A long-distance infrared target recognition method based on propagation state field and cross-scale evidence preservation

Abstract

This invention belongs to the fields of infrared intelligent sensing, long-range target recognition, infrared image processing, and intelligent monitoring, and relates to a long-range infrared target recognition method based on propagation state field and cross-scale evidence preservation. The method acquires a temporal infrared image sequence affected by atmospheric turbulence and performs time synchronization, spatial coarse registration, and radiometric normalization. Based on local displacement changes, edge position changes, and local grayscale fluctuations between adjacent time phases, a local drift field and a radiative scintillation field are constructed, respectively. Furthermore, turbulent state units are constructed by combining edge stability, local sharpness, and temporal consistency, and stable structural anchors are extracted. Based on the stable structural anchors, path consistency constraints are established to perform local drift compensation, radiative scintillation correction, detail degradation compensation, and thermal radiation consistency preservation on the infrared image. The compensated infrared image, along with turbulence intensity level, structural stability, thermal radiation preservation degree, and compensation reliability, are output.

Description

Technical Field

[0001] This invention belongs to the fields of infrared intelligent sensing, long-range target recognition, infrared image processing and intelligent monitoring technology, and relates to a long-range infrared target recognition method based on propagation state field and cross-scale evidence preservation. Background Technology

[0002] Infrared imaging can acquire target information by utilizing the difference in thermal radiation between the target and the background. It still has good perception capabilities under conditions such as nighttime, low light, obscured backgrounds, and complex weather, and is therefore widely used in scenarios such as border monitoring, port surveillance, forest patrol, low-altitude sensing, long-range early warning, and intelligent security. Compared with visible light imaging, infrared imaging does not depend on external lighting conditions and can obtain the thermal radiation response of targets at long distances, providing important information for target detection, target identification, and status assessment.

[0003] In long-range infrared observation scenarios, there is usually a long propagation path between the target and the imaging equipment. During propagation, the target's thermal radiation signal is easily affected by factors such as atmospheric absorption, scattering, path disturbance, thermal diffusion, imaging jitter, and local turbulence. The same target may appear as a small, weak area with blurred edges, low local contrast, and unstable thermal response in an infrared image. For objects such as vehicles, ships, personnel, low-altitude flying objects, abnormal heat sources in forest areas, or moving targets in ports, long-range infrared imaging results often fail to fully preserve the target's thermal structure, edge contours, and category-related details. Relying solely on a single frame image for identification can easily lead to missed detections, false detections, or category confusion.

[0004] Background thermal clutter in complex, long-range scenes can also affect the stability of infrared target recognition. In ports, forests, border areas, sea surfaces, urban edges, and low-altitude scenes, the background may contain numerous local bright spots, reflective areas, environmental heat sources, motion interference, and non-stationary thermal disturbances. These background responses may resemble real targets in terms of grayscale intensity, local morphology, or temporal variations. Traditional methods based on threshold segmentation, local contrast enhancement, single-frame detection networks, or ordinary image enhancement struggle to reliably distinguish the thermal response of real targets from background thermal clutter, easily misidentifying local abnormal responses in the background as targets.

[0005] Existing infrared target recognition methods typically input long-range infrared images directly into detection or classification models, or perform preprocessing such as denoising, enhancement, and super-resolution reconstruction before detection. While these methods can improve image visual quality or target response to some extent, they usually do not explicitly describe the relationship between long-range propagation status and the reliability of identification evidence, nor do they fully utilize the persistence of thermal anomalies, motion continuity, and spatial consistency of targets across continuous time phases. When a target only exhibits weak thermal anomalies in a single frame image, single-frame features are often insufficient to form a stable basis for target identification.

[0006] In multi-scale feature processing, distant, small targets are prone to identity information loss. Due to the small target scale, after downsampling, scale transformation, feature fusion, or candidate region selection, the target's original thermal structure, edge contours, and category-related responses may be overwhelmed by background textures or noise responses. Existing multi-scale fusion methods typically focus more on feature enhancement or detection accuracy, lacking the ability to maintain the consistency of the same candidate target's identity at different scales. This results in the candidate target being detected, but its category, distance range, spatial state, and recognition confidence remain unstable.

