Infrared target detection method based on multidimensional degraded spectrum estimation and adaptive enhancement

Abstract

This invention discloses an infrared target detection method based on multidimensional degradation spectrum estimation and adaptive enhancement. The method generates a global degradation spectrum and a local degradation spectrum through a degradation spectrum estimation network, and fuses them to obtain a multidimensional degradation spectrum tensor that represents the overall degradation trend, local degradation intensity, and spatial degradation distribution. Then, enhancement control parameters are generated based on the multidimensional degradation spectrum. Finally, the enhanced infrared image is input into the target detection branch to obtain the target bounding box, category confidence, target heatmap, and multi-scale detection features. A detection feedback weight map is generated, and this weight map is used to apply weighted constraints to the enhancement loss, target high-frequency preservation loss, local contrast preservation loss, and teacher-student feature distillation loss, so that the enhancement module and the detection network focus on the target region, small target region, weak target region, low-confidence difficult region, and target edge region.

Description

Technical Field

[0001] This invention belongs to the fields of infrared image processing, image enhancement and reconstruction, and intelligent target detection technology. Specifically, it relates to an infrared target detection method based on multidimensional degradation spectrum estimation and adaptive enhancement, which can be applied to the detection of small and weak targets under complex weather conditions, long-distance observation, complex thermal backgrounds, and low contrast conditions. Background Technology

[0002] Infrared imaging equipment can form images based on the difference in thermal radiation between a target and the background, and has important application value in scenarios such as night surveillance, unmanned platform sensing, border inspection, industrial thermal anomaly detection, and security early warning. However, actual infrared images are usually affected by a combination of factors during the imaging process, including sensor noise, non-uniform response, atmospheric attenuation, thermal diffusion, motion disturbance, imaging distance, and complex background thermal interference. These factors often result in various degradation phenomena such as noise, stripes, low contrast, blurring, thermal saturation, insufficient resolution for small targets, and strong background clutter.

[0003] Existing infrared target detection methods typically employ either enhancement-before-detection or direct end-to-end detection. Enhancement-before-detection methods often use fixed denoising, destriating, histogram equalization, sharpening, or super-resolution processing workflows. While these can improve some visual effects, they cannot dynamically adjust the processing strategy based on the actual degradation state of the current image, easily leading to problems such as over-enhancement, under-enhancement, smoothing of weak targets, or amplification of background noise. Direct end-to-end detection methods, while simplifying the process, are more sensitive to changes in degradation type under conditions of cross-device operation, weather conditions, and complex thermal backgrounds.

[0004] More importantly, most existing enhancement algorithms prioritize pixel reconstruction error or visual sharpness as optimization targets, failing to adequately consider whether the enhancement results are beneficial for target localization, target classification, and weak target recall. When small infrared targets only exhibit weak bright spots, local temperature differences, or high-frequency edge responses, conventional denoising and reconstruction processes may mistakenly interpret them as noise suppression. Conversely, when background thermal interference is strong, conventional contrast enhancement may amplify false targets in the background. Therefore, how to adaptively enhance infrared images based on degradation type and spatial location, while recognizing the multidimensional degradation states of infrared images, and using detection results to inversely constrain the enhancement process, is a pressing issue in the field of infrared target detection. Summary of the Invention

[0005] This invention proposes an infrared target detection method based on multidimensional degraded spectrum estimation and adaptive enhancement. The overall technical approach of this method is as follows:

[0006]

[0007] in, This represents the original infrared image input. This represents the multidimensional degradation spectrum estimated from the original infrared image. This represents the infrared image after adaptive enhancement. This represents the target detection result. The invention mainly comprises three technical modules: a multidimensional degradation spectrum estimation module, a degradation spectrum-driven adaptive enhancement module, and a detection feedback constraint and feature distillation module.

[0008] The technical solution of the present invention is as follows:

[0009] The infrared target detection method based on multidimensional degenerate spectrum estimation and adaptive enhancement has the following steps:

[0010] Step 1: Construct a multidimensional degradation spectrum estimation. Normalize the input raw infrared image and extract various degradation features from the image, such as noise degradation, stripe or non-uniformity degradation, low contrast degradation, blurring degradation, thermal saturation degradation, insufficient resolution degradation, and complex background thermal interference. Generate global degradation spectrum and local degradation spectrum through the degradation spectrum estimation network, and fuse the two to obtain a multidimensional degradation spectrum tensor that can characterize the overall degradation trend, local degradation intensity, and spatial degradation distribution of the infrared image.

[0011] The second step involves constructing an adaptive enhancement driven by the degradation spectrum. Enhancement control parameters are generated based on the multidimensional degradation spectrum, and these parameters are used to dynamically control the activation level and fusion weight of multiple enhancement sub-branches, such as denoising, stripe removal, local contrast enhancement, deblurring, thermal saturation repair, small target super-resolution enhancement, and edge high-frequency enhancement. Simultaneously, a spatial gating map is generated based on the local degradation spectrum, enabling the enhancement process to be targeted according to the degradation type and intensity at different spatial locations. Finally, an enhanced infrared image is generated through residual enhancement to maintain the original infrared thermal radiation structure and reduce the risk of false target generation.

