Dual-band infrared image decision fusion detection method based on monocular Cassegrain structure

Abstract

The invention discloses a dual-band infrared image decision fusion detection method of a monocular Cassegrain structure, including building a binocular vision platform of medium-wave infrared and Cassegrain structure long-wave infrared; obtaining medium-wave and Cassegrain structure long-wave UAV target images; constructing a KAN_yolov10 pre-detection head; using a multi-scale feature fusion module to fuse the medium-wave and Cassegrain structure long-wave UAV target images to obtain a fused image of the UAV target; inputting the fused image into the KAN_yolov10 pre-detection head, and outputting the detection result of the UAV target. The method of the invention uses hardware equipment to enlarge the region of interest, enhance the characteristics of the UAV target, and improve the detection accuracy of weak targets in complex backgrounds through decision-level fusion without the need for completely strict pixel alignment.

Description

Technical Field

[0001] The invention belongs to the technical field of unmanned aerial vehicle detection, and specifically relates to a dual-band infrared image decision fusion detection method of a monocular Cassegrain structure. Background Art

[0002] "Low, slow and small" drone targets are easily submerged in the complex background of the city and are easily affected by factors such as lighting and weather, making anti-drone tracking in the complex background of the city a very challenging task; the observation performance and machine vision perception capabilities of a single sensor will decrease under conditions of severe external interference; multi-source target detection has a large amount of calculation and poor stability, and the infrared energy of the Cassegrain structure is weaker and the field of view is smaller. Infrared imaging systems have excellent night vision capabilities and good anti-interference capabilities, and are increasingly widely used in the military field. Multi-band fusion detection greatly increases the probability of target detection by utilizing the redundancy and complementarity of information in different bands.

[0003] The mainstream drone detection technologies now include radar detection, radio frequency detection, audio detection, photoelectric detection, infrared thermal imaging detection, and detection of multiple sensor fusions. The detection cost of drones based on photoelectric detection is low, and visible light images have rich texture information, but they are easily affected by buildings, clouds, vegetation, etc., and cannot work at night; drones based on infrared detection have long detection distances, have target highlight information, and perform better at night. With the development of night vision technology, a single sensor cannot be used in all occasions, especially in urban backgrounds, where a single infrared camera drone target has only point features, so it is difficult to detect weak drone targets; and in scenes with direct sunlight, the contrast of medium-wave infrared is higher, and long-wave infrared has light scattering and refraction, so the detected image information is more. When the solar radiation is low at night, the texture details of medium-wave infrared images are more, and the contrast of long-wave infrared is better, so airborne optoelectronic platforms equipped with medium-wave and long-wave infrared cameras can work day and night, and the two contain a lot of complementary information. The joint detection of medium- and long-wave infrared improves the accuracy of detection.

[0004] Moreover, the existing fusion detection algorithm first performs registration and then fusion, and then performs subsequent advanced image processing, such as target detection and tracking. The amount of computation is large, and it is difficult to meet the real-time requirements. In addition, the drone target is small and appears as a point target in the infrared image at a long distance. It lacks edge features, which makes it very difficult to extract features and match the image. Summary of the invention

[0005] The purpose of the present invention is to address the above-mentioned deficiencies in the prior art and to provide a dual-band infrared image decision fusion detection method with a monocular Cassegrain structure, so as to solve the problem that the existing UAV detection is easily affected by the environment, resulting in poor detection results.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] A dual-band infrared image decision fusion detection method for a monocular Cassegrain structure comprises the following steps:

[0008] S1. Build a binocular vision platform for medium-wave infrared and Cassegrain structure long-wave infrared, and calibrate the binocular camera on the binocular vision platform;

[0009] S2. Based on the binocular vision platform, obtain the medium-wave and Cassegrain structure long-wave UAV target images, and pre-process the UAV target images;

[0010] S3, construct KAN_yolov10 pre-detection head based on KAN algorithm;

[0011] S4, using a multi-scale feature fusion module to fuse the medium-wave and Cassegrain structure long-wave UAV target images to obtain a fused image of the UAV target;

[0012] S5. Introduce Manhattan distance into the KAN_yolov10 pre-detection head, then input the fused image into the KAN_yolov10 pre-detection head, and output the detection result of the drone target.

