Infrared unmanned aerial vehicle target recognition method for complex environment

Abstract

The present application relates to the technical field of infrared target recognition, and particularly relates to an infrared unmanned aerial vehicle target recognition method for complex environment, comprising: acquiring an infrared image and synchronous environment parameters, and constructing an atmospheric point spread function; constructing a sparse non-negative constraint radiation inversion target function and solving; calculating an inversion residual and generating a spatial attention weight; residual attention guided multi-scale feature recalibration; recalibration feature inputting a detection network to complete target recognition.In the scheme, first, natural edge information is automatically extracted from the infrared image to estimate the atmospheric point spread function, and a layered correction model is established in combination with the environment parameters acquired by the airborne sensor to convert single-point measurement values into path integral equivalent values, so as to accurately describe the image degradation process.On this basis, a sparse non-negative constraint target function is constructed and a split Bregman iteration algorithm is used to solve, so that the original radiation distribution of the target is recovered, background noise is effectively suppressed, and the edge details of the weak and small target are enhanced.

Description

Technical Field

[0001] This invention relates to the field of infrared target recognition technology, and in particular to an infrared UAV target recognition method for complex environments. Background Technology

[0002] Unmanned aerial vehicle (UAV) remote sensing technology is playing an increasingly important role in environmental monitoring, target reconnaissance, and disaster assessment. Among these technologies, infrared imaging, with its all-weather capability, lack of susceptibility to lighting conditions, and excellent smoke penetration, is gradually becoming a core tool for UAVs to detect targets in complex environments. The essence of infrared imaging is to convert the thermal radiation signals of the target and background into grayscale images, and then identify and locate targets of interest through image processing and analysis. With the miniaturization and increasing intelligence of UAV platforms, higher demands are placed on the adaptability of infrared target recognition technology in complex environments. Especially in scenarios with significant atmospheric attenuation, such as fog, smoke, rain, and snow, maintaining stable target recognition performance has become a critical issue that urgently needs to be addressed in this field.

[0003] Existing infrared UAV target recognition methods typically employ deep learning structures such as feature pyramid networks to extract multi-scale features, which are then input into a detection network for target classification and localization. However, in complex environments, the scattering and absorption of infrared radiation by atmospheric particles leads to image blurring, decreased contrast, and edge diffusion. The apparent size and radiation intensity of the target in the pixel domain are significantly distorted, originally clear edges are broadened, and details of small targets are obscured by noise. Traditional methods directly extract features from degraded images, ignoring the physical mechanisms of image degradation. This results in a mismatch between the features extracted from each layer of the feature pyramid and the actual target scale. High-level semantic features are insufficient to represent degraded small targets, while low-level detail features are easily affected by background noise. Ultimately, this leads to a significant increase in the false negative and false positive rates, failing to meet the practical application requirements of UAV remote sensing monitoring in complex environments. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an infrared UAV target recognition method for complex environments, thereby solving the technical problems mentioned in the background section.

[0005] To achieve the above objectives, the present invention provides the following technical solution: An infrared UAV target recognition method for complex environments includes the following steps: S1: Infrared images are acquired by an infrared thermal imager mounted on a drone, and environmental parameters are obtained simultaneously by airborne sensors; based on the environmental parameters, an atmospheric stratification correction model is used to convert single-point measurements into path integral equivalent values, and the atmospheric point spread function is estimated using natural edge information extracted from the infrared images. S2: Based on the atmospheric point spread function and environmental parameters, establish an infrared radiation transmission degradation model, introduce the sparsity and non-negativity priors of infrared targets, construct an objective function, and obtain the restored radiation distribution map through iterative optimization. S3: Calculate the absolute difference between the original degraded image and the restored observation based on the degradation model, generate an inversion residual map, normalize the inversion residual map and combine it with a learnable scaling parameter to construct a spatial attention mask; S4: Adjust the spatial attention mask to the same spatial size as the feature maps of each layer of the feature pyramid network, and recalibrate the multi-scale feature maps by element-wise multiplication to achieve feature enhancement in areas with severe atmospheric degradation. S5: Input the recalibrated multi-scale feature map into the target detection network, and output the target class probability and bounding box offset through the classification branch and regression branch respectively to complete the infrared UAV target recognition in complex environments.