[0012] Step 3: Construct detection feedback constraints and feature distillation optimization. The enhanced infrared image is input into the target detection branch to obtain the target bounding box, category confidence, target heatmap, and multi-scale detection features. A detection feedback weight map is generated based on the detection results. This detection feedback weight map is used to apply weighted constraints to the enhancement loss, target high-frequency preservation loss, local contrast preservation loss, and teacher-student feature distillation loss. This ensures that the enhancement module and detection network focus on target regions, small target regions, weak target regions, low-confidence difficult regions, and target edge regions. As a result, the enhancement process directly serves target localization, target classification, and weak target recall. Finally, the category, location, and confidence of the infrared target are output.

[0013] The first step, multidimensional degradation spectrum estimation, is specifically performed as follows:

[0014] 1.1 Generation of Global and Local Degradation Spectra

[0015] Infrared images are typically affected not by a single degradation factor during the imaging process, but by a combination of factors. This invention first models the degradation state of the input infrared image, representing the actual acquired infrared image as the result of a composite degradation process on an ideal infrared image:

[0016]

[0017] in, This represents the actual raw infrared image acquired. This represents an ideal, non-degradable infrared image. This represents the composite degradation process in infrared images. Represents a multidimensional degradation spectrum. This represents random noise, sensor disturbances, or other unmodelable interference.

[0018] To describe the impact of different degradation factors on infrared images, this invention defines the multidimensional degradation spectrum as follows:

[0019]

[0020] in, Indicates the intensity of noise degradation. Indicates the intensity of stripe or non-uniform degradation. Indicates the intensity of low contrast degradation. Indicates the intensity of fuzzy degradation. Indicates the intensity of thermal saturation degradation. Indicates the intensity of degradation due to insufficient resolution. This indicates the intensity of high-frequency thermal interference at the edges. This degradation spectrum does not simply determine whether the image quality is good or bad, but rather identifies where the infrared image is defective and how severe each type of degradation is.

[0021] 1.2 Multidimensional Degradation Spectrum Fusion

[0022] This section fuses the global degradation spectrum with the local degradation spectrum to obtain a multidimensional degradation spectrum tensor. This tensor simultaneously characterizes the overall degradation type, degradation intensity, and spatial distribution location of the infrared image, providing a control basis for subsequent adaptive enhancement. To obtain the aforementioned multidimensional degradation spectrum, the original infrared image is... Input Degeneracy Spectrum Estimation Network Degradation features are extracted from infrared images, and a global degradation spectrum is obtained through pooling and mapping:

[0023]

[0024] in, This indicates a global average pooling operation. This represents the mapping function of a multilayer perceptron. (·) express An activation function is used to normalize the intensity of each degradation dimension to the [0,1] interval. Since degradation in infrared images exhibits spatial inhomogeneity—for example, stripes may be concentrated in local areas, thermal saturation may only occur near high-temperature targets, and complex background thermal interference may only exist in certain areas—this invention further generates a local degradation spectrum:

[0025]

[0026] in, This represents a locally degenerate spectral estimation network. Indicates the number of degradation types. and These represent the height and width of the infrared image, respectively. Indicates the first Class degradation at pixel position The local degradation intensity at a given location. To utilize both global and local degradation information simultaneously, this invention fuses the two to obtain the final multidimensional degradation spectrum tensor used to guide enhancement:

[0027]

[0028] in, Indicates the first The final degradation intensity of the class at position (x,y). Indicates the first in the global degradation spectrum The intensity of class degradation, This represents the adjustment coefficient of the local degradation spectrum. Therefore, the final degradation spectrum considers both the overall degradation trend of the entire image and the local degradation differences at different spatial locations.

[0029] During the training phase, if degenerate labels exist, a supervised approach is used to constrain the degenerate spectrum estimation results:

[0030]

[0031] in, Indicates the global degradation spectrum label, whether real or synthesized. Indicates the label of the local or multidimensional degenerate spectrum tensor. This represents the weighting coefficient for local degradation spectrum loss.

[0032] The second step, degradation spectrum-driven adaptive enhancement, is implemented as follows:

[0033] 2.1 Enhanced Control Parameter Generation

[0034] Based on the global degradation information in the multidimensional degradation spectrum, enhancement control parameters are generated to dynamically determine the activation degree and fusion weight of enhancement sub-branches such as denoising, destriding, local contrast enhancement, deblurring, thermal saturation repair, small target super-resolution enhancement, and edge high-frequency enhancement.

[0035] After obtaining the multidimensional degradation spectrum, this invention does not employ a fixed enhancement strategy, but rather dynamically adjusts the activation level and fusion weight of different enhancement sub-branches based on the degradation type and intensity of the current infrared image. First, based on the global degradation spectrum... Generate enhanced control parameters:

[0036]

[0037] in, Represents the enhanced control parameter vector. and Represents the learnable linear mapping parameters. Represents a non-linear activation function. (·) denotes the normalization function. Enhanced control parameters are expressed as:

[0038]

[0039] in, This indicates the denoising enhancement weights. This indicates the weighting of stripes removal. Indicates the weight of local contrast enhancement. This indicates deblurring and weight enhancement. Indicates the heat saturation repair weight. This indicates the weights for super-resolution enhancement of small targets. This indicates the edge high-frequency enhancement weight.