[0013] Furthermore, S1 includes the following sub-steps:

[0014] S11. Build a binocular vision platform for medium-wave infrared and Cassegrain structure long-wave infrared on the gimbal, and image the medium-wave optical path and the Cassegrain structure long-wave optical path on the same plane;

[0015] S12. Use Zhang Zhengyou calibration method to obtain the internal parameter matrix of the binocular camera:

[0016] ;

[0017] Among them, M is the internal parameter matrix; f is the focal length of the binocular camera; , They represent the width and height of a pixel in physical size respectively; is the equivalent focal length expressed in pixel width and length on the u-axis in the pixel coordinate system, is the focal length parameter on the v axis in the pixel coordinate system; , is the principal point coordinate, which indicates the position of the optical center of the image in the pixel coordinate system;

[0018] S13. Obtain the mapping relationship between the pixel coordinate system and the world coordinate system according to the intrinsic parameter matrix:

[0019] ;

[0020] in, is the homography matrix, which represents the mapping relationship between the pixel coordinate system on the image and the world coordinate system; is the degree of inclination between the x-axis and the y-axis in the image coordinate system; is the rotation matrix, t is the translation vector, is the rigid body transformation matrix; s is the scale factor;

[0021] S14. Convert the medium wave and Cassegrain structure long wave UAV target images according to the homography matrix:

[0022] ;

[0023] in, is a point on the medium-wave UAV target image, is a point on the Cassegrain structure long-wave UAV target image.

[0024] Furthermore, S2 pre-processes the drone target image, including:

[0025] Perform Gaussian filtering on the drone target image:

[0026] ;

[0027] in, is a Gaussian function; x and y are pixel coordinates; is the standard deviation of the Gaussian function; is the base of natural logarithms;

[0028] Use the Tophat algorithm to suppress the background part in the drone target image:

[0029] ;

[0030] in, is the Tophat operation; I is the grayscale image; b is the morphological structure operator; For the opening operation.

[0031] Furthermore, S3 specifically includes:

[0032] The KAN algorithm is used to construct the C2f_KAN structure, and the C2f_KAN structure is used to replace the c2f structure of yolov 10n to construct the KAN_yolov10 pre-detection head.

[0033] Furthermore, the KAN algorithm is:

[0034] ;

[0035] in, is the output result after the input m is processed by the KAN algorithm; , , and They are the univariate function matrices of the L-1th layer, L-2th layer, 1st layer, and 0th layer networks respectively; It is a function matching operator, which means that the output of one function is used as the input of the next function; is the vector of input variables.

[0036] Furthermore, an efficient multi-scale attention mechanism is introduced before the KAN_yolov10 pre-detection head, and a fusion quality evaluation factor is introduced after the KAN_yolov10 pre-detection head;

[0037] The fusion quality evaluation factor is:

[0038] ;

[0039] in, is the fusion quality evaluation factor, is the index of the pixel coordinates; is a hyperparameter used to balance the importance of localization quality and classification quality; To locate the quality; For classification quality.

[0040] Further, S4 includes the following sub-steps:

[0041] S41, using the GRDB module to extract the depth features of the medium-wave UAV target image and the Cassegrain structure long-wave UAV target image in the ROI area;

[0042] S42, using resnet50 residual network to extract shallow features of medium-wave UAV target images and Cassegrain structure long-wave UAV target images in the ROI area;

[0043] S43, firstly fusing the shallow features of the medium-wave UAV target image with the shallow features of the Cassegrain structure long-wave UAV target image, and then fusing the deep features of the medium-wave UAV target image with the deep features of the Cassegrain structure long-wave UAV target image;

[0044] S44, send the fused shallow features and the fused deep features into the convolution layer for fusion to obtain the fused image of the drone target:

[0045] ;

[0046] in, To fuse the images; is a convolution operation of size 1x1; is the shallow feature after fusion; is the deep feature after fusion; It is the concatenation operation of feature maps.