[0006] In one possible implementation, in step S1, the specific method for estimating the atmospheric point spread function is as follows: detect the edges of ground features with step characteristics in the infrared image, extract the grayscale profile along the edge normal direction, obtain the edge spread function by fitting the error function, and then derive the standard deviation of the atmospheric point spread function; the atmospheric point spread function is approximated by a two-dimensional Gaussian model, and its standard deviation is obtained by taking the average of multiple edge fitting results.

[0007] In one possible implementation, in step S1, the atmospheric stratification correction model is specifically as follows: the radiation transmission path is divided into several layers of equal thickness according to the UAV's flight altitude and the average distance of the scene. Based on the measurement values ​​of the airborne sensor, the temperature, humidity and pressure values ​​of each layer are calculated in combination with the standard atmospheric vertical lapse rate. The equivalent path average extinction coefficient is obtained by integrating the extinction coefficient of each layer.

[0008] In one possible implementation, in step S2, the infrared radiation transmission degradation model is that the original degradation image is equal to the original radiation distribution convolved with the atmospheric point spread function, multiplied by the atmospheric transmittance, and noise is added; the atmospheric transmittance is calculated based on environmental parameters using the Beer-Lambert law.

[0009] In one possible implementation, in step S2, the objective function is: in This represents the restored radiation distribution to be solved. This represents the value of the independent variable that minimizes the objective function. For the original degraded image, The original radiation distribution, Here is the atmospheric point spread function. Atmospheric transmittance, Represents the image gradient. and The regularization parameter is used; the objective function is solved using the split Bregman iterative algorithm to obtain the restored radiation distribution map. .

[0010] In one possible implementation, in step S3, the inversion residual map Defined as: in For the original degraded image in coordinates The grayscale value at that location; Will Normalization yields the normalized residual plot. Then through learnable scaling parameters Generate spatial attention mask: in Spatial attention mask in coordinates The value at that location, To normalize the residual plot in coordinates The value at that location.

[0011] In one possible implementation, in step S4, the feature pyramid network contains multiple feature layers with different resolutions. The spatial attention mask is adjusted to the same size as the feature map of each layer by interpolation. The recalibrated feature map is the element-wise product of the original feature map and the attention weights at the corresponding positions.

[0012] In one possible implementation, in step S5, the target detection network uses a one-stage detection head, followed by a parallel classification sub-network and a regression sub-network after each feature layer, to predict the target category probability and the offset of the bounding box relative to the anchor point, respectively; during the inference stage, redundant detection boxes are removed by non-maximum suppression to obtain the final recognition result.

[0013] Beneficial effects compared to existing technologies: 1. In this scheme, a degraded target inversion and recognition strategy based on the cumulative effect of atmospheric transmission path radiation is constructed. First, natural edge information is automatically extracted from infrared images to estimate the atmospheric point spread function. Then, a hierarchical correction model is established by combining environmental parameters obtained from airborne sensors, converting single-point measurements into path integral equivalent values, thereby accurately describing the image degradation process. On this basis, a sparse non-negative constraint objective function is constructed and solved using the split Bregman iterative algorithm to recover the original radiation distribution map of the target. This effectively suppresses background noise and enhances the edge details of weak targets, providing a high-quality input image for subsequent feature extraction. 2. In this scheme, a spatial attention mask is generated using the residual between the inversion restoration result and the original degraded image. This attention mask has a clear physical meaning, directly indicating the region with the most severe atmospheric degradation. Applying the attention mask to the feature maps of each layer of the feature pyramid network enables adaptive recalibration of multi-scale features. This allows the network to dynamically focus on regions with severe information loss, obtaining enhanced feature responses in degraded regions while maintaining normal feature representation in background regions. This significantly improves the network's ability to extract features from multi-scale targets in complex environments. 3. In this scheme, an online hard example mining strategy incorporating physical priors is introduced during the training phase of the target detection network. Regions with severe atmospheric degradation are identified using residual maps, and the losses of corresponding anchor points are weighted and ranked, making these difficult-to-identify regions more likely to be selected as hard example samples for gradient updates. Simultaneously, it is explicitly stated that the weighted loss is only used for ranking and selection and does not participate in gradient calculation, ensuring the effectiveness of hard example selection while avoiding gradient distortion. This strategy guides the network to focus on regions difficult to identify due to atmospheric degradation during training, effectively improving the model's sensitivity to detecting small targets in complex environments. Attached Figure Description

