Low-altitude small target recognition method and system based on full stokes polarization imaging

Abstract

The application discloses a low-altitude small target recognition method and system based on full Stokes polarization imaging, and the method comprises the following steps: collecting a plurality of full Stokes low-altitude polarization data, and analyzing and obtaining target polarization parameters; constructing a full polarization dataset based on the target polarization parameters and a CPFNet polarization fusion network; constructing an improved YOLOv8-FPZ network, and training the improved YOLOv8-FPZ network by using the full polarization dataset, wherein the improved YOLOv8-FPZ network is obtained by replacing the C2F module in the neck network with a DBB convolution module, replacing the Detect detection head with an FPZHead detection head, and optimizing the loss function to be Focaler-GioU on the basis of the existing YOLOv8 network; and performing low-altitude small target recognition and outputting on the full Stokes polarization data to be detected, so that the problems of low small target detection precision, weak environmental interference suppression capability and insufficient real-time performance in a low-altitude complex scene are solved.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of optical imaging and computer vision, and in particular to a method and system for identifying small targets at low altitudes based on full Stokes polarization imaging. Background Technology

[0002] With the rapid development of UAV technology, its applications in military reconnaissance, environmental monitoring, and smart security are becoming increasingly widespread. However, target detection technology in complex low-altitude scenarios still faces severe challenges: targets are typically small in size, have low signal-to-noise ratios, and move rapidly, and are easily affected by background interference (such as cloud reflection, ground clutter, and dynamic lighting), leading to a significant decrease in the detection accuracy of traditional visible light imaging methods under low visibility conditions. Existing technologies mainly rely on light intensity information for target identification, but intensity imaging is easily affected by ambient scattered and reflected light, making it difficult to effectively distinguish subtle differences between targets and complex backgrounds. Especially under adverse conditions such as fog, haze, and backlighting, the false detection rate and false negative rate increase sharply.

[0003] To overcome the aforementioned bottlenecks, polarization imaging technology has gradually become a research hotspot. This technology, by analyzing the polarization characteristics of light, can reflect the physical properties of a target's surface, such as material and geometry, thereby enhancing the contrast between the target and the background. However, traditional polarization imaging methods are mostly limited to extracting linear polarization information, neglecting the circular polarization component, which has important characterizing capabilities for target contours and microstructures. This incomplete information acquisition makes it difficult for existing technologies to fully explore the deep correlation between polarization features and target physical properties, making them susceptible to noise interference in complex scenes and lacking robustness in detection. For example, existing solutions use multi-angle linear polarization light fusion or encoded image input, which can partially improve imaging quality in low-light environments, but cannot solve core problems such as blurred edges of small targets and sparse features.

[0004] Currently, Stokes linear polarization technology is relatively mature and widely used. Chinese patent application CN202311239303.4 proposes an underwater target detection and identification method based on projected polarization distance characteristics. This method uses a linear polarization detection device to capture the polarization information of underwater targets, processes it to extract linear polarization images containing the target's shape and texture, and then stitches these polarization feature images together before inputting them into the YOLO target recognition network for underwater target detection and identification training and inference, outputting the target's location and classification information. However, linear polarization images only capture the linear polarization component, ignoring the circular polarization component. This makes it unable to detect the circularly polarized light characteristics present in materials. Furthermore, in scattering media (such as fog and atmospheric turbulence), light may depolarize and produce circularly polarized components. Linear polarization images cannot effectively analyze such scenarios, while a full Stokes system can quantify circular polarization characteristics through S3 parameters, improving robustness under environmental interference. Therefore, due to the lack of information dimensions, linear polarization images may have analytical blind spots in scientific research and industrial applications, while full Stokes polarization images provide more comprehensive detection capabilities through complete polarization state description.

[0005] In recent years, all-Stokes polarization imaging technology has been considered a significant technological breakthrough. By simultaneously acquiring linear and circular polarization information, it constructs a complete four-dimensional polarization feature space, providing a more comprehensive data foundation for analyzing target characteristics in complex scenes. Research shows that the circular polarization component has a unique response to target edges and curvature changes, significantly enhancing the contour features of small targets. Meanwhile, the introduction of deep learning technology has opened up new avenues for the efficient utilization of polarization information: through end-to-end network models, the potential correlation between target characteristics and noise can be learned directly from raw polarization data, replacing traditional physical model-based solution methods, thereby achieving joint optimization of noise suppression, feature enhancement, and target detection.

[0006] At the algorithmic level, existing target detection technologies exhibit a polarized contradiction: while two-stage algorithms based on candidate boxes (such as Faster R-CNN) offer high accuracy, their computational complexity is high, making it difficult to meet the real-time requirements of UAV platforms; while single-stage algorithms (such as the YOLO series) offer significant speed advantages, their feature extraction capabilities for small low-altitude targets are limited, and they are easily affected by background interference. Particularly in the field of polarization imaging, existing research largely focuses on the separate optimization of traditional optical systems and detection algorithms, lacking a collaborative design that considers the characteristics of full Stokes data and the architecture of deep learning models. This results in multi-dimensional polarization information not being fully translated into improved detection performance.

