Radioactive contamination area identification method and system based on image segmentation

Abstract

The invention relates to the technical field of image analysis, in particular to a radioactive contamination area identification method and system based on image segmentation. The method comprises the following steps: acquiring multiband image data; carrying out false contamination highlight area detection and false contamination dark area identification on the multiband image data to respectively obtain false contamination highlight area data and false contamination dark area data; generating a pseudo-contamination graph according to the pseudo-contamination highlight area data and the pseudo-contamination dark area data to obtain pseudo-contamination graph data; interference suppression processing is carried out according to the pseudo-contamination graph data to obtain graph data after interference suppression; according to the image data after interference suppression, performing false contamination area shielding to obtain contamination area data; and carrying out pollution area fine segmentation on the pollution area data to obtain pollution area identification data. According to the method, by constructing artifact elimination and attention guidance recognition, the radioactive contamination area can be extracted under illumination and reflection conditions, false detection interference is reduced, and the method is suitable for high-interference environments such as nuclear emergency.

Description

Technical Field

[0001] This invention relates to the field of image analysis technology, and in particular to a method and system for identifying radioactive contamination areas based on image segmentation. Background Technology

[0002] In scenarios such as nuclear emergency detection and radioactive material leak response, the on-site environment is complex and the location of contamination sources is uncertain. Relying on manual investigation is not only inefficient and risky, but also limited by the subjective judgment of the human eye. In recent years, image recognition technology has been gradually introduced into radioactive contamination detection tasks, using visible light or multi-band images to assist in the identification of contaminated areas. However, due to interference from multiple light sources, equipment reflections, damp ground, and structural shadows often present at outdoor nuclear emergency sites, traditional image segmentation methods are prone to misidentifying uncontaminated areas as contaminated targets, resulting in a high false alarm rate. At the same time, some truly contaminated areas are missed by existing methods due to weak texture and similar brightness, affecting the overall reliability of identification. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention proposes a method and system for identifying radioactive contamination areas based on image segmentation, thereby resolving at least one of the aforementioned technical issues.

[0004] This application provides a method for identifying radioactive contamination regions based on image segmentation, comprising the following steps: S1: Acquire multi-band image data; perform false contamination highlight area detection and false contamination dark area identification on the multi-band image data to obtain false contamination highlight area data and false contamination dark area data respectively; generate a false contamination map based on the false contamination highlight area data and false contamination dark area data to obtain false contamination map data. S2: Perform interference suppression processing on the pseudo-contamination map data to obtain the interference-suppressed map data; S3: Based on the data after interference suppression, perform pseudo-contamination area shielding to obtain contaminated area data; S4: Perform fine segmentation of the polluted area data to obtain polluted area identification data.

[0005] This invention constructs a pre-processing interference removal link, consisting of pseudo-contaminated bright areas, pseudo-contaminated dark areas, and a pseudo-contaminated map, to reduce background noise interference in contaminated area identification under outdoor nuclear emergency environments such as lighting conditions, metal reflection, water accumulation, and multi-source shadow superposition. Through illumination field estimation, reflection suppression, and texture compensation processing in S2, the illumination differences in the input image are effectively balanced, preventing misidentification or missed detection by the segmentation model under extreme lighting conditions. S3, through data masking and cross-band verification based on the pseudo-contaminated map, effectively filters out artifact responses in non-contaminated areas, ensuring segmentation is performed only on reliable regions, improving the model's convergence and stable output capability. S4's attention-based fine segmentation and edge optimization enhance the discernibility of contaminated area boundaries, maintaining stable segmentation results even under weak color rendering, fine-grained diffusion, or irregular diffusion patterns.

[0006] Preferably, the detection of false contamination in bright areas specifically involves: Preliminary brightness field data is obtained by performing preliminary brightness field screening on multi-band image data; Specular reflection discrimination processing is performed based on the preliminary brightness field data to obtain specular reflection data; Color saturation and reflection are verified based on preliminary brightness field data to obtain color saturation data; Infrared reflectance data was obtained by conducting infrared reflectance comparison screening based on preliminary brightness field data; Brightness edge diffusion is detected based on preliminary brightness field data to obtain brightness edge diffusion data; By fusing specular reflection data, color saturation data, infrared reflection data, and brightness edge diffusion data, pseudo-contamination highlight area data is obtained.

[0007] This invention effectively distinguishes between bright pseudo-contamination areas and genuine contamination areas in outdoor environments. Brightness field screening provides an initial candidate range; specular reflection discrimination identifies specular bright spots on metal surfaces or water accumulation; color saturation verification filters out colorless saturated bright spots caused by overexposure; infrared reflection comparison eliminates pseudo-contamination points that appear as bright spots in visible light but have no temperature difference signal in the infrared band. Brightness edge diffusion detection identifies brightness halos caused by reflection, preventing them from interfering with subsequent segmentation. Through the fusion of these multi-dimensional features, the system can effectively eliminate bright artifacts in the early stages, improving the input quality of the contamination identification process, reducing the false detection rate, and enhancing the model's stability in multi-light source, strong reflection, and material-related scenarios.

[0008] Preferably, the identification of false contamination dark areas specifically involves: The brightness attenuation region data of the multi-band image data is initially screened to obtain the brightness attenuation region data. Shadow directionality is determined from the brightness attenuation region data to obtain shadow pseudo-dark area data; Texture energy attenuation detection is performed based on the brightness attenuation region data to obtain wet artifact data; Multi-band difference verification is performed based on the brightness attenuation region data to obtain infrared pseudo-dark area data; By fusing and judging shadow false dark area data, wet artifact data and infrared false dark area data, false contamination dark area data is obtained.

[0009] This invention effectively identifies pseudo-contamination dark areas caused by factors such as shadows, humidity, or material reflection. Shadow directionality discrimination can distinguish between single-source projection and multi-source superposition interference, texture energy attenuation detection helps identify textured blurring areas caused by water accumulation, oil films, etc., while infrared band verification can eliminate non-contaminated areas that appear dark but have stable thermal characteristics. Through multi-source pseudo-dark area data fusion and judgment, the system can reduce the false judgment rate of weakly colored contamination areas, providing a cleaner image basis and more stable candidate region boundaries for contamination area identification.

[0010] Preferably, the shadow directionality determination specifically involves: Multi-directional gradient vectors are extracted from the brightness attenuation region data to obtain multi-directional gradient data; The directional entropy is calculated from the multi-directional gradient data to obtain the directional entropy data; The directional entropy data is divided into value ranges to obtain directional entropy map data; Multi-source shadow peak decomposition is performed on the directional entropy map data to obtain multi-source shadow data; Directional temporal detection is performed based on multi-source shadow data to obtain shadow pseudo-dark area data.

[0011] This invention extracts multi-directional gradient vectors from brightness attenuation regions and combines this with directional entropy for refined discrimination, enabling accurate identification of multi-source shadows and false dark areas in a scene. Multi-directional gradient extraction provides a structured foundation for directional pattern construction, while directional entropy calculation quantifies the complexity of gradient distribution within a region, effectively distinguishing between single-source shadows, dual-source superimposed shadows, and false dark areas caused by non-directional texture attenuation. Value range partitioning maps directional entropy features to explicit spatial regions, giving shadow structures visual boundaries. Peak decomposition identifies multi-directional main peak features caused by multiple light sources, and combined with directional temporal detection to extract drift patterns of dynamic shadows such as robotic arm movement and swaying leaves, reliably eliminating false dark areas in outdoor conditions such as multi-source intersections, dynamic occlusion, and vegetation texture interference.

[0012] Preferably, the pseudo-contamination map is generated as follows: Based on the data of the highlighted and dark areas of the pseudo-contamination, the pseudo-trace regions are binarized and aligned to obtain the pseudo-trace region aligned data. The artifact region alignment data is merged to obtain the merged region data. Morphological noise removal is performed on the merged regional data to obtain regional noise-reduced data; Cross-modal stabilization artifact processing is performed on the regional noise reduction data to obtain pseudo-contamination map data.

[0013] This invention achieves consistent spatial representation among artifacts from multiple bands, different scales, and different interference sources by binarizing, aligning, and uniformly registering the bright and dark areas of pseudo-contamination. This avoids boundary misalignment or artifact omission caused by differences in data resolution. The region merging step logically fuses the two types of artifacts, resulting in a unified representation of the interference region composed of reflections, bright noise, and dark artifacts such as shadows and dampness. Morphological noise removal effectively eliminates isolated small patches and scattered reflections, improving the connectivity and structural integrity of the artifact region. Cross-modal stabilization artifact processing, combined with multi-band features (such as infrared stability), filters out stable artifacts with consistent anomalous patterns across multiple bands.

