Infrared image multi-feature representation method based on physical property decoupling

Abstract

This invention discloses a multivariate representation method for infrared images based on physical property decoupling, belonging to the field of infrared image processing and intelligent sensing technology. The method includes: extracting appearance representations from the original raw image, truncating the effective grayscale range and mapping it to a preset interval to generate a first representation expressing the brightness and contours of the scene; establishing a grayscale-thermal radiation response mapping based on the blackbody radiation law, and generating a second representation expressing the combined effect of temperature and emissivity through nonlinear inverse transformation; and extracting and reconstructing high-frequency components through multi-scale frequency domain decomposition to generate a third representation expressing edges and local changes. The three representations are input into an encoding network for independent feature encoding, and after feature space alignment to ensure a unified spatial size and channel dimension, they are complementary and fused to generate a unified infrared multivariate representation. This invention provides a comprehensive and physically meaningful basic representation for infrared target detection in complex scenes.

Description

Technical Field

[0001] This invention relates to the field of infrared image processing and intelligent sensing technology, and more specifically to a multivariate representation method for infrared images based on physical attribute decoupling. Background Technology

[0002] Currently, infrared images can reflect the difference in thermal radiation between a target and the background, possessing advantages such as nighttime imaging, low-light adaptability, and strong perception capabilities under complex weather conditions. They are widely used in scenarios such as road monitoring, unmanned platform sensing, and search and rescue in complex environments. Unlike visible light images, single-band infrared images typically provide only one grayscale observation value. This observation value is not a direct expression of a single physical factor, but rather formed by the combined effects of factors such as object surface temperature, material emissivity, environmental radiation, atmospheric transmission, and sensor response. Therefore, similar or identical grayscale responses may correspond to different temperatures, materials, and imaging conditions, leading to the problem that infrared images contain interrelated physical information and their physical meaning is difficult to express directly.

[0003] Existing infrared target detection methods typically process infrared images as ordinary single-channel grayscale images, or copy them into three channels and then use feature extraction methods designed for RGB images. Furthermore, for ease of display and processing, high-bit-width raw infrared images usually need to be compressed to a lower grayscale range, which can easily weaken subtle response differences between the target and the background. These methods mainly utilize the target's apparent grayscale and contour information, failing to fully leverage the relationship between grayscale response and thermal radiation patterns, and also struggling to simultaneously consider regional thermal properties and local structural information such as edges and textures. In challenging scenarios such as low contrast, complex background interference, and distant, weak targets, problems such as insufficient target saliency, blurred structural boundaries, missed detections, false detections, and positioning errors are prone to occur.

[0004] Furthermore, determining physical factors such as temperature, material emissivity, and environmental radiation solely based on single-band infrared images is an underdetermined problem, making it difficult to obtain unique and reliable results. The characterizations obtained by different processing methods also show significant differences in physical meaning, response range, distribution scale, and feature dimensions, making it difficult to fully leverage their complementary effects when directly spliced ​​or superimposed.

[0005] Therefore, there is an urgent need for a method that, based on the laws of infrared physical imaging, constructs apparent representations, thermophysical property representations, and structural detail representations from infrared images containing the original thermal radiation intensity, and performs feature space alignment and complementary fusion of different representations, thereby forming a multi-dimensional representation of infrared images with clearer physical meaning and more comprehensive information. Summary of the Invention

[0006] In view of the above problems, the present invention proposes a multivariate characterization method for infrared images based on physical property decoupling to overcome or at least partially solve the above problems.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] In a first aspect, embodiments of the present invention provide a multivariate characterization method for infrared images based on physical property decoupling, comprising the following steps: S1. Perform appearance representation extraction on the original infrared Raw image, truncate the effective grayscale range of the original infrared Raw image and map it to a preset grayscale range to generate a first representation for expressing the relative brightness and overall contour information of the scene. S2. Perform thermophysical property characterization extraction on the original infrared Raw image, establish the mapping relationship between the pixel gray value of the Raw image and the thermal radiation response based on the blackbody radiation law, and perform nonlinear inverse transformation on the mapping result to generate a second characterization to express the information of the combined effect of temperature and material emissivity. S3. Perform structural detail characterization extraction on the original infrared Raw image, extract high-frequency components at each scale through multi-scale frequency domain decomposition and reconstruct them to generate a third characterization to express target edge and local change information. S4. Input the first representation, the second representation, and the third representation into the corresponding encoding network for independent feature encoding to obtain the first encoding feature, the second encoding feature, and the third encoding feature; perform feature space alignment processing on the first encoding feature, the second encoding feature, and the third encoding feature to make the three have a unified spatial size and channel dimension; S5. The first, second, and third coding features, after feature space alignment, are complementary and fused to generate a unified infrared multivariate representation.