[0007] Low-altitude small target identification is a core technology of intelligent sensing systems, and its performance directly determines the mission effectiveness of drones in scenarios such as border patrol, counter-drone operations, and disaster relief. However, the existing technology system has the following key shortcomings:

[0008] (1) Traditional linear polarization imaging has insufficient information dimensions and cannot effectively characterize the polarization characteristics of small targets;

[0009] (2) The fusion mechanism of full Stokes data and deep learning models is not yet perfect, resulting in low feature utilization.

[0010] (3) The detection algorithm lacks a balance between speed and accuracy, making it difficult to adapt to dynamic low-altitude scenarios. Therefore, it is urgent to construct an efficient target detection method for full Stokes polarization imaging, and to break through the bottlenecks in accuracy and real-time performance of small target detection in complex low-altitude environments through the collaborative innovation of deep polarization information mining and lightweight network design. Summary of the Invention

[0011] Technical Objective: To address the problems of missing polarization information, poor adaptability to low-altitude scenes, and insufficient accuracy in small target detection in existing polarization imaging technologies, this invention discloses a low-altitude small target recognition method and system based on full Stokes polarization imaging, which solves the problems of low accuracy in small target detection, weak environmental interference suppression, and insufficient real-time performance in complex low-altitude scenes.

[0012] Technical solution: To achieve the above technical objectives, the present invention adopts the following technical solution.

[0013] A method for identifying small low-altitude targets based on full Stokes polarization imaging, the method comprising:

[0014] Collect several full Stokes low-altitude polarization data and analyze them to obtain the target polarization parameters;

[0015] A full Stokes polarization dataset is constructed based on the target polarization parameters combined with the CPFNet polarization fusion network;

[0016] An improved YOLOv8-FPZ network was constructed and trained using the full Stokes polarization dataset to obtain the trained improved YOLOv8-FPZ network. The improved YOLOv8-FPZ network is based on the existing YOLOv8 network, replacing the C2F module in the neck network with a DBB convolutional module, replacing the Detect head with an FPZHead head, and optimizing the loss function to Focaler-GIoU.

[0017] The full Stokes polarization data to be detected is input into the improved YOLOv8-FPZ network after training to perform low-altitude small target recognition and output, and output the small target detection results, including the small target category, location and confidence level.

[0018] A low-altitude small target recognition system based on full Stokes polarization imaging is used to implement the low-altitude small target recognition method based on full Stokes polarization imaging described above. The system includes a four-channel polarization imaging unit, a target polarization parameter acquisition unit, a dataset generation unit, an embedded processing unit, and a result output unit connected in sequence.

[0019] The four-channel polarization imaging unit is used to acquire full Stokes low-altitude polarization data;

[0020] The target polarization parameter acquisition unit is used to receive all-Stokes low-altitude polarization data and parse it to obtain the target polarization parameters;

[0021] The dataset generation unit is used to generate a large number of polarization images based on the CPFNet polarization fusion network to create a full Stokes polarization dataset; the CPFNet polarization fusion network is used to fuse circular polarization information.

[0022] Embedded processing unit for deploying the trained, improved YOLOv8-FPZ network;

[0023] The result output unit is used to input the full Stokes polarization data to be detected into the trained and improved YOLOv8-FPZ network to perform low-altitude small target recognition and output, and output the small target detection results, including the small target category, location and confidence level.

[0024] Beneficial effects:

[0025] (1) Full polarization feature enhancement: By fusing linear polarization and circular polarization components through the CPFNet polarization fusion network, the contrast between the target and the background of the polarization image under complex low-altitude background is significantly improved, and the edge positioning accuracy is improved.

[0026] (2) Network performance optimization: The FPZHead design reduces the number of model parameters and improves inference speed, enabling real-time detection on low-computing-power platforms (such as UAV onboard equipment);

[0027] (3) Improved environmental robustness: Feature extraction based on polarization images effectively suppresses atmospheric scattering and specular reflection interference, and the detection accuracy under haze and backlight conditions is significantly improved compared with the traditional RGB method;

[0028] (4) Breakthrough in low-altitude small target detection: Simulation results show that the recognition accuracy of low-altitude small targets in complex environments is improved to 74.6% mAP@0.5 / %, and mAP@0.5:0.95 / % is improved to 30.0%. Experimental results prove that the overall performance of the YOLOv8-FPZ network is good.

