A UAV detection method and system combined with infrared non-uniform deviation field correction

Abstract

The invention discloses a UAV detection method combined with infrared non-uniform deviation field correction, which belongs to the field of degraded image target detection. The method comprises: acquiring a degraded infrared image; inputting the degraded infrared image into a unified framework of infrared image non-uniform correction and UAV detection to perform non-uniform correction and UAV target detection, and outputting a UAV target detection result; the unified framework of infrared image non-uniform correction and UAV detection comprises a deviation field correction module and a UAV target detection module, and the deviation field correction module comprises a deviation field prediction network and a deviation field clearing structure; inputting the degraded infrared image into the deviation field correction module, performing deviation field prediction through the deviation field prediction network to generate a predicted deviation field, inputting the predicted deviation field and the degraded infrared image into the deviation field clearing structure to clear the predicted deviation field in the degraded infrared image to obtain a corrected image; transmitting the corrected image to the UAV target detection module to perform UAV recognition to obtain a UAV target in the corrected image.

Description

Technical Field

[0001] The present invention belongs to the field of degraded image target detection, and in particular relates to a drone detection method and system combined with infrared non-uniform deviation field correction. Background Art

[0002] In existing infrared drone monitoring systems, the low-frequency non-uniformity effect caused by temperature often leads to a decrease in image quality, which seriously affects the detection performance of drone targets. Existing methods often fail to effectively improve the accuracy and reliability of detection by performing target detection after performing non-uniformity correction. Most existing studies handle non-uniformity correction and target detection separately, and fail to effectively solve the problem of the mutual influence between the two.

[0003] Current non-uniformity correction methods are mainly divided into two categories: model-based and deep learning-based. Model-based methods usually correct image deviations through bivariate polynomial fitting, but they are prone to overfitting or poor convergence when dealing with complex or severely degraded non-uniformity scenes. Existing model-based correction methods are highly dependent on expert experience, which can easily lead to over- or under-correction, especially when facing complex or severely degraded images, it is difficult to obtain satisfactory results. Deep learning-based methods rely on complex network structures and a large amount of high-quality training data, but in practical applications it is difficult to obtain real degraded and clear image pairs, resulting in limited application effects.

[0004] More importantly, existing research focuses more on improving image visual quality, but less on the impact of non-uniformity correction on target detection. The lack of a unified framework that can solve these two problems at the same time means that drone target detection in complex environments still faces great challenges. Summary of the invention

[0005] The purpose of the present invention is to overcome the problem that the non-uniform effect in the existing infrared UAV image processing and the UAV target detection cannot be processed simultaneously, and a UAV detection method and system combined with infrared non-uniform deviation field correction are proposed.

[0006] In order to achieve the above object, the present invention adopts the following technical scheme:

[0007] In a first aspect, the present invention provides a method for detecting a drone combined with infrared non-uniform deviation field correction, comprising the following steps:

[0008] Obtain a degraded infrared image, input the degraded infrared image into the unified framework of infrared image non-uniform correction and UAV detection to perform non-uniform correction and UAV target detection, and output the UAV target detection result;

[0009] The unified framework of infrared image non-uniform correction and UAV detection includes a deviation field correction module and a UAV target detection module, and the deviation field correction module includes a deviation field prediction network and a deviation field clearing structure;

[0010] The step of inputting the degraded infrared image into the unified framework of infrared image non-uniform correction and drone detection to perform non-uniform correction and drone target detection includes the following steps:

[0011] The degraded infrared image is input into the deviation field correction module, the deviation field is predicted by the deviation field prediction network to generate a predicted deviation field, the predicted deviation field and the degraded infrared image are input into the deviation field clearing structure to clear the predicted deviation field in the degraded infrared image to obtain a corrected image;

[0012] The corrected image is transmitted to the UAV target detection module for UAV recognition to obtain the UAV target in the corrected image.

[0013] Further, the bias field prediction network includes downsampling, global and local feature encoders, convolution, full connection and bias field prediction models;

[0014] The step of inputting the degraded infrared image into the deviation field prediction module to perform deviation field prediction and generate a predicted deviation field comprises the following steps:

[0015] The degraded infrared image is downsampled to obtain a downsampled degraded infrared image, and the downsampled degraded infrared image is input into the global and local feature encoder for feature extraction to obtain the non-uniformity feature of the degraded infrared image; the non-uniformity feature of the degraded infrared image is input into the convolution for convolution operation, and then input into the fully connected layer for full connection operation to obtain the integrated estimated feature; the integrated estimated feature is input into the deviation field prediction model for deviation field prediction to obtain the predicted deviation field;

[0016] The global and local feature encoder includes a global encoder and a local encoder. The global encoder extracts the global non-uniformity features of the downsampled degraded infrared image, and the local encoder extracts the local non-uniformity features of the downsampled degraded infrared image. The global non-uniformity features of the downsampled degraded infrared image are fused with the local non-uniformity features of the downsampled degraded infrared image to obtain the non-uniformity features of the degraded infrared image.