[0009] Furthermore, S1 specifically includes: S11. Based on the grayscale distribution statistical characteristics, scene dynamic range, and sensor response range of the original infrared Raw image, determine the low grayscale truncation boundary. and high grayscale cutoff boundary ; S12, according to Raw image original pixel values D Perform an outward cutoff at the boundary; S13, truncate the effective grayscale range [ , Map the image to a preset display grayscale range to obtain the first representation; in, This is the truncated response value. This is the clipping function.

[0010] Furthermore, S2 specifically includes: S21. Obtain prior calibration parameters, including a low-temperature reference point. and high temperature reference point ,in, and These are the pixel response values ​​in the Raw image corresponding to the low-temperature reference point and the high-temperature reference point, respectively. S22. According to the Stefan-Boltzmann radiation law, from the reference temperature... and Calculate the corresponding blackbody radiation response values ​​respectively. L l = σT l 4 and L h = σT h 4 ,in σ It is the Stefan-Boltzmann constant; S23. Based on the aforementioned low-temperature reference point and high-temperature reference point, establish the Raw image pixel response value. D With thermal radiation response value L Linear mapping relationship between them: L = aD + b The slope bias ; S24. Calculate the response values ​​of each pixel in the original infrared raw image. D By substituting the linear mapping relationship pixel by pixel, the thermal radiation response map is obtained. L ( D ), and the thermal radiation response map L ( D Apply nonnegative constraints: L + =max( L ( D ),0); S25, Response diagram to nonnegative thermal radiation L + Perform a fourth root transformation: To obtain the second characterization M .

[0011] Furthermore, the second characterization M Satisfying M≈ ,in T The surface temperature of the object. εThe second characterization is used to uniformly characterize the combined contribution of object temperature and material emissivity to infrared radiation, where is the surface emissivity of the object.

[0012] Furthermore, S3 specifically includes: S31. Perform multi-scale wavelet decomposition on the original infrared Raw image, and in the... i Extracting high-frequency components in the horizontal direction on a scale H i Vertical high-frequency components V i and diagonal high frequency components D i ; S32, the horizontal high-frequency components at each scale H i Vertical high-frequency components V i and diagonal high frequency components D i To restore the original image spatial resolution, the restored horizontal components at the same scale are obtained. H i ', Vertical component V i 'and diagonal components D i '; S33, in the i In terms of scale, according to Synthesize the detail response map at this scale E i ; S34. Combine detail response maps of different scales according to preset weights. E i The third representation is generated.

[0013] Furthermore, the third characterization ,in, , α and β All are non-negative weights. and These are the detail response maps for the first and second scales, respectively.

[0014] Furthermore, in step S4, feature space alignment processing is performed on the first coded feature, the second coded feature, and the third coded feature, specifically including: S41. Perform appearance coding, thermophysical property coding, and structural coding on the first characterization, the second characterization, and the third characterization respectively to obtain the first coding feature, the second coding feature, and the third coding feature; S42. Perform feature value range mapping on the first coding feature, the second coding feature and the third coding feature respectively, so that the response value distribution ranges of the three are mutually adapted; S43. The first coding feature, the second coding feature and the third coding feature are respectively adjusted in terms of spatial scale and channel dimension mapping so that the three have a unified spatial size and the same number of channels.

[0015] Furthermore, S5 specifically includes: S51. The first encoded feature after alignment is used as the main feature to maintain the target main structure and scene spatial relationship; S52. The aligned second coding feature is used as supplementary information for thermophysical properties to provide a joint basis for judging temperature and material emissivity; S53. The aligned third encoded features are used as supplementary structural information to provide target contours, edges and local texture details; S54, according to Weighted fusion is performed, where, The fusion weight coefficients for the second encoded features. The fusion weight coefficients for the third coding feature. F For the unified infrared multivariate characterization, F G This is the first coding feature. F M This is the second coding feature. F S This is the third coding feature.