[0029] (5) Application areas: This invention can be widely applied to scenarios such as UAV autonomous inspection, border security monitoring, and aviation obstacle warning, providing a high-precision and high-real-time solution for low-altitude intelligent sensing systems. Attached Figure Description

[0030] Figure 1 This is a flowchart of a method according to an embodiment of the present invention;

[0031] Figure 2 Flowchart for full Stokes polarization image acquisition and polarization parameter analysis;

[0032] Figure 3 The CPFNet network structure diagram is shown in the embodiment.

[0033] Figure 4 This is a multi-algorithm fusion diagram for an embodiment;

[0034] Figure 5 This is a DBB structure design diagram for an embodiment;

[0035] Figure 6 This is a schematic diagram of the RepVGG convolutional architecture as an example.

[0036] Figure 7 This is a structural diagram of the YOLOv8-FPZ example.

[0037] Figure 8 This is a diagram showing the cable connections for a full Stokes camera as an example.

[0038] Figure 9 This is an example of the detection results for an embodiment. Detailed Implementation

[0039] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0040] Example

[0041] As attached Figure 1 As shown in this embodiment, a low-altitude small target identification method based on full Stokes polarization imaging includes the following steps:

[0042] S1: Collect several full Stokes low-altitude polarization data and analyze them to obtain the target polarization parameters;

[0043] like Figure 2As shown, in this embodiment, a four-channel polarization imaging unit acquires the full Stokes polarization image of the target, and the target polarization parameters are obtained by analyzing the original polarization map in four directions. The full Stokes image includes full Stokes polarization images of aircraft and pedestrians acquired under low-altitude shooting conditions, different weather environments, and different object surface properties. The target polarization parameters are a multi-dimensional polarization image set containing Stokes parameters S0, S1, S2, and S3, where S0 is the total light intensity, S1 and S2 are the intensity difference signals of different linear polarization directions, and S3 is the intensity difference between right-handed and left-handed circularly polarized light.

[0044] S2: Construct a full Stokes polarization dataset based on target polarization parameters combined with the CPFNet polarization fusion network;

[0045] The CPFNet polarization fusion network is used to enhance the features of the target polarization parameters; as shown in the attached image. Figure 3 As shown, the CPFNet polarization fusion network includes an FCN layer, convolutional layers, three ResBlock layers, three multi-scale feature extraction layers, four upsampling modules, and the last two polarization parameter resolution layers. This enables the computation and generation of a large number of fused polarization images and the creation of the full Stokes polarization dataset. The FCN layer employs a fully convolutional structure, and each multi-scale feature extraction layer includes a dual-directional CGabor convolutional kernel. Its calculation formula is as follows:

[0046]

[0047] Among them, W θ (x,y) is the output of the dual-directional CGabor convolution kernel, where x,y are spatial coordinates, θ is the direction angle, σ and ψ are parameters, and parameter ψ is set as the initial phase; x′ and y′ are the new coordinates after rotating θ around the origin; λ is the frequency parameter, and φ(x′,y,θ) is the rotation phase factor;

[0048] As can be seen from this calculation formula, the dual-directional CGabor convolution kernel includes two branches, the first branch... It is a Gaussian function used to introduce spatial smoothness to the convolution kernel. The parameter σ controls the standard deviation of the Gaussian function, which in turn controls the width of the convolution kernel. The smaller σ is, the more concentrated the response of the convolution kernel; the larger σ is, the more broad the response of the convolution kernel.

[0049] The second branch is the cosine function, a sine wave function that controls both the frequency and phase of the convolution kernel, directly determining its core performance. The frequency parameter λ controls the periodic changes of the kernel, while the parameter ψ is set as the initial phase. This crucial component allows the convolution kernel to keenly capture periodic features in the image. By adjusting the value of λ, we can precisely control the kernel's ability to perceive image details. Different λ values ​​control the periodicity of the kernel, and the value of λ can be flexibly adjusted to focus on texture features at different scales. The rotation phase factor φ(x′,y,θ) is the most distinctive part of this convolution kernel, controlling its rotation characteristics. This rotation phase factor is a unique innovation of this convolution kernel. It is controlled through a specific mathematical formula, allowing the kernel's rotation angle to be adjusted according to actual needs during the convolution operation, thereby enhancing the kernel's ability to extract complex image features. The formula for the rotation phase factor is as follows:

[0050]

[0051] This term simulates the rotation of circularly polarized light. In circularly polarized light, the rotation of the electric field vector is position-dependent, so this term reflects the rotational property of circularly polarized light and can describe its propagation. The direction of rotation of the polarized light changes as x′ and y′ change, and the introduction of this term helps the convolution kernel capture this rotational characteristic.

[0052] This part is closely related to the angle θ, controlling the directionality of the convolution kernel. This invention uses θ to rotate the convolution kernel, allowing it to respond to information from different directions in the image. Through this, the invention embeds directional features from the image into the design of the convolution kernel, enabling it to adapt to different orientations within the image. The positive and negative signs are used to distinguish between right-handed and left-handed circularly polarized light; this ensures the convolution kernel can adapt to different types of circularly polarized light.