[0017] Furthermore, the deviation field prediction model is as follows:

[0018]

[0019]

[0020] in, B ( xi , yj) represents the deviation field image, ( xi , yj ) represents the coordinate position value in the image, D represents the polynomial order, represents a single coefficient, t and s is the intermediate variable, Represents the coefficient vector composed of all polynomial coefficients; FC represents a fully connected operation, Conv5 represents 5 convolution operations, Ffused Represents the non-uniformity features of the degraded infrared image obtained by global and local feature encoders.

[0021] Furthermore, the drone target detection module includes a multi-layer convolutional neural network, loss optimization and a detection head, wherein the loss optimization includes target enhancement and background suppression loss and bias robust loss;

[0022] The step of transmitting the corrected image to the drone target detection module for drone recognition to obtain the drone target in the corrected image includes the following steps:

[0023] The corrected image is input into a multi-layer convolutional neural network to extract target features of different scales in each feature extraction layer to obtain a corrected image feature map. The corrected image feature map is jointly optimized using target enhancement, background suppression loss and bias robust loss to obtain multi-level features after background noise interference suppression. The multi-level features after background noise interference suppression are feature fused and the UAV target detection result is output through the detection head. The target detection result includes the location information and confidence of the UAV.

[0024] Furthermore, the calculation formula of the target enhancement and background suppression loss is as follows:

[0025]

[0026]

[0027] in, L TEBS represents the target enhancement and background suppression loss, L CE (·,·) represents the cross entropy loss, M A binary mask representing the background box, x represents the horizontal coordinate of the image, y represents the image ordinate, F i Represents the first layer of multi-layer convolutional neural network i The feature map of the stage, the target region represents the target area, and the background region represents the background area.

[0028] Furthermore, the calculation formula of the bias robust loss is as follows:

[0029]

[0030] in, represents the bias robust loss, Represents the first layer of multi-layer convolutional neural network i The cosine similarity of the stages.

[0031] In the second aspect, the present invention provides a drone detection device with combined infrared non-uniform deviation field correction, which utilizes the aforementioned drone detection method with combined infrared non-uniform deviation field correction, including a drone monitoring device, the drone monitoring device including an infrared camera and an embedded processor, the infrared camera and the embedded processor are connected, the embedded processor is provided with an image processing unit, and the image processing unit includes a deviation field correction module and a target detection module.

[0032] In a third aspect, the present invention provides a UAV detection system combined with infrared non-uniform deviation field correction, using the UAV detection method combined with infrared non-uniform deviation field correction, comprising:

[0033] An infrared image acquisition module, used for acquiring a degraded infrared image;

[0034] A drone detection module is used to input the degraded infrared image into a unified framework of infrared image non-uniform correction and drone detection to perform non-uniform correction and drone target detection, and output a drone target detection result; the unified framework of infrared image non-uniform correction and drone detection includes a deviation field correction module and a drone target detection module, and the deviation field correction module includes a deviation field prediction network and a deviation field clearing structure;

[0035] The step of inputting the degraded infrared image into the unified framework of infrared image non-uniform correction and drone detection to perform non-uniform correction and drone target detection includes the following steps:

[0036] The degraded infrared image is input into the deviation field correction module, the deviation field is predicted by the deviation field prediction network to generate a predicted deviation field, the predicted deviation field and the degraded infrared image are input into the deviation field clearing structure to clear the predicted deviation field in the degraded infrared image to obtain a corrected image;

[0037] The corrected image is transmitted to the UAV target detection module for UAV recognition to obtain the UAV target in the corrected image.

[0038] In a fourth aspect, the present invention provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the aforementioned method for detecting a drone with combined infrared non-uniform deviation field correction when executing the computer program.

[0039] In a fifth aspect, the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for detecting a drone combined with infrared non-uniform deviation field correction is implemented.

[0040] Compared with the prior art, the present invention has the following beneficial technical effects:

[0041] The present invention proposes a UAV detection method for combined infrared non-uniform deviation field correction, which realizes the joint optimization of non-uniformity correction and UAV target detection through a unified framework. It includes two parts, a deviation field correction module and a UAV target detection module, which are respectively used to solve the low-frequency non-uniformity correction problem of infrared images and the accurate detection problem of UAV targets. The present invention realizes end-to-end optimization of the two through a combined correction and detection framework, which can adapt to images with different degrees of degradation and improve the real-time and accuracy of detection. Through this new joint correction and detection framework, not only the correction quality of infrared images is improved, but also the performance of target detection is enhanced, thereby providing a more reliable UAV monitoring solution in complex practical application scenarios. Compared with the traditional step-by-step processing method, the lightweight correction modeling and unified processing framework of the present invention improve the overall processing effect while reducing the amount of calculation.

[0042] Furthermore, by using the lightweight design of the deviation field correction module and the dual encoder structure for the non-uniformity correction of the degraded infrared image, the present invention can effectively correct the low-frequency non-uniformity deviation in the infrared image, so that the image quality originally affected by factors such as temperature can be significantly improved. The corrected image has higher contrast and clearer details, providing a higher quality input for the drone target detection module, thereby improving the accuracy of subsequent detection.