[0047] Furthermore, the GRDB module is used in S41 to extract deep features, specifically:

[0048] ;

[0049] in; is the output of the GRDB module, representing the extracted deep features; is a grouped residual dense block; It is an n-level cascade convolution layer; It is the input of the GRDB module, which represents the input medium-wave UAV target image or Cassegrain structure long-wave UAV target image; is the gradient operator.

[0050] Furthermore, the shallow feature fusion and deep feature fusion of the drone target image in S43 are as follows:

[0051] ;

[0052] ;

[0053] ;

[0054] in, is the fused feature; It is the characteristics of medium wave UAV target image; The characteristics of the long-wave UAV target image for the Cassegrain structure; Obtained for medium wave UAV target image The weight after the maximum and minimum normalization of the graph; The Cassegrain structure long-wave UAV target image is obtained The weights after maximum and minimum normalization of the graph.

[0055] Furthermore, S5 introduces the Manhattan distance in the KAN_yolov10 pre-detection head, and uses the Manhattan distance to measure the distance between the candidate position of the drone target to be detected in the fused image and the center or corner of the reference truth box, specifically:

[0056] ;

[0057] in, is the Manhattan distance; is the position of the center of the reference truth box, Represents the Manhattan distance from the sample point to the center of the benchmark truth box, represents the Manhattan distance from the center of the reference truth box to any corner point, w and h represent the width and height of the reference truth box respectively; is the candidate position center of the target to be detected.

[0058] The dual-band infrared image decision fusion detection method of the monocular Cassegrain structure provided by the present invention has the following beneficial effects:

[0059] 1. The present invention builds a binocular vision platform of a Cassegrain structure long-wave infrared camera and an ordinary medium-wave infrared camera, increases the detection accuracy of weak drone targets by designing target detection and image fusion, designs a pre-detection head based on yolov10, and replaces the original c2f structure with the KAN (Kolmogorov–Arnold Networks) algorithm to reduce the number of model parameters, increase the speed of reasoning, and quantize and miniaturize the model, taking the detection result as prior information, and giving the aiming frame position in the other mode corresponding to each mode according to the corresponding relationship between the two modes, and replacing the original center degree measurement according to the Manhattan distance of the two aiming frames, which can solve the spatial offset problem of the two modes. In feature fusion, the resnet50 residual network with reduced convolution layers is used to extract shallow features, the GRDB module is used to extract deep information, and feature fusion is performed separately, and the final fused image is obtained according to the fusion quality evaluation factor, and finally the enhanced fused image is sent to the pre-detection head to obtain the final detection result.

[0060] 2. The binocular vision platform constructed by the present invention enables the binocular images to be aligned in hardware, performs monocular distortion correction, eliminates image distortion, and aligns the binocular images in software after calibration, providing hardware support for realizing dual-band fusion UAV target detection.

[0061] 3. The present invention utilizes the complementary advantages of Cassegrain structure long-wave infrared and medium-wave infrared. Under different weather conditions, different bands have different performances on the same object. Among them, medium-wave infrared has a longer detection distance than Cassegrain structure long-wave infrared in hot and humid weather. According to the radiation characteristics of different objects, medium-wave infrared is also suitable for detecting aircraft engines and tail flames; in haze weather, Cassegrain structure long-wave infrared has a longer detection distance than medium-wave infrared and is also suitable for detecting aircraft skin.

[0062] 4. The decision-level small target detection algorithm of the present invention is based on dual-band image fusion. After detection, the target ROI area and the background area of ​​the drone target image are fused and restored into a new target enhanced image. The fused and enhanced image is then subjected to target detection, and the three detection results are weighted averaged to obtain the final result. The present invention can adapt to the complex and changeable urban environment, weather and lighting, and enhances the robustness in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 The present invention is a flow chart of the dual-band infrared image decision fusion detection method of the monocular Cassegrain structure.