[0007] Therefore, there is a need for an infrared target identification method that can simultaneously consider long-distance propagation degradation, weak target energy accumulation, background non-stationarity suppression, and cross-scale target identity evidence preservation, so that it can stably output target category, spatial location, distance range, identification confidence level, and target discernibility level under conditions of low signal-to-noise ratio, strong background thermal clutter, small target size, and significant changes in propagation state. Summary of the Invention

[0008] The purpose of this invention is to provide a long-range infrared target recognition method based on propagation state field and cross-scale evidence preservation. In cases where long-range infrared images have small target scale, weak thermal structure, strong background thermal clutter, obvious propagation degradation, and insufficient single-frame recognition evidence, the method can construct target recognition evidence that reflects the propagation state, target energy persistence, and cross-scale identity consistency, and outputs target category, spatial location, distance range, recognition confidence, and target discernibility level.

[0009] This invention does not simply treat long-range infrared target recognition as a single-frame image detection or enhanced classification problem. Instead, it first constructs a propagation state field based on multi-temporal infrared images to describe the degree of propagation degradation and the reliability of target evidence in different spatial regions. Then, it uses the propagation state field to perform reliability weighting and energy aggregation on weak thermal anomaly responses in continuous temporal phases to form candidate target evidence. Furthermore, it distinguishes between target energy, background energy, and uncertain mixed energy in the candidate regions to suppress non-stationary background interference. Then, it maintains the thermal structure, edge contour, motion trend, and category-related response of candidate targets at different scales through a cross-scale identity evidence preservation process. Finally, it performs a joint decision based on distance, category, and confidence, thereby improving the stability and interpretability of long-range infrared weak target recognition.

[0010] The technical solution of the present invention is as follows: A long-range infrared target recognition method based on propagation state field and cross-scale evidence preservation includes the following steps: Step 1, Acquisition of multi-temporal long-range infrared images. Acquire a sequence of multi-temporal long-range infrared images of the area to be observed, denoted as: ; in, Indicates the first Infrared images of each phase, The number of time phases is indicated. The infrared images can be acquired by short-wave infrared imaging equipment, mid-wave infrared imaging equipment, long-wave infrared imaging equipment, or multi-band infrared imaging equipment. The multi-time phase long-distance infrared image sequence is subjected to time synchronization, spatial registration, and radiometric normalization processing to obtain the infrared image sequence to be identified.

[0011] Step 2, Propagation State Field Construction. Based on the local contrast, background thermal clutter intensity, target scale response, inter-frame stability, image sharpness, and propagation degradation estimate in the infrared image sequence to be identified, a propagation state field is constructed. The propagation state field is used to characterize the degree of propagation degradation and the reliability of target evidence in different spatial regions during long-range infrared imaging.

[0012] Step 2.1, Construction of propagation state components. Based on the infrared image sequence to be identified, local contrast components are constructed respectively. Background thermal clutter components Target scale response components Inter-frame stability components Image sharpness component and propagation degradation estimation components The aforementioned propagation state components collectively characterize the degradation state of a long-range infrared target during propagation and serve as inputs for the propagation state field fusion calculation.

[0013] Step 2.2, Propagation State Field Fusion Calculation. Based on the propagation state components obtained in Step 2.1, the propagation state field is constructed using the propagation state fusion function: ; in, Indicates the first Spatial position in each phase The propagation state field response at the location, This represents the propagation state fusion function.

[0014] The propagation state fusion function adopts a weighted fusion form, and its specific propagation state field is represented as follows: ; in, Indicates the first propagation state components, This represents the corresponding weighting coefficient. This indicates the number of propagation state components. Propagation state components include one or more of the following: local contrast component, background thermal clutter component, scale response component, temporal stability component, image sharpness component, and propagation degradation component.

[0015] Step 3, Weak Target Energy Aggregation. Based on the propagation state field, the thermal anomaly responses in the infrared image sequence to be identified are spatiotemporally aggregated to extract candidate energy trajectories that exhibit thermal anomaly persistence, motion continuity, and spatial consistency in continuous temporal phases, forming a target candidate evidence set.

[0016] Step 3.1 Construction of the candidate target evidence set. Using high-response regions, local thermal anomaly responses, and temporal continuity responses in the propagation state field, a candidate target evidence set is constructed: ; in, Indicates the first Evidence of target candidates This indicates the number of candidate pieces of evidence for the target.