[0040] 2.2 Spatial Gating Enhancement Processing

[0041] A spatially gated map is generated based on the local degradation spectrum, enabling the enhancement module to perform differentiated processing on the degradation type and intensity of different regions in the image. For example, it can enhance destriating in areas with obvious stripes, enhance high-frequency and contrast preservation in areas with weak targets, and suppress noise and false targets in complex background areas.

[0042] To achieve targeted processing for different degradation types, this invention sets up multiple enhancement sub-branches. First, the basic features of the input infrared image are obtained through a shallow feature extraction network:

[0043]

[0044] in, This represents a shallow feature extraction network used to extract basic texture, edge, and thermal response features from an input infrared image. Subsequently, the basic features... By inputting different enhancement sub-branches, we obtain the enhancement features corresponding to different degradation types. Enhanced features of the output of each enhanced sub-branch:

[0045]

[0046] in, Corresponding to the denoising branch, Corresponding to the stripe removal branch, Corresponding to the contrast enhancement branch, Corresponding to the defuzzing branch, Corresponding to the heat saturation repair branch, Corresponding to the super-resolution enhancement branch, Corresponding to the high-frequency enhancement branch at the edge.

[0047] To enable the enhancement process to not only know which type of degradation to enhance, but also which spatial locations to enhance, this invention further utilizes local degradation spectra. Generate a spatial gating graph and apply it, along with global augmentation weights, to each augmentation sub-branch:

[0048]

[0049] in, This indicates the enhanced features after fusion. Indicates the first Global fusion weights for each enhanced sub-branch ( ) indicates that by the first Spatial gating maps generated by local degradation spectra, This indicates pixel-wise or position-wise multiplication. This method achieves joint enhancement of degradation type adaptation and spatial location adaptation.

[0050] 2.3 Multi-branch feature fusion and residual enhancement

[0051] The enhanced features output from each enhancement sub-branch are fused according to global weights and a spatial gating map, and a residual compensation map is generated through convolutional mapping. The residual compensation map is then added to the original infrared image to obtain the enhanced infrared image, thus preserving the original thermal radiation structure and reducing the risk of false target generation.

[0052] After obtaining the fused and enhanced features, an enhanced infrared image is generated using a residual method:

[0053]

[0054] in, This represents the convolution mapping function, used to fuse and enhance features. Convert to a residual compensation map with the same size as the input infrared image; This represents the final enhanced infrared image. The significance of using the residual form in this formula is that the enhancement module only learns the degradation details that need to be compensated, rather than completely regenerating the entire image. This helps to preserve the original infrared thermal radiation structure and reduces the risk of generating false targets.

[0055] Furthermore, to prevent small, weak, and low-contrast infrared targets from being smoothed during enhancement or denoising, this invention introduces a high-frequency preservation mechanism. For any infrared image... Its high-frequency response is defined as:

[0056]

[0057] in, The first-order gradient response represents the weighting coefficients. The second-order gradient or Laplacian response is used to describe local bright spots and abrupt changes, while the first-order gradient response is used to describe target edges and contours. Based on this high-frequency response, a high-frequency preservation loss for small targets is constructed:

[0058]

[0059] in, This loss can be obtained from training bounding boxes, candidate box scales, or detection feedback weight maps, and is used to identify small target regions. This loss constrains the enhanced image to preserve bright spots, edges, contours, and local temperature differences within small target regions. Besides high-frequency information, infrared weak targets also rely on the local contrast between the target and the surrounding background. For target candidate regions... Its local contrast is defined as:

[0060]

[0061] To prevent the enhanced difference between the target and the background from being weakened, this invention constructs a local contrast preservation loss:

[0062]

[0063] This loss means that if the enhanced target-background contrast is lower than the original image, a penalty is imposed; if the enhanced contrast is higher than or equal to the original image, no penalty is imposed.

[0064] The third step, detection feedback constraints and feature distillation, is specifically operated as follows:

[0065] 3.1 Target Detection and Feedback Weight Map Generation

[0066] The enhanced infrared image is input into the target detection branch to obtain the target bounding box, category confidence, target heatmap, and multi-scale detection features. A detection feedback weight map is generated based on the target bounding box location, detection confidence, small target scale, target edges, and heatmap response. After adaptive enhancement, the enhanced infrared image... The input is given to the object detection branch, and the detection results are obtained:

[0067]

[0068] in, This refers to an object detection network or object detection branch. This represents the learnable parameters of the detection network. Detection results include the target candidate box set B, class confidence or target confidence P, target center heatmap H, detection head classification or regression response Z, and multi-scale detection features { }

[0069] 3.2 Enhanced Training Based on Detection Feedback Constraints

[0070] By using a detection feedback weight map to weight the enhancement loss, small target high frequency preservation loss, and local contrast preservation loss, the enhancement module focuses on the target region, small target region, weak target region, low confidence difficult region, and target edge region, thereby making the enhancement results more conducive to target localization and recall.