[0064] Figure 2 This is the image of the small target in the medium wave camera of the present invention and its gray space distribution diagram, where: Figure 2 a1 in the figure is the original image taken by a medium wave camera; Figure 2 a2 in the figure is the grayscale space distribution diagram corresponding to the original image taken by the medium wave camera; Figure 2 The red boxes in a1 and a2 in the figure represent the enlarged weak UAV targets.

[0065] Figure 3 This is a diagram of a small target in the image of a Cassegrain structure long-wave camera and its grayscale spatial distribution, where: Figure 3 a1 in the figure is the original image taken by the Cassegrain structure long-wave camera; Figure 3 a2 in the figure is the grayscale space distribution diagram corresponding to the original image taken by the Cassegrain structure long-wave camera; Figure 3 The red boxes in a1 and a2 in the figure represent the enlarged weak UAV targets.

[0066] Figure 4 This is a real picture taken by a UAV after the medium-wave camera of the present invention and the Cassegrain structure long-wave infrared camera are registered and weighted averaged.

[0067] Figure 5 The figure is a flow chart of the image data processing of the present invention.

[0068] Figure 6 This is a comparison diagram before and after the image data enhancement of the present invention; wherein, Figure 6 a1 in the figure is the original image taken by a medium wave camera; Figure 6 a2 in the Figure 6 A1 enhanced medium-wave infrared image; Figure 6 b1 in the figure is the original image taken by the Cassegrain structure long-wave camera; Figure 6 b2 in the pair Figure 6 The long-wave infrared image after b1 enhancement; among them, Figure 6 The red boxes in a2 and b2 represent the enhanced weak UAV targets.

[0069] Figure 7 This is a structural diagram of the dual-band image decision fusion detection algorithm of the present invention. DETAILED DESCRIPTION

[0070] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0071] Example 1

[0072] The dual-band infrared image decision fusion detection method of the monocular Cassegrain structure of this embodiment refers to Figure 1 , which specifically includes the following:

[0073] Step S1: In conventional camera imaging, when the long-wave infrared image can be seen clearly, the medium-wave infrared image is too small to extract the corner points on the calibration plate, so the traditional infrared binocular calibration method is no longer applicable. Based on this, this embodiment builds a binocular vision platform for medium-wave infrared and Cassegrain structure long-wave infrared, and calibrates the binocular camera on the binocular vision platform. The binocular camera includes a medium-wave camera and a Cassegrain structure long-wave camera. Figure 2 and Figure 3 , which specifically includes the following sub-steps:

[0074] Step S11, building a binocular vision platform for medium-wave infrared and Cassegrain structure long-wave infrared on the gimbal, and imaging the medium-wave optical path and the Cassegrain structure long-wave optical path on the same plane;

[0075] Step S12: Due to the limitation of Cassegrain structure long-wave infrared, only obvious corners in nature can be found. In this embodiment, the four corners of the building are selected at four positions in the image coordinate system. According to Zhang Zhengyou's calibration method, the rotation matrix and translation vector of the left camera (Cassegrain structure long-wave infrared camera) to the right camera (medium-wave infrared camera) coordinate system are calculated, and images are taken at different positions, and the corresponding corner points of the building in the image are matched. Figure 4 The picture shown is a real picture taken by a drone after simple weighted averaging of the overlapping parts after calibration.

[0076] Among them, the Zhang Zhengyou calibration method is used to obtain the internal parameter matrix of the binocular camera:

[0077] ;

[0078] Among them, M is the internal parameter matrix; f is the focal length of the binocular camera; , They represent the width and height of a pixel in physical size, respectively, and define the transformation relationship from the world coordinate system to the camera coordinate system; is the equivalent focal length expressed in pixel width and length on the u-axis in the pixel coordinate system, is the focal length parameter on the v axis in the pixel coordinate system; , is the principal point coordinate, which indicates the position of the optical center of the image in the pixel coordinate system.

