenuation direction, remains at 185.4. This demonstrates the numerical attenuation mechanism of heterogeneous physical field constraints on orthogonal chaotic thermal gradients. The system decomposes the reconstructed structure tensor field, extracts the principal feature values ​​and principal feature vectors of each pixel position, and the processor extracts the local maxima of the gradient magnitude in the direction of the principal feature vector based on the image matrix, outputting the target contour features.

[0036] The measured data shows that the extraction accuracy of the output contour of the sample group of the present invention reaches 96.8% with a false alarm rate of 2.1%. The false alarm rate of the first comparison sample group increased to 45.7% due to the lack of radiation field projection constraints. The extraction accuracy of the second comparison sample group decreased to 68.5% due to contour smoothing caused by exceeding the limit window. The parameter gradient test was carried out by adjusting the intensity of the injected Gaussian white noise. As the noise intensity setting value increased from 10dB to 30dB, the measured extraction accuracy of the sample group of the present invention remained between 92.6% and 94.1%. When the noise intensity setting value exceeded 35dB, the tensor eigenvalue decomposition was sharply aggravated by the background random polarization interference, and the measured extraction accuracy showed a nonlinear decline. The accuracy dropped to 74.2%; this gradient data confirms that 20dB to 30dB constitutes the preferred working range of this tensor projection algorithm; the data difference between the first comparative sample group and the sample group of this invention confirms that the radiation gradient distribution has a synergistic effect on the unidirectional modulation constraint of the thermal morphology tensor, and the removal of the multi-physics field cross constraint leads to the failure of single-mode feature extraction; the experimental data confirms that the second-order projection operator based on the radiation gradient pointing constraint can separate the radiation homogeneous geometric boundary from the background containing heat source interference, and uses the matrix orthogonal attenuation law to solve the frequency domain mismatch between the high spatial resolution of infrared and the low spatial resolution of radiation, maintaining the extraction accuracy of heterogeneous image features within the determined noise threshold range.

[0037] Example 3: The system handles industrial inspection tasks where the background temperature of thermal radiation fluctuates periodically. The infrared grayscale data output by the infrared imaging sensor has time-varying global contrast characteristics, and random electromagnetic interference in the environment superimposes high-frequency white noise into the image matrix. The system acquires spatially registered infrared grayscale data and radiation intensity distribution images in an environment with a temperature higher than 15℃, extracts a preset Gaussian derivative convolution kernel, and uses this Gaussian derivative convolution kernel to perform sliding window convolution operations on the pixel matrix of the infrared grayscale data and the radiation intensity distribution image, respectively, to calculate the smooth grayscale gradient at each pixel position, generating a noise-resistant radiation gradient vector field and an infrared gradient vector field. The processor normalizes the magnitude of each pixel in the radiation gradient vector field and outputs the unit direction field. .

[0038] The processor is based on the formula A second-order projection operator P is constructed, and the matrix product between P and the original structure tensor composed of the infrared gradient vector field is calculated to generate a reconstructed structure tensor field. The system decomposes this reconstructed structure tensor field and extracts the principal eigenvalues ​​and principal eigenvectors of each pixel position. The processor counts the principal eigenvalues ​​of all pixels in the image, constructs a one-dimensional distribution histogram, and uses a sliding grayscale reference to divide all pixels into edge and background classes, calculating the inter-class variance of the pixel distributions of the two classes. The processor iterates through all available eigenvalue grayscale levels and establishes the specific eigenvalue that maximizes the inter-class variance as... The high threshold is determined by multiplying it by a set scaling factor of 0.4, which is then used as the low threshold. The system extracts local maxima of the gradient magnitude along the principal feature vector direction from the image matrix, filters out pixels with principal feature values ​​lower than the low threshold, marks pixels with principal feature values ​​higher than the high threshold as strong edge points, and retains pixels with spatial connectivity to strong edge points among pixels with principal feature values ​​in the double threshold range. The processor filters out non-target edges caused by background temperature drift based on the above criteria and outputs the target contour features representing the geometric boundary of the distribution of radioactive materials.

[0039] Example 4: When the system is deployed at a facility site with geometric field of view difference and unknown background scattered radiation, the infrared imaging sensor and gamma camera experience spatial parallax due to the physical installation distance. The background scattering signal causes the magnitude of the radiation gradient vector to diverge. Before acquiring images, the system controls the inspection terminal to move to a set calibration station. This calibration station is equipped with a characteristic target with a surface temperature of 80°C and encapsulated with a known activity standard radioactive source. The sensor simultaneously acquires infrared grayscale data and radiation intensity distribution images covering the characteristic target. The processor extracts the geometric center coordinates of the target's thermal edge from the infrared grayscale data and the coordinates of the target's potential energy extremum point from the radiation intensity distribution image. Based on the pixel offset of the two sets of coordinates and the field of view scaling factor, the affine transformation matrix M is quantized and derived, where M is a parameter matrix representing the spatial mapping relationship. The affine transformation matrix M is written into memory. During operation, the processor uses the affine transformation matrix M to resample the acquired radiation intensity distribution image, unifying the pixel spatial correspondence between the infrared grayscale data and the radiation intensity distribution image.