[0016] Furthermore, the fusion weight coefficients and The value can be dynamically determined based on the response intensity distribution of the second and third coding features, or preset to a fixed value based on the target application scenario.

[0017] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a multivariate characterization method for infrared images based on physical property decoupling, which has the following beneficial effects: Based on the principles of infrared radiation imaging, this invention constructs appearance representation, thermophysical property representation, and structural detail representation, respectively. Through feature encoding, range mapping, and spatial dimension alignment, these three features are mapped to a unified feature representation space and then fused complementaryly. This results in a unified infrared multivariate representation that considers scene appearance, thermophysical properties, and local detail information. It overcomes the limitations of single grayscale representation and the coupling of multiple physical information, providing a more comprehensive and physically meaningful basic representation for the detection and identification of complex backgrounds, low contrast, and distant, small targets. Attached Figure Description

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

[0019] Figure 1 This is a flowchart of the infrared image multivariate characterization method provided in the embodiments of the present invention; Figure 2 This is a schematic diagram of the infrared image multivariate characterization method provided in the embodiments of the present invention; Figure 3 This is a schematic diagram illustrating the process of obtaining the appearance characterization, thermophysical property characterization, and structural detail characterization provided in the embodiments of the present invention; Figure 4 This is a schematic diagram of the feature space alignment and fusion process of the three-way representation provided in an embodiment of the present invention. Detailed Implementation

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

[0021] This invention discloses a multivariate characterization method for infrared images based on physical property decoupling, such as... Figure 1 As shown, it includes the following steps: S1. Perform appearance representation extraction on the original infrared Raw image, truncate the effective grayscale range of the original infrared Raw image and map it to a preset grayscale range to generate a first representation for expressing the relative brightness and overall contour information of the scene. S2. Perform thermophysical property characterization extraction on the original infrared Raw image, establish the mapping relationship between the pixel gray value of the Raw image and the thermal radiation response based on the blackbody radiation law, and perform nonlinear inverse transformation on the mapping result to generate a second characterization to express the information of the combined effect of temperature and material emissivity. S3. Perform structural detail characterization extraction on the original infrared Raw image, extract high-frequency components at each scale through multi-scale frequency domain decomposition and reconstruct them to generate a third characterization to express target edge and local change information. S4. Input the first representation, the second representation, and the third representation into the corresponding encoding network for independent feature encoding to obtain the first encoding feature, the second encoding feature, and the third encoding feature; perform feature space alignment processing on the first encoding feature, the second encoding feature, and the third encoding feature to make the three have a unified spatial size and channel dimension; S5. The first, second, and third coding features, after feature space alignment, are complementary and fused to generate a unified infrared multivariate representation.

[0022] The following is a detailed description of each of the above steps; Step S1: Perform appearance characterization extraction on the raw infrared RAW image; This step aims to extract the apparent representation from the Raw infrared image that reflects the relative brightness and overall contour information of the scene, as the basis for subsequent fusion. In this invention, "apparent representation" refers to the grayscale representation obtained by truncating the grayscale response of the Raw infrared image with an effective grayscale range and grayscale mapping, which is used to express the relative brightness and overall contour information of the scene. It is denoted as G, i.e., the first representation.

[0023] Specifically, such as Figure 2 As shown in the schematic diagram, apparent characteristics are extracted from the original infrared Raw image. G Characterization of thermophysical properties M and structural detail representation S Subsequently, the various representations are unified and complementary in the multi-dimensional feature space to ultimately generate a unified infrared multi-dimensional representation.

[0024] The unification of multi-feature space includes range adaptation, scale matching, and channel mapping.

[0025] like Figure 3 As shown, step S1 includes the following sub-steps: S11. Determining the grayscale range boundary; The low grayscale boundary is determined based on the grayscale distribution, scene dynamic range, and sensor response range of the raw infrared image. and high grayscale boundaries ; For example, low and high quantiles can be selected based on the cumulative gray-level distribution of the Raw image, and the corresponding gray-level response values ​​can be used as low and high gray-level boundaries, respectively.

[0026] S12. Truncate the response outside the boundary:

[0027] in, This is the response of raw pixel values ​​from a raw infrared image. Dc This is the truncated response value. This is the clipping function.

[0028] S13. Mapping to a preset grayscale range to obtain the apparent representation. G ; For example, under 8-bit image processing conditions, the preset grayscale range can be set to [0, 255].