[0079] Step S13: Obtain the mapping relationship between the pixel coordinate system and the world coordinate system according to the intrinsic parameter matrix:

[0080] ;

[0081] in, is the homography matrix, which represents the mapping relationship between the pixel coordinate system on the image and the world coordinate system; It is the tilt between the x-axis and the y-axis in the image coordinate system, usually 0; is the rotation matrix, t is the translation vector, is the rigid body transformation matrix; s is the scale factor.

[0082] Step S14: convert the medium wave and Cassegrain structure long wave UAV target images according to the homography matrix:

[0083] ;

[0084] in, is a point on the medium-wave UAV target image, is a point on the Cassegrain structure long-wave UAV target image.

[0085] Step S2: Reference Figure 5 Based on the binocular vision platform, medium-wave and Cassegrain structure long-wave UAV target images are obtained. After data cleaning, data balancing and data segmentation, qualified data are selected from the approximately 100,000 sets of image data obtained by shooting, and morphological operations are performed after Gaussian filtering. The processed and enhanced images are as follows Figure 6 As shown;

[0086] Specifically, this embodiment preprocesses the drone target image, wherein data cleaning is to screen out data that meets the requirements, including deduplication, format and size standardization; data balancing refers to ensuring that the number of images in various scenarios is relatively balanced to avoid model deviation caused by certain categories; data segmentation refers to dividing the data into a training set, a validation set, and a test set.

[0087] Furthermore, Gaussian filtering is performed on the drone target image:

[0088] ;

[0089] in, is a Gaussian function; x and y are pixel coordinates; is the standard deviation of the Gaussian function; is the base of natural logarithms;

[0090] Use the Tophat algorithm to suppress the background part in the drone target image:

[0091] ;

[0092] in, For The Tophat operation at ; I is the grayscale image; b is the morphological structure operator; For the opening operation.

[0093] Step S3: Reference Figure 7 , the C2f_KAN structure is constructed using the KAN algorithm, and the C2f_KAN structure is used to replace the c2f structure of yolov 10n to construct the KAN_yolov10 pre-detection head;

[0094] The pre-detection head of this embodiment is improved on the basis of yolov10. Yolov10 introduces a consistent dual allocation strategy, which combines one-to-many and one-to-one label allocation methods. During the training process, one-to-many allocation is used to provide rich supervision signals, which helps the model (yolov10) learn more features. During the inference process, the model uses one-to-one allocation to avoid the use of NMS, thereby improving the inference speed and efficiency. This dual allocation strategy not only improves the training effect, but also reduces the inference delay, allowing yolov10 to achieve end-to-end deployment without additional inference costs. However, yolov10 sacrifices some accuracy in order to achieve real-time target detection, so the Kolmogorov–Arnold Networks (KAN) algorithm is used in the backbone network to construct a C2f_KAN structure to replace the original c2f structure, which can improve the accuracy and parameter efficiency of the model without increasing too much computational burden. The KAN algorithm is:

[0095] ;

[0096] in, The output result after the KAN algorithm processes the input m; , , and They are the univariate function matrices of the L-1th layer, L-2th layer, 1st layer, and 0th layer networks respectively; It is a function matching operator, which means that the output of one function is used as the input of the next function; is the vector of input variables.

[0097] The KAN algorithm is a new neural network architecture that uses learnable functions to replace traditional fixed activation functions, thereby eliminating the reliance on linear weight matrices. This design makes KAN superior to traditional MLP in terms of accuracy and interpretability, and achieves better performance than MLP with fewer parameters. By introducing the design concept of KAN, the C2f_KAN structure further improves the performance and efficiency of the model while reducing the number of parameters and computation.