[0017] Step 3.2 Energy polymerization response calculation. For the first step obtained in step 3.1... Evidence of candidate targets The energy trajectory of the candidate energy trajectory in continuous time phases is established. The aggregated response of the candidate energy trajectory is expressed as: ; in, Indicates the first The aggregated response of candidate energy trajectories Indicates the first The candidate energy trajectory in the first Spatial support region in each temporal phase This represents the reliability weight determined by the propagation state field. Indicates the first A temporal infrared image in spatial location The pixel grayscale value or normalized thermal radiation response value at that location. The reliability weight can be obtained by normalizing the propagation state field response, or it can be jointly determined by the propagation state field response and the local thermal anomaly response.

[0018] Step 4, Background Non-stationarity Suppression. The target candidate evidence set is subjected to background non-stationarity suppression processing.

[0019] Step 4.1 Energy Domain Delineation. For the energy domain obtained in Step 3.2... Converged response of candidate energy trajectories Combined with the propagation state field constructed in step 2 Calculate the energy attribution probability of the candidate region: ; in, Indicates spatial location The probability of belonging to the target energy domain. Indicates spatial location The probability of belonging to the background energy domain. Indicates spatial location The probability of belonging to an uncertain mixed domain, satisfying: ; Based on the energy attribution probability, the candidate region is divided into a target energy attribution region, a background energy attribution region, and an uncertain mixed region: ; in, Indicates the first Candidate regions corresponding to each target candidate piece of evidence Indicates the target energy domain. Indicates the background energy domain. This represents an uncertain mixed domain. Step 4.1 yields the target energy domain. Background energy attribution domain and uncertain mixed domain .

[0020] Step 4.2 Background Nonstationary Suppression. Based on the target energy domain, background energy domain, and uncertain mixing domain obtained in Step 4.1, background nonstationary suppression is performed on the energy response in the candidate region to obtain the target response after background suppression: ; in, This represents the target response after background nonstationarity suppression. This represents the background nonstationary suppression function. This represents the energy aggregation response obtained in step 3.2. The background nonstationarity suppression adjusts the energy response in the candidate region through the target energy domain, the background energy domain, and the uncertain mixing domain to obtain the target response after background suppression. The target response obtained in step 4.2 This serves as the input for step 5, cross-scale identity evidence preservation.

[0021] Step 5, cross-scale identity evidence preservation. Candidate targets, after background non-stationary suppression, undergo cross-scale processing to preserve their thermal structure, edge contours, motion trends, and category-related responses at different scales.

[0022] Step 5.1 Construction of the thermal radiation evidence chain. Based on the target response obtained in Step 4.2. Extracting target-related responses at multiple scales to construct a chain of evidence for thermal radiation: ; in, Indicates the first The thermal radiation evidence chain for each candidate target Indicates the first Target identity evidence at various scales Indicates the number of scales. Step 5.1 yields the chain of evidence for thermal radiation. This serves as the input for establishing cross-scale identity evidence preservation constraints in step 5.2.

[0023] Step 5.2 Maintaining Constraints on Cross-Scale Identity Evidence. Based on the thermal radiation evidence chain obtained in Step 5.1. Establish cross-scale identity evidence preservation constraints: ; in, This indicates that cross-scale identity evidence remains constrained. Indicates the first Target identity evidence at various scales This represents the scaling function.

[0024] The cross-scale identity evidence preservation constraint is used to maintain the consistency of target identity evidence for the same candidate target at different scales. Step 5.2 obtains the thermal radiation evidence chain after cross-scale constraint. And serve as the input for the joint decision in step 6.

[0025] Step 6: Joint decision based on distance, category, and confidence level. Based on the thermal radiation evidence chain obtained in Step 5, a joint decision based on distance, category, and confidence level is made for the candidate target to obtain the long-range infrared target identification result.

[0026] Step 6.1 Generation of Joint Judgment Result. Based on the thermal radiation evidence chain obtained in Step 5. Step 2 constructs the propagation state field Based on the target energy attribution results obtained in step 4, a joint decision result for candidate targets is generated: ; in, Indicates the first The joint judgment result of the candidate targets, This represents the joint decision function. Step 6.1 yields the joint decision result for the candidate targets. This information is used as input for calculating the target discernibility score in step 6.2.