[0053] In this embodiment, the structural connection relationships of the CPFNet polarization fusion network include:

[0054] The first layer is the FCN layer. The input of the FCN layer is the target polarization parameter. The FCN layer is a fully convolutional structure. In this embodiment, the FCN layer is a network structure designed specifically for pixel-level prediction tasks (such as semantic segmentation, object detection, etc.). It is designed as a fully convolutional structure to replace the fully connected layer of the traditional CNN, supporting pixel-level prediction of arbitrary input size.

[0055] The second layer is a convolutional layer, whose input is the output of the first FCN layer. This convolutional layer is a 1×1 layer; the output scale of the convolutional layer is... Where H is the height, W is the width, and 16 represents the number of channels;

[0056] The third layer connects the first ResBlock layer and the first multi-scale feature extraction layer; the input to the first ResBlock layer is the output of the convolutional layer, and the output scale of the first multi-scale feature extraction layer is...

[0057] The fourth layer is a connection between the second ResBlock layer and the second multi-scale feature extraction layer; the input of the second ResBlock layer is the output of the first multi-scale feature extraction layer, and the output scale of the second multi-scale feature extraction layer is...

[0058] The fifth layer is a connection between the third ResBlock layer and the third multi-scale feature extraction layer; the input of the third ResBlock layer is the output of the second multi-scale feature extraction layer, and the output scale of the third multi-scale feature extraction layer is...

[0059] Layers 6 through 9 are four upsampling layers connected in sequence. The input of the first upsampling layer is the output of the third multi-scale feature extraction layer. The input of the second upsampling layer is the output of the first upsampling layer and the output of the second multi-scale feature extraction layer. The input of the third upsampling layer is the output of the second upsampling layer and the output of the first multi-scale feature extraction layer. The input of the fourth upsampling layer is the output of the third upsampling layer and the output of the convolutional layer.

[0060] The last two layers, the tenth and eleventh layers, are polarization parameter resolution layers, used to calculate and generate a large number of fused polarization images and create the full Stokes polarization dataset. The formula for calculating the degree of polarization (DOP) is as follows:

[0061]

[0062] Among them, the three ResBlock layers are used to solve the gradient vanishing and gradient exploding problems that occur during the training of deep neural networks, making it easier to train deeper networks; the three multi-scale feature extraction layers are used to effectively reduce noise interference by filtering and compressing redundant information in the input image, thereby improving the accuracy and reliability of image classification, object detection and object recognition; the four upsampling layers are used to restore the size of the image and gradually enhance the expressive power of features.

[0063] The three-layer multi-scale feature extraction layer jointly extracts local and global features. Each layer contains a dual-directional CGabor convolution kernel designed in this invention, used to extract orientation-sensitive features and circular polarization rotation characteristics in the image. Each multi-scale feature extraction layer includes a max pooling layer, a first CGabor convolution kernel, a batch normalization layer, a second CGabor convolution kernel, a stacking layer, a connecting layer, and a ReLU activation function layer connected in sequence. The input of the stacking layer is the input of the multi-scale feature extraction layer and the output of the second CGabor convolution kernel.

[0064] By integrating circularly polarized light channels through the CPFNet network, complete light field polarization information is captured. This includes the dual-directional CGabor convolution kernel designed in this invention, enhancing image contrast and detail representation capabilities. A large number of polarized images are fused, and a full Stokes polarization dataset is created. The target polarization parameters obtained in S1 are input into the CPFNet polarization image fusion network proposed in this invention, as follows: Figure 3 As shown, the final fused full Stokes polarization image is compared with the fusion results of GIDC, VGenNet, and ILNet methods as follows: Figure 4 As shown, the polarization images fused by the CPFNet polarization image fusion network clearly show the boundaries and wing shapes of the aircraft, highlighting the realism of the target. Based on this, a full Stokes polarization small target dataset was created.

[0065] S3: Construct an improved YOLOv8-FPZ network and train it using the full Stokes polarization dataset to obtain the trained improved YOLOv8-FPZ network; see attached. Figure 7 As shown, the improved YOLOv8-FPZ network is based on the existing YOLOv8 network, replacing the C2F module in the neck network with a DBB convolutional module, replacing the Detect head with an FPZHead head, and optimizing the loss function to Focaler-GIoU; the implementation steps of the improved YOLOv8-FPZ network include:

[0066] S31. Replace the C2F modules in the neck network with DBB convolutional modules, including: replacing the C2F modules in layers 15, 18, and 21 of the neck network with DBB convolutional modules; the structure of the DBB convolutional module is as follows:

[0067] The first branch consists of connected 1×1 convolutional layers and batch normalization layers;

[0068] The second branch includes connected 1×1 convolutional layers, batch normalization layers, K×K convolutional layers (K being the size of different convolutions), and batch normalization layers; where K is selected according to the actual situation;

[0069] The third branch includes connected 1×1 convolutional layers, batch normalization layers, average pooling layers, and batch normalization layers;

[0070] The fourth branch includes connected K×K convolutional layers and batch normalization layers;

[0071] An overlay layer is used to receive the output from four branches;

[0072] Nonlinear layers are used to perform nonlinear processing on the output of the stacked layers before outputting the final output.