[0014] Preferably, S2 specifically comprises: Illumination field data is obtained by estimating the illumination field based on the pseudo-contamination map data; Based on the illumination field data, reflection artifacts are suppressed in the pseudo-contamination map data to obtain artifact suppression map data; Texture compensation for wet areas is performed based on artifact suppression map data to obtain area compensation data; Shading compensation is applied to the regional compensation data to obtain the interference-suppressed image data.

[0015] This invention estimates the illumination field of pseudo-contamination map data, enabling the construction of a near-realistic global illumination distribution model in outdoor environments. This allows for the quantification of the impact of localized illumination unevenness, strong reflection points, and occlusion shadows on the image signal. Reflection artifact suppression based on illumination field data effectively reduces the effects of metallic reflections, water surface specular components, and overexposed bright areas, making the texture information of potentially contaminated areas clearer and more discernible after brightness normalization. The texture compensation stage for wet areas addresses texture loss caused by dampness and water accumulation by improving region separability through local texture restoration, avoiding misjudgments in dark areas. Subsequent shadow compensation of the compensated region data restores local grayscale shifts caused by multiple light sources or occlusion, resulting in a more consistent overall image brightness structure.

[0016] Preferably, S3 specifically comprises: Background artifact masking data is obtained by masking background artifacts based on the data after interference suppression. Contamination texture features are extracted from the background artifact masking data to obtain contamination texture feature data; Based on the pollution texture feature data, pollution candidate regions are generated to obtain preliminary region data; Cross-band verification was performed on the preliminary regional data to obtain data on the polluted areas.

[0017] This invention, based on the interference-suppressed image data, effectively removes residual reflective areas, bright edges, and wet stain shadows through background artifact masking. This ensures that subsequent feature extraction is performed only within the reliable region, reducing the probability of false positives. The contamination texture feature extraction step captures texture gradients, diffusion edges, and local anomaly distribution features related to radioactive contamination in the denoised and normalized image. Contamination candidate regions formed based on these texture features can initially locate contaminated areas, and a continuous spatial representation is established through region growing and density aggregation. Cross-band verification verifies the consistency of thermal features of candidate regions through infrared or other spectral channels, effectively eliminating false contamination fragments caused by uneven surface materials, environmental noise, or natural stains.

[0018] Preferably, S4 specifically comprises: Attention-based fine segmentation is performed on the polluted area data to obtain finely segmented data. Edge optimization is performed on the finely segmented data to obtain the boundary data of the polluted area; Artifact inversion verification is performed based on the boundary data of the contaminated area to obtain the contaminated area identification data.

[0019] This invention performs attention-based fine segmentation on contaminated area data, adaptively enhancing the salient features of real contaminated areas during the model inference stage while suppressing residual responses of background artifacts in the feature space. This ensures that the segmentation output maintains good structural stability even under conditions of weak color rendering, blurred boundaries, or fragmented textures. The edge optimization step after fine segmentation, combined with local gradient and spatial consistency, refines and smooths the boundaries of contaminated areas, effectively eliminating edge jaggedness, gaps, and breaks caused by imaging noise or local artifacts, resulting in higher geometric accuracy of the contaminated area contours. Based on the optimized boundary data, artifact inverse verification is performed to identify and exclude pseudo-contaminated boundaries caused by reflection afterimages, shadow drift, or abrupt changes in background texture, thereby improving the reliability of the identification results.

[0020] Preferably, attention-based fine segmentation specifically includes: Deep texture features and edge features are extracted from the polluted area data to obtain deep texture data and edge feature data, respectively. Channel attention weights and spatial attention weights are calculated on deep texture data and edge feature data to obtain channel attention weight data and spatial attention weight data. Based on the channel attention weight data and spatial attention weight data, artifact weight fusion suppression is performed on the polluted area data to obtain artifact weight data. The decoder performs fine segmentation on the data contaminated by the artifact weight data to obtain finely segmented data.

[0021] This invention extracts deep texture and edge features from contaminated area data, enabling the capture of core structural information of radioactively contaminated areas across multiple scales, including diffusion textures, local gradients, and irregular edge trends. The joint calculation of channel attention weights and spatial attention weights allows the model to dynamically allocate weights based on the response intensity and spatial importance of different feature channels, thereby highlighting the key structures of the real contaminated area and suppressing artifacts such as reflective debris, damp textures, and shadowed edges. Artifact weight fusion suppression, combining the attention mechanism with artifact priors, effectively suppresses the diffusion effect of background interference during the decoding stage, ensuring the segmentation model maintains feature purity against the background. Guided execution of the decoder based on artifact weights enables fine segmentation, obtaining contaminated area masks with higher boundary accuracy and structural integrity, improving the usability and robustness of the segmentation results, and meeting the requirements for refined contamination identification in nuclear emergency scenarios.

[0022] Preferably, this application also provides a radioactive contamination region identification system based on image segmentation, used to perform the radioactive contamination region identification method based on image segmentation as described above, the radioactive contamination region identification system based on image segmentation includes: The pseudo-contamination map generation module is used to acquire multi-band image data; perform pseudo-contamination highlight area detection and pseudo-contamination dark area identification on the multi-band image data to obtain pseudo-contamination highlight area data and pseudo-contamination dark area data respectively; generate a pseudo-contamination map based on the pseudo-contamination highlight area data and pseudo-contamination dark area data to obtain pseudo-contamination map data. The interference suppression processing module is used to perform interference suppression processing on the pseudo-contamination map data to obtain interference-suppressed map data. The pseudo-contamination area shielding module is used to shield pseudo-contamination areas based on the interference suppression image data to obtain contamination area data. The fine segmentation module for contaminated areas is used to perform fine segmentation of contaminated area data to obtain contaminated area identification data.

[0023] The beneficial effects of this invention are as follows: By constructing a hierarchical processing chain, this invention effectively solves the problem of unstable identification of contaminated areas caused by factors such as multi-source reflection, damp water accumulation, shadow superposition, and complex materials in nuclear emergency scenarios. In S1, the detection of false-contaminated bright areas and the identification of false-contaminated dark areas in multi-band images can separate typical artifacts such as reflections, overexposure, shadows, and wet stains from the early stages. The generation of false-contaminated maps achieves a unified expression of bright and dark area artifacts, providing a clear spatial prior for interference suppression. S2, through illumination field estimation, reflection suppression, and texture compensation, can restore the brightness consistency and texture coherence of images in outdoor environments, allowing potential contamination features to stand out against a noisy background. S3, the combination of false-contaminated area masking, texture feature extraction, and cross-band verification ensures that the system performs contamination identification only in trustworthy candidate areas, significantly reducing the risk of false detection. S4, through attention-guided fine segmentation and boundary correction, achieves high-precision extraction of contaminated area boundaries and eliminates artifact interference. Attached Figure Description

[0024] Other features, objects, and advantages of this application will become more apparent from the following detailed description of the non-limiting embodiments, taken with reference to the accompanying drawings: Figure 1 A flowchart illustrating the steps of a radioactive contamination region identification method based on image segmentation according to an embodiment is shown. Figure 2 A flowchart illustrating the steps of a method for detecting false contamination in bright areas according to an embodiment is shown. Figure 3 A flowchart illustrating the steps of an interference suppression processing method according to one embodiment is shown. Figure 4 A flowchart illustrating the steps of a pseudo-contaminated area shielding method according to an embodiment is shown. Figure 5 A flowchart illustrating the steps of a method for fine segmentation of contaminated areas according to an embodiment is shown. Detailed Implementation

[0025] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0026] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. Functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0027] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0028] Please see Figures 1 to 5 This application provides a method for identifying radioactive contamination regions based on image segmentation, comprising the following steps: S1: Acquire multi-band image data; perform false contamination highlight area detection and false contamination dark area identification on the multi-band image data to obtain false contamination highlight area data and false contamination dark area data respectively; generate a false contamination map based on the false contamination highlight area data and false contamination dark area data to obtain false contamination map data. Specifically, the system synchronously acquires multi-band image data of the target area using mobile monitoring equipment, including at least visible light and infrared images, and preferably also low-light enhanced images. After acquisition, the system performs distortion correction and spatial registration on the images of each band, and resamples them to the same image resolution and coordinate system, thereby forming a multi-band aligned image dataset.