[0027] Step 6.2 Target Distinguishing Score Calculation. Based on the joint decision result obtained in Step 6.1. The target's thermal structural integrity, target-background separation, temporal stability, and propagation state reliability are calculated, and a target discernibility score is constructed. ; in, Indicates the first Target discernibility score for each candidate target Indicates the integrity of the target thermal structure. Indicates the target-background separation degree. Indicates timing stability, Indicates the reliability of the propagation state. These are the weighting coefficients. Step 6.2 yields the target discernibility score. In conjunction with the joint judgment results Generate target category, target spatial location, target distance range, recognition confidence level, and target discernibility level.

[0028] Step 7: Output the recognition results and provide feedback updates. Output the target category, target spatial location, target distance range, recognition confidence level, and target discernibility level. Based on the target discernibility level and recognition confidence level, update the propagation state field construction parameters, weak target energy aggregation parameters, background non-stationary suppression parameters, and cross-scale identity evidence preservation parameters to support continuous recognition in subsequent time phases.

[0029] The beneficial effects of this invention are: First, by constructing a propagation state field, this invention introduces propagation degradation, background thermal clutter, and image stability factors in long-range infrared imaging into the target recognition process, enabling the recognition strategy to adaptively adjust according to observation conditions.

[0030] Second, by aggregating the energy of weak targets, this invention transforms unstable and incomplete weak target responses in a single frame into multi-temporal continuous target evidence, thereby improving the identifiability of distant weak targets under low signal-to-noise ratio conditions.

[0031] Third, by dividing the target energy domain, background energy domain, and uncertain mixed domain, this invention can suppress non-stationary interference in complex thermal backgrounds and reduce misidentification caused by background bright spots, noise points, and thermal clutter.

[0032] Fourth, this invention, through a cross-scale identity evidence preservation mechanism, can maintain the target's thermal structure, edge contour, and category-related response during multi-scale enhancement and feature fusion, thereby reducing the loss of identity information of distant small targets during scale changes.

[0033] Fifth, this invention outputs target category, spatial location, distance range, identification confidence level, and target discernibility level, realizing the expansion from a single detection result to a comprehensive judgment of target status, and is more suitable for application scenarios such as long-range infrared early warning, port surveillance, forest patrol, border monitoring, and low-altitude perception. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the overall process of a long-range infrared target recognition method based on propagation state field and cross-scale evidence preservation in an embodiment of the present invention; Figure 2 This is a schematic diagram of the multi-temporal long-distance infrared image acquisition and preprocessing process in an embodiment of the present invention; Figure 3 This is a schematic diagram of the propagation state field construction process in an embodiment of the present invention; Figure 4 This is a schematic diagram of the weak target energy aggregation process in an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the division of the target energy domain, background energy domain, and uncertain mixed domain in an embodiment of the present invention; Figure 6 This is a schematic diagram illustrating cross-scale identity evidence preservation and thermal radiation evidence chain construction in an embodiment of the present invention; Figure 7 This is a schematic diagram of the joint decision output of distance, category, and confidence in an embodiment of the present invention; Figure 8 This is a schematic diagram of the feedback update process in an embodiment of the present invention; Figure 9 The images shown are infrared target recognition results and annotation diagrams in this embodiment of the invention. Detailed Implementation

[0035] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and technical solutions. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. Without departing from the concept of the present invention, those skilled in the art can make substitutions or modifications to the feature extraction method, fusion method, decision method, and parameter settings in each step. Example

[0036] A long-range infrared target recognition method based on propagation state field and cross-scale evidence preservation, such as Figure 1 As shown, it includes multi-temporal long-range infrared image acquisition, propagation state field construction, weak target energy aggregation, background non-stationarity suppression, cross-scale identity evidence preservation, distance category confidence joint decision-making, and feedback update.

[0037] Step 1: Acquisition of multi-temporal long-range infrared images.

[0038] like Figure 2 As shown, an infrared imaging device is used to continuously observe the area to be observed, obtaining a multi-temporal long-range infrared image sequence. The infrared imaging device can be a short-wave infrared, mid-wave infrared, long-wave infrared, or multi-band infrared imaging device. The area to be observed can be a long-range observation scene such as a port, forest, border, sea surface, low altitude, or urban edge. The target to be identified can be a vehicle, ship, person, low-altitude flying object, abnormal heat source in forest area, or moving target in port.