[0073] As attached Figure 5 As shown, traditional convolutional layers struggle to fully capture these complex multi-scale and multi-directional features, while the core idea of ​​DBB is to extract features through four parallel branches and achieve diverse feature representations during the training phase.

[0074] Assume the input is a tensor The output is a tensor Among them, C in C out These represent the number of input and output channels, respectively; the output calculation formula for the DBB convolution module is:

[0075]

[0076] Where N is the total number of branches in the DBB, and its value is 4; W i Let be the convolution kernel of the i-th branch, with shape as follows Where k i b is the size of the convolution kernel in the i-th branch; i Let be the bias term of the convolution kernel in the i-th branch, with shape . The convolution kernel size is 1×1, denoted as W. 1×1 However, during fusion, it will be filled with an equivalent 3×3; the bias is b. 1×1 The convolution output is:

[0077] Y 1×1 =W 1×1 *X+b 1×1

[0078] S32. Replace the Detect head with the FPZHead head; the FPZHead head includes a CBS module, a DFL-FPZ module, a RepVGG convolution module, and a loss module connected in sequence. The DFL-FPZ module is used to model polarization characteristics through a multi-scale architecture of four-channel polarization distribution mapping and weight sharing, and integrates the RepVGG convolution module to improve inference efficiency through structural reparameterization. The DFL-FPZ module is designed by combining a distributed feature learning module with a full Stokes polarization parameter depth and introducing an improved distributed regression mechanism to achieve fine-grained modeling and efficient learning of light wave polarization state characteristics.

[0079] The DFL-FPZ module comprises a four-channel polarization distribution mapping module, a multi-channel polarization feature processing module, and a polarization regularization module connected in sequence. The multi-channel polarization feature processing module uses parallel 1×1, 3×3, and 5×5 convolutions, which are then stacked and output through overlay layers. By introducing a four-channel polarization distribution mapping, a multi-scale feature processing architecture with shared weights, and polarization-aware regularization loss, the DFL-FPZ module significantly improves the algorithm's performance in optical feature extraction and complex target detection. For the input Stokes parameters, i.e., polarization parameters S0, S1, S2, and S3, the polarization parameters of each channel are first represented in a distributed manner:

[0080]

[0081] in, Represents the polarization parameter S i The probability distribution of A i and d i The weights and biases are learnable, softmax is a function of the probability distribution, and x is the input feature vector. The output of each channel is a probability distribution used to describe the spatial variation of polarization characteristics. The final polarization regression operation yields:

[0082]

[0083] in, For the regression results, N is the number of segments in the discretized distribution. Because the polarization parameter has a wide range of values, with a maximum of 16, more segments are needed to cover the variations. n ∈{1,2,...,N} represents the discretization interval. The above formula realizes the transformation of polarization characteristics from distributed to continuous numerical values, and can capture complex changes in polarization information. This invention designs a multi-channel polarization feature processing module for the multidimensional characteristics of full Stokes polarization images. It includes a 1×1 convolution to enhance the feature interaction capability of the channel dimension; a 3×3 convolution to display local feature extraction in a small receptive field; and a 5×5 convolution module to demonstrate global context capture in a large receptive field. Its core formula is as follows:

[0084]

[0085] Where F(S) is the output of the multi-channel polarization feature processing module, and W a,j and b a,j Here, represents the convolution kernel and bias, respectively. The subscript 'a' indicates the parameter group identifier, and the subscript 'j' indicates the input channel index. This enables multi-channel feature fusion. By sharing weights, this module can simultaneously extract the correlation and independence between polarization parameters. To further enhance the expressive power of polarization characteristics, a polarization characteristic-aware regularization loss is introduced into the polarization regularization module, as shown in the following formula:

[0086]

[0087] Among them, L pol This is the output of the polarization regularization module. Polarization parameter S i The gradient of the polarization image measures the rate of change of the polarization image in space. This is a balancing coefficient used to constrain the spatial smoothness of the polarization parameter. Relying on this regularization term, the network gains the ability to capture more spatial information, thereby improving its adaptability to complex optical scenes and enabling it to better handle various complex optical conditions.

[0088] For fully Stokes polarization images, this enhanced expressive power can significantly improve the accuracy of target detection. In this invention, FPZHead also features a RepVGG convolution module to enhance the feature extraction capability for small, sheet-polarized targets in complex environments; such as... Figure 6 As shown, the RepVGG convolutional module consists of a stack of 3×3 convolutions and ReLU, which enhances the diversity of feature extraction during the training phase through a multi-branch structure. However, during the inference phase, it achieves efficient inference through structural reparameterization techniques, while maintaining the high representational power of the training phase.