[0014] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0015] Figure 1 This is a schematic diagram of the infrared UAV target recognition method of the present invention. Detailed Implementation

[0016] The present invention will now be described in detail with reference to specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. Those skilled in the art, based on their understanding of the inventive concept, can make various modifications and variations to the present invention, and these modifications and variations should also fall within the scope of protection of the present invention. The technical solutions in this application are designed to address the problems described in the background, and are generally as follows: Example: Please refer to Figure 1 This embodiment introduces an infrared UAV target recognition method for complex environments. More specifically, this embodiment addresses the challenges of infrared image degradation, target scale distortion, and low recognition rate caused by atmospheric attenuation in complex environments by proposing a degraded target inversion and recognition strategy based on the cumulative radiation effect of atmospheric transmission paths. The steps are described in detail below.

[0017] S1: Acquire infrared images and synchronize environmental parameters, and construct the atmospheric point spread function. This step first involves acquiring infrared images of the target area using an infrared thermal imager mounted on the drone, while simultaneously using the onboard sensor system to obtain environmental parameter vectors corresponding to the current frame in real time, including atmospheric temperature. relative humidity Atmospheric pressure and the average distance of the scene measured by the laser rangefinder .in , , Affects atmospheric absorption and scattering coefficients. Used to estimate the length of the radiation transmission path. To ensure the accuracy of parameter synchronization, the sensor data acquisition frequency was set to 10 Hz, consistent with the frame rate of the infrared camera, and timestamp alignment was used to ensure that each frame of the image corresponds to real-time environmental parameters.

[0018] Considering that airborne sensors measure environmental parameters at the current location of the UAV, and atmospheric conditions along the radiation transmission path may exhibit spatial non-uniformity (especially in complex environments such as fog, haze, and smoke), this embodiment introduces an atmospheric stratification correction model to convert single-point measurements into path integral equivalent values. Specifically, based on the UAV's flight altitude... (Acquired in real time by airborne altitude sensors) and average scene distance By combining standard atmospheric vertical profiles, the path is divided into several layers of equal thickness, using measurements taken at the current altitude of the UAV. , , Based on this, the temperature, humidity, and pressure values ​​of each layer are calculated using the vertical lapse rate specified by international standard atmospheric standards (temperature lapse rate approximately 6.5 K / km, pressure lapse rate follows an exponential law, and humidity decreases exponentially with increasing altitude). Then, the extinction coefficient of each layer is integrated to obtain the equivalent path-averaged extinction coefficient. When the flight altitude is below 200 m and the atmosphere is uniformly mixed, the airborne measurements can be directly used as representative values ​​for the entire path, in which case the error is negligible. Through the above processing, the representativeness of environmental parameters for the entire transmission path is ensured.

[0019] Complex environments (such as haze, smoke, rain, and snow) primarily affect infrared imaging through the scattering and absorption of target radiation by atmospheric particles, leading to image blurring, decreased contrast, and edge diffusion. To quantitatively describe this degradation process, an atmospheric point spread function (PSF) needs to be constructed. This step, based on atmospheric radiative transfer theory, estimates the PSF using natural edge information automatically extracted from infrared images. Specifically, it detects edges of features with step characteristics (such as building outlines, horizons, and ridgelines) in infrared images, extracts grayscale profiles of these edges, and obtains the edge spread function by fitting an error function to the profiles, thus deriving the atmospheric PSF. In practice, the Canny operator is first used to extract image edges, then edge grayscale profiles are sampled along the edge normal direction. The profile data is then fitted with the error function model using least squares to obtain the variance parameter of the edge spread function, which is directly related to the standard deviation of the atmospheric PSF. To improve estimation accuracy, sub-pixel edge localization technology is employed, using interpolation to obtain finer grayscale profiles, thereby reducing quantization errors. The atmospheric point spread function is approximated using a two-dimensional Gaussian model: in Indicates the atmospheric point spread function at pixel coordinates The value at that location, This represents the standard deviation of the point spread function, the value of which is determined in real time by fitting the edge spread curve. This standard deviation reflects the severity of atmospheric attenuation: that is... The larger the value, the more severe the image blur. To ensure the accuracy of the estimation, this step selects multiple edges with different directions and locations for fitting, and takes the average of the edge estimation results as the final value. This reduces the impact of single-edge noise. Furthermore, for images without obvious edges, a blind estimation method based on image patch statistics can be used as a backup, but edge-based estimation is preferred to ensure clear physical meaning.