[0039] The acquired multi-temporal long-range infrared image sequence is denoted as: ; in, Indicates the first Infrared images of each phase, Indicates the number of time phases. Time synchronization of image sequences ensures a consistent temporal relationship between different frames; spatial registration of image sequences reduces positional shifts caused by platform jitter, slight changes in field of view, or changes in imaging device attitude; radiometric normalization of image sequences reduces the impact of overall grayscale drift and response variations between different frames.

[0040] Step 2, construct the propagation state field.

[0041] like Figure 3 As shown, after obtaining the infrared image sequence to be identified, a propagation state field is constructed for each temporal image. The propagation state field is used to describe the degree of propagation degradation and the reliability of target evidence in different spatial regions under the current long-range infrared imaging conditions. Unlike directly performing uniform enhancement on the entire image, this embodiment provides differentiated reliability descriptions for different regions through the propagation state field.

[0042] For the Infrared images of each phase Calculate the local contrast respectively Background thermal clutter intensity Target-scale response Inter-frame stability Image clarity and propagation degradation estimator Among them, local contrast reflects the grayscale difference between the target and the local background; background thermal clutter intensity describes the complexity of the thermal response of non-targets in the background; target scale response describes the response intensity of targets of different sizes at the current location; inter-frame stability describes the stability of the thermal response at the current location in continuous time phases; image sharpness describes the degree of preservation of local edge, texture or gradient information; and propagation degradation estimate describes the degradation effects caused by atmospheric propagation, path disturbance, imaging blur or thermal diffusion.

[0043] The propagation state field is represented as: ; in, It can be a weighted summation function, a nonlinear mapping function, a rule fusion function, or a feature fusion function implemented by a neural network.

[0044] The propagation state field is calculated using a weighted fusion method: ; in, Indicates the first propagation state components, This indicates the corresponding weight. Weight It can be determined based on preset experience, or it can be adaptively adjusted according to different observation scenarios.

[0045] Step 3, weak target energy aggregation.

[0046] like Figure 4 As shown, distant targets may only appear as weak thermal anomalies or weak thermal anomaly patches in a single-frame infrared image. To reduce the impact of single-frame noise, bright background spots, and random thermal disturbances on the recognition results, this embodiment aggregates the thermal anomaly responses in continuous time phases based on the propagation state field.

[0047] Specifically, firstly, local thermal anomaly responses are extracted from each temporal image; then, reliability weights are assigned to each thermal anomaly response based on the propagation state field; finally, based on temporal persistence, spatial proximity, and motion consistency, thermal anomaly responses with consistent changing relationships in consecutive temporal phases are connected to form candidate energy trajectories. These candidate energy trajectories form the target candidate evidence set. ; in, Indicates the first One target candidate piece of evidence. The aggregated response of the candidate energy trajectories is: ; in, Indicates the first The candidate energy trajectory in the first Spatial support region in each temporal phase This represents the reliability weight determined by the propagation state field. If a thermal anomaly region has high local contrast, low background clutter, stable inter-frame response, and high propagation state reliability, then the reliability weight corresponding to that region is large; conversely, if the region may belong to random noise or non-stationary background disturbance, then the reliability weight is small.

[0048] Step 4: Suppress background nonstationarity.

[0049] like Figure 5 As shown, background nonstationarity suppression is applied to the target candidate evidence set obtained after aggregating weak target energy. For the ... There are 1 target candidate evidence, and their candidate regions are denoted as . In this embodiment, the energy response within the candidate region is divided into a target energy domain, a background energy domain, and an uncertain mixed domain: ; in, Indicates the target energy domain. Indicates the background energy domain. This represents an uncertain mixed domain.

[0050] The energy attribution probability of a candidate region is expressed as: ; in, , and Representing spatial location The probability of belonging to the target energy domain, the background energy domain, and the uncertain mixture domain, satisfying: ; For the target energy attribution domain, its thermal response is preserved and its distinction from the surrounding background is enhanced; for the background energy attribution domain, suppression weights are applied to reduce false identifications caused by background thermal clutter and random bright spots; for the uncertain mixed domain, its undiscriminated attributes are preserved and further judgment is made in conjunction with the subsequent cross-scale identity evidence preservation process. This approach avoids directly deleting weak target edge regions while reducing false alarms caused by complex thermal backgrounds.