[0043] Furthermore, the non-uniformity correction is modeled as a deviation field prediction model through a high-order bivariate polynomial, which is an estimation problem of a small number of parameters. Combining prior knowledge and data-driven methods, the correction process can be more lightweight and accurate.

[0044] Furthermore, in the unified framework UAV target detection module, the present invention introduces an auxiliary loss mechanism, Target Enhancement with Background Suppression Loss (TEBSLoss), which effectively reduces the interference of background noise through multi-layer feature extraction and background suppression mechanism. By adding supervision at different levels in the detection network, the network's ability to distinguish small UAV targets is enhanced, avoiding background interference, thereby improving the accuracy of detection. This module can more accurately identify small UAV targets in infrared images, improve the accuracy and robustness of target positioning, and is particularly suitable for UAV monitoring tasks in complex scenes.

[0045] Furthermore, by adopting a lightweight network structure and a multi-level feature extraction strategy, the system has strong real-time processing capabilities while maintaining high accuracy, which is suitable for real-time monitoring tasks of drones. In addition, the system can adaptively adjust the correction and detection parameters in different scenarios, has strong environmental adaptability, and can be applied to drone target detection with different bias field degradation.

[0046] Furthermore, the comparative experiments show that the present invention is superior to the existing methods in terms of comprehensive performance of non-uniformity correction and target detection. The correction module achieves higher PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index) on images with different degrees of degradation, with PSNR of 31.961 and SSIM of 0.9827; the detection module achieves higher P (Precision) and R (Recall) in the UAV detection task, with P of 0.999 and R of 0.822; and maintains a high FPS (Frames PerSecond), with 32 frames per second, meeting the needs of real-time monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The drawings described herein are for explanation purposes only and are not intended to limit the scope of the present invention in any way. In addition, the shapes and proportional dimensions of the components in the drawings are only for illustration purposes to help understand the present invention, and are not intended to specifically limit the shapes and proportional dimensions of the components of the present invention. In the drawings:

[0048] Figure 1 The present invention is a flow chart of a UAV detection method combined with infrared non-uniform deviation field correction.

[0049] Figure 2 This is a structural diagram of a UAV detection system combined with infrared non-uniform deviation field correction according to the present invention.

[0050] Figure 3 This is an electronic device diagram of a UAV detection method combined with infrared non-uniform deviation field correction in the present invention.

[0051] Figure 4 The present invention is a principle block diagram of a UAV detection method combined with infrared non-uniform deviation field correction. DETAILED DESCRIPTION

[0052] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0053] Embodiment 1

[0054] See also Figure 1 , a UAV detection method combined with infrared non-uniform deviation field correction, comprising the following steps:

[0055] Acquire degraded infrared images;

[0056] The degraded infrared image is input into the unified framework of infrared image non-uniform correction and UAV detection for non-uniform correction and UAV target detection, and the UAV target detection result is output; the unified framework of infrared image non-uniform correction and UAV detection includes a bias field correction module and a UAV target detection module, and the bias field correction module includes a bias field prediction network and a bias field clearing structure;

[0057] The degraded infrared image is input into the unified framework of infrared image non-uniform correction and drone detection for non-uniform correction and drone target detection, which includes the following steps:

[0058] The degraded infrared image is input into the deviation field correction module, the deviation field is predicted by the deviation field prediction network to generate a predicted deviation field, the predicted deviation field and the degraded infrared image are input into the deviation field clearing structure to clear the predicted deviation field in the degraded infrared image to obtain a corrected image;

[0059] The corrected image is transmitted to the UAV target detection module for UAV recognition to obtain the UAV target in the corrected image.

[0060] In this embodiment, by analyzing the pixel value distribution and characteristics of the degraded infrared image, the deviation field correction module can estimate a deviation field, which reflects the correction amount for each pixel in the image. Then, the original image is corrected using this deviation field to generate a higher quality corrected image. Detect and analyze the features in the corrected image; by matching and identifying these features, the drone visual inspection module can accurately determine the position of the drone in the image; integrate non-uniform correction and target detection into a unified framework, simplifying the processing flow and improving processing efficiency; through non-uniform correction, the quality of the infrared image is improved, thereby improving the accuracy of drone target detection; it can process infrared images of various degrees of degradation and has strong adaptability and robustness.

[0061] The bias field prediction network includes downsampling, global and local feature encoders, convolution, full connection and bias field prediction models;

[0062] Inputting the degraded infrared image into the deviation field prediction module to perform deviation field prediction and generate a predicted deviation field includes the following steps:

[0063] The degraded infrared image is downsampled to obtain a downsampled degraded infrared image, and the downsampled degraded infrared image is input into the global and local feature encoder for feature extraction to obtain the non-uniformity feature of the degraded infrared image; the non-uniformity feature of the degraded infrared image is input into the convolution for convolution operation, and then input into the fully connected layer for full connection operation to obtain the integrated estimated feature; the integrated estimated feature is input into the deviation field prediction model for deviation field prediction to obtain the predicted deviation field;

[0064] The global and local feature encoder includes a global encoder and a local encoder. The global encoder extracts the global non-uniformity features of the downsampled degraded infrared image, and the local encoder extracts the local non-uniformity features of the downsampled degraded infrared image. The global non-uniformity features of the downsampled degraded infrared image are fused with the local non-uniformity features of the downsampled degraded infrared image to obtain the non-uniformity features of the degraded infrared image.