[0071] To ensure that the enhancement process directly serves the detection task, this invention generates a detection feedback weight map based on the target bounding box, confidence score, target heatmap, small target scale information, and edge response information output by the detection branch.

[0072]

[0073] in, This represents the target region weight map generated from the target candidate boxes. Represents the heat map of the target center. Weighted plot representing regions of low confidence or high uncertainty This represents a weighted map of small target regions. This represents the weight map of the target edge region. The weights of low-confidence or high-uncertainty regions can be obtained inversely from the detection confidence:

[0074]

[0075] in, Indicates pixel position or detection response position Belongs to the Confidence level of the target class; Indicates the target category index; This indicates the highest detection confidence for that location across all categories; This indicates the weight of low confidence or high uncertainty corresponding to that location. The formula means that the lower the detection confidence, the more likely the region is to be a weak target, a low-contrast target, or a difficult sample; therefore, its uncertainty weight is higher, and more attention should be paid to it in subsequent enhancement and distillation. For small target regions, the relative area of ​​the candidate boxes can be used for judgment.

[0076]

[0077] in, Indicates the first The area ratio of each candidate box relative to the entire image; and They represent the first The width and height of each candidate box. When Less than the preset threshold At that time, the first The region corresponding to each candidate box is determined as a small target region and written into... After obtaining the detection feedback weight map, this invention incorporates it into the enhancement loss, causing the enhancement module to prioritize processing detection-related regions:

[0078]

[0079] in, It can be a high-quality infrared image, a pseudo-high-quality image generated by a teacher network, or a reference image constructed under self-supervised supervision. This indicates the loss in detection and perception enhancement. This represents the adjustment coefficient of the detection feedback weight map to the enhancement loss. This loss gives higher weights to target regions, weak target regions, small target regions, and edge regions in the enhancement loss, so that the enhancement module no longer only pursues pixel similarity across the entire image, but focuses on recovering image information useful for detection.

[0080] To prevent the enhancement module from generating false targets in the background region, this invention also introduces a detection consistency constraint between the original image and the enhanced image. Detection is performed on the original infrared image and the enhanced infrared image respectively, yielding:

[0081]

[0082] And construct the detection consistency loss:

[0083]

[0084] in, This represents the detection consistency loss between the original image and the enhanced image. and These represent the target heatmaps corresponding to the original infrared image and the enhanced infrared image, respectively. The weighting coefficients representing the bounding box consistency loss; Indicates generalized intersection, union, and comparison; This represents the true labeled box, the teacher's network prediction box, or the matched reference candidate box; This represents the candidate bounding boxes obtained from image augmentation. This loss allows the augmented image to improve the response to the real target, but limits the augmentation process from generating high-confidence targets out of thin air in the background region.

[0085] 3.3 Teacher-Student Feature Distillation and Joint Optimization

[0086] Building upon the detection feedback constraint, this invention further introduces a teacher-student feature distillation mechanism. A teacher detection network and a student detection network are introduced to extract multi-scale detection features, target heatmaps, and detection head responses, respectively. The distillation loss is weighted using a detection feedback weight map, enabling the student network to focus on learning the feature representations of the teacher network in infrared weak target regions and difficult regions. Furthermore, a joint loss function simultaneously optimizes degradation spectrum estimation, image enhancement, and target detection results.

[0087] The teacher detection network is used to extract high-quality detection features, while the student detection network is used for practical deployment. Both the teacher and student networks extract multi-scale detection features respectively.

[0088]

[0089] in, This indicates a teacher assessment network. This indicates that students are testing the network; and This indicates the teacher network and the student network. Detection features at various scales; This represents the total number of multi-scale feature layers participating in distillation. Since the number of feature channels may differ between the teacher and student networks, a feature adaptation layer is introduced.

[0090]

[0091] in, Indicates the first A feature adaptation layer at each scale is used to map student network features to the same number of channels or feature dimensions as teacher network features.

[0092] To focus distillation on target and difficult regions, rather than averaging distillation across the entire image, this invention uses a detection feedback weight map to weight multi-scale feature distillation:

[0093]

[0094] in, Indicates the first A set of spatial locations on a scale feature map This indicates the number of spatial locations in the feature map at this scale. Represents a spatial location on the feature map. This indicates scaling the detection feedback weight map to the [number]th [level]. After each feature scale, at the location The weight value at that location, This represents the adjustment coefficient of the detection feedback weights on the distillation loss. The formula indicates that the target region, small target region, and low-confidence difficult region have higher weights during the distillation process, thus enabling the student detection network to focus on learning the key feature representations of the teacher network in infrared weak target detection. In addition to intermediate features, this invention also distills the target response heatmap to construct:

[0095]

[0096] in, This represents the distillation loss in the target response heatmap. and These represent the target heatmaps output by the student detection network and the teacher detection network, respectively. Furthermore, this invention can also incorporate detection head response distillation:

[0097]

[0098] in, This indicates the distillation loss in response to the detection head; and These represent the detection head classification or regression responses of the teacher detection network and the student detection network, respectively. This represents the normalized exponential function; Indicates the distillation temperature coefficient; This represents the Kullback-Leibler divergence, used to measure the difference between the teacher response distribution and the student response distribution; Represents the weighting coefficients of the bounding box distillation term; and Let represent the matching candidate boxes predicted by the teacher network and the student network, respectively. Through the distillation constraints described above, the student network learns not only the teacher network's judgment of the target category, but also the teacher network's ability to locate the target's position and boundaries.