[0051] Step 5, cross-scale identity evidence preservation.

[0052] like Figure 6 As shown, cross-scale identity evidence preservation processing is performed on candidate targets after background non-stationarity suppression. Long-range infrared targets exhibit different characteristics at different scales; at smaller scales, they mainly appear as thermal anomaly centers, while at larger scales, they may appear as contours, edges, or motion trends. If multi-scale fusion is performed directly, the identity information of weak targets is easily obscured by background textures or noise. Therefore, this embodiment preserves the target's thermal structure, edge contours, motion trends, and category-related responses during multi-scale processing.

[0053] For the 1 candidate target, constructing a chain of evidence for thermal radiation: ; in, Indicates the first Target identity evidence at various scales Indicates the number of scales. Evidence of target identity may include one or more of the following: target thermal structure evidence, edge contour evidence, motion trend evidence, and category response evidence.

[0054] To maintain the consistency of target identity across different scales, a scale consistency constraint is introduced: ; in, This represents the scaling function, used to scale the number of elements in a scaled matrix. Target identity evidence at one scale is mapped to the first... Each scale. This constraint ensures that the thermal structure, edge contours, and category-related responses of the same target remain consistent across different scales, thereby reducing target identity drift caused by scale transformations.

[0055] Step 6: Joint decision based on distance, category, and confidence.

[0056] like Figure 7 As shown, a joint decision is made based on the thermal radiation evidence chain, outputting the target category, target spatial location, target distance range, identification confidence level, and target discernibility level. Unlike methods that only output the detection box and category, this embodiment jointly models the distance, category, and confidence level, enabling the output results to better reflect the comprehensive state of long-range infrared targets.

[0057] No. The joint decision result of the candidate targets is expressed as: ; in, The joint decision function can be implemented by a classifier, a regressor, a rule-based decision module, a neural network decision module, or a combination thereof.

[0058] The target discernibility score is expressed as follows: ; in, Indicates the integrity of the target thermal structure. Indicates the target-background separation degree. Indicates timing stability, Indicates the reliability of the propagation state. These are the weighting coefficients. According to... The size of the target can be used to classify it into highly distinguishable, moderately distinguishable, lowly distinguishable, or unstable distinguishable levels.

[0059] Step 7: Output the recognition results and provide feedback for updates.

[0060] like Figure 8As shown, after outputting the target identification result, the aforementioned process is updated based on the target discernibility level and identification confidence. If the candidate target has a low identification confidence but a high target discernibility score, the category discrimination weight can be increased; if the identification confidence is low and the target discernibility score is low, the multi-temporal energy aggregation intensity can be increased, or the background non-stationary suppression intensity can be improved; if the candidate target maintains a stable response in subsequent temporal phases, its thermal radiation evidence chain is updated to support continuous identification.

[0061] By updating through feedback, the method can adaptively adjust the recognition strategy according to changes in scene, propagation state, and target response during continuous observation, thereby improving the stability and robustness of long-range infrared target recognition.

[0062] Example 2: This embodiment uses the identification of long-distance transport vehicles in a port setting as an example to illustrate the specific implementation process of the method of the present invention.

[0063] Step 1: Continuous observation of the port area was conducted using a long-wave infrared imaging device. The infrared imaging device operated in the 8μm–14μm band, with an image resolution of 640×512, a lens focal length of 200mm, a sampling frequency of 25fps, and an observation distance of approximately 3.5km. Twenty consecutive long-range infrared images were acquired, resulting in an infrared image sequence. ; The infrared image sequence is then subjected to time synchronization, spatial registration, and radiometric normalization to obtain the infrared image sequence to be identified.

[0064] Step 2: Construct the propagation state field based on the infrared image sequence to be identified. Calculate the local contrast component, background thermal clutter component, target scale response component, temporal stability component, image sharpness component, and propagation degradation estimation component, and then use a weighted fusion method to construct the propagation state field. ; in , , , , , These correspond to the local contrast component, background thermal clutter component, target scale response component, temporal stability component, image sharpness component, and propagation degradation estimation component, respectively.