[0065] This embodiment uses a global and local feature encoder that combines the advantages of a global encoder and a local encoder, and can simultaneously capture the overall and local features of the input data. The global and local feature encoder structure can be used for tasks such as image recognition, segmentation, and generation. This embodiment uses a global encoder to extract the overall features of the image. Unlike the global encoder, this embodiment uses a local encoder to focus on the local features or detail information of the input data. The local encoder extracts the unique features of these areas or fragments by focusing on the local areas or fragments of the data, and the local encoder extracts the local features of the image. The global encoder extracts the overall features of the image, while the local encoder extracts the local detail features of the image. The combination of the two can achieve a more sophisticated image processing effect.

[0066] The deviation field prediction model is as follows:

[0067]

[0068]

[0069] in, B ( xi , yj ) represents the deviation field image, ( xi , yj ) represents the coordinate position value in the image, D represents the polynomial order, represents a single coefficient, t and s is the intermediate variable, Represents the coefficient vector composed of all polynomial coefficients; FC represents a fully connected operation, Conv5 represents 5 convolution operations, Ffused Represents the non-uniformity features of the degraded infrared image obtained by global and local feature encoders.

[0070] High-order bivariate polynomial fitting to correct image deviation is an image processing technique that combines the principles and methods of polynomial fitting to correct deviations in images. Polynomial fitting is a mathematical method used to find a polynomial function that can approximate a given set of data points as closely as possible. These deviations can cause inaccurate or distorted information in the image, thereby affecting the quality and usability of the image. High-order bivariate polynomial fitting can correct these deviations. This embodiment uses a global and local feature encoder structure to extract non-uniformity features from the original image. The features extracted by the global and local feature encoder structure are input into a series of convolutional layers, which further refine the features, capture subtle changes and patterns in the image, and extract the essential features of the deviation field through multi-layer cascades using convolution kernels. After being refined by the convolutional layer, the feature map is flattened and input into the fully connected layer, which can integrate features from different convolutional layers and learn the complex relationships between them. By stacking the connection layers, a powerful feature representation is constructed for subsequent correction tasks. After the features are integrated by the fully connected layer, these features are used to estimate the deviation field. A model can be trained to learn the mapping from features to deviation fields. Once the deviation field is obtained, the original degraded infrared image can be corrected and the deviation can be corrected by high-order bivariate polynomial fitting.

[0071] By properly selecting the order of the polynomial and the fitting method, the deviation in the image can be effectively corrected, improving the quality and usability of the image. The value of the deviation field is applied to each pixel of the image, adjusting the properties and thus eliminating the non-uniformity.

[0072] The drone target detection module includes a multi-layer convolutional neural network, loss optimization, and a detection head. The loss optimization includes target enhancement and background suppression loss and bias robust loss.

[0073] The corrected image is transmitted to the drone target detection module for drone recognition to obtain the drone target in the corrected image, including the following steps:

[0074] The corrected image is input into a multi-layer convolutional neural network to extract target features of different scales in each feature extraction layer to obtain a corrected image feature map. The corrected image feature map is jointly optimized using target enhancement, background suppression loss and bias robust loss to obtain multi-level features after background noise interference suppression. The multi-level features after background noise interference suppression are feature fused and the UAV target detection result is output through the detection head. The target detection result includes the UAV's location information and confidence.

[0075] The calculation formula for target enhancement and background suppression loss is as follows:

[0076]

[0077]

[0078] in, L TEBS represents the target enhancement and background suppression loss, L CE (·,·) represents the cross entropy loss, M A binary mask representing the background box, x represents the horizontal coordinate of the image, y represents the image ordinate, F i Represents the first layer of multi-layer convolutional neural network i The feature map of the stage, the target region represents the target area, and the background region represents the background area.

[0079] In target detection tasks, the complexity of background information often has a negative impact on the performance of the model. The background may contain features similar to the target, or there may be noise and interference that are unrelated to the target, which will increase the difficulty of the model to identify and locate the target. Therefore, through background suppression technology, the interference of background information can be reduced and the accuracy of the model's recognition and positioning of the target can be improved. Multi-scale feature extraction improves the detection ability of drone targets of different sizes and distances. Background noise suppression reduces the impact of background noise on the detection results through TEBS Loss and improves the accuracy of detection. Feature fusion combines feature information at different levels to enhance the robustness and generalization ability of the model. It is suitable for drone monitoring and detection tasks in complex backgrounds. In these scenarios, drone targets may appear at different distances, angles, and lighting conditions, and the background may also contain a lot of noise and interference. By combining multi-layer convolutional neural networks and TEBS Loss, accurate drone target detection can be achieved.

[0080] The calculation formula of bias robust loss is as follows:

[0081]

[0082] in, represents the bias robust loss, Represents the first layer of multi-layer convolutional neural network i The cosine similarity of the stages.