[0029] In terms of detecting false contamination in bright areas, the system constructs a brightness reference map in the visible light image by means of local mean filtering or Gaussian blurring, and performs preliminary screening based on the brightness offset value. Areas with brightness higher than the surrounding level are identified as bright candidate areas, such as when the brightness offset value is greater than the sum of the local average brightness value and the local standard deviation of brightness. For these candidate regions, the system identifies artifact types through the following methods: 1. The system calculates the brightness gradient amplitude and direction characteristics of the bright areas. If there is a concentrated distribution of high amplitude gradients and the brightness is close to the upper limit of saturation, it is judged as a specular reflection artifact, commonly seen in scenes such as metal reflection or water surface reflection; 2. The system converts the image to the HSV color space and checks the saturation characteristics of the bright areas. If the brightness value is high but the saturation is far below the set threshold (e.g., 0.1), it indicates that the area is a pseudo-bright area formed due to overexposure and is marked as an exposure artifact; 3. The system searches for pixels in the infrared image that correspond to the visible light bright candidate areas. If the temperature value at this location is very small compared to the surrounding background (e.g., less than 1 degree), it indicates that there is no actual thermal radiation anomaly in this area, but only a pseudo-bright spot caused by optical reflection. This part is classified as an infrared reflection artifact; 4. The system analyzes the brightness distribution structure through edge detection and identifies areas with the characteristics of bright center and rapid edge decay, which are judged as halo-type bright artifacts. The artifact layers generated by the above multiple highlight artifact discrimination processes are merged through a logical "OR" operation to form pseudo-contaminated highlight area data.

[0030] In the process of identifying false-contaminated dark areas, the system performs preliminary screening of areas with brightness significantly lower than the local reference value based on the normalization of brightness in the visible light image. By analyzing the ratio of image brightness to the brightness reference map (e.g., less than 0.6), areas with brightness attenuation are identified and designated as candidate dark areas. The system extracts image gradient direction information within the candidate dark areas and constructs a local direction distribution histogram. By calculating the direction entropy, the system determines the characteristics of the region's illumination source: if the direction is concentrated and the entropy value is low (less than 0.7), it indicates that the region is mainly caused by the projected shadow formed by a single light source; if the direction distribution is diverse and the entropy value is high (greater than 0.9), it can be identified as a composite shadow caused by the overlap of multiple light sources, or as stripe occlusion caused by structures such as branches and leaves. Such areas are identified as false dark areas. The system performs texture energy analysis on candidate dark areas, including calculating indicators such as gray-level co-occurrence matrix energy, wavelet high-frequency energy, and edge energy. If the texture energy of this area is lower than that of the surrounding normal area, and there are a few bright reflection points (the ratio exceeds 2.0 and the number of reflection points accounts for more than 5% of the total number of pixels in the area), it indicates that the surface texture is obscured or the reflection is enhanced due to moisture or water accumulation, and this area is marked as a moisture artifact area. The system combines infrared images to perform multi-band information comparison and check the temperature performance of candidate dark areas in infrared images. If the temperature value of this area in the infrared band is not significantly different from the surrounding background, but appears as a dark area in the visible light image, it indicates that the phenomenon is caused by material characteristics or changes in illumination, and is not a true temperature anomaly; this type of area is identified as an infrared pseudo-dark area. The system fuses the identified shadow pseudo-dark areas, moisture artifact areas, and infrared pseudo-dark area data to generate pseudo-contaminated dark area data.

[0031] During the generation of the pseudo-contamination map, the system binarizes the identified bright and dark artifact regions separately and aligns them spatially in a unified coordinate system. The system then merges these two types of artifact regions using a logical OR operation to generate the pseudo-contamination map data.

[0032] S2: Perform interference suppression processing on the pseudo-contamination map data to obtain the interference-suppressed map data; Specifically, in the background image after removing false contamination areas, the system performs large-scale smoothing or low-frequency illumination modeling operations on the visible light image, such as using polynomial surface fitting or Retinex-type methods, to obtain global and local illumination field information. For false contamination highlight areas, the system normalizes or compresses the corresponding brightness areas based on the aforementioned illumination field, restoring the brightness of specular reflection areas and overexposed areas to the average level of their neighborhoods. For identified halo-type highlight artifacts, the system suppresses the brightness difference between the center and edges through local contrast adjustment, generating artifact-suppressed image data. For texture loss areas in false contamination dark areas caused by moisture or water accumulation, the system uses surrounding uncontaminated reference areas as a basis, employing block matching or nonlocal means algorithms to reconstruct the texture of the damaged areas, restoring a local texture structure consistent with similar materials. For dark area interference caused by shadows, the system compensates for brightness and color based on the illumination field information, enhancing its grayscale value and color saturation, making the visual effect of the shadow area closer to the normal area. For shadows caused by multiple light sources, the system analyzes the brightness gradient trend along each main direction and selects the direction with the most consistent brightness for compensation. After the above processing, the output image with suppressed interference has a more uniform illumination distribution, reducing the impact of artifacts on image analysis.

[0033] S3: Based on the data after interference suppression, perform pseudo-contamination area shielding to obtain contaminated area data; Specifically, the system uses the generated pseudo-contamination map data as a mask to mask the interference-suppressed image data, marking all locations marked as bright or dark artifacts as unusable areas for contamination analysis, thus generating background artifact masking data. Within the effective image area of ​​the non-background artifact masking data, the system extracts various texture and structural features related to radioactive contamination, including the amplitude and direction distribution of local gradients, local contrast and roughness of the image, energy, contrast, entropy, and other indicators in the gray-level co-occurrence matrix, as well as shape features of small-scale patches, such as compactness, aspect ratio, and boundary complexity. These features are combined into a unified contamination texture feature vector. Based on these features, the system uses cluster analysis or sets a discrimination threshold to classify pixels or regions that deviate from the background in texture into contamination candidate points or candidate patches. For these contamination candidate points or candidate patches, the system uses region growing or connected component analysis to fuse regions with similar spatial locations or texture features, generating contamination candidate regions. The system performs feature comparison analysis on candidate contamination regions in infrared images and other available spectral band images to examine whether these regions exhibit consistency in multispectral performance, such as whether they are accompanied by a slight increase in temperature or changes in reflectance characteristics. If some regions are abnormal only in visible light images but show no significant difference in infrared images, they are excluded; while those regions that show certain abnormal characteristics in both visible light and infrared images are retained to generate contamination region data.

[0034] S4: Perform fine segmentation of the polluted area data to obtain polluted area identification data.

[0035] Specifically, the system constructs an input feature tensor containing multi-source information for image patches corresponding to contaminated area data. This tensor can include the weight information of the image after interference suppression, the contaminated area data, and the artifact regions. The system extracts deep texture and edge features of the image through convolutional neural networks or other deep feature extraction modules. The system performs feature optimization processing through attention, that is, it uses channel attention weighting calculation to weight the input data of each feature channel (deep texture and edge features). Subsequently, spatial attention is used to assign weights to the feature response distribution at different locations in the image. Based on channel and spatial attention optimization, the system integrates the weight information of artifact regions into the feature map to suppress the influence of residual artifacts on the segmentation results. Layer-by-layer upsampling and multi-scale feature fusion output are performed through the decoder structure to generate fine segmentation data. The system optimizes the boundary regions in the fine segmentation data, combines local gradient distribution and edge detection results to reconstruct the boundaries, and uses conditional random fields, graph cut algorithms, or narrowband morphological operations to correct jaggedness, discontinuity, or blurred edges, generating contaminated area boundary data. The system performs reverse verification on the boundary data of the identified contaminated areas, and verifies each area by combining artifact maps, illumination field data and multi-band images: if some area boundaries are found to highly overlap with artifact areas and do not show abnormal contamination characteristics in other spectral channels, they are judged as artifacts and removed; otherwise, the areas that pass multiple rounds of verification and their boundary information are retained as contaminated area identification data.

[0036] Preferably, the detection of false contamination in bright areas specifically involves: S11: Perform preliminary brightness field screening on multi-band image data to obtain preliminary brightness field data; Specifically, in the initial brightness field screening stage, the system performs local brightness analysis on the visible light image. The system uses Gaussian filtering to generate a smooth brightness reference surface in each local region of the image. The system calculates the difference between each pixel in the original multi-band image data and its corresponding brightness reference value, forming a brightness residual map. For each pixel, the system combines the brightness fluctuation of its local area to determine whether the point is a brightness anomaly. When the brightness of a pixel is higher than the average brightness fluctuation range of its neighborhood (e.g., exceeding one and a half times the standard deviation), it is considered that the point has a brightness anomaly. The system traverses the entire image, performing the above determination pixel by pixel, to form preliminary brightness field data.