[0020] In this way, this step can estimate the degradation kernel function corresponding to the current atmospheric environment directly from a single frame image without relying on additional weather stations, thus ensuring that the subsequent inversion and restoration has an accurate physical model. Simultaneously, the acquired environmental parameters... , , , This will be used in step S2 to calculate atmospheric transmittance. The transmittance and the point spread function together constitute a complete degradation model.

[0021] S2: Construct and solve the radiative inversion objective function with sparse nonnegativity constraints. Obtaining the atmospheric point spread function Next, the infrared image acquired in step S1, i.e., the original degraded image, needs to be processed. The original radiation distribution of the target was recovered. Original degraded image It can be modeled as primitive radiation With point spread function The convolution multiplied by atmospheric transmittance And superimposed noise .

[0022] Atmospheric transmittance The environmental parameters obtained in step S1 are calculated according to the Beer-Lambert law, specifically as follows: ,in The atmospheric extinction coefficient can be determined according to... , , The atmospheric extinction coefficient can be estimated by referring to a table or using empirical formulas. This embodiment uses the following empirical formula to calculate the atmospheric extinction coefficient. : in As the reference extinction coefficient, , , This is an empirical coefficient, obtained through prior calibration using standard radiative transfer software (such as MODTRAN) under different atmospheric conditions. In this embodiment, we take... =0.15km -1 , =0.02, =0.005, =-0.0001. Inversion This is a typical ill-conditioned inverse problem, and direct deconvolution will amplify the noise. In order to obtain a stable and physically reasonable solution, this invention introduces two physical priors unique to infrared targets: sparsity and nonnegativity.

[0023] In infrared images, the target typically occupies only a small number of pixels, while the background occupies the vast majority, thus the original radiation distribution is... It exhibits sparsity; simultaneously, the radiation intensity value is always non-negative. Furthermore, the radiation distribution within the target should be continuous and smooth; therefore, a total variational regularization term is introduced to constrain smoothness. Based on the above priors, the following objective function is constructed: in This represents the restored radiation distribution to be solved. This represents the value of the independent variable that minimizes the objective function. Represents the L2 norm. Describing the L1 norm, Represents the image gradient. and This is a regularization parameter that can be adaptively set according to the image signal-to-noise ratio, typically... Take a value between 0.01 and 0.1. A value between 0.05 and 0.2 is used to balance the weights of the data fidelity, sparsity, and smoothness terms. This embodiment employs an adaptive adjustment strategy: First, the signal-to-noise ratio (SNR) of the image is estimated. The estimation method is as follows: Select the 10% pixel region in the image with the most gradual gray-level changes (determined by the minimum local variance criterion), and calculate the variance of the pixel gray-level values ​​within this region as the noise variance. Simultaneously, the 5% pixel region with the largest gradient magnitude in the image (as the potential signal region) is selected, and the mean gray value of the pixels within this region is calculated. Then the signal power is approximately: The signal-to-noise ratio (SNR) is defined as SNR = / Then let =0.05 / SNR, =0.1 / SNR, which provides stronger regularization for low signal-to-noise ratio images. The first term of this objective function ensures that the restored result is consistent with the observed data, while the second term forces... To be as sparse as possible (i.e., most pixel values ​​are zero), the third term makes The gradient is sparse, thus preserving the target edge while suppressing background noise.