[0083] The bias robust loss uses a similarity measurement method to measure the similarity of features between the clear image and the corrected image, maximizes the similarity between the features of the clear image and the corrected image, and makes the features of the corrected image as consistent as possible with the features of the clear image, thereby improving the fidelity of the corrected image in the feature space and ensuring that it retains the basic features that are beneficial to detection.

[0084] Embodiment 2

[0085] A drone detection device combined with infrared non-uniform deviation field correction, using a drone detection method combined with infrared non-uniform deviation field correction described in Example 1, includes a drone monitoring device, the drone monitoring device includes an infrared camera and an embedded processor, the infrared camera is connected to the embedded processor, the embedded processor is provided with an image processing unit, and the image processing unit includes a deviation field correction module and a target detection module.

[0086] The deviation field correction module in the image processing unit of this embodiment performs non-uniform correction on the captured infrared image to generate a corrected image, which is input into the target detection module. The module uses deep learning technology or traditional image processing algorithms to extract and classify features of the image, identify the drone target, and output its location information and confidence. The detection results can be displayed or transmitted through the output interface of the embedded processor, such as a display screen, a network interface, etc. By integrating the infrared camera and the image processing unit, a high-performance embedded processor is used for real-time image processing and analysis. The infrared camera is not affected by light and can monitor at night or in bad weather conditions. High-precision, long-distance, and all-weather monitoring of the monitored target is achieved. The infrared camera is used to capture infrared images of the monitored area and has the ability to see at night and penetrate smoke and haze. The image processing unit is responsible for preprocessing and enhancing the images captured by the infrared camera, such as denoising, contrast adjustment, etc. It is possible to detect, track and identify targets on the image and extract key information. The image of the monitored area is captured by the infrared camera and transmitted to the ground station or the cloud in real time for display and analysis. The image processing unit is used to process and analyze the image to automatically detect and track the target. It can identify the type, location, speed and other information of the target to provide support for decision-making. The high-performance embedded processor is responsible for running the image processing algorithm and real-time data processing. It has the characteristics of low power consumption and high performance, and can meet the dual requirements of real-time performance and power consumption. The combination of high-performance embedded processor and image processing unit improves the accuracy and reliability of target recognition.

[0087] Embodiment 3

[0088] See also Figure 2 A UAV detection system combined with infrared non-uniform deviation field correction, using a UAV detection method combined with infrared non-uniform deviation field correction described in Example 1, comprising:

[0089] An infrared image acquisition module, used for acquiring a degraded infrared image;

[0090] The UAV detection module is used to input the degraded infrared image into the infrared image non-uniform correction and UAV detection unified framework for non-uniform correction and UAV target detection, and output the UAV target detection result; the infrared image non-uniform correction and UAV detection unified framework includes a deviation field correction module and a UAV target detection module, and the deviation field correction module includes a deviation field prediction network and a deviation field clearing structure;

[0091] The degraded infrared image is input into the unified framework of infrared image non-uniform correction and drone detection for non-uniform correction and drone target detection, which includes the following steps:

[0092] The degraded infrared image is input into the deviation field correction module, the deviation field is predicted by the deviation field prediction network to generate a predicted deviation field, the predicted deviation field and the degraded infrared image are input into the deviation field clearing structure to clear the predicted deviation field in the degraded infrared image to obtain a corrected image;

[0093] The corrected image is transmitted to the UAV target detection module for UAV recognition to obtain the UAV target in the corrected image.

[0094] This embodiment uses a deviation field correction module to perform non-uniform correction on the degraded infrared image, effectively eliminating the deviation field in the image, thereby improving the clarity and quality of the image. The corrected image is closer to the real scene, reducing the detection error caused by image degradation. The drone detection module uses the corrected image to detect drone targets, which can more accurately identify drone targets, improve the accuracy and reliability of drone target detection, and reduce the probability of false alarms and missed alarms. Unifying infrared image non-uniform correction and drone detection in a framework achieves seamless connection and collaborative work between the two, simplifies the system structure, and improves the system's integration and maintainability.

[0095] Embodiment 4

[0096] See also Figure 3 , an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for detecting a drone combined with infrared non-uniform deviation field correction described in Embodiment 1 is implemented.

[0097] Embodiment 5

[0098] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for detecting a drone combined with infrared non-uniform deviation field correction described in Embodiment 1 is implemented.

[0099] Embodiment 6

[0100] The following is a detailed description of a drone detection method combined with infrared non-uniform deviation field correction in conjunction with the accompanying drawings: a unified framework of infrared image non-uniform correction and drone detection is adopted, which mainly includes a deviation field correction module and a drone target detection module to perform drone target detection on degraded infrared images, which are used to solve the low-frequency non-uniformity correction problem of infrared images and the accurate detection problem of drone targets, respectively.

[0101] See also Figure 4,The overall framework is divided into the upper layer of bias field correction module and the lower layer of UAV target detection module. The system input is a degraded infrared image with low-frequency ,nonuniformity effect. After passing through the bias field correction module, a ,corrected image is generated. The corrected image is further transmitted to the target ,detection module, which identifies the UAV target in the image and outputs the ,detection result.