[0037] S12: Perform specular reflection discrimination processing based on the preliminary brightness field data to obtain specular reflection data; Specifically, in the specular reflection discrimination process, the system analyzes the edge and direction features of the initial brightness field data. The system uses gradient operators (such as the Sobel operator) to process candidate regions, extracting the gradient magnitude and direction distribution information of each pixel, and calculating the gradient consistency of local regions, i.e., assessing whether there is a clear dominant direction and whether the edge intensity is higher than the average level of the overall image. The system combines the following discrimination conditions to determine whether a region belongs to a specular reflection artifact: First, the brightness value of the candidate region must be close to the upper limit of image saturation; second, the gradient direction within the region must show a high degree of concentration, i.e., most edge directions tend to be consistent, which can be quantified by indicators such as direction entropy or the proportion of the dominant direction, such as direction entropy less than 0.7; third, the area of ​​the region must exceed a set minimum threshold, such as 20 pixels. When the above conditions are met simultaneously, the system identifies the region as a specular reflection artifact and generates corresponding specular reflection data.

[0038] S13: Perform color saturation and reflection verification based on the preliminary brightness field data to obtain color saturation data; Specifically, the system focuses on identifying areas with brightness values ​​close to the saturation limit but significantly low color saturation. These areas appear as grayish-white or nearly colorless reflective surfaces, found on glass, white-painted surfaces, or highly reflective floor tiles. The system counts the percentage of pixels within a candidate area that meet the criteria of extremely high brightness (greater than 240) and extremely low saturation (less than 0.15). If the percentage of such pixels exceeds a preset threshold (e.g., exceeding 60% of the total number of pixels in the area), the area is identified as a color saturation artifact. The system marks areas meeting these characteristics as color saturation artifacts, i.e., color saturation data.

[0039] S14: Perform infrared reflectance comparison screening based on preliminary brightness field data to obtain infrared reflectance data; Specifically, for each bright candidate pixel, the system obtains the temperature value of its corresponding location in the infrared image and performs a difference analysis based on the average temperature of its surrounding neighborhood. When the temperature value corresponding to a bright spot in the infrared image differs little from the surrounding background temperature, failing to reach the set temperature significance threshold (e.g., less than 0.5 degrees Celsius), it can be determined that the bright spot does not exhibit obvious thermal radiation characteristics. Such areas are not thermal anomalies caused by radioactive contamination, but rather brightness artifacts caused by factors such as light reflection and mirror materials. The system marks all bright areas that meet the above conditions as infrared reflection artifact areas and generates corresponding infrared reflection data.

[0040] S15: Perform brightness edge diffusion detection based on the preliminary brightness field data to obtain brightness edge diffusion data; Specifically, during the brightness edge diffusion detection process, the system uses the Laplace-Gaussian (LoG) algorithm to conduct in-depth analysis of the brightness structure of candidate bright areas identified in the initial brightness anomaly screening to identify diffusion-type halo artifacts. The system sequentially judges the following three characteristics: the center of the region exhibits a strong brightness response, possessing a core area with concentrated high brightness; there should be a clear decreasing brightness trend around the core, forming a brightness edge that gradually transitions outward from the center; and near the boundary of the entire region, a "halo" structure should be present, i.e., a three-layer distribution of bright-dark-bright consisting of a central bright area, an outer dark ring, and a further outer brightness transition zone. If a candidate region simultaneously possesses the above characteristics, the system determines it to be a reflection diffusion artifact and marks such regions as brightness edge diffusion data.

[0041] S16: Perform regional fusion of specular reflection data, color saturation data, infrared reflection data, and brightness edge diffusion data to obtain pseudo-contaminated highlight area data.

[0042] Specifically, the system uses a logical "OR" operation to fuse the four types of artifact mask data, forming a preliminary merged layer covering all highlight artifact types. The system performs morphological opening operations on the merged result, using small-sized structuring elements to remove edge burrs and isolated noise, smoothing region boundaries. Connected regions are then filtered by area, retaining only valid regions with an area greater than a set threshold (e.g., 10 pixels) to obtain the pseudo-contaminated highlight region data.

[0043] Preferably, the identification of false contamination dark areas specifically involves: The brightness attenuation region data of the multi-band image data is initially screened to obtain the brightness attenuation region data. Specifically, in the initial screening stage of brightness attenuation regions, the system performs local brightness ratio analysis on the visible light image to identify areas with abnormal dark regions. The system constructs a brightness reference surface using a sliding window Gaussian filter, preferably with a large size (e.g., a 21×21 pixel window) and setting appropriate blur parameters (e.g., a value of 4), thereby obtaining a brightness reference value. The system calculates the ratio of the brightness value of each pixel in the original image to its corresponding brightness reference value to analyze the degree of brightness attenuation. When the brightness of a pixel is lower than its local brightness reference value, and the brightness ratio is lower than a preset threshold (e.g., less than 0.65), the pixel is marked as a brightness attenuation anomaly. The system performs the above judgment on all pixels across the entire image and clusters pixels that meet the conditions using connected component analysis to extract brightness attenuation region data.

[0044] Shadow directionality is determined from the brightness attenuation region data to obtain shadow pseudo-dark area data; Specifically, the system uses gradient operators (such as the Sobel operator) to process each candidate dark region pixel in the image, extracting its gradient direction and magnitude. Within each brightness decay region, the system calculates a local gradient direction histogram, dividing the pixel's gradient direction into several angular intervals (e.g., 8 to 12 directional segments). Based on this direction histogram, the system calculates the directional consistency entropy index, i.e. , For directional consistency entropy, For the gradient direction angle segment, For in The proportion of pixels appearing within a directional segment to the total number of pixels in the current region is the normalized frequency of the directional histogram. The system sets discrimination rules based on the value of directional entropy: when the directional entropy is low (less than 1.0), it indicates a high concentration of gradient directions, corresponding to a projected shadow caused by a single light source; when the directional entropy is in the middle range (greater than 1.0 to less than or equal to 2.0), it indicates the existence of multiple main directions within the region, caused by overlapping multiple light sources or occlusion by complex structures; if the directional entropy is significantly high (greater than 2.0), it indicates a relatively dispersed directional distribution, representing texture interference or contaminated edge areas, and should not be judged as a shadow. The system combines the number of directional peaks with spatial distribution patterns for screening (statistically counting the number of peaks (local maximum values) in the directional histogram; if there are more than 3 peaks and the proportions of each peak are similar (e.g., the difference in proportion is less than 10%), it tends to be considered a non-shadow structure; the pixel positions corresponding to the peak directions are visualized as a directional distribution map; if these pixels are linearly or bandedly clustered, it indicates a shadow; if they are distributed in a grid-like or dotted pattern, it is considered interference), eliminating interference caused by high-frequency textures or edge overlap. Regions that match the shadow characteristics are marked, and shadow pseudo-dark area data is output.

[0045] Texture energy attenuation detection is performed based on the brightness attenuation region data to obtain wet artifact data; Specifically, the system moves pixel-by-pixel in the image using a preset sliding window (preferably 11×11), calculating indicators such as Gray-Level Co-occurrence Matrix (GLCM) energy, Laplacian response variance, or local brightness variance at each position. The system calculates the global mean and standard deviation of these texture energy indicators across the entire image as a baseline. The system determines whether the local texture energy of a certain brightness decay area is too low. If the texture energy of this area is lower than the global average level (e.g., below 60% of the average), and the brightness of this area remains low, without exhibiting normal reflective or shadow characteristics, then this area is considered to have blurred textures due to factors such as surface dampness, water accumulation, or contaminant coverage, and is considered an interference factor not related to actual contamination. The system marks such areas as damp artifact areas and outputs the corresponding damp artifact data.

[0046] Multi-band difference verification is performed based on the brightness attenuation region data to obtain infrared pseudo-dark area data; Specifically, the system extracts the brightness value of each brightness attenuation candidate region in the infrared image and compares it with the infrared background temperature of the surrounding region to evaluate its difference under thermal imaging. If the temperature value of a brightness attenuation region in the infrared image is basically the same as that of its surrounding background region, i.e., no temperature difference change is observed (less than 0.5), the system determines that although the region appears dark in the visible light image, it has no abnormal thermal characteristics. Such regions experience brightness reduction due to factors such as uneven illumination, surface blackening, or strong light absorption of the material, rather than substantial darkening caused by radioactive contamination or other heat sources. The system marks these brightness attenuation regions without thermal anomaly characteristics as infrared pseudo-dark areas and generates corresponding infrared pseudo-dark area data.