[0024] This step employs the split Bregman iterative algorithm to efficiently solve the aforementioned optimization problem. By introducing auxiliary variables, the complex optimization is decomposed into several easily solvable subproblems, which are iterated alternately until convergence. Specifically, auxiliary variables are introduced... and The original problem is transformed into an equivalent constraint form, and then updated alternately. , and Bregman multipliers The solution is gradually approached from the optimal solution. During the iteration process, The subproblem can be solved using the Fast Fourier Transform. The subproblems are solved using soft thresholding. The iteration count is capped, or the iteration stops when the difference between two consecutive iterations is less than a threshold. The final result is the restored super-resolution radiative distribution map. , Composed of the restored radiance values ​​of each pixel, its spatial resolution is the same as the original image, but it restores the original radiance and edge sharpness lost due to atmospheric attenuation, especially the details of small targets that were originally blurred and erased. The signal is enhanced, and background noise is effectively suppressed.

[0025] S3: Calculate the inversion residuals and generate spatial attention weights Although the restored image obtained in step S2 Theoretically, it approximates the true radiation distribution, but due to noise and model errors, the restoration process is not perfect. To fully utilize the restoration results and guide the neural network to focus on the regions with the most severe information loss, this step calculates the inversion residual between the original degraded image and the restored image. The residual map is defined. as follows: in For the original degraded image in coordinates grayscale value at that location The radiation distribution obtained in step S2 is restored. The point spread function estimated in step S1, Atmospheric transmittance. Residual plot. Each pixel value represents the absolute difference between the original observation and the restored observation at that location. The larger the difference, the more severe the degradation of that area during imaging, and the more original information is lost. Therefore, it should be a key area of ​​focus for subsequent recognition networks.

[0026] To incorporate residual information into deep networks, this invention uses residual maps... After normalization, it is used as spatial attention weight. The specific operation is as follows: First, for Perform maximum-minimum normalization to obtain the normalized residual plot. Its range is Then, through a learnable scaling parameter... Adjusting the attention intensity yields the final spatial attention mask: in Spatial attention mask in coordinates The value at that location, To normalize the residual plot in coordinates The value at that location is a learnable scaling parameter. Initially set to 1.0, and automatically adjusted through backpropagation during network training, its function is to adaptively control the magnitude of attention enhancement in different scenarios, avoiding over-enhancement leading to noise amplification or under-enhancement leading to information loss. In this embodiment... A momentum update method was used, with a learning rate of 0.001, and training was performed synchronously with the backbone network. This attention mask... The feature maps output from each layer of the feature pyramid network are multiplied element-wise to enhance the feature responses in severely degraded areas and suppress less degraded background areas. The introduction of attention masks allows the network to dynamically focus on areas that have suffered significant information loss due to atmospheric attenuation. These areas are often where weak targets are located, thus effectively improving the sensitivity of small target detection.

[0027] S4: Residual Attention-Guided Multiscale Feature Recalibration This step will use the attention mask generated in step S3. This approach is applied to feature pyramid networks to achieve adaptive recalibration of multi-scale features. Feature pyramid networks typically contain multiple feature layers of different resolutions, corresponding to targets of different scales. High-level features (low resolution) are rich in semantic information and suitable for large target detection; low-level features (high resolution) are rich in detail information and suitable for small target detection. The feature pyramid network used in this step is a classic structure, consisting of a bottom-up backbone network (such as ResNet) and top-down lateral connections, generating feature layers at four scales. , , , These correspond to downsampling of the original image by 4, 8, 16, and 32 times, respectively. The backbone network can use lightweight MobileNet or EfficientNet to adapt to the UAV platform. In this embodiment, ResNet-18 is selected as the backbone to achieve a balance between accuracy and speed.

[0028] Before applying attention masks to each layer, it is necessary to... Adjust to the same spatial dimensions as the feature maps of each layer. For low-level features, directly... Downsampling is performed to match the feature map size; for high-level features, then... Upsampling is performed. Specifically, bilinear interpolation is used. Scaling is performed to obtain an attention map with the same spatial size as the feature maps of each layer. ,in This represents the feature layer number. To maintain relative consistency in attention intensity, the interpolation process keeps the total numerical sum constant; that is, average pooling is used for downsampling, and bilinear interpolation is used for upsampling followed by multiplication by a scaling factor. Then, the adjusted attention mask is multiplied element-wise with the feature map: in Indicates the first The feature map output by the layer feature pyramid network in coordinates The feature vector at that location, Indicates adjustment to the number Attention mask of layer size in The value at that location, This indicates the recalibrated feature map in The feature vector at that location. This operation enables feature maps with the same weights to be spatially enhanced in a differentiated manner: in severely degraded regions, the feature response is amplified, which helps the network capture weak target signals; in slightly degraded regions, the feature response remains unchanged or is slightly suppressed, avoiding interference from background noise.