[0089] Define the input features of FPZHead as follows The output consists of multiple prediction branches from the detection task, including: classification branch expressions and regression branch expressions. Where f cls It is a classification feature extraction function, f reg These are regression feature extraction functions, implemented by a multi-branch network based on the RepVGG core convolutional module, with the following output:

[0090] P cls =f cls (F)

[0091] P reg =f reg(F)

[0092] Among them, P cls For the output of the classification branch, P reg This is the output of the regression branch;

[0093] S33. Optimize the loss function to Focaler-GIoU, and improve detection robustness by dynamically adjusting sample weights.

[0094] This loss function improves the detector's performance across different detection tasks by focusing on different regression samples. Finally, comparative experiments were conducted using existing advanced detectors and regression methods for different detection tasks, and the detection performance was further improved using the method proposed in this invention. The current improved bounding box regression formula is as follows:

[0095]

[0096] Where [d,u]∈[0,1], IoU is the original IoU value, IoU focaler To obtain the new IoU value after reconstruction, the IoU value is dynamically adjusted by comparing it with the original IoU value. focaler Different regression samples were obtained. The IoU of the regression samples was then calculated. focaler Applying this to the existing IoU-based bounding box regression loss function, the resulting regression loss function is:

[0097] L Focaler-GIoU =L GIoU +IoU-IoU focaler

[0098] Among them, L Focaler-GIoU L is the output of the Focaler-GIoU loss function. GIoU For generalized intersection and comparison of losses;

[0099] The improved YOLOv8-FPZ network training and deployment process in this invention includes:

[0100] The YOLOv8-FPZ network, improved through end-to-end training on the full Stokes polarization dataset, employs a three-stage progressive transfer learning strategy. The initial stage replaces the training with DBB convolutions, followed by embedding an FPZHead detection head and jointly optimizing detection and polarization regression tasks. Finally, the loss function is optimized, and a balancing mechanism automatically adjusts the weights of each task to ensure model stability in multi-target scenarios. The model is deployed on low-computing platforms, automatically reducing GPU frequency during periods without target detection, and maintaining overall system stability at low power consumption, meeting the long-term operational requirements of real-time polarization target detection in industrial scenarios.

[0101] This invention innovatively combines full polarization with a deep learning architecture to solve problems such as low resolution of weakly polarized signals, blurred target features in low-altitude scenes, and complex interference. Experiments show that this method has higher detection accuracy than traditional single-linear polarization imaging techniques and can efficiently serve aircraft target recognition and environmental perception systems, demonstrating high precision and real-time performance. The method described in this invention significantly improves detection accuracy under fog and backlight conditions compared to traditional RGB methods, and noticeably enhances target-background contrast. It can provide a real-time detection method for small targets in low-altitude vision systems such as UAV inspections and intelligent security systems.

[0102] S4: Input the full Stokes polarization data to be detected into the improved YOLOv8-FPZ network after training to perform low-altitude small target recognition and output, and output the small target detection results, including the small target category, location and confidence level.

[0103] This embodiment performed simulation verification, selecting over 3000 full Stokes polarization images for annotation. The data was allocated proportionally, resulting in a training set of 3021 images, a validation set of 300 images, and a test set of 225 images. The experiment used a multi-aperture full Stokes polarization imaging device, connected as follows: Figure 8 As shown, the entire Stokes polarization dataset is input into the improved YOLOv8-FPZ network of this invention, and the recognition performance and accuracy of polarization images are evaluated using existing YOLOv8 standard evaluation metrics. The final simulation results are as follows:

[0104] The standard YOLOv8 evaluation metric is used, with mean precision (mAP) as the representative, which is the average of all APs. AP is a core evaluation metric in object detection tasks, used to measure the accuracy of the model in object detection at different confidence thresholds. AP is calculated based on the area under the Precision-Recall curve and is closely related to the values ​​of precision and recall.

[0105] Precision indicates the percentage of true positive samples predicted by the model, where TP is the number of correctly predicted positive samples and FP is the number of incorrectly predicted positive samples. Recall indicates the proportion of true positive samples detected by the model, where FN is the number of incorrectly predicted negative samples.

[0106] Calculation of average precision:

[0107] (1) Calculate the Precision and Recall of the model under different confidence thresholds.

[0108] (2) Construct a PR curve based on all points.

[0109] (3) Perform numerical integration on the PR curve and calculate the area under the curve. For multi-class detection tasks, AP will calculate the area under the curve for each class separately and then take the average value, where N is the number of classes. i Let AP be the AP for the i-th category. The formula above is as follows:

[0110]

[0111] In practice, AP is typically evaluated based on IoU thresholds, with mAP@0.5 being the default. mAP@0.5 represents the average accuracy when the IoU threshold is fixed at 0.5; when IoU ≥ 0.5, the predicted bounding box is considered a correct positive sample. Compared to mAP@0.5, mAP@0.5:0.95 is more stringent because it covers multiple IoU thresholds, comprehensively reflecting the model's performance.