[0098] Furthermore, this embodiment introduces an efficient multi-scale attention mechanism (EMA) before the KAN_yolov10 pre-detection head. EMA processes the feature map of the model by weighted averaging to extract more useful feature information. It effectively captures spatial information of different scales through feature grouping, parallel sub-networks and cross-space learning, enhances the model's perception of details, and can enhance the detection success rate of occluded drone targets; a fusion quality evaluation factor is introduced after the KAN_yolov10 pre-detection head;

[0099] Among them, the fusion quality evaluation factor is:

[0100] ;

[0101] in, is the fusion quality evaluation factor, is the index of the pixel coordinates; is a hyperparameter used to balance the importance of localization quality and classification quality; For the positioning quality, the positioning quality (Hybrid LocationQuality, HLQ) uses the scale-invariant Wasserstein distance (Scale-Invariant WassersteinDistance, SIWD); For classification quality, the classification confidence is used, which is the probability that the model predicts that the pixel belongs to the target category.

[0102] Step S4, using a multi-scale feature fusion module to fuse the medium-wave and Cassegrain structure long-wave UAV target images to obtain a fused image of the UAV target;

[0103] Since the Cassegrain structure is a reflective telescope with a very small field of view, a focal length of 300mm, and a field of view that is part of the medium-wave infrared, the relative position relationship of the two cameras can be obtained through the calibration of the binocular camera in step S1. However, what is obtained is not two strictly aligned images, but a roughly aligned position relationship. Therefore, the detection results are spatially associated. As long as one mode can detect the target, the relative position area in the other mode will be obtained, further increasing the robustness of the model. After the target detection network, the detection result, that is, the ROI area to be fused, is cropped and sent to the fusion network. In addition, the long wave of the Cassegrain structure has an amplifying effect, and the edge information is richer relative to the medium wave. Based on this, this step includes the following sub-steps:

[0104] Step S41, using the GRDB module to respectively extract the depth features of the medium-wave UAV target image and the Cassegrain structure long-wave UAV target image in the ROI area;

[0105] ;

[0106] in; is the output of the GRDB module, representing the extracted deep features; is a grouped residual dense block; It is an n-level cascade convolution layer; It is the input of the GRDB module, which represents the input medium-wave UAV target image or Cassegrain structure long-wave UAV target image; is the gradient operator.

[0107] Step S42, using the resnet50 residual network to respectively extract the shallow features of the medium-wave UAV target image and the Cassegrain structure long-wave UAV target image in the ROI area;

[0108] Step S43, firstly fuse the shallow features of the medium-wave UAV target image with the shallow features of the Cassegrain structure long-wave UAV target image, and then fuse the deep features of the medium-wave UAV target image with the deep features of the Cassegrain structure long-wave UAV target image;

[0109] The shallow feature fusion and deep feature fusion of the drone target image in this embodiment can be expressed by the following formula:

[0110] ;

[0111] ;

[0112] ;

[0113] ;

[0114] ;

[0115] ;

[0116] in, is the fused feature; It is the characteristics of medium wave UAV target image; The characteristics of the long-wave UAV target image for the Cassegrain structure; is the maximum IQEF value between the feature point and all ground truth (GT) GT boxes, that is, is the IQEF of the point; Mapping medium waves for predicted IQEF value, Mapping long waves for predicted IQEF value;

[0117] Given the predicted IQEF map wave value and long wave value ; for and The minimum value of for and The maximum value of Obtained for medium wave UAV target image The weight after the maximum and minimum normalization of the graph; The Cassegrain structure long-wave UAV target image is obtained The weights after maximum and minimum normalization of the graph.

[0118] Step S44: Send the fused shallow features and the fused deep features to the convolution layer for fusion to obtain a fused image of the drone target:

[0119] ;

[0120] in, To fuse the images; is a convolution operation of size 1x1; is the shallow feature after fusion; is the deep feature after fusion; It is the concatenation operation of the feature map; the shallow features after fusion and the fused deep features; All are obtained by calculation in step S43.