[0029] The attention-recalibrated multi-scale feature maps retain the scale diversity of the original feature pyramid while incorporating degenerate prior information provided by the physical inversion process, thus offering more discriminative feature representations for subsequent object detection. Furthermore, the recalibrated feature maps maintain their original channel count and structure, allowing direct input into the detection head without modifying subsequent network design, demonstrating good compatibility.

[0030] S5: Recalibrate the feature input detection network to complete target recognition. The recalibrated multi-scale feature map output from step S4 is input into the detection head of the target detection network. The bounding box coordinates and size of the target are predicted through the regression branch, and the target category and its confidence level are output through the classification branch, thus completing the infrared target recognition in complex environments.

[0031] Specifically, the detection head consists of a series of convolutional and fully connected layers, predicting multiple candidate boxes for each anchor point position on each feature layer. This embodiment uses a one-stage detection head, where two parallel sub-networks follow each feature layer: a classification sub-network and a regression sub-network. The classification sub-network predicts the probability of each anchor point corresponding to each category, while the regression sub-network predicts the offset of the bounding box relative to the anchor point. The classification sub-network uses four 3×3 convolutional layers followed by one 3×3 convolutional layer to output category scores; the regression sub-network has a similar structure, outputting four coordinate offsets. Batch normalization and ReLU activation functions are used after all convolutional layers.

[0032] For each candidate bounding box, the regression branch outputs its offset relative to the anchor point, and the classification branch outputs the probability of belonging to each category. During the training phase, a multi-task loss function is used to jointly optimize the regression and classification performance; during the inference phase, non-maximum suppression is used to remove redundant detection boxes to obtain the final target recognition result.

[0033] To further improve the performance of small object detection, an online hard example mining strategy can be employed during training, focusing on samples corresponding to regions with high responses in the residual map. The specific implementation of the online hard example mining strategy is as follows: In each training batch, the classification loss and regression loss of all anchor points are first calculated via forward propagation. Let the... The classification loss for each anchor point is The regression loss is Then the total loss at the anchor point is ,in As a balance factor, this embodiment takes =2. Then, all anchor points are adjusted according to the total loss. Sort by value from largest to smallest, and select the top values ​​with the highest loss values. One anchor point is used as a difficult example sample, among which Set to 30% of the total number of anchor points in the current batch. For these difficult samples, calculate their gradients and participate in parameter updates; set the gradients of the remaining simple samples to zero and do not participate in updates.

[0034] To further incorporate physical priors, this strategy introduces residual plots. This serves as a supplementary basis for selecting difficult examples. For each anchor point, based on its location... Query residual plot value at that point .like Greater than the preset threshold (This embodiment takes) If the anchor point is located in a region with severe atmospheric degradation, it is considered to have priority in the difficult case screening. In practice, before loss ranking, the total loss of anchor points located in high residual regions is multiplied by a weighting factor. =1.2, which is the weighted enhancement loss. Then use the weighted enhanced loss. The sorting is performed, but it's important to note that this weighted loss is only used for sorting difficult examples, not for gradient calculation; the original total loss is still used for gradient updates. The purpose of this approach is to make anchor points located in high residual regions more likely to be selected as hard examples, thus giving them more opportunities for optimization during training, while avoiding gradient distortion caused by artificially amplifying the loss value. This weighted ranking method makes anchor points corresponding to regions that are difficult to identify due to atmospheric degradation more likely to be selected as hard examples, thereby guiding the network to focus on regions with severe information loss. During training, the above online hard example mining is performed simultaneously with regular forward and backward propagation, updating the network parameters once per batch.