[0047] By fusing and judging shadow false dark area data, wet artifact data and infrared false dark area data, false contamination dark area data is obtained.

[0048] Specifically, the system uses a region merging operation to overlay and fuse the three types of artifact layers mentioned above, generating a unified pseudo-contamination dark area layer. During the merging process, the system prioritizes marking regions that simultaneously meet multiple artifact characteristics. For example, if a region exhibits both obvious directional clutter and no significant temperature difference in the infrared image, the system will consider it a non-contaminated pseudo-dark area and assign it a higher artifact confidence score. (For each region, the system extracts its features across three layers, including principal direction stability and projected contour density in the shadow layer; texture energy entropy and brightness compression coefficient in the humid layer; and regional temperature difference distribution and infrared brightness gradient in the infrared layer. If all three features point to a certain type (e.g., all exhibiting occlusion, low texture, and infrared stability), the system assigns a higher artifact confidence score (e.g., 0.8-1) to the region using a feature consistency scoring function (e.g., simple weighted sum or cross-consistency decision tree). If the three features conflict (e.g., significant temperature rise in the infrared image while the visible light is dark), the system sets the confidence score to a lower level (e.g., 0.3-0.5) or places it in the manual review queue.) This yields pseudo-contaminated dark area data.

[0049] Preferably, the shadow directionality determination specifically involves: Multi-directional gradient vectors are extracted from the brightness attenuation region data to obtain multi-directional gradient data; Specifically, in the multi-directional gradient vector extraction process, the system analyzes the edge direction information of the brightness attenuation region data using the original visible light image. The system uses the Sobel operator to process the image, extracting the brightness changes in the horizontal and vertical directions to obtain local gradient direction information in both directions. The system calculates the gradient direction angle for each pixel, standardizes and discretizes this value, dividing it into eight principal directions (e.g., 0 degrees, 45 degrees, 90 degrees, etc.), and constructs a local direction histogram. For each brightness attenuation connected region with an area greater than a set threshold, the system statistically analyzes the gradient direction distribution of all pixels within it, obtaining multi-directional gradient data.

[0050] The directional entropy is calculated from the multi-directional gradient data to obtain the directional entropy data; Specifically, in the process of calculating directional entropy, the system performs consistency analysis on the directional histogram of each brightness attenuation region based on the extracted multi-directional gradient data to represent the gradient direction distribution characteristics of that region. The system constructs a directional distribution histogram by statistically analyzing the proportion of each main direction within each region and calculates its directional consistency entropy index. A smaller directional entropy value indicates a more concentrated gradient direction in the region, corresponding to a clear shadow formed by a single light source; a moderate directional entropy value indicates interference from multiple light sources, overlapping building structures, or composite shadows caused by boundaries; a high directional entropy value indicates chaotic textures, blurred boundaries, or even the presence of actual pollution areas within the region. The system calculates the directional entropy value for each brightness attenuation region separately and generates corresponding directional entropy layer data to obtain the directional entropy data.

[0051] The directional entropy data is divided into value ranges to obtain directional entropy map data; Specifically, when the directional entropy value of a certain region is less than 1.0, it indicates that the gradient direction within that region is highly concentrated, exhibiting obvious unidirectional characteristics. The system classifies this region as a single-source shadow area and assigns it the label "Shadow_1". If the directional entropy value is between 1.0 and 2.0, it indicates that multiple main directions exist within the region, possibly due to interference from multiple light sources or the intersection of building structures. Such regions are labeled "Shadow_2", representing composite shadows or overlapping occlusions. When the directional entropy value is greater than or equal to 2.0, it indicates that the directional distribution of the region is extremely chaotic and cannot be attributed to a single lighting structure. The system labels this region as "Non_shadow", considering it a texture pseudo-dark area or a real pollution interference area. The system classifies each brightness attenuation connected region according to the above classification criteria and maps the obtained classification results to the image space, outputting directional entropy map data.

[0052] Multi-source shadow peak decomposition is performed on the directional entropy map data to obtain multi-source shadow data; Specifically, in the multi-source shadow peak decomposition process, the system analyzes the gradient direction histogram of the region marked "Shadow_2" in the direction entropy map, i.e., the region where the direction consistency entropy is in the middle range, to determine whether the region is affected by the superposition of multiple light sources. The system performs peak detection on the direction histogram of each region, preferably using a method of first-order difference combined with a set threshold to extract the dominant direction with a prominent proportion. When there are more than two dominant directions with a significant proportion in the direction distribution of a certain region (e.g., each dominant peak accounts for more than 20% of all directions), the system determines that the region is a composite shadow caused by the superposition of multiple light sources. If the angle between multiple dominant directions exceeds a certain angle (e.g., greater than 45°), it is considered that there is a light source direction conflict phenomenon in the region, caused by the mixing of natural light and artificial light sources or multiple reflections. The system checks the symmetry characteristics of the dominant direction distribution. If it finds that a certain region has strong dominant peaks in relative directions (e.g., 0° and 180°), it indicates that the region is periodically shaded by strip-shaped structures, such as striped shadows caused by vegetation gaps, fences, or grid structures. The system integrates the above judgment results and outputs a multi-source shadow data layer to obtain multi-source shadow data, recording the number of main directions, directional intensity and distribution structure characteristics in each region.

[0053] Directional temporal detection is performed based on multi-source shadow data to obtain shadow pseudo-dark area data.

[0054] Specifically, the system tracks the changes in the dominant gradient direction of each region with multi-directional attributes across consecutive frames. If a large angular shift (greater than 15 degrees) in the dominant direction of the region is detected across multiple time frames, and the change does not exhibit a clear periodic pattern, the system classifies it as a temporary projected shadow caused by a moving object (such as a robotic arm, vehicle, or pedestrian). If the direction change exhibits short-period oscillations, such as a direction switch every two to three frames, it is considered as swaying occlusion caused by lightweight objects such as leaves or branches blown by the wind. Conversely, if the dominant direction remains stable for a longer period, it indicates that the region belongs to a static shadow cast by a fixed structure (such as a building or wall). Based on the above analysis results, the system marks dynamic projection areas (including temporary projected shadows and swaying occlusions) as pseudo-dark areas of the variable occlusion type and incorporates them into the shadow pseudo-dark area data. The system outputs the shadow pseudo-dark area data.

[0055] Preferably, the pseudo-contamination map is generated as follows: Based on the data of the highlighted and dark areas of the pseudo-contamination, the pseudo-trace regions are binarized and aligned to obtain the pseudo-trace region aligned data. Specifically, the system performs uniform thresholding on both the highlight artifact layer and the dark artifact layer, marking artifact pixels as regions with a value of 1 and setting the remaining pixels to 0, forming a binary mask layer. If the two types of artifact data originate from images of different resolutions or have undergone processing in different bands, the system uses bilinear interpolation or nearest neighbor interpolation to uniformly resample them to the original image acquisition resolution, such as 640×480 or the standard size output by the actual sensor. When there are geometric deviations or optical distortions in images of different bands, spatial registration between images is performed using preset camera intrinsic matrix and distortion correction parameters, or through feature point matching algorithms (such as ORB, SIFT), calculating the alignment transformation relationship, and applying it to the two artifact mask layers to complete the spatial alignment of the artifact images. The system then merges the aligned highlight artifact and dark artifact masks to generate a unified artifact region alignment data layer, obtaining the artifact region alignment data.

[0056] The artifact region alignment data is merged to obtain the merged region data. Specifically, the system merges two artifact mask layers using a pixel-by-pixel logical "OR" operation to form a preliminary unified artifact region. The system can optionally perform neighborhood expansion processing, such as using standard structuring elements (e.g., 3×3 or 5×5) to perform image dilation on the preliminarily merged artifact layer, expanding the artifact boundary range, filling in edge breakage areas or blurred boundaries, and ensuring a safe redundancy band during processing. The dilation operation iterates 1 to 2 times. The system integrates fragmented artifact regions. If the centroid distance between two artifact regions is less than a set threshold (e.g., less than 10 pixels), the system determines that they originate from the same artifact source and performs a merging operation; conversely, if the area of ​​a certain artifact block is much smaller than the adjacent region (e.g., the area difference is greater than 20 times), the small region will be marked as isolated noise. The system outputs a region merged data layer, obtaining the region merged data.