[0099] In the training of the object detection network, the three modules mentioned above are not trained independently, but are jointly optimized under a unified objective function. The overall loss function is expressed as:

[0100]

[0101] in, This indicates the loss in multidimensional degradation spectrum estimation. This indicates the loss of detection and perception enhancement. This indicates that the loss is maintained at a high frequency for small targets. This indicates a loss of local contrast retention. Indicates the target detection loss. This represents the distillation loss of multi-scale detection features. This represents the distillation loss in the target response heatmap. This indicates the distillation loss in response to the detection head. This represents the consistency loss between the original image and the enhanced image detection. The object detection loss can be expressed as:

[0102]

[0103] The bounding box regression loss can be further expressed as:

[0104]

[0105] The beneficial effects of this invention are as follows:

[0106] 1. Strong multi-dimensional degradation perception capability of infrared images. This invention uses joint modeling of global and local degradation spectra to uniformly characterize various degradation factors in infrared images, such as noise, stripes, low contrast, blur, thermal saturation, insufficient resolution, and thermal interference from complex backgrounds. This enables the model to not only judge the quality of images but also identify specific degradation types, degradation intensities, and spatial distribution locations, which is beneficial for improving the image adaptability under complex weather conditions, long-distance observation, and complex thermal background conditions.

[0107] 2. More precise adaptive enhancement. This invention dynamically generates enhancement control parameters based on a multidimensional degradation spectrum and controls multiple enhancement sub-branches such as denoising, stripe removal, local contrast enhancement, deblurring, thermal saturation restoration, small target super-resolution enhancement, and edge high-frequency enhancement through spatial gating graphs. This allows the enhancement process to perform differentiated processing for different degradation types and different image regions, avoiding problems such as over-enhancement, under-enhancement, background noise amplification, and loss of target details caused by fixed enhancement methods.

[0108] 3. Improved preservation of weak targets. This invention introduces constraints on high-frequency preservation and local contrast preservation of small targets during the enhancement process, enabling the enhanced infrared image to retain target bright spots, edge contours, local temperature differences, and contrast differences between the target and the background. This reduces the risk of false suppression of weak targets during denoising, reconstruction, or enhancement, and helps improve the detection recall rate of small, weak, and low-contrast targets.

[0109] 4. Enhanced results are more conducive to target detection. This invention utilizes the target bounding box, confidence score, target heatmap, and multi-scale detection features output by the detection branch to generate a detection feedback weight map, which is then used for enhancement loss and teacher-student feature distillation constraints. This allows the enhancement module to focus on recovering effective detection information in target areas, small target areas, low-confidence difficult areas, and edge areas, thereby avoiding the enhancement process from pursuing only visual clarity while ignoring the requirements of the detection task and improving the robustness of infrared target localization, classification, and detection in complex backgrounds. Attached Figure Description

[0110] Figure 1 This is a flowchart of the infrared target detection method based on multidimensional degradation spectrum estimation and adaptive enhancement according to the present invention.

[0111] Figure 2 This is a rendering of an embodiment of the present invention. Detailed Implementation

[0112] The specific embodiments of the present invention will now be described with reference to the accompanying drawings. 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 equivalent substitutions or combinations regarding the degradation spectrum dimension, enhancement sub-branch, detection network structure, and distillation loss form.

[0113] Example 1: An Infrared Target Detection Method Based on Multidimensional Degradation Spectrum Estimation and Adaptive Enhancement

[0114] like Figure 1 As shown, in this embodiment, the infrared image to be detected is first acquired. Then, it is scaled to the size required for the detection network, while intensity normalization is performed. Subsequently, Input the multidimensional degradation spectrum estimation module, output the global degradation spectrum d from the degradation spectrum estimation network, and output the local degradation spectrum from the local degradation spectrum estimation network. Then, a multidimensional degenerate spectrum tensor is generated according to the multidimensional degenerate spectrum fusion formula. .

[0115] Then, An adaptive enhancement module driven by the input degradation spectrum is used. This module first generates enhancement control parameters a based on the global degradation spectrum d, and then generates a spatial gating map based on the local degradation spectrum. If the fringe degradation is strong, the weight of the defringe branch is increased; if the low-contrast degradation is strong, the weight of the local contrast enhancement branch is increased; if the resolution of small targets is insufficient, the weights of the super-resolution enhancement branch and the edge high-frequency enhancement branch are increased. The outputs of each enhancement sub-branch are fused with the global weights and the spatial gating map, and an enhanced infrared image is generated in the form of residuals. .