[0029] The truncated effective grayscale range is mapped to a preset grayscale range to obtain an apparent representation of the relative brightness of the target and background, the overall outline, and the scene contrast. G Apparent representation G It provides basic scene perception information, which serves as a foundation for subsequent scene perception representation in fusion.

[0030] Step S2: Perform thermophysical property characterization extraction on the original infrared Raw image; Based on the blackbody radiation law, this step establishes a mapping relationship between the pixel grayscale values ​​of the Raw image and the thermal radiation response, and generates a second characterization reflecting the combined effect of temperature and material emissivity through nonlinear inverse transformation. In this invention, "thermophysical property characterization" refers to the characterization used to express the information of the combined effect of temperature and material emissivity obtained by establishing a mapping relationship between the pixel grayscale values ​​of the Raw image and the thermal radiation response based on the blackbody radiation law and performing a nonlinear inverse transformation on the mapping result. It is denoted as M, i.e., the second characterization.

[0031] like Figure 3 As shown, S2 specifically includes: S21. Reference point calibration; Based on the prior temperature of common objects, a low-temperature reference point is selected. and high temperature reference point .

[0032] Furthermore, both the low-temperature and high-temperature reference points are selected within the effective operating range of the infrared imaging equipment and determined based on the specific sensor and application scenario. As an optional calibration method, taking a road monitoring scenario as an example, the low-temperature reference point can be selected from a stable low-temperature background in the scenario, such as a low-temperature atmosphere calibrated at 273.15K (0 degrees Celsius); the high-temperature reference point can be selected from a stable high-temperature heat source, such as a human body surface calibrated at 308.15K (35 degrees Celsius). The reference temperature can be determined through prior knowledge of the temperature in the application scenario, or by measuring the temperature of the calibration object using a temperature measuring device, and by reading the grayscale response value of the corresponding position in the Raw image to form two sets of calibration reference points.

[0033] S22. According to the Stefan-Boltzmann radiation law, determine the corresponding radiation response at the reference temperature. and :

[0034] in, σ is the Stefan-Boltzmann constant.

[0035] S23, Solving for linear parameter conversion; Assuming the grayscale response of a raw image D Response to thermal radiation received by the sensor L There is a linear relationship between them; the slope can be solved using two sets of calibration points. a and bias b :

[0036] in, and These are the pixel response values ​​in the Raw image corresponding to the low-temperature reference point and the high-temperature reference point, respectively.

[0037] Thus, a conversion equation from pixel grayscale to thermal radiation response value is established:

[0038] S24. Substitute the obtained linear conversion relationship to obtain the thermal radiation response diagram L(D); To ensure physical validity, the calculated radiative response is non-negatively processed:

[0039] S25, to L + Performing a fourth root transformation yields the thermophysical property characterization. M :

[0040] Since the actual surface radiation of an object needs to be multiplied by the material's emissivity based on the blackbody radiation. (0 < ≤1), therefore the above transformation result can be approximated as:

[0041] Right now M Incorporating the actual temperature of objects T With material emissivity The combined contribution of these factors eliminated the strong nonlinear effects of the fourth power of temperature, making subsequent processing more conducive to reflecting the differences in thermophysical properties.

[0042] Step S3: Perform structural detail characterization extraction on the original infrared Raw image; This step extracts and reconstructs high-frequency components at each scale through multi-scale frequency domain decomposition to generate a third representation that expresses the target edge and local change information. In this invention, "structural detail representation" refers to the representation obtained by extracting and reconstructing high-frequency components at each scale from the Raw infrared image through multi-scale frequency domain decomposition, which is used to express the target edge and local change information. It is denoted as S, i.e., the third representation.

[0043] like Figure 3 As shown, S3 specifically includes: S31. Multi-scale wavelet decomposition obtains multi-directional high-frequency components; Raw infrared images D Multi-scale wavelet decomposition is performed. At each scale, two-dimensional wavelet decomposition is used to obtain the low-frequency subband and the high-frequency subbands in the horizontal, vertical, and diagonal directions, respectively. The low-frequency subband is then further decomposed at the next scale. Specifically, at the... i At different scales, the high-frequency coefficients in the horizontal direction were obtained. H i Vertical high frequency coefficient V i and diagonal high frequency coefficient D i .