[0065] In this embodiment, the propagation state field response range is normalized to the [0,1] interval. After calculation, the average propagation state field response of the target area is 0.82, and the average propagation state field response of the background area is 0.31.

[0066] Step 3: Spatiotemporal aggregation of thermal anomaly responses in continuous time phases based on the propagation state field.

[0067] First, candidate target evidence was extracted using high-response regions in the propagation state field, resulting in 12 candidate target pieces of evidence: ; Then, the energy trajectories of each candidate target in continuous time phases are established, and the energy convergence response is calculated: ; The reliability weight is based on the propagation state field response: ; The calculation yielded the following results: , , , The aggregate response of the remaining candidate targets was all below 120.

[0068] This embodiment sets the aggregation response threshold: ; Therefore, candidate targets are retained. , as well as Proceed to the follow-up processing stage.

[0069] Step 4: Perform background nonstationarity suppression on the candidate target evidence.

[0070] Set the target domain determination threshold: Set the background field determination threshold: .

[0071] Based on the energy attribution probability, the candidate region is divided into the target energy attribution region, the background energy attribution region, and the uncertain mixed region.

[0072] The target energy domain is retained with a weight of 1.0; the uncertain mixed domain is retained with a weight of 0.5; and the background energy domain is retained with a weight of 0.

[0073] The target response is obtained after background suppression: The average response in the target area increased by approximately 27%, while the average response in the background area decreased by approximately 64%.

[0074] The average response in the target area increased by approximately 27%, while the average response in the background area decreased by approximately 64%.

[0075] This embodiment uses three scales: 32×32, 64×64, and 128×128, to extract the target thermal structure response, edge contour response, and motion trend response, respectively, and construct a thermal radiation evidence chain: ; The cross-scale identity evidence preservation constraint parameters are set as follows: After optimization, a consistent chain of thermal radiation evidence across scales was obtained.

[0076] Step 6: Perform a joint decision on candidate targets based on distance, category, and confidence level. The joint decision yields the following target category probabilities: transport vehicles: 0.91, ships: 0.06, other targets: 0.03. Simultaneously, the target's thermal structural integrity is also obtained. Target-background separation: Timing stability: Reliability of propagation status: .set up: According to the target discernibility score calculation formula: ; get In this embodiment, it is identified as a highly distinguishable target.

[0077] Step 7, as follows Figure 9 As shown, the recognition results are output.

[0078] The final output is as follows: Target category: Transport vehicles; Target distance range: 3km to 4km; Target center location: (421, 267); Recognition confidence level: 0.91; Target visibility level: High.

[0079] Subsequently, the propagation state field parameters, energy aggregation parameters, and cross-scale identity evidence preservation parameters are updated based on the target discernibility level and identification confidence level for continuous identification in subsequent time phases.

Claims

1. A method for long-range infrared target recognition based on propagation state field and cross-scale evidence preservation, characterized in that, The steps are as follows: Step 1: Acquisition of multi-temporal long-range infrared images; A multi-temporal long-range infrared image sequence of the area to be observed is acquired, denoted as: ； in, Indicates the first Infrared images of each phase, The number of time phases is indicated; the multi-time phase long-range infrared image sequence is subjected to time synchronization, spatial registration and radiometric normalization to obtain the infrared image sequence to be identified; Step 2, construct the propagation state field; Based on the local contrast, background thermal clutter intensity, target scale response, inter-frame stability, image sharpness, and propagation degradation estimate in the infrared image sequence to be identified, a propagation state field is constructed. Step 3, gathering weak target energy; Based on the propagation state field, the thermal anomaly response in the infrared image sequence to be identified is spatiotemporally aggregated, and candidate energy trajectories with thermal anomaly persistence, motion continuity and spatial consistency in continuous time phases are extracted to obtain the target candidate evidence set. Step 4, background nonstationarity suppression; Background nonstationarity suppression is applied to the target candidate evidence set, and the energy response within the candidate region is divided into the target energy attribution domain, the background energy attribution domain, and the uncertain mixed domain. Step 5, preserving identity evidence across scales; Cross-scale processing is performed on the target's energy attribution domain and uncertain mixed domain to preserve the target's thermal structure, edge contour, motion trend, and category-related response at different scales, thus constructing a thermal radiation evidence chain. Step 6: Joint decision based on distance, category, and confidence. Based on the aforementioned thermal radiation evidence chain, the category probability, spatial location, distance range, identification confidence level, and target discernibility level of the candidate target are calculated to obtain the long-range infrared target identification result. Step 7, Feedback Update; Based on the target discernibility level and identification confidence, the parameters for constructing the propagation state field, the parameters for aggregating the energy of weak targets, the parameters for suppressing background nonstationarity, and the parameters for maintaining cross-scale identity evidence are updated.