[0057] Morphological noise removal is performed on the merged regional data to obtain regional noise-reduced data; Specifically, in the region denoising process, the system uses morphological methods and connected component analysis to remove invalid noise points based on the merged artifact layer data. The system calculates the area of ​​all connected regions of the artifacts. If the pixel area of ​​a region is less than a preset minimum area threshold (10 to 25 pixels), the region is identified as isolated noise or false positives and removed from the artifact layer. After area filtering, the system performs morphological opening operations on the remaining artifact layers, i.e., erosion followed by dilation, using a standard 3×3 structuring element. For larger artifact regions, the system can optionally perform morphological closing operations, using a 5×5 structuring element for dilation and erosion. Through these processes, the system outputs region denoising data.

[0058] Cross-modal stabilization artifact processing is performed on the regional noise reduction data to obtain pseudo-contamination map data.

[0059] Specifically, the system calculates the brightness variance in the visible light image and the temperature variance in the infrared image for each artifact block in the regional noise reduction data. It then constructs a cross-modal consistency index using preset weights, i.e., a weighted regression of the brightness and temperature variances to obtain the modal consistency index. If a region exhibits minimal temperature variation in the infrared image (e.g., below 0.3℃) but significant brightness fluctuations in the visible light image, the system determines that the region is an artifact caused by strong reflection. If the variations in both modes are relatively stable, it belongs to a fixed-structure artifact, such as surface textures, device edges, or inherent dark patterns in the material. The system combines temporal information to compare the position and area of ​​artifact regions in consecutive image frames. If an artifact block highly overlaps in position (overlap greater than 90%) in two frames and its area changes by no more than 10%, it is identified as a stable artifact region and assigned a higher artifact weight. Optionally, the system can also construct a artifact weight map based on the aforementioned cross-modal consistency index, assigning an artifact weight value between 0 and 1 to each pixel in the image. This can also be regarded as an artifact confidence map, that is, normalizing the cross-modal consistency index corresponding to the artifact block to [0,1]; and using this value as the unified confidence of all pixels in the region. The system outputs structured artifact contamination map data, including an artifact region mask map, an artifact consistency index map / artifact confidence map / artifact weight map reflecting artifact confidence, and attribute information of artifact connected regions (such as area, main direction, shape features, etc.).

[0060] Preferably, S2 specifically comprises: S21: Estimate the illumination field based on the pseudo-contamination map data to obtain the illumination field data; Specifically, the system uses pseudo-contamination map data / artifact masks to mask identified artifact areas such as bright reflective areas and shadow areas in the original acquired image, ensuring that illumination estimation is performed only in non-artifact areas. In the remaining effective image area, the system uses background fitting methods to model the brightness channel as a whole. For example, large-scale Gaussian filtering techniques are used to smooth the image brightness through large convolution kernels, thereby obtaining a low-frequency brightness distribution map of the entire image. For scenes with undulating terrain and large variations in illumination gradients, the system can also use polynomial surface fitting methods (such as quadratic surfaces) to fit and model the image brightness. The system outputs a continuous, smooth brightness baseline map / illumination field data, representing the illumination structure of the image under ideal, interference-free conditions.

[0061] S22: Based on the illumination field data, reflectance artifact suppression is performed on the pseudo-contamination map data to obtain artifact suppression map data; Specifically, the system locates specular reflection regions with high brightness characteristics in the pseudo-contamination map data. This subset, obtained through brightness thresholding or structural feature filtering, represents a subset of the pseudo-contamination map data / artifact mask. The system performs illumination normalization on pixels within these regions, normalizing the ratio of brightness values ​​in the pseudo-contamination map data to the estimated illumination field baseline. For regions where brightness abrupt changes still exist at some specular reflection edges, the system optionally performs local contrast limiting. By calculating the difference between a pixel value and its neighborhood average, pixels with differences exceeding a set threshold are corrected, bringing their brightness back to the neighborhood average level. The system outputs the image after reflection suppression, serving as the artifact suppression map data.

[0062] S23: Perform texture compensation for wet areas based on artifact suppression map data to obtain area compensation data; Specifically, the system extracts artifact blocks marked as damp areas from the pseudo-contamination map. These areas are characterized by lower brightness and reduced texture energy (by comparing the average brightness of pixels within the damp area (e.g., the average grayscale value in the grayscale image) with adjacent normal areas; if the average brightness is lower than the reference area by a certain percentage (e.g., 20%), it can be considered as having lower brightness; the Laplacian variance is less than a threshold). These areas are prone to local texture blurring due to water accumulation, wet marks, or contaminants. The system sets a certain width (e.g., 20 pixels) around each damp area. After excluding artifact areas / artifact suppression map data within this extended range, it selects neighboring areas with continuous structure and consistent material as reference samples. Texture feature vectors are extracted from the reference areas, which may include parameters such as the Gray-Level Co-occurrence Matrix (GLCM) index, Local Binary Pattern (LBP), and Laplacian variance. In the compensation stage, the system uses the Non-Local Means (NLM) algorithm or the PatchMatch method based on block matching to migrate the texture structure in the selected reference area to the corresponding damp artifact area. The system outputs the texture-compensated image data, obtaining the region compensation data.

[0063] S24: Perform shadow compensation on the regional compensation data to obtain the interference-suppressed image data.

[0064] Specifically, the system locates artifact regions / shadow pseudo-dark areas of all shadow types and processes each region separately. For each shadow block, the system extracts the average brightness of its surrounding non-artifact regions as a reference background brightness, while simultaneously calculating the average brightness of the current shadow block / shadow pseudo-dark area data itself, and uses this to apply linear gain compensation to the region, ensuring that its overall brightness level is consistent with the surrounding environment. If color shifts are detected in certain shadow areas (the system calculates the mean difference between the shadow area and its surrounding reference area in each RGB channel; if the mean difference of any channel exceeds a set threshold (e.g., >15% or a set standard deviation multiple), it is considered a color shift), the system can switch to the HSV color space and adjust the luminance channel (V channel) separately. Specifically, it calculates the average V channel values ​​of the shadow area and its surrounding reference area. If there is a significant difference, it calculates the ratio of the two as a luminance gain factor. The system applies this luminance gain factor to the V channel pixel values ​​of the shadow area, linearly stretching or scaling them to make the luminance level of the shadow area consistent with the surrounding area. After adjustment, the system retains the original image's hue and saturation channels, only recombining the modified luminance channel with the original hue and saturation channels, converting it back to an RGB image, and outputting the luminance-coordinated image data. For areas with severely suppressed brightness and a graying trend, the system can optionally reconstruct their color information using an edge-aware interpolation algorithm or an image restoration model based on a generative adversarial network (GAN). The system performs gradient blending processing in the shadow edge areas, using Gaussian weights or... The system employs a blending method to achieve a smooth transition in brightness. The system outputs image data after shadow compensation and fusion, i.e., the image data after interference suppression.

[0065] Preferably, S3 specifically comprises: S31: Based on the data after interference suppression, perform background artifact masking to obtain background artifact masking data; Specifically, the system directly applies the pseudo-contamination map data / artifact mask to the interference-suppressed image data / interference-suppressed image data. Pixels marked as artifacts in the mask are set as "undetectable regions" or assigned special identifiers (such as null values, ignore flags, etc.), thereby eliminating this artifact interference at the image data level and retaining only non-artifact regions as the basis for subsequent input. The system can optionally perform a morphological operation on the edges of the artifact regions, namely dilation followed by erosion, to construct a softly transitioned masking edge band. The system outputs background artifact masking data.

[0066] S32: Extract contamination texture features from the background artifact masking data to obtain contamination texture feature data; Specifically, the system constructs a fixed-size local window (using 11×11 pixels) around each valid pixel, and extracts multiple types of texture and structural features within this window. This includes, but is not limited to, using gradient operators (such as Sobel or Scharr) to obtain the local gradient magnitude and direction information of the region; calculating gray-level co-occurrence matrix related indicators, such as energy, contrast, and entropy; evaluating edge density using Laplacian variance to determine the region's sharpness; constructing a histogram using Local Binary Patterns (LBP) to capture micro-texture patterns; and simultaneously extracting the local brightness mean and standard deviation. The system also records structural features such as boundary compactness and shape stretching for each candidate region. The system aggregates these features according to spatial location into a multi-dimensional feature vector set, constituting the contamination texture feature data.