[0121] This embodiment restores the obtained fused features to the scene, uses deep feature fusion and shallow feature fusion to predict boundary results, performs convolution and upsampling on these features and connects them in the channel dimension, and finally restores the fused and enhanced features to the medium wave scene, which can provide richer background and context information.

[0122] Step S5: Since the traditional centerness measurement method will cause a sharp drop in the centerness score when the predicted box deviates slightly from the center of the reference truth box, in order to overcome the above problem, the Manhattan distance is introduced, that is, the Manhattan distance is introduced in the KAN_yolov10 pre-detection head, and then the fused image is input into the KAN_yolov10 pre-detection head to output the detection result of the drone target;

[0123] The Manhattan distance is introduced in the KAN_yolov10 pre-detection head. The Manhattan distance is used to measure the distance between the candidate position of the drone target to be detected in the image and the center or corner point of the reference truth box. Specifically:

[0124] ;

[0125] in, is the Manhattan distance; is the position of the center of the reference truth box, Represents the Manhattan distance from the sample point to the center of the benchmark truth box, Represents the Manhattan distance from the center of the reference truth box to any corner point, w and h represent the width and height of the reference truth box respectively; sample point coordinates is the candidate location center of the target to be detected; the detection result includes the location coordinates and classification results of the drone.

[0126] Although the specific implementation of the invention is described in detail in conjunction with the drawings, it should not be understood as limiting the scope of protection of this patent. Within the scope described in the claims, various modifications and variations that can be made by those skilled in the art without creative work still fall within the scope of protection of this patent.

Claims

1. A dual-band infrared image decision fusion detection method with a monocular Cassegrain structure, characterized in that: The following steps are involved: S1. Build a binocular vision platform for medium-wave infrared and Cassegrain structure long-wave infrared, and calibrate the binocular camera on the binocular vision platform; S2. Based on the binocular vision platform, obtain the medium-wave and Cassegrain structure long-wave UAV target images, and pre-process the UAV target images; S3, construct KAN_yolov10 pre-detection head based on KAN algorithm; S4, using a multi-scale feature fusion module to fuse the medium-wave and Cassegrain structure long-wave UAV target images to obtain a fused image of the UAV target; S5, introduce Manhattan distance into KAN_yolov10 pre-detection head, input the fused image into KAN_yolov10 pre-detection head, and output the detection result of UAV target; The S4 comprises the following sub-steps: S41, using the GRDB module to extract the depth features of the medium-wave UAV target image and the Cassegrain structure long-wave UAV target image in the ROI area; S42, using resnet50 residual network to extract shallow features of medium-wave UAV target images and Cassegrain structure long-wave UAV target images in the ROI area; S43, firstly fusing the shallow features of the medium-wave UAV target image with the shallow features of the Cassegrain structure long-wave UAV target image, and then fusing the deep features of the medium-wave UAV target image with the deep features of the Cassegrain structure long-wave UAV target image; S44, send the fused shallow features and the fused deep features into the convolution layer for fusion to obtain the fused image of the drone target: ； in, To fuse the images; is a convolution operation of size 1x1; is the shallow feature after fusion; is the deep feature after fusion; It is the concatenation operation of feature maps.

2. The dual-band infrared image decision fusion detection method of the monocular Cassegrain structure according to claim 1 is characterized in that: The S1 comprises the following sub-steps: S11. Build a binocular vision platform for medium-wave infrared and Cassegrain structure long-wave infrared on the gimbal, and image the medium-wave optical path and the Cassegrain structure long-wave optical path on the same plane; S12. Use Zhang Zhengyou calibration method to obtain the internal parameter matrix of the binocular camera: ； Among them, M is the internal parameter matrix; f is the focal length of the binocular camera; , They represent the width and height of a pixel in physical size respectively; is the equivalent focal length expressed in pixel width and length on the u-axis in the pixel coordinate system, is the focal length parameter on the v axis in the pixel coordinate system; , is the principal point coordinate, which indicates the position of the optical center of the image in the pixel coordinate system; S13. Obtain the mapping relationship between the pixel coordinate system and the world coordinate system according to the intrinsic parameter matrix: ； in, is the homography matrix, which represents the mapping relationship between the pixel coordinate system on the image and the world coordinate system; is the degree of inclination between the x-axis and the y-axis in the image coordinate system; is the rotation matrix, t is the translation vector, is the rigid body transformation matrix; s is the scale factor; S14. Convert the medium wave and Cassegrain structure long wave UAV target images according to the homography matrix: ； in, is a point on the medium wave UAV target image, is a point on the Cassegrain structure long-wave UAV target image.