[0102] See also Figure 4 ,The deviation field correction module is responsible for the non-uniformity correction of the degraded image, and its specific steps are as follows:

[0103] The degraded infrared image is input into the system as the input of the deviation field correction.

[0104] The module adopts a global and local feature encoder structure, including a global encoder and a local encoder. The global encoder extracts the overall feature information of the image to capture the global non-uniformity deviation of the image; the local encoder extracts local detail features to accurately capture small-scale non-uniformity.

[0105] The features extracted by the encoder are combined through a series of convolutional and fully connected layers to further refine the non-uniformity characteristics of the image to ensure the accuracy of the deviation field.

[0106] The deviation field is predicted through a lightweight network structure, and the general structure of the degraded image is combined with a correction method based on a high-order bivariate polynomial to model the deviation field, thereby generating a corresponding corrected image. The corrected image is used as the input of the target detection module to provide high-quality detection data.

[0107] The general structure of a degraded image is: degraded image = clear image + bias field;

[0108] Therefore, the correction idea for degraded infrared images is to remove the bias field in the degraded images to obtain corrected images.

[0109] The bias field is modeled using a high-order bivariate polynomial as follows:

[0110]

[0111] in, B ( xi , yj ) represents the deviation field image, ( xi , yj ) represents the coordinate position value in the image, D represents the polynomial order, represents a single coefficient, t and s is an intermediate variable;

[0112]

[0113] Represents the coefficient vector composed of all polynomial coefficients; FC represents a fully connected operation, Conv5 represents 5 convolution operations, F fused Represents the non-uniformity features of the degraded infrared image obtained by global and local feature encoders.

[0114] Although the present embodiment adopts a correction method based on a high-order bivariate polynomial, other types of models may also be used to achieve similar non-uniformity correction effects. For example, a Fourier transform model may be used to extract low-frequency noise and perform correction, or a wavelet transform model may be used to separate image components of different frequencies. These models may also achieve similar effects when removing the deviation field, and may replace the polynomial model used in the present embodiment.

[0115] The modular lightweight front-end and back-end encoder structure achieves efficient non-uniformity correction through the estimation of a small number of parameters, which can effectively reduce the amount of calculation and adapt to different degrees of image degradation.

[0116] In addition to the currently used lightweight transformer model, i.e., the global encoder, other lightweight deep learning networks can be used for parameterized prediction, such as convolutional neural networks or long short-term memory networks. These networks can learn the characteristics of the bias field and achieve automatic correction, achieving similar results to the current method.

[0117] Different prior constraints can also be introduced, such as gradient domain constraints, edge preservation constraints, etc., to replace the current prior form: high-order polynomials. These constraints can also ensure the stability and accuracy of the correction, so that the system can obtain reliable correction effects in different image scenes.

[0118] Based on the rectified image, the drone target detection module analyzes the image through a multi-layer convolutional neural network structure to locate and identify the drone target. The specific steps are as follows:

[0119] The rectified image is input and features at different levels are extracted through multiple layers of convolution units. Each layer of convolution kernel focuses on extracting target features of different scales, especially the detailed features of small targets such as drones.

[0120] In order to avoid falling into the local optimal solution, this implementation designs a self-supervised loss, namely BR loss (bias robust loss), to maximize the consistency of the feature representation of the clear image and the corrected image, thereby improving the fidelity of the corrected image in the feature space and ensuring that it retains the basic features that are beneficial to detection.

[0121] The calculation formula of BR loss is as follows:

[0122]

[0123] in, represents the bias robust loss, Represents the first layer of multi-layer convolutional neural network i The cosine similarity of the stages.

[0124]

[0125] Represents the feature map of the i-th stage of the clear image, represents the feature map of the ith stage of the corrected image, i =1, 2, 3, 4.

[0126] Here, the clear image is only used for joint training of correction and detection to evaluate the alignment of the corrected image and should not be used for correction training on real infrared images.

[0127] TEBS Loss is added to different feature extraction layers to further optimize the performance of the feature extraction network. TEBS Loss suppresses the interference of background noise, enhances the network's attention to the drone target, and improves the accuracy of detection. The calculation formula for target enhancement and background suppression loss is shown in the following formula:

[0128]

[0129]

[0130] in, L TEBS represents the target enhancement and background suppression loss, L CE (·,·) represents the cross entropy loss, M A binary mask representing the background box, x represents the horizontal coordinate of the image, y represents the image ordinate, F i Represents the first layer of multi-layer convolutional neural network i The feature map of the stage, the target region represents the target area, and the background region represents the background area.

[0131] Based on the combination of multi-level features, the target detection results are finally generated in the output layer, including the location information and confidence of the drone.

[0132] The drone target detection module introduces TEBS Loss as an auxiliary loss, enabling the network to effectively distinguish targets from backgrounds, thereby improving the accuracy and robustness of drone detection.