[0067] S33: Generate pollution candidate regions based on pollution texture feature data to obtain preliminary region data; Specifically, the system standardizes the multidimensional texture feature vector of each pixel and determines whether it is a feature outlier based on empirical thresholds or statistical distribution rules (such as Z-score). If a pixel exhibits behavior deviating from the average level in multiple feature dimensions simultaneously (e.g., at least two dimensions exceed the set outlier threshold), the system marks it as a potential contamination outlier. The system organizes and classifies these outliers using spatial aggregation, preferably using a density-based clustering algorithm (such as DBSCAN). By setting a proximity distance threshold and a minimum number of points, spatially close and feature-similar outliers are merged into several contamination candidate blocks. As an alternative, a morphological region growing method can be used, based on the 8-neighborhood connection rule, to expand the regions of outliers and form coherent contamination candidate blocks. The system performs area filtering on the generated contamination candidate blocks, deleting tiny patches with areas smaller than a set threshold (e.g., less than 20 pixels) to avoid misclassifying isolated outliers as contamination areas. The system outputs preliminary contamination candidate region data / preliminary region data, including a unique number, center position, pixel area, and shape factor for each candidate block.

[0068] S34: Perform cross-band verification on the preliminary regional data to obtain the polluted area data.

[0069] Specifically, the system maps the mask layer corresponding to the candidate / preliminary region data to the infrared image space, ensuring precise spatial correspondence between the candidate regions in images across different spectral bands. After registration, the system extracts the temperature sub-map of each candidate region in the infrared image and calculates the average temperature value within the region. The system sets an extended neighborhood of a certain width (e.g., 10 pixels) around the region. By comparing the internal temperature of the region with the surrounding background temperature, the system evaluates whether the candidate region possesses obvious thermal anomaly characteristics. If the temperature difference is insufficient (e.g., less than 0.3℃), it indicates that the region does not exhibit thermal response characteristics in the infrared image, and the visual anomalies are caused by non-thermal interference factors such as surface reflection, oil stains, and water stains; the system will discard such regions. Conversely, if the temperature difference is significant, and the region appears as a clear cold or hot spot in the infrared image, consistent with texture anomalies in the visible light image, the system considers it a target region with genuine contamination. The system integrates and outputs all candidate regions that have passed infrared verification, generating a contaminated region data layer to obtain the contaminated region data.

[0070] Preferably, S4 specifically comprises: S41: Perform attention-based fine segmentation on the polluted area data to obtain finely segmented data; Specifically, the system takes the image patch corresponding to the contaminated area and its mask as input, and optionally introduces an artifact intensity map as auxiliary information. The input includes the image data after interference suppression and an optional artifact weight map. The feature extraction module uses a multi-layer convolutional neural network structure, such as ResNet or a lightweight UNet encoder, to generate feature maps at multiple scales to fully capture the texture and morphological features of the contaminated area. The system calculates channel attention and spatial attention separately. For example, based on the channel attention step / module, global average pooling and a multi-layer perceptron network are used to obtain the importance weight of each feature channel, and the channel intensity is adjusted accordingly. , This is a channel attention weight map. The Sigmoid activation function is used. This is the second-layer channel weight mapping matrix. It is a linear correction function. This is the first-layer channel weight mapping matrix, with preset values. For global average pooling, The input data is used for spatial attention; the spatial attention module, based on the spatial fusion result of average pooling and max pooling, constructs an attention map for each pixel location through convolution operations, i.e. , This is a spatial attention weight map. The Sigmoid activation function is used. For convolution, For channel average pooling, For channel splicing, For channel max pooling, The input data is used as input data. After obtaining the channel and spatial attention weights, the system fuses feature maps at multiple scales to generate feature maps with a dual attention mechanism. If the system provides an artifact intensity map, it suppresses severely disturbed region features through inverse weight suppression of artifact regions, preventing false positives caused by artifacts. The system gradually restores the spatial resolution of the image through a multi-layer decoding network and outputs a probability map / contamination probability map of contaminated regions by combining the skip connection mechanism in the encoding stage. The system applies an adaptive or fixed threshold strategy to binarize the probability map to obtain finely segmented data.

[0071] S42: Perform edge optimization on the finely segmented data to obtain the boundary data of the polluted area; Specifically, the system performs edge extraction on the mask layer corresponding to the finely segmented data, optionally using the Canny algorithm or gradient difference-based methods to obtain a preliminary boundary map, while combining the brightness gradient data of the original image. The system constructs a boundary optimization model, including a Conditional Random Field (CRF) or Graph Cut model. The function in the model consists of two parts: firstly, constructing region confidence based on the brightness information and contamination probability map of the original image; secondly, establishing smoothing constraints through gradient consistency between adjacent pixels. In the specific execution process, if the CRF model is used, the system constructs a fully connected dense-CRF structure, treating each pixel as a node and defining a function. The function contains two parts: a region term, representing the confidence that the current pixel belongs to a contaminated region based on the contamination probability map and brightness difference; and a smoothing term, penalizing regions with discontinuous or messy boundaries based on the consistency of brightness or color difference between adjacent pixels. If the Graph Cut model is used, the system uses pixels as graph nodes, with edges connecting them to represent the similarity weights between pixel pairs, and constructs a minimum cut to delineate the boundary between the contaminated region and the background region. After optimization, the system applies a morphological closing operation (such as a 5×5 structuring element) to the boundary results and uses a contour fitting algorithm (such as Ramer-Douglas-Peucker) to simplify and smooth the boundary shape. The system outputs the boundary data of the contaminated area.

[0072] S43: Perform artifact reverse verification based on the contaminated area boundary data to obtain contaminated area identification data.

[0073] Specifically, for each finely segmented contaminated region, the system extracts several feature information for artifact detection within its boundary band (e.g., extending outwards by 3 to 5 pixels). These features include: whether the region's boundary has a high degree of overlap with the edge of an existing artifact mask (e.g., exceeding 80%); whether the boundary is located at a position with a large change in brightness gradient in the original image, and whether its central region does not exhibit infrared thermal anomalies (temperature difference below a preset threshold); whether the texture features within the region are too uniform (e.g., the mean feature based on a local binary pattern exceeds a set threshold); and whether the region appears repeatedly in multiple frames of data but does not exhibit temporal expansion, i.e., its morphology is highly stable in the time series. Based on these features, the system sets a judgment rule: if a region simultaneously satisfies two or more artifact features, it can be determined as a false contaminated region. For applications with a large amount of labeled data, a pre-trained shallow neural network model can be introduced for binary classification to learn artifact features for discrimination. When samples are insufficient or the model is unstable, rule-based judgment can be used as an alternative. For regions identified as artifacts, the system removes them from the fine segmentation mask using mask subtraction; alternatively, these regions can be marked as "regions to be reviewed." The system outputs the contaminated region identification results, yielding contaminated region identification data.

[0074] Preferably, attention-based fine segmentation specifically includes: Deep texture features and edge features are extracted from the polluted area data to obtain deep texture data and edge feature data, respectively. Specifically, the system constructs an input tensor. Input data includes interference-suppressed image data and contaminated area data, with optional additions of pseudo-contamination image data and infrared temperature difference maps as additional channels. The constructed input tensor has multiple channels, with sizes of 1 batch, 3 to 5 channels, and a spatial resolution of 256×256. The system employs a dual-branch network structure to process texture and edge features separately. The texture branch serves as the main path, utilizing a convolutional neural network with deep semantic capabilities (such as ResNet18 or MobileNet) to extract semantic texture features at multiple levels, resulting in deep texture feature data with a large receptive field. The edge branch serves as an auxiliary path, employing a shallow, lightweight network structure—a three-layer convolutional network—and combining image edge enhancement priors (such as Sobel, Laplacian, or Canny operators) to strengthen edge information in the image and extract edge features. The system outputs deep texture feature data and edge feature data respectively. The former possesses strong target semantic expression capabilities, suitable for characterizing the texture differences of contaminated areas; the latter highlights the spatial localization capability of region boundaries, facilitating further structure-sensitive identification and correction operations.

[0075] Channel attention weights and spatial attention weights are calculated on deep texture data and edge feature data to obtain channel attention weight data and spatial attention weight data. Specifically, during the channel attention weight calculation, the system performs global average pooling and global max pooling operations on each feature map channel (containing deep texture data and edge feature data) to extract the average and peak response values ​​of that channel across the entire spatial range. The two pooling results are concatenated and input into a multilayer perceptron (MLP) consisting of two fully connected neural networks. After processing by a non-linear activation function, the channel attention weight data is obtained. During the spatial attention weight calculation, the system performs average pooling and max pooling operations on the channel dimensions of the feature map (containing deep texture data and edge feature data) to obtain two single-channel spatial response maps. These are concatenated and input into a 7×7 convolutional module. After processing by an activation function, a spatial attention weight map is generated. The system outputs both channel attention weight data and spatial attention weight data.