2. The long-range infrared target recognition method based on propagation state field and cross-scale evidence preservation as described in claim 1, characterized in that, In step 2, the propagation state field is composed of one or more of the following: local contrast component, background thermal clutter component, target scale response component, inter-frame stability component, image sharpness component, and propagation degradation component. Its propagation state field is expressed as: ； in, Indicates local contrast; Indicates the intensity of background thermal clutter; Indicates the target scale response; Indicates inter-frame stability; Indicates image sharpness; This represents the propagation degradation estimate; This represents the propagation state fusion function.

3. The long-range infrared target identification method based on propagation state field and cross-scale evidence preservation as described in claim 2, characterized in that, The propagation state field is obtained using a weighted fusion method: ； in, Indicates the first propagation state components; This represents the corresponding weight coefficient; Indicates the number of propagation state components.

4. The long-range infrared target identification method based on propagation state field and cross-scale evidence preservation as described in claim 1, characterized in that, Step 3, the aggregation of weak target energy, includes: Local thermal anomaly responses are extracted from continuous time phases. These responses are then weighted by reliability based on the propagation state field. Thermal anomaly responses satisfying temporal continuity, spatial proximity, and motion consistency are connected to form candidate energy trajectories. The aggregated response of these candidate energy trajectories is expressed as: ； in, Indicates the first Aggregate response of candidate energy trajectories; Indicates the first The candidate energy trajectory in the first Spatial support region in each temporal phase; Indicates reliability weight; This represents the pixel grayscale value or normalized thermal radiation response value at the corresponding location.

5. The long-range infrared target identification method based on propagation state field and cross-scale evidence preservation as described in claim 1, characterized in that, In step 4, background nonstationarity suppression includes: Energy attribution relationships are established based on the thermal response intensity, response duration, propagation state field response, and temporal stability within the candidate region, dividing the candidate region into a target energy attribution domain, a background energy attribution domain, and an uncertain mixed domain. ； The energy attribution probability of a candidate region is expressed as: ； And satisfy: 。 6. The long-range infrared target identification method based on propagation state field and cross-scale evidence preservation as described in claim 1, characterized in that, In step 5, cross-scale identity evidence preservation includes: For candidate targets, thermal structure evidence, edge contour evidence, motion trend evidence and category response evidence are extracted at at least two scales, and the identity consistency of the same candidate target is maintained at different scales through scale consistency constraints. The chain of evidence for thermal radiation is represented as follows: ； in, Indicates the first Target identity evidence at various scales; Indicates the number of scales.

7. The long-range infrared target identification method based on propagation state field and cross-scale evidence preservation as described in claim 6, characterized in that, The scale consistency constraint is: ； in, This indicates that cross-scale identity evidence remains constrained; This represents the scaling function.

8. The long-range infrared target identification method based on propagation state field and cross-scale evidence preservation as described in claim 1, characterized in that, In step 6, the joint judgment result is expressed as follows: ； in, Indicates the first The joint judgment result of the candidate targets; This represents the joint decision function; the joint decision result includes at least the target category, target spatial location, target distance interval, identification confidence level, and target discernibility level.

9. The long-range infrared target identification method based on propagation state field and cross-scale evidence preservation as described in claim 8, characterized in that, The target discernibility score is: ； in, Indicates the integrity of the target thermal structure; Indicates the separation degree between the target and the background; Indicates timing stability; Indicates the reliability of the propagation state; This represents the weighting coefficient.

10. The long-range infrared target identification method based on propagation state field and cross-scale evidence preservation as described in claim 1, characterized in that, In step 7, the propagation state field construction parameters, weak target energy aggregation parameters, background non-stationary suppression parameters, and cross-scale identity evidence preservation parameters are updated based on the target discernibility level and recognition confidence level to support continuous target recognition in subsequent time phases.

Images