[0133] The target detection module currently used in this embodiment can be replaced with other known detection network architectures, such as the YOLO (You Only Look Once, single gaze target detection algorithm) series, Faster R-CNN (Faster Region-based Convolutional Neural Network, faster region-based convolutional neural network) and other detection methods. By replacing different detection architectures, the system can select the most suitable detection model according to different scene requirements. These networks can improve adaptability in different scenes while maintaining detection accuracy.

[0134] The detection module can also be adjusted in an adaptive way, that is, the detection parameters or the number of network layers can be adaptively adjusted according to the different conditions of image quality and deviation field. For example, by introducing a preprocessing module, the confidence threshold of the detection result can be dynamically adjusted according to the quality of the image. This adaptive strategy can improve the flexibility of detection in complex scenarios.

[0135] The framework of this embodiment supports end-to-end joint optimization. The deviation field correction module and the target detection module are trained in a unified network, so that correction and detection can optimize each other and improve the overall performance of the system. In addition, the deviation field correction module has adaptive correction capabilities and can automatically adjust the correction parameters according to the different degrees of image degradation to meet the needs of different scenarios.

[0136] This embodiment can be integrated into existing drone monitoring equipment, mainly including an infrared camera and an image processing unit. The deviation field correction and target detection module can be implemented on a high-performance embedded processor to support real-time processing. When processing degraded images, the system first performs non-uniformity correction and then performs target detection to ensure the accuracy of target detection.

[0137] This embodiment is suitable for all-weather monitoring of drones, especially at night or in bad weather conditions, and can accurately detect and track drone targets in complex backgrounds. The system can be used in military, security and other fields, which require high accuracy and real-time performance of target detection.

[0138] This embodiment achieves efficient processing of infrared images in complex environments through a unified non-uniformity correction and UAV target detection framework, significantly improving the detection accuracy and real-time performance of UAV monitoring.

[0139] Although this embodiment uses a unified framework to achieve joint optimization of non-uniformity correction and target detection, it is also possible to adopt a method of correction first and then detection or parallel processing. For example, image correction is first performed to improve image quality, and then the detection algorithm is run on the corrected image. In this way, the correction and detection modules can be separated, which is suitable for scenes with low real-time requirements. Although this method is slightly different in efficiency, it can still achieve similar technical effects.

[0140] In addition to the end-to-end integration solution, the framework can also be integrated in a modular way. For example, the correction module can be embedded into the existing infrared detection system as a plug-in module to enable the system to have correction and detection functions. In this way, modular configuration can be performed according to needs, with higher system compatibility.

[0141] The following experimental verification results further illustrate the drone detection method of this embodiment combined with infrared non-uniform deviation field correction. The first traditional method, the second traditional method and the drone detection method of this embodiment combined with infrared non-uniform deviation field correction are used to detect drone targets in degraded infrared images. The first traditional method is to directly apply the drone detection model to the non-uniform degraded image, and the second traditional method is to use the correction model as a preprocessing step to perform correction before the image is passed to the detector. Both traditional methods can only optimize one task separately. The drone detection method of this embodiment combined with infrared non-uniform deviation field correction realizes the joint optimization of non-uniform correction and target detection through a unified framework, and the correction quality and detection accuracy are improved at the same time. The drone target detection result in the degraded infrared image obtained using the method of this embodiment is clearer and more reliable.

[0142] PSNR and SSIM are used as objective evaluation indicators for the deviation field correction module in the unified framework of infrared image non-uniform correction and drone detection provided in this embodiment. P and R are used to evaluate the detection performance of the drone target detection module in the unified framework of infrared image non-uniform correction and drone detection provided in this embodiment. Finally, FPS is used as the evaluation indicator for the real-time performance of the unified framework of infrared image non-uniform correction and drone detection. The public data set and real data are corrected and detected using the traditional method and the method of this embodiment, and the comparison results are shown in Table 1.

[0143] Table 1 Comparison results between two traditional methods and the method of this embodiment

[0144]

[0145] It can be seen from the experimental results that this embodiment is superior to the traditional method in terms of the comprehensive performance of non-uniformity correction and target detection. The correction module achieves higher PSNR and SSIM on images with different degrees of degradation, and the detection module achieves higher P and R in the drone detection task, and maintains a high FPS to meet the needs of real-time monitoring. Because the first traditional method refers to directly performing drone target detection without performing image non-uniformity deviation field correction operations, the first traditional method has no value for the indicators that measure the correction effect, that is, the PSNR and SSIM indicators of the first traditional method in Table 1 are blank.

[0146] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0147] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0148] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0149] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention is described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation of the present invention can still be modified or replaced by equivalents, and any modification or equivalent replacement that does not deviate from the spirit and scope of the present invention should be included in the protection scope of the present invention.