[0076] Based on the channel attention weight data and spatial attention weight data, artifact weight fusion suppression is performed on the polluted area data to obtain artifact weight data. Specifically, the system calls a pre-calculated artifact weight map, which represents the confidence level of each pixel location as an artifact region, ranging from 0 to 1, with higher values ​​indicating a greater likelihood of artifacts. The system combines the original contaminated region feature map with channel attention, spatial attention, and the artifact weight map to construct a fused image. Each channel feature is weighted according to the channel attention weight; the intensity at different locations in the image is adjusted / weighted based on the spatial attention weight; and the inverse weighting of the artifact weight map suppresses the response amplitude of high-confidence artifact regions, thus highlighting contaminated regions and weakening artifact regions. The system outputs artifact weight data.

[0077] The decoder performs fine segmentation on the data contaminated by the artifact weight data to obtain finely segmented data.

[0078] Specifically, the system employs a lightweight UNet-style decoding network consisting of multiple decoding layers. Each decoding layer includes upsampling, feature concatenation, convolution, batch normalization, and non-linear activation. A skip connection mechanism concatenates the shallow feature maps extracted in the encoder stage with the output of the current decoder layer. In the decoder's output stage, the system generates a probability map of each pixel being a contaminated region. Based on this map, the system sets a fixed or adaptive threshold (e.g., 0.5) to determine the prediction result for each pixel, constructing a fine-segmentation mask. Pixels with values ​​higher than the threshold are identified as contaminated regions, while those lower are excluded. Through this decoder fine-segmentation process, the system obtains finely segmented data.

[0079] Preferably, this application also provides a radioactive contamination region identification system based on image segmentation, used to perform the radioactive contamination region identification method based on image segmentation as described above, the radioactive contamination region identification system based on image segmentation includes: The pseudo-contamination map generation module is used to acquire multi-band image data; perform pseudo-contamination highlight area detection and pseudo-contamination dark area identification on the multi-band image data to obtain pseudo-contamination highlight area data and pseudo-contamination dark area data respectively; generate a pseudo-contamination map based on the pseudo-contamination highlight area data and pseudo-contamination dark area data to obtain pseudo-contamination map data. The interference suppression processing module is used to perform interference suppression processing on the pseudo-contamination map data to obtain interference-suppressed map data. The pseudo-contamination area shielding module is used to shield pseudo-contamination areas based on the interference suppression image data to obtain contamination area data. The fine segmentation module for contaminated areas is used to perform fine segmentation of contaminated area data to obtain contaminated area identification data.

[0080] Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended application documents rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of the equivalents of the application documents be incorporated into the invention.

[0081] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A method for identifying radioactive contamination regions based on image segmentation, characterized in that, Includes the following steps: S1: Acquire multi-band image data; perform false contamination highlight area detection and false contamination dark area identification on the multi-band image data to obtain false contamination highlight area data and false contamination dark area data respectively; generate a false contamination map based on the false contamination highlight area data and false contamination dark area data to obtain false contamination map data. S2: Perform interference suppression processing on the pseudo-contamination map data to obtain the interference-suppressed map data; S3: Based on the data after interference suppression, perform pseudo-contamination area shielding to obtain contaminated area data; S4: Perform fine segmentation of the polluted area data to obtain polluted area identification data.

2. The method according to claim 1, characterized in that, The specific steps for detecting false contamination in highlighted areas are as follows: Preliminary brightness field data is obtained by performing preliminary brightness field screening on multi-band image data; Specular reflection discrimination processing is performed based on the preliminary brightness field data to obtain specular reflection data; Color saturation and reflection are verified based on preliminary brightness field data to obtain color saturation data; Infrared reflectance data was obtained by conducting infrared reflectance comparison screening based on preliminary brightness field data; Brightness edge diffusion is detected based on preliminary brightness field data to obtain brightness edge diffusion data; By fusing specular reflection data, color saturation data, infrared reflection data, and brightness edge diffusion data, pseudo-contamination highlight area data is obtained.

3. The method according to claim 1, characterized in that, The specific steps for identifying false contamination dark areas are: The brightness attenuation region data of the multi-band image data is initially screened to obtain the brightness attenuation region data. Shadow directionality is determined from the brightness attenuation region data to obtain shadow pseudo-dark area data; Texture energy attenuation detection is performed based on the brightness attenuation region data to obtain wet artifact data; Multi-band difference verification is performed based on the brightness attenuation region data to obtain infrared pseudo-dark area data; By fusing and judging shadow false dark area data, wet artifact data and infrared false dark area data, false contamination dark area data is obtained.

4. The method according to claim 3, characterized in that, The shadow directionality determination is specifically as follows: Multi-directional gradient vectors are extracted from the brightness attenuation region data to obtain multi-directional gradient data; The directional entropy is calculated from the multi-directional gradient data to obtain the directional entropy data; The directional entropy data is divided into value ranges to obtain directional entropy map data; Multi-source shadow peak decomposition is performed on the directional entropy map data to obtain multi-source shadow data; Directional temporal detection is performed based on multi-source shadow data to obtain shadow pseudo-dark area data.

5. The method according to claim 1, characterized in that, The pseudo-contamination map is generated as follows: Based on the data of the highlighted and dark areas of the pseudo-contamination, the pseudo-trace regions are binarized and aligned to obtain the pseudo-trace region aligned data. The artifact region alignment data is merged to obtain the merged region data. Morphological noise removal is performed on the merged regional data to obtain regional noise-reduced data; Cross-modal stabilization artifact processing is performed on the regional noise reduction data to obtain pseudo-contamination map data.

6. The method according to claim 1, characterized in that, S2 specifically refers to: Illumination field data is obtained by estimating the illumination field based on the pseudo-contamination map data; Based on the illumination field data, reflection artifacts are suppressed in the pseudo-contamination map data to obtain artifact suppression map data; Texture compensation for wet areas is performed based on artifact suppression map data to obtain area compensation data; Shading compensation is applied to the regional compensation data to obtain the interference-suppressed image data.

7. The method according to claim 1, characterized in that, S3 specifically refers to: Background artifact masking data is obtained by masking background artifacts based on the data after interference suppression. Contamination texture features are extracted from the background artifact masking data to obtain contamination texture feature data; Based on the pollution texture feature data, pollution candidate regions are generated to obtain preliminary region data; Cross-band verification was performed on the preliminary regional data to obtain data on the polluted areas.

8. The method according to claim 1, characterized in that, S4 specifically refers to: Attention-based fine segmentation is performed on the polluted area data to obtain finely segmented data. Edge optimization is performed on the finely segmented data to obtain the boundary data of the polluted area; Artifact inversion verification is performed based on the boundary data of the contaminated area to obtain the contaminated area identification data.

9. The method according to claim 8, characterized in that, Attention-based fine segmentation specifically refers to: Deep texture features and edge features are extracted from the polluted area data to obtain deep texture data and edge feature data, respectively. Channel attention weights and spatial attention weights are calculated on deep texture data and edge feature data to obtain channel attention weight data and spatial attention weight data. Based on the channel attention weight data and spatial attention weight data, artifact weight fusion suppression is performed on the polluted area data to obtain artifact weight data. The decoder performs fine segmentation on the data contaminated by the artifact weight data to obtain finely segmented data.

10. A system for identifying radioactive contamination regions based on image segmentation, characterized in that, For performing the image segmentation-based radioactive contamination region identification method as described in claim 1, the image segmentation-based radioactive contamination region identification system comprises: The pseudo-contamination map generation module is used to acquire multi-band image data; perform pseudo-contamination highlight area detection and pseudo-contamination dark area identification on the multi-band image data to obtain pseudo-contamination highlight area data and pseudo-contamination dark area data respectively; generate a pseudo-contamination map based on the pseudo-contamination highlight area data and pseudo-contamination dark area data to obtain pseudo-contamination map data. The interference suppression processing module is used to perform interference suppression processing on the pseudo-contamination map data to obtain interference-suppressed map data. The pseudo-contamination area shielding module is used to shield pseudo-contamination areas based on the interference suppression image data to obtain contamination area data. The fine segmentation module for contaminated areas is used to perform fine segmentation of contaminated area data to obtain contaminated area identification data.

Images