[0112] The final all-Stokes polarization target recognition scheme and results are shown in Tables 1 and 2:

[0113] Table 1 Ablation experimental protocol for the improved YOLOv8-FPZ model

[0114]

[0115] Table 2 Results of ablation experiments using the improved YOLOv8-FPZ model

[0116]

[0117]

[0118] As shown in Tables 1 and 2, compared with the original model, the addition of the DBB convolutional module improved the accuracy of aircraft and humans in mAP@0.5 by 0.3% and 2.2%, respectively; the overall mAP@0.5 improved by 1.2%, and the overall mAP@0.5:0.95 improved by 1.5%. After replacing it with the FPZHead detection head, the accuracy of aircraft in mAP@0.5 improved by 5.7%, the overall mAP@0.5 improved by 2.7%, and the overall mAP@0.5:0.95 improved by 0.7%. Finally, after optimizing the loss function, the accuracy of aircraft and humans in mAP@0.5 improved by 1.1% and 1.2%, respectively. The detection results are as follows: Figure 9 As shown.

[0119] These findings demonstrate that this invention addresses the technical challenges of multi-target detection in complex low-altitude environments by constructing a small target image dataset that integrates full Stokes polarization information, overcoming the perception bottleneck of traditional RGB vision in strong scattering and high-dynamic scenes. A YOLOv8-FPZ deep learning framework is proposed, which effectively suppresses environmental interference such as atmospheric turbulence and specular reflection through full polarization feature analysis. Its core mechanism lies in the fact that the polarization difference characteristics of the full Stokes vector can decouple the target's intrinsic radiation from background noise in the frequency domain, thereby improving the signal-to-noise ratio. Ultimately, the overall detection accuracy (mAP@0.5) is improved by 5.1%, and the detection accuracy of Person and Airplane targets is improved by 3.2% and 7.1% respectively compared to the baseline model. This invention establishes a paradigm innovation for full Stokes polarization enhanced perception in low-altitude vision systems, providing a theoretically complete and engineering-feasible solution for applications such as UAV inspection and intelligent security.

[0120] This embodiment also discloses a low-altitude small target recognition system based on full Stokes polarization imaging, including a four-channel polarization imaging unit, a target polarization parameter acquisition unit, a dataset generation unit, an embedded processing unit, and a result output unit connected in sequence.

[0121] The four-channel polarization imaging unit includes four polarization imaging acquisition paths, each of which includes a connected polarization analysis device and a detector; it synchronously acquires full Stokes polarization information through four sets of linear polarizers with polarization directions of 0°, 45°, and 90° and circular polarization phase delay devices; the four-channel polarization imaging unit is used to acquire full Stokes low-altitude polarization data.

[0122] The target polarization parameter acquisition unit is used to receive all-Stokes low-altitude polarization data and parse it to obtain the target polarization parameters;

[0123] The dataset generation unit is used to generate a large number of polarization images based on the CPFNet polarization fusion network to create a full Stokes polarization dataset; the CPFNet polarization fusion network is used to fuse circular polarization information.

[0124] Embedded processing unit for deploying the trained, improved YOLOv8-FPZ network;

[0125] The result output unit is used to input the full Stokes polarization data to be detected into the trained and improved YOLOv8-FPZ network to perform low-altitude small target recognition and output, and output the small target detection results, including the small target category, location and confidence level.

[0126] The terms "first" and "second" mentioned in the embodiments of this application are merely used as name identifiers and do not represent first or second in terms of order or importance.

[0127] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that all or part of the steps in the methods of the above embodiments can be implemented by means of software plus a general-purpose hardware platform. Based on this understanding, the technical solution of this application can be embodied in the form of a software product. This computer software product can be stored in a storage medium. The memory can be various types of memory, such as random access memory, read-only memory, flash memory, etc., such as read-only memory (ROM) / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which can be a personal computer, server, or network communication device such as a router) to execute the methods described in various embodiments or some parts of the embodiments of this application.