Claims

1. A UAV detection method combined with infrared non-uniform deviation field correction, characterized in that: The following steps are involved: Obtain a degraded infrared image, input the degraded infrared image into the unified framework of infrared image non-uniform correction and UAV detection to perform non-uniform correction and UAV target detection, and output the UAV target detection result; The unified framework includes a deviation field correction module and a UAV target detection module; The bias field correction module includes a bias field prediction network and a bias field cleaning structure. The bias field prediction network includes downsampling, global and local feature encoders, convolution, full connection and bias field prediction models. The drone target detection module includes loss optimization, which includes target enhancement and background suppression loss and bias robust loss; the calculation formula for target enhancement and background suppression loss is: in, is the cross entropy loss, is the binary mask of the background box, is the image horizontal coordinate, is the image ordinate; The degraded infrared image is input into a unified framework for non-uniform correction and UAV target detection, including: The degraded infrared image is input into the deviation field correction module, and the deviation field is predicted through the deviation field prediction network to generate a predicted deviation field, including: downsampling the degraded infrared image to obtain a downsampled degraded infrared image, inputting the downsampled degraded infrared image into the global and local feature encoder for feature extraction, the global and local feature encoder includes a global encoder and a local encoder, and obtaining the non-uniformity characteristics of the degraded infrared image; inputting the non-uniformity characteristics of the degraded infrared image into the convolution for convolution operation, and then inputting the fully connected layer for full connection operation to obtain integrated estimated features; inputting the integrated estimated features into the deviation field prediction model for deviation field prediction to obtain a predicted deviation field; the global encoder extracts the global non-uniformity characteristics of the downsampled degraded infrared image, the local encoder extracts the local non-uniformity characteristics of the downsampled degraded infrared image, and the global non-uniformity characteristics of the downsampled degraded infrared image are fused with the local non-uniformity characteristics of the downsampled degraded infrared image to obtain the non-uniformity characteristics of the degraded infrared image; The predicted deviation field and the degraded infrared image are input into a deviation field clearing structure to clear the predicted deviation field in the degraded infrared image to obtain a corrected image; the corrected image is transmitted to a UAV target detection module for UAV recognition to obtain a UAV target in the corrected image.

2. The method for detecting unmanned aerial vehicles combined with infrared non-uniform deviation field correction according to claim 1 is characterized in that: The deviation field prediction model is as follows: in, B ( x i , y j ) represents the deviation field image, ( x i , y j ) represents the coordinate position value in the image, D represents the polynomial order, represents a single coefficient, t and s is the intermediate variable, Represents the coefficient vector composed of all polynomial coefficients; FC represents a fully connected operation, Conv5 represents 5 convolution operations, F fused Represents the non-uniformity features of the degraded infrared image obtained by global and local feature encoders.

3. The method for detecting unmanned aerial vehicles combined with infrared non-uniform deviation field correction according to claim 1 is characterized in that: The drone target detection module includes a multi-layer convolutional neural network and a detection head; The step of transmitting the corrected image to the drone target detection module for drone recognition to obtain the drone target in the corrected image includes the following steps: The corrected image is input into a multi-layer convolutional neural network to extract target features of different scales in each feature extraction layer to obtain a corrected image feature map. The corrected image feature map is jointly optimized using target enhancement, background suppression loss and bias robust loss to obtain multi-level features after background noise interference suppression. The multi-level features after background noise interference suppression are feature fused and the UAV target detection result is output through the detection head. The target detection result includes the location information and confidence of the UAV.

4. The method for detecting unmanned aerial vehicles combined with infrared non-uniform deviation field correction according to claim 1 is characterized in that: The calculation formula of the bias robust loss is as follows: in, represents the bias robust loss, Represents the first layer of multi-layer convolutional neural network i The cosine similarity of the stages.

5. A UAV detection device combined with infrared non-uniform deviation field correction, using a UAV detection method combined with infrared non-uniform deviation field correction as described in any one of claims 1 to 4, characterized in that: It includes a drone monitoring device, which includes an infrared camera and an embedded processor. The infrared camera is connected to the embedded processor, and the embedded processor is provided with an image processing unit. The image processing unit includes a deviation field correction module and a target detection module.

6. A UAV detection system combined with infrared non-uniform deviation field correction, using a UAV detection method combined with infrared non-uniform deviation field correction as described in any one of claims 1 to 4, characterized in that: include: An infrared image acquisition module, used for acquiring a degraded infrared image; A drone detection module is used to input the degraded infrared image into a unified framework of infrared image non-uniform correction and drone detection to perform non-uniform correction and drone target detection, and output a drone target detection result; the unified framework of infrared image non-uniform correction and drone detection includes a deviation field correction module and a drone target detection module, and the deviation field correction module includes a deviation field prediction network and a deviation field clearing structure; The step of inputting the degraded infrared image into the unified framework of infrared image non-uniform correction and drone detection to perform non-uniform correction and drone target detection includes the following steps: The degraded infrared image is input into the deviation field correction module, the deviation field is predicted by the deviation field prediction network to generate a predicted deviation field, the predicted deviation field and the degraded infrared image are input into the deviation field clearing structure to clear the predicted deviation field in the degraded infrared image to obtain a corrected image; The corrected image is transmitted to the UAV target detection module for UAV recognition to obtain the UAV target in the corrected image.

7. An electronic device, characterized in that: The invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, a method for detecting a drone combined with infrared non-uniform deviation field correction as described in any one of claims 1 to 4 is implemented.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for detecting a drone combined with infrared non-uniform deviation field correction as described in any one of claims 1 to 4 is implemented.

Images