[0035] Finally, it should be noted that the above embodiments are merely examples for clearly illustrating the present invention and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for infrared unmanned aerial vehicle (UAV) target recognition in complex environments, characterized in that, Includes the following steps: S1: Infrared images are acquired by an infrared thermal imager mounted on a drone, and environmental parameters are obtained simultaneously by airborne sensors; based on the environmental parameters, an atmospheric stratification correction model is used to convert single-point measurements into path integral equivalent values, and the atmospheric point spread function is estimated using natural edge information extracted from the infrared images. S2: Based on the atmospheric point spread function and environmental parameters, establish an infrared radiation transmission degradation model, introduce the sparsity and non-negativity priors of infrared targets, construct an objective function, and obtain the restored radiation distribution map through iterative optimization. S3: Calculate the absolute difference between the original degraded image and the restored observation based on the degradation model, generate an inversion residual map, normalize the inversion residual map and combine it with a learnable scaling parameter to construct a spatial attention mask; S4: Adjust the spatial attention mask to the same spatial size as the feature maps of each layer of the feature pyramid network, and recalibrate the multi-scale feature maps by element-wise multiplication to achieve feature enhancement in areas with severe atmospheric degradation. S5: Input the recalibrated multi-scale feature map into the target detection network, and output the target class probability and bounding box offset through the classification branch and regression branch respectively to complete the infrared UAV target recognition in complex environments.

2. The infrared UAV target recognition method for complex environments as described in claim 1, characterized in that, In step S1, the specific method for estimating the atmospheric point spread function is as follows: detect the edges of ground features with step characteristics in the infrared image, extract the grayscale profile along the edge normal direction, obtain the edge spread function by fitting the error function, and then derive the standard deviation of the atmospheric point spread function; the atmospheric point spread function is approximated by a two-dimensional Gaussian model, and its standard deviation is obtained by taking the average of multiple edge fitting results.

3. The infrared UAV target recognition method for complex environments as described in claim 1, characterized in that, In step S1, the atmospheric stratification correction model is specifically as follows: the radiation transmission path is divided into several layers of equal thickness according to the UAV flight altitude and the average distance of the scene. Based on the measurement value of the airborne sensor, the temperature, humidity and pressure values ​​of each layer are calculated in combination with the standard atmospheric vertical lapse rate. The equivalent path average extinction coefficient is obtained by integrating the extinction coefficient of each layer.

4. The infrared UAV target recognition method for complex environments as described in claim 1, characterized in that, In step S2, the infrared radiation transmission degradation model is that the original degradation image is equal to the convolution of the original radiation distribution and the atmospheric point spread function, multiplied by the atmospheric transmittance, and noise is added; the atmospheric transmittance is calculated based on environmental parameters using the Beer-Lambert law.

5. The infrared UAV target recognition method for complex environments as described in claim 1, characterized in that, In step S2, the objective function is: in This represents the restored radiation distribution to be solved. This represents the value of the independent variable that minimizes the objective function. For the original degraded image, The original radiation distribution, Here is the atmospheric point spread function. Atmospheric transmittance, Represents the image gradient. and The regularization parameter is used; the objective function is solved using the split Bregman iterative algorithm to obtain the restored radiation distribution map. .

6. The infrared UAV target recognition method for complex environments as described in claim 1, characterized in that, In step S3, the inversion residual map Defined as: in For the original degraded image in coordinates The grayscale value at that location; Will Normalization yields the normalized residual plot. Then through learnable scaling parameters Generate spatial attention mask: in Spatial attention mask in coordinates The value at that location, To normalize the residual plot in coordinates The value at that location.

7. The infrared UAV target recognition method for complex environments as described in claim 1, characterized in that, In step S4, the feature pyramid network contains multiple feature layers with different resolutions. The spatial attention mask is adjusted to the same size as the feature map of each layer by interpolation. The recalibrated feature map is the element-wise product of the original feature map and the attention weights at the corresponding positions.

8. The infrared UAV target recognition method for complex environments as described in claim 1, characterized in that, In step S5, the target detection network uses a one-stage detection head, followed by parallel classification and regression sub-networks after each feature layer to predict the target category probability and the offset of the bounding box relative to the anchor point, respectively; during the inference stage, redundant detection boxes are removed by non-maximum suppression to obtain the final recognition result.

Images