[0040] After completing spatial resampling, the system continuously acquires a sequence of background radiation intensity images for a preset duration of 100 seconds. The processor calculates the average radiation count at each pixel location in the sequence and establishes a baseline matrix characterizing the static scattering background at the site. ,in To record the numerical matrix of ambient background radiation intensity, during real-time operation, the processor employs a sliding window strategy with a 10-second cycle. If the average radiation count of the entire image in the current cycle deviates by more than 5% compared to the previous cycle, a memory overwrite instruction for the baseline matrix is ​​automatically executed, using the new count value as the background compensation benchmark for the current cycle. The system acquires a real-time radiation intensity distribution image during operation, and the processor calculates the relationship between this real-time radiation intensity distribution image and the baseline matrix. The difference is used to filter out the ambient background scattering flux. The processor calculates the gray-level gradient at each pixel location in the difference image and calculates the modulus normalization to generate a unit orientation field. The processor is based on the formula Construct a second-order projection operator P, where P is a second-order directional projection matrix. The unit direction field vector, As the transpose of the unit direction field vector, the processor uses the second-order projection operator P to perform component projection operations on the original structure tensor constructed from infrared grayscale data. The spatial mapping and baseline subtraction steps transform sensor assembly errors and static environmental interference into compensation parameters, providing a consistent data source required for extracting target contour features.

[0041] Example 5: When the system is adapted to a radiation detection hardware platform equipped with a heterogeneous scintillator crystal, in the case where the radiation intensity distribution image exhibits a physically diffuse boundary due to the detector's point spread effect, the system initiates a hardware spatial scale calibration procedure before performing Gaussian filtering of the infrared grayscale data; the system controls the test fixture to place a standard point-shaped radiation source in a shielded room maintained at a constant temperature of 20°C, and uses the gamma camera to be calibrated to acquire the radiation intensity distribution image of the point-shaped radiation source, extracts the two-dimensional spatial distribution matrix of the radiation count in the image, and intercepts a one-dimensional distribution profile outward from the center of the radiation count peak. Based on this one-dimensional distribution profile, the pixel span distance d at the position where the radiation count decays to half of its extreme value is determined; the processor then applies the formula... The Gaussian scaling parameter σ is calculated, where the constant 2.355 is the half-width transformation coefficient of the normal distribution function. The system uses this Gaussian scaling parameter σ to instantiate a two-dimensional Gaussian function and calculates the first-order partial derivative of the two-dimensional Gaussian function to generate a Gaussian derivative convolution kernel. This parameter mapping step ensures that the filtering scale of the infrared structure tensor matches the physical spatial resolution of the radiation detector.

[0042] During continuous operation, the system continuously acquires real-time radiation intensity distribution images and infrared grayscale data, calculates the global grayscale mean of the radiation intensity distribution image in real time, and initiates a background radiation dynamic compensation procedure when the mean exceeds a preset baseline drift threshold. The baseline drift threshold is set as the initial baseline matrix. The processor extracts the spatial variance distribution of the infrared grayscale gradient within the current field of view after initiating the procedure, which is 1.5 times the global mean. It then selects the connected region with the smallest spatial variance as the dynamic compensation window, calculates the background radiation count increment within the window, and updates the initial baseline matrix with this increment. The system establishes the background baseline for the current period, and the processor calculates the difference between the real-time radiation intensity distribution image and the background baseline for the current period to generate a background-removed image. The system then calculates the gray-level gradient at each pixel location in the background-removed image and performs normalization on the pixel magnitude to generate a unit orientation field. According to the formula Construct a second-order projection operator P, where Given the transpose of the unit direction field vector, the system uses the second-order projection operator P to perform component projection transformation on the original structure tensor obtained by processing infrared grayscale data based on Gaussian derivative convolution kernels; the above hardware scale calibration parameters and dynamic baseline compensation parameters are simultaneously applied to the matrix operation process to output the target contour features characterizing the geometric boundary of the distribution of radioactive materials.