[0128] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for identifying small low-altitude targets based on full Stokes polarization imaging, characterized in that the method... include: Collect several full Stokes low-altitude polarization data and analyze them to obtain the target polarization parameters; A full Stokes polarization dataset is constructed based on the target polarization parameters combined with the CPFNet polarization fusion network; An improved YOLOv8-FPZ network was constructed and trained using the full Stokes polarization dataset to obtain the trained improved YOLOv8-FPZ network. The improved YOLOv8-FPZ network is based on the existing YOLOv8 network, replacing the C2F module in the neck network with a DBB convolutional module, replacing the Detect head with an FPZHead head, and optimizing the loss function to Focaler-GioU. The full Stokes polarization data to be detected is input into the trained and improved YOLOv8-FPZ network for low-altitude small target recognition and output. The output of small target detection results includes the target category, location, and confidence score. The CPFNet polarization fusion network includes an FCN layer, convolutional layers, three ResBlock layers, three multi-scale feature extraction layers, four upsampling modules, and the last two polarization parameter resolution layers. This enables the calculation and generation of a large number of fused polarization images and the creation of a full Stokes polarization dataset. The FCN layer adopts a fully convolutional structure, and the input of the FCN layer is the target polarization parameters. Each multi-scale feature extraction layer includes a dual-directional CGabor convolution kernel. The calculation formula for the dual-directional CGabor convolution kernel is as follows: ， in, The output of the dual-directional CGabor convolution kernel, where x and y are spatial coordinates. For direction and angle, σ, ψ For parameters, parameters ψ Set as the initial phase; and Rotation around the origin The new coordinates after; λ For frequency parameters, For rotational phase factor; the dual-directional CGabor convolution kernel consists of two branches, the first branch of which is... The first branch is a Gaussian function used to introduce spatial smoothness to the convolution kernel; the second branch is a cosine function used to control the frequency and phase of the convolution kernel, and also includes a rotation phase factor. ; The formula for calculating the rotation phase factor is as follows: ， in, The item simulates the rotation of circularly polarized light. The directionality of the convolution kernel is controlled, and the positive and negative signs are used to distinguish between right-handed and left-handed circularly polarized light.

2. The low-altitude small target identification method based on full Stokes polarization imaging according to claim 1, characterized in that: The output calculation formula of the DBB convolution module is: ， Where X is the input tensor of the DBB convolution module; N is the total number of branches of DBB, with a value of 4; The convolution kernel for the i-th branch has the following shape: ,in Let be the size of the convolution kernel in the i-th branch. These represent the number of input and output channels, respectively. The bias term of the convolution kernel in the i-th branch has the following shape: The convolution kernel size is 1×1, represented as However, during fusion, it will be filled with an equivalent 3×3; the bias is... The convolution output is: .

3. The low-altitude small target identification method based on full Stokes polarization imaging according to claim 1, characterized in that: The FPZHead detection head includes a DFL-FPZ module; the DFL-FPZ module comprises a four-channel polarization distribution mapping analysis, a multi-channel polarization feature processing, and a polarization characteristic sensing regularization, connected in sequence. The calculation process of the DFL-FPZ module includes: For the input Stokes parameters, which are polarization parameters First, the polarization parameters of each channel are represented in a distributed manner: ， in, Representing polarization parameters The probability distribution, and For learnable weights and biases, Let x be a function of the probability distribution, and let x be the input feature vector; the final polarization regression operation yields: ， in, For the regression results, N is the number of segments in the discretized distribution. The discretized interval; the calculation formula of the multi-channel polarization feature processing module includes: ， in, The output of multi-channel polarization feature processing, and These are the convolution kernel and bias, respectively, and their subscripts are... a This represents the parameter group identifier, with the subscript j indicating the input channel index. The polarization regularization module formula is as follows: ， in, This is the output of polarization regularization. Polarization parameter gradient, This is the balance coefficient.

4. The low-altitude small target identification method based on full Stokes polarization imaging according to claim 1, characterized in that: The FPZHead detection head also features a RepVGG convolution module. The input feature of the FPZHead detection head is F, and the output consists of multiple prediction branches for the detection task, including classification branch expressions and regression branch expressions, specifically: ， in, For the output of the classification branch, This is the output of the regression branch.

5. The method for identifying low-altitude small targets based on full Stokes polarization imaging according to claim 1, characterized in that: The regression loss function is: ， in, The output of the Focaler-GIoU loss function, For generalized intersection and comparison of losses, For regression samples, This is the original value.

6. A low-altitude small target recognition system based on full Stokes polarization imaging, used to implement the low-altitude small target recognition method based on full Stokes polarization imaging as described in any one of claims 1-5, characterized in that: It includes a four-channel polarization imaging unit, a target polarization parameter acquisition unit, a dataset generation unit, an embedded processing unit, and a result output unit connected in sequence; The four-channel polarization imaging unit is used to acquire full Stokes low-altitude polarization data; The target polarization parameter acquisition unit is used to receive all-Stokes low-altitude polarization data and parse it to obtain the target polarization parameters; The dataset generation unit is used to generate a large number of polarization images based on the CPFNet polarization fusion network to create a full Stokes polarization dataset; the CPFNet polarization fusion network is used to fuse circular polarization information. Embedded processing unit for deploying the trained, improved YOLOv8-FPZ network; The result output unit is used to input the full Stokes polarization data to be detected into the trained and improved YOLOv8-FPZ network to perform low-altitude small target recognition and output, and output the small target detection results, including the small target category, location and confidence level.

Images