[0116] Finally, Inputting the object detection branch yields bounding boxes, confidence scores, object heatmaps, and multi-scale detection features. The system then generates a detection feedback weight map based on these detection results. and utilize Weighted loss algorithms are applied to the enhancement loss, multi-scale feature distillation loss, heatmap distillation loss, and detector head response distillation loss, ensuring that the enhancement module and student detection network focus on target regions, small target regions, low-confidence regions, and target edge regions. After training, the inference phase only requires a single frame of infrared image as input to sequentially complete degradation spectrum estimation, adaptive enhancement, and infrared target detection output.

[0117] Example 2: Infrared Target Detection in Complex Urban Night Scenes

[0118] Step 1: Input the acquired nighttime urban infrared images into the multidimensional degradation spectrum estimation module. Normalize the images and estimate the degradation intensity of noise, stripes, low contrast, blur, insufficient resolution, and complex background thermal interference. Generate global degradation spectrum and local degradation spectrum, and fuse them to obtain the multidimensional degradation spectrum tensor.

[0119] Step 2: Input the multidimensional degradation spectrum into the adaptive enhancement module. Dynamically adjust the weights of branches such as denoising, stripe removal, local contrast enhancement, small target super-resolution enhancement, and edge high-frequency enhancement according to the degradation type. Combine this with the spatial gating map to perform differentiated enhancement on different regions, generating an enhanced infrared image.

[0120] Step 3 inputs the enhanced infrared image into the target detection branch, outputting the category, location, and confidence level of pedestrians and vehicles. For example... Figure 2 As shown, compared with the original degraded image, the present invention can effectively suppress stripe noise and background clutter, enhance the edge and local contrast of distant and weak targets, and improve the recall rate and localization accuracy of infrared target detection.

Claims

1. An infrared target detection method based on multidimensional degenerate spectrum estimation and adaptive enhancement, characterized in that, The steps are as follows: Step 1: Construct a multidimensional degradation spectrum estimation. Normalize the input raw infrared image and extract multiple degradation features from the image, including noise degradation, stripe or non-uniformity degradation, low contrast degradation, blurring degradation, thermal saturation degradation, insufficient resolution degradation, and thermal interference from complex backgrounds. Generate global degradation spectrum and local degradation spectrum through the degradation spectrum estimation network, and fuse the two to obtain a multidimensional degradation spectrum tensor that can characterize the overall degradation trend, local degradation intensity, and spatial degradation distribution of the infrared image. The second step involves constructing an adaptive enhancement driven by the degradation spectrum. Enhancement control parameters are generated based on the multidimensional degradation spectrum, and these parameters are used to dynamically control the activation level and fusion weight of multiple enhancement sub-branches, including denoising, stripe removal, local contrast enhancement, deblurring, thermal saturation repair, small target super-resolution enhancement, and edge high-frequency enhancement. Simultaneously, a spatial gating map is generated based on the local degradation spectrum, enabling the enhancement process to be targeted according to the degradation type and intensity at different spatial locations. Finally, an enhanced infrared image is generated through residual enhancement to maintain the original infrared thermal radiation structure and reduce the risk of false target generation. The third step involves constructing detection feedback constraints and feature distillation. The enhanced infrared image is input into the target detection branch to obtain the target bounding box, category confidence, target heatmap, and multi-scale detection features. A detection feedback weight map is then generated based on the detection results. This weight map is used to apply weighted constraints to the enhancement loss, target high-frequency preservation loss, local contrast preservation loss, and teacher-student feature distillation loss. This ensures that the enhancement module and detection network focus on target regions, small target regions, weak target regions, low-confidence difficult regions, and target edge regions. Consequently, the enhancement process directly serves target localization, target classification, and weak target recall, ultimately outputting the category, location, and confidence level of the infrared target.