[0044] S32. Restore each high-frequency coefficient to the original image spatial size using inverse wavelet transform to obtain the horizontal component with the same resolution as the original image. H i ', Vertical component V i 'and diagonal components D i '.

[0045] S33. Combine the three directional components into a detail response for that scale. E i :

[0046] This synthesis method can integrate edge strength from different directions, avoiding the loss of information from a single direction.

[0047] S34, Multi-scale detail blending; By fusing detail responses at different scales using non-negative weights, a structural detail representation for expressing the target contour, edges, and local variations is obtained. S :

[0048] The weights can be adjusted based on the target size distribution or scene characteristics; for example, for large targets, the upper-level scale can be appropriately increased. E 2) The weighting of this factor can highlight lower-level scales for smaller objectives.E 1).

[0049] The above is one implementation method for second-level scale wavelet decomposition; Without loss of generality, when using wavelet decomposition of level three or higher, the weighted combination of the detailed responses at each scale can be performed in the same manner:

[0050] Among them, the decomposition series Choose according to the required level of detail. The result... S It highlights the target's outline, edges, and local gray-scale abrupt changes, providing detailed structural features for subsequent fusion.

[0051] Step S4: Input the first representation, the second representation, and the third representation into the corresponding encoding networks for independent feature encoding, and perform feature space alignment; Because the physical meaning, response range, and data distribution of the three representations differ significantly, direct fusion is insufficient to effectively utilize their complementary information. This step maps the three representations to a unified feature representation space through independent encoding and feature space alignment.

[0052] like Figure 4 As shown, S4 specifically includes: S41, Independent Feature Coding; Impressions G Characterization of thermophysical properties M and structural detail representation S Three independent encoding networks are input, and their respective high-level semantic features are extracted. The appearance encoding network focuses on extracting the global layout and light-dark relationships of the scene; the thermal physical property encoding network focuses on extracting the distribution of thermal physical properties related to temperature and material; and the structure encoding network focuses on extracting local structural patterns such as edges and corners. The structures of the three encoding networks can be the same or different, and their parameters are optimized to adapt to the characteristics of different inputs.

[0053] As an optional implementation, the three types of representation encoding can employ convolutional coding modules with identical structures but independent parameters. For example, this convolutional coding module includes two layers of standard convolutional structures with Conv3×3, BN, and ReLU, and maps the number of output channels to a preset value C, so that all three encoded features have a feature dimension of C×H×W.

[0054] S42, Feature range adaptation; To address the differences in numerical range and distribution scale among the three encoded features, range mapping is performed separately. Specifically, batch normalization, layer normalization, or adaptive scaling and translation operations can be used to ensure that the mean and variance of each encoded feature are of similar magnitude, thus preventing one feature from dominating during subsequent fusion due to numerical differences.

[0055] S43. Spatial dimensions and passageway dimensions are unified; By interpolating and mapping convolutional channels, the three encoded features are made to correspond to the same scene region at the same spatial location and have a unified spatial size and channel dimension, resulting in apparent encoded features aligned to a unified feature representation space. F G Thermophysical property coding features F M and structural detail encoding features F S ,in: .

[0056] Step S5: Characterization fusion; This step emphasizes the information from the three features, with apparent features as the main body, and introduces complementary information from thermophysical properties and structural details, performs weighted fusion, and outputs a unified infrared multivariate characterization.

[0057] S5 specifically includes: S51-S53, Determine the division of labor based on information focus; Based on the information emphasis of the three-way coding features after alignment, the apparent coding features are used. F G Preserve the light / dark relationship between the target subject and the scene, and encode features using thermal physical properties. F M Supplementing temperature and material information to encode features in structural details. F S Complete the target outline, edges, and local structure; S54, Weighted Complementary Integration; Using appearance coding features as the main component, complementary fusion weights are introduced from thermophysical property coding features and structural detail coding features respectively, and weighted fusion is performed:

[0058] in, F This represents the unified infrared multivariate characterization after fusion. and These represent the fusion weights of the thermophysical property encoding features and the structural detail encoding features, respectively, and are determined based on the response state or application scenario of these two representations.

[0059] As an optional implementation, the fusion weights can be dynamically determined based on the response states of each representation. The thermophysical property encoding features are statistically analyzed separately. and structural detail encoding features The response intensity is used to obtain the corresponding average response intensity. and The fusion weights are determined based on the relative magnitudes of the two components.