3. The dual-band infrared image decision fusion detection method of the monocular Cassegrain structure according to claim 1 is characterized in that: The S2 pre-processes the drone target image, including: Perform Gaussian filtering on the drone target image: ； in, is a Gaussian function; x and y are pixel coordinates; is the standard deviation of the Gaussian function; is the base of natural logarithms; Use the Tophat algorithm to suppress the background part in the drone target image: ； in, is the Tophat operation; I is the grayscale image; b is the morphological structure operator; For the opening operation.

4. The dual-band infrared image decision fusion detection method of the monocular Cassegrain structure according to claim 1 is characterized in that: The S3 specifically includes: The KAN algorithm is used to construct the C2f_KAN structure, and the C2f_KAN structure is used to replace the c2f structure of yolov 10n to construct the KAN_yolov10 pre-detection head.

5. The dual-band infrared image decision fusion detection method of the monocular Cassegrain structure according to claim 4 is characterized in that: The KAN algorithm is: ； in, is the output result after the input m is processed by the KAN algorithm; , , and They are the univariate function matrices of the L-1th layer, L-2th layer, 1st layer, and 0th layer networks respectively; It is a function matching operator, which means that the output of one function is used as the input of the next function; is the vector of input variables.

6. The dual-band infrared image decision fusion detection method of the monocular Cassegrain structure according to claim 1 is characterized by: An efficient multi-scale attention mechanism is introduced before the KAN_yolov10 pre-detection head, and a fusion quality evaluation factor is introduced after the KAN_yolov10 pre-detection head; The fusion quality evaluation factor is: ； in, is the fusion quality evaluation factor, is the index of the pixel coordinates; is a hyperparameter used to balance the importance of localization quality and classification quality; To locate the quality; For classification quality.

7. The dual-band infrared image decision fusion detection method of the monocular Cassegrain structure according to claim 1 is characterized by: In S41, the GRDB module is used to extract deep features, specifically: ； in; is the output of the GRDB module, representing the extracted deep features; is a grouped residual dense block; It is an n-level cascade convolution layer; It is the input of the GRDB module, which represents the input medium-wave UAV target image or Cassegrain structure long-wave UAV target image; is the gradient operator.

8. The dual-band infrared image decision fusion detection method of the monocular Cassegrain structure according to claim 1 is characterized by: The shallow feature fusion and deep feature fusion of the drone target image in S43 are specifically: ； ； ； in, is the fused feature; It is the characteristics of medium wave UAV target image; The characteristics of the long-wave UAV target image for the Cassegrain structure; Obtained for medium wave UAV target image The weights after the maximum and minimum normalization of the graph; The Cassegrain structure long-wave UAV target image is obtained The weights after maximum and minimum normalization of the graph.

9. The dual-band infrared image decision fusion detection method of the monocular Cassegrain structure according to claim 1 is characterized by: In S5, the Manhattan distance is introduced in the KAN_yolov10 pre-detection head, and the Manhattan distance is used to measure the distance between the candidate position of the drone target to be detected in the fused image and the center or corner point of the reference truth frame, specifically: ； in, is the Manhattan distance; is the position of the center of the reference truth box, Represents the Manhattan distance from the sample point to the center of the benchmark truth box, represents the Manhattan distance from the center of the reference truth box to any corner point, w and h represent the width and height of the reference truth box respectively; is the candidate position center of the target to be detected.

Images