[0043] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A feature extraction method based on image processing, characterized in that, The method includes the following steps: Step S101: Acquire spatially registered infrared grayscale data and radiation intensity distribution image. The infrared grayscale data and radiation intensity distribution image have the same pixel resolution. The ambient temperature when acquiring the infrared grayscale data is not lower than 15℃. Step S102: Calculate the gray-level gradient of each pixel position in the radiation intensity distribution image, generate a radiation gradient vector field, and perform normalization processing on the magnitude of each pixel in the radiation gradient vector field to generate a unit direction field characterizing the directionality of the radiation gradient. Step S103: Calculate the gradient vector of each pixel position in the infrared grayscale data, and construct the original structure tensor pixel by pixel based on the tensor product of the gradient vector and its transpose vector. The original structure tensor is used to define the gradient distribution characteristics of the infrared grayscale data in the local neighborhood. Step S104: Construct a second-order projection operator based on the direction vector of each pixel position in the unit direction field, and use the second-order projection operator to perform component projection transformation on the original structure tensor. The component projection transformation reconstructs the spatial components of the original structure tensor according to the directionality of the direction vector, generating a reconstructed structure tensor field constrained by the directionality of the radiation gradient. Step S105: Perform eigenvalue decomposition on the reconstructed structural tensor field to extract principal eigenvalues ​​and principal eigenvectors, and determine the target contour features characterizing the geometric boundary of the distribution of radioactive materials based on the principal eigenvalues ​​and principal eigenvectors.

2. The feature extraction method based on image processing according to claim 1, characterized in that, Step S104 specifically includes: performing matrix multiplication on the second-order directional projection matrix corresponding to the second-order projection operator and the corresponding original structure tensor; defining an anisotropic projection space parallel to the gradient direction of the radiation distribution through the second-order directional projection matrix, so that the original structure tensor is reconstructed in the anisotropic projection space, suppressing the gradient components in the original structure tensor that are orthogonal to the gradient direction of the radiation distribution, thereby filtering out the background noise edges in the reconstructed structure tensor field.

3. The feature extraction method based on image processing according to claim 1, characterized in that, The reconstructed structural tensor field has an anisotropy metric value that characterizes the local geometry. The method further includes: calculating the anisotropy metric value of each pixel position in the reconstructed structural tensor field, and determining the structural element attributes of morphological operations based on the anisotropy metric value; if the anisotropy metric value is lower than a preset shape threshold, the structural element is adjusted to an isotropic circular structure; if the anisotropy metric value is not lower than the shape threshold, the structural element is adjusted to a linear structural element along the extension direction of the principal feature vector, so as to suppress random edge signals while enhancing the target contour features.

4. The feature extraction method based on image processing according to claim 1, characterized in that, The construction of the projection operator P in step S104 follows these rules: Where P is the projection operator. Let be the direction vector of the unit direction field at the corresponding pixel coordinates, and It is the transpose of the direction vector.

5. The feature extraction method based on image processing according to claim 1, characterized in that, The method also includes: performing spatial consistency smoothing on the reconstructed structural tensor field; and using the dominant flow direction information determined by neighboring pixels to perform weighted direction compensation at feature break points, so that the reconstructed structural tensor field maintains the consistency of local topological extension at singular point locations, in order to repair discontinuous pixels in the target contour features.

6. The feature extraction method based on image processing according to claim 1, characterized in that, The calculation of the gray-level gradient of each pixel position in the radiation intensity distribution image in step S102 specifically includes: performing Gaussian filtering and first derivative operation on the radiation intensity distribution image to generate a radiation gradient vector field; the normalization process includes mapping the vector magnitude of each pixel in the radiation gradient vector field to 1 to eliminate the fluctuation of radiation intensity values ​​in the radiation intensity distribution image and extracting spatial pointing features.

7. The feature extraction method based on image processing according to claim 1, characterized in that, The specific steps for determining the target contour features in step S105 include: extracting the local maxima of the gradient magnitude in the direction of the principal feature vector in the reconstructed structural tensor field, performing edge tracking in combination with the preset dual threshold criteria, and outputting the target structural contour after verification by the unit direction field.

8. The feature extraction method based on image processing according to claim 3, characterized in that, The structural element properties include the ratio of the major axis to the minor axis and the deflection angle; the anisotropy metric is positively correlated with the ratio of the major axis to the minor axis, and the deflection angle is consistent with the direction of the principal eigenvector of the reconstructed structural tensor field.

9. The feature extraction method based on image processing according to claim 1, characterized in that, When performing component projection transformation in step S104, a gain coefficient is also introduced. The gain coefficient is adaptively updated based on the mutual information between the infrared grayscale data and the radiation intensity distribution image within the local window.

10. The feature extraction method based on image processing according to claim 1, characterized in that, The method also includes: superimposing the extracted target contour features onto infrared grayscale data for fusion display, and using the target contour features to mark the boundary of the characteristic distribution area representing the radioactive material leakage area on the infrared grayscale data.

Images