[0060] in, A stabilizing term is set up to avoid the denominator being zero, thereby adaptively adjusting its contribution to the fusion result based on the relative response state of the two types of representations in the current infrared image.

[0061] As another optional implementation, the fusion weights can be preset according to the target application scenario. For scenarios where there is a difference in thermal radiation distribution between the target and the background, and the apparent grayscale contrast is low, the fusion weight of the thermophysical property encoding features can be relatively increased; for scenarios where distant, weak targets, target outlines, or local details are easily lost, the fusion weight of the structural detail encoding features can be relatively increased. The specific values ​​of the above fusion weights can be determined according to different sensors and target types.

[0062] Through the above steps, the unified infrared multivariate characterization generated by this invention F Simultaneously, it carries information on the scene's apparent brightness and darkness, thermophysical properties, and structural details, overcoming the shortcomings of traditional single-grayscale representations such as physical information coupling, limited contrast, and loss of detail. This representation can be directly used for downstream infrared target detection, recognition, or segmentation tasks, and is particularly suitable for challenging scenarios such as low contrast, complex background interference, and distant, weak targets, providing a more comprehensive and physically meaningful feature foundation for intelligent sensing systems.

[0063] To verify the effectiveness of the proposed multivariate representation method for infrared images based on physical property decoupling, a comparative experiment was conducted using the FLIR infrared dataset. The target detection performance was compared between the traditional single grayscale representation method and the proposed multivariate representation method. The comparative experiment used the same data partitioning, YOLOv8n target detection model, and training hyperparameters, only changing the representation method of the input image. The mean accuracy (mAP) was calculated on the same test set. Experimental results show that the average accuracy is significantly improved after using the proposed multivariate representation method. These results demonstrate that this method, by effectively fusing appearance, thermophysical properties, and texture detail information, can effectively improve the feature representation capability of infrared targets, providing more accurate and richer feature representations for subsequent detection and recognition.

[0064] Example: Taking actual road monitoring application as an example; An infrared imaging device capable of acquiring and outputting raw infrared images is placed above roads, at traffic intersections, or on vehicle platforms. This device acquires single-band raw infrared images of the road scene and transmits the image data to an image processor. The image processor performs the appearance characterization extraction, thermophysical property characterization extraction, structural detail characterization extraction, feature encoding alignment, and characterization fusion steps described in this embodiment to generate a unified infrared multivariate characterization. This unified infrared multivariate characterization can then be input into a target detector to achieve accurate detection and identification of targets such as vehicles, pedestrians, and road signs. The above processing method is also applicable to unmanned sensing platforms equipped with infrared imaging devices and search and rescue scenarios in complex environments.

[0065] The above description of the disclosed embodiments enables those skilled in the art to make or use 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 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 disclosed herein.

Claims

1. A multivariate representation method for infrared images based on physical property decoupling, characterized in that, Includes the following steps: S1. Perform appearance representation extraction on the original infrared Raw image, truncate the effective grayscale range of the original infrared Raw image and map it to a preset grayscale range to generate a first representation for expressing the relative brightness and overall contour information of the scene. S2. Perform thermophysical property characterization extraction on the original infrared Raw image, establish the mapping relationship between the pixel gray value of the Raw image and the thermal radiation response based on the blackbody radiation law, and perform nonlinear inverse transformation on the mapping result to generate a second characterization to express the information of the combined effect of temperature and material emissivity. S3. Perform structural detail characterization extraction on the original infrared Raw image, extract high-frequency components at each scale through multi-scale frequency domain decomposition and reconstruct them to generate a third characterization to express target edge and local change information. S4. Input the first representation, the second representation, and the third representation into the corresponding encoding network for independent feature encoding to obtain the first encoding feature, the second encoding feature, and the third encoding feature; perform feature space alignment processing on the first encoding feature, the second encoding feature, and the third encoding feature to make the three have a unified spatial size and channel dimension; S5. The first, second, and third coding features, after feature space alignment, are complementary and fused to generate a unified infrared multivariate representation.

2. The infrared image multivariate characterization method based on physical property decoupling as described in claim 1, characterized in that, S1 specifically includes: S11. Based on the grayscale distribution statistical characteristics, scene dynamic range, and sensor response range of the original infrared Raw image, determine the low grayscale truncation boundary. and high grayscale cutoff boundary ; S12, according to Raw image original pixel values D Perform an outward cutoff at the boundary; S13, truncate the effective grayscale range [ , Map the image to a preset display grayscale range to obtain the first representation; in, This is the truncated response value. This is the clipping function.