2. The infrared target detection method based on multidimensional degraded spectrum estimation and adaptive enhancement as described in claim 1, characterized in that, The first step, multidimensional degradation spectrum estimation, is specifically performed as follows: Step 1.1 Generation of Global and Local Degradation Spectra The degradation state of the input infrared image is modeled, and the actual acquired infrared image is represented as the result of the ideal infrared image after composite degradation: in, This represents the actual raw infrared image acquired. This represents an ideal, non-degradable infrared image. This represents the composite degradation process in infrared images. Represents a multidimensional degradation spectrum. This represents random noise, sensor disturbances, or other unmodelable interference. The multidimensional degradation spectrum is defined as: in, Indicates the intensity of noise degradation. Indicates the intensity of stripe or non-uniform degradation. Indicates the intensity of low contrast degradation. Indicates the intensity of fuzzy degradation. Indicates the intensity of thermal saturation degradation. Indicates the intensity of degradation due to insufficient resolution. Indicates the intensity of high-frequency thermal interference at the edge; Step 1.2 Multidimensional Degradation Spectrum Fusion Original infrared image Input Degeneracy Spectrum Estimation Network Degradation features are extracted from infrared images, and a global degradation spectrum is obtained through pooling and mapping: in, This indicates a global average pooling operation. This represents the mapping function of a multilayer perceptron. (·)express An activation function is used to normalize the intensity of each degradation dimension to the [0,1] interval; Generate a locally degraded spectrum: in, This represents a locally degenerate spectral estimation network. Indicates the number of degradation types. and These represent the height and width of the infrared image, respectively. Indicates the first Class degradation at pixel position The intensity of local degradation at the location; Multidimensional Degenerate Spectral Tensor: in, Indicates the first The final degradation intensity of the class at position (x,y). Indicates the first in the global degradation spectrum The intensity of class degradation, The adjustment coefficient representing the local degradation spectrum; During the training phase, if degenerate labels exist, a supervised approach is used to constrain the degenerate spectrum estimation results: in, Indicates the global degradation spectrum label, whether real or synthesized. Indicates the label of the local or multidimensional degenerate spectrum tensor. This represents the weighting coefficient for local degradation spectrum loss.

3. The infrared target detection method based on multidimensional degraded spectrum estimation and adaptive enhancement as described in claim 1, characterized in that, The second step, degradation spectrum-driven adaptive enhancement, is implemented as follows: Step 2.1 Enhanced control parameter generation According to the global degradation spectrum Generate enhanced control parameters: in, Represents the enhanced control parameter vector. and Represents the learnable linear mapping parameters. Represents a non-linear activation function. (·) represents the normalization function; the enhanced control parameters are expressed as: in, This indicates the weights for denoising and enhancing the signal. This indicates weighting to remove stripes. Indicates the weight of local contrast enhancement. This indicates deblurring and weight enhancement. Indicates the heat saturation repair weight. This indicates the weights for super-resolution enhancement of small targets. Indicates the high-frequency enhancement weight at the edge; Step 2.2 Spatial Gating Enhancement Processing Multiple enhancement sub-branches are set up; firstly, the basic features of the input infrared image are obtained through a shallow feature extraction network: in, This represents a shallow feature extraction network used to extract basic texture, edge, and thermal response features from an input infrared image; subsequently, the basic features are... By inputting different enhancement sub-branches, we obtain the enhancement features corresponding to different degradation types. Enhanced features of the output of each enhanced sub-branch: in, Corresponding to the denoising branch, Corresponding to the stripe removal branch, Corresponding to the contrast enhancement branch, Corresponding to the defuzzing branch, Corresponding to the heat saturation repair branch, Corresponding to the super-resolution enhancement branch, Corresponding edge high-frequency enhancement branch; Using local degradation spectrum Generate a spatial gating graph and apply it, along with global augmentation weights, to each augmentation sub-branch: in, This indicates the enhanced features after fusion. Indicates the first The global fusion weights of each enhanced sub-branch ( ) indicates that by the first Spatial gating maps generated by local degradation spectra. This indicates pixel-wise or position-wise multiplication; Step 2.3 Multi-branch feature fusion and residual enhancement The enhanced features output from each enhanced sub-branch are fused according to global weights and spatial gating graphs, and a residual compensation map is generated through convolutional mapping. After obtaining the fused and enhanced features, an enhanced infrared image is generated using a residual method: in, This represents the convolution mapping function, used to fuse and enhance features. Convert to a residual compensation map with the same size as the input infrared image; This represents the final enhanced infrared image.