3. The infrared image multivariate characterization method based on physical property decoupling as described in claim 1, characterized in that, S2 specifically includes: S21. Obtain prior calibration parameters, including a low-temperature reference point. and high temperature reference point ,in, and These are the pixel response values ​​in the Raw image corresponding to the low-temperature reference point and the high-temperature reference point, respectively. S22. According to the Stefan-Boltzmann radiation law, from the reference temperature... and Calculate the corresponding blackbody radiation response values ​​respectively. L l = σT l 4 and L h = σT h 4 ,in σ It is the Stefan-Boltzmann constant; S23. Based on the aforementioned low-temperature reference point and high-temperature reference point, establish the Raw image pixel response value. D Linear mapping relationship between the thermal radiation response value L: L = aD + b The slope bias ; S24. Substitute the response value D of each pixel in the original infrared raw image into the linear mapping relationship pixel by pixel to obtain the thermal radiation response map. L ( D ), and the thermal radiation response map L ( D Apply nonnegative constraints: L + =max( L ( D ),0); S25, Response diagram to nonnegative thermal radiation L + Perform a fourth root transformation: To obtain the second characterization M .

4. The infrared image multivariate characterization method based on physical property decoupling as described in claim 3, characterized in that, The second characterization M satisfy M ≈ ,in T The surface temperature of the object. ε The second characterization is used to uniformly characterize the combined contribution of object temperature and material emissivity to infrared radiation, where is the surface emissivity of the object.

5. The infrared image multivariate characterization method based on physical property decoupling as described in claim 1, characterized in that, S3 specifically includes: S31. Perform multi-scale wavelet decomposition on the original infrared Raw image, and in the... i Extracting high-frequency components in the horizontal direction on a scale H i Vertical high-frequency components V i and diagonal high frequency components D i ; S32, the horizontal high-frequency components at each scale H i Vertical high-frequency components V i and diagonal high frequency components D i To restore the original image spatial resolution, the restored horizontal components at the same scale are obtained. H i ', Vertical component V i 'and diagonal components D i '; S33, in the i In terms of scale, according to Synthesize the detail response map at this scale E i ; S34. Combine detail response maps of different scales according to preset weights. E i The third representation is generated.

6. The infrared image multivariate characterization method based on physical property decoupling as described in claim 5, characterized in that, The third characterization ,in, , α and β All are non-negative weights. and These are the detail response maps for the first and second scales, respectively.

7. The infrared image multivariate characterization method based on physical property decoupling as described in claim 1, characterized in that, In step S4, feature space alignment processing is performed on the first coded feature, the second coded feature, and the third coded feature, specifically including: S41. Perform appearance coding, thermophysical property coding, and structural coding on the first characterization, the second characterization, and the third characterization respectively to obtain the first coding feature, the second coding feature, and the third coding feature; S42. Perform feature value range mapping on the first coding feature, the second coding feature and the third coding feature respectively, so that the response value distribution ranges of the three are mutually adapted; S43. The first coding feature, the second coding feature and the third coding feature are respectively adjusted in terms of spatial scale and channel dimension mapping so that the three have a unified spatial size and the same number of channels.

8. The infrared image multivariate characterization method based on physical property decoupling as described in claim 1, characterized in that, S5 specifically includes: S51. The first encoded feature after alignment is used as the main feature to maintain the target main structure and scene spatial relationship; S52. The aligned second coding feature is used as supplementary information for thermophysical properties to provide a joint basis for judging temperature and material emissivity; S53. The aligned third encoded features are used as supplementary structural information to provide target contours, edges and local texture details; S54, according to Weighted fusion is performed, where, The fusion weight coefficients for the second encoded features. The fusion weight coefficients for the third coding feature. F For the unified infrared multivariate characterization, F G This is the first coding feature. F M This is the second coding feature. F S This is the third coding feature.

9. The infrared image multivariate characterization method based on physical property decoupling as described in claim 8, characterized in that, The fusion weight coefficient and The value can be dynamically determined based on the response intensity distribution of the second and third coding features, or preset to a fixed value based on the target application scenario.

Images