4. The infrared target detection method based on multidimensional degenerate spectrum estimation and adaptive enhancement as described in claim 1, characterized in that, The third step, detection feedback constraints and feature distillation, is specifically operated as follows: Step 3.1 Target Detection and Feedback Weight Map Generation The infrared image is input into the target detection branch to obtain the target bounding box, category confidence, target heatmap, and multi-scale detection features; A detection feedback weight map is generated based on the target bounding box location, detection confidence, small target scale, target edges, and heatmap response; after adaptive enhancement, the infrared image is enhanced. The input is given to the object detection branch, and the detection results are obtained: in, This indicates an object detection network or object detection branch. This represents the learnable parameters of the detection network; the detection results include the target candidate box set B, class confidence or target confidence P, target center heatmap H, detection head classification or regression response Z, and multi-scale detection features { }; Step 3.2 Detection Feedback Constraint Reinforcement Training Based on the target bounding box, confidence score, target heatmap, small target scale information, and edge response information output by the detection branch, a detection feedback weight map is generated: in, This represents the target region weight map generated from the target candidate boxes. Represents the heat map of the target center. This represents a weighted map indicating regions of low confidence or high uncertainty. This represents a weighted map of small target regions. This represents the weight map of the target edge region; the weights of low-confidence or high-uncertainty regions are derived inversely from the detection confidence: in, Indicates pixel position or detection response position Belongs to the Confidence level of the target class; Indicates the target category index; This indicates the highest detection confidence for that location across all categories; This indicates the weight of low confidence or high uncertainty corresponding to that position; For small target regions, the determination is based on the relative area of ​​the candidate boxes: in, Indicates the first The area ratio of each candidate box relative to the entire image; and They represent the first The width and height of each candidate box; when Less than the preset threshold At that time, the first The region corresponding to each candidate box is determined as a small target region and written into... After obtaining the detection feedback weight map, an enhancement loss is introduced to make the enhancement module prioritize processing detection-related regions. in, It can be a high-quality infrared image, a pseudo-high-quality image generated by a teacher network, or a reference image constructed under self-supervised supervision. This indicates the loss in detection and perception enhancement. This represents the adjustment coefficient of the detection feedback weight map on the enhancement loss; Step 3.3 Teacher-Student Feature Distillation and Joint Optimization Teacher detection network and student detection network are introduced to extract multi-scale detection features, target heatmap and detection head response respectively; the distillation loss is weighted by the detection feedback weight map, so that the student network focuses on learning the feature representation of the teacher network in infrared weak target area and difficult area, and the degradation spectrum estimation, image enhancement and target detection results are optimized simultaneously by the joint loss function; The teacher detection network is used to extract high-quality detection features, while the student detection network is used for actual deployment; both the teacher and student networks extract multi-scale detection features respectively. in, This indicates a teacher assessment network. This indicates that students are testing the network; and This indicates the teacher network and student network. Detection features at various scales; This represents the total number of multi-scale feature layers participating in distillation; since the number of feature channels in the teacher network and the student network may differ, a feature adaptation layer is introduced: in, Indicates the first A feature adaptation layer at each scale is used to map student network features to the same number of channels or feature dimensions as teacher network features.

5. The infrared target detection method based on multidimensional degenerate spectrum estimation and adaptive enhancement as described in claim 4, characterized in that, In step 3.2, a consistency constraint between the detection of the original image and the enhanced image is introduced; the original infrared image and the enhanced infrared image are detected separately to obtain: And construct the detection consistency loss: in, This represents the detection consistency loss between the original image and the enhanced image. and These represent the target heatmaps corresponding to the original infrared image and the enhanced infrared image, respectively. The weighting coefficients representing the bounding box consistency loss; Indicates generalized intersection, union, and comparison; This represents the true labeled box, the teacher's network prediction box, or the matched reference candidate box; This represents the candidate bounding boxes obtained from enhanced image detection.

6. The infrared target detection method based on multidimensional degenerate spectrum estimation and adaptive enhancement as described in claim 4, characterized in that, In step 3.3, the multi-scale feature distillation is weighted using a detection feedback weight map: in, Indicates the first A set of spatial locations on a scale feature map This indicates the number of spatial locations in the feature map at this scale. Represents a spatial location on the feature map. This indicates scaling the detection feedback weight map to the [number]th [level]. After each feature scale, at the location The weight value at that point, This represents the adjustment coefficient of the detection feedback weight on distillation loss; Distill the target response heatmap to construct: in, This represents the distillation loss in the target response heatmap. and These represent the target heatmaps output by the student detection network and the teacher detection network, respectively.

7. The infrared target detection method based on multidimensional degenerate spectrum estimation and adaptive enhancement as described in claim 4, characterized in that, In step 3.3, the detection head response distillation is introduced: in, This indicates the distillation loss in response to the detection head; and These represent the detection head classification or regression responses of the teacher detection network and the student detection network, respectively. This represents the normalized exponential function; Indicates the distillation temperature coefficient; This represents the Kullback-Leibler divergence, used to measure the difference between the teacher response distribution and the student response distribution; Represents the weight coefficients of the bounding box distillation term; and These represent the predicted matching candidate boxes for the teacher network and the student network, respectively.

8. The infrared target detection method based on multidimensional degenerate spectrum estimation and adaptive enhancement as described in claim 1, characterized in that, In the training of the object detection network, the overall loss function is expressed as: in, This indicates the loss in multidimensional degradation spectrum estimation. This indicates the loss of detection and perception enhancement. This indicates that the loss is maintained at a high frequency for small targets. This indicates a loss of local contrast retention. Indicates the target detection loss. This represents the distillation loss of multi-scale detection features. This represents the distillation loss in the target response heatmap. This indicates the distillation loss in response to the detection head. The loss represents the consistency loss between the original image and the enhanced image detection; the object detection loss is expressed as: The bounding box regression loss is expressed as: 。 9. The infrared target detection method based on multidimensional degenerate spectrum estimation and adaptive enhancement as described in claim 3, characterized in that, Step 2.3 introduces a high-frequency hold mechanism; for any infrared image Its high-frequency response is defined as: in, The first-order gradient response represents the weighting coefficients. The second-order gradient or Laplacian response is used to describe local bright spots and abrupt changes, while the first-order gradient response is used to describe target edges and contours. Based on this high-frequency response, a high-frequency retention loss for small targets is constructed: in, It can be obtained from training bounding boxes, candidate box scales, or detection feedback weight maps, and is used to identify small target regions; For the target candidate region Its local contrast is defined as: Constructing a local contrast preservation loss: 。

Images