Unmanned aerial vehicle target detection system and method based on multi-scale feature fusion

Abstract

The application relates to the technical field of image processing, in particular to a UAV target detection system and method based on multi-scale feature fusion. In the system, average pooling downsampling operation and global multi-scale attention enhancement operation are introduced on the basis of the YOLOv11 architecture, the acquisition of detail features is further optimized, and the positioning ability of the system for key feature information is improved. A feature fusion sub-module is introduced, the sensitivity of the system to multi-scale UAV targets is further improved, and the detection precision is improved. An aggregated non-maximum suppression module is constructed after the head network, so that the system can be more adapted to complex and dense scenes. In the method, a loss function combining a bounding box regression loss and a Gaussian distribution distance loss is used to train the target detection system, the shape and scale of the bounding box are comprehensively considered, and the similarity of the real frame and the predicted frame distribution is further measured, so that the respective advantages of the two are fully utilized, and the recognition accuracy of the UAV target is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of image processing, and particularly relates to a UAV target detection system and method based on multi-scale feature fusion. BACKGROUND

[0002] In the field of computer vision, the anti-UAV detection technology based on a target detection model is one of the widely applied and popular research directions, and is specifically applied to an intelligent processing core component module in an anti-UAV system. The anti-UAV detection task based on a target detection model is mainly to accurately detect and identify UAV target instances in image, video and video stream data through computer vision related technologies and to perform positioning processing thereon. Essentially, the task belongs to a classification and regression process. In actual life, the UAV target detection and identification technology plays an important role in air traffic management, public safety guarantee in airports, enterprise park security and the like and has high application value.

[0003] The UAV target detection algorithm based on deep learning can be divided into a single-stage detection algorithm and a two-stage detection algorithm according to a detection idea. Among them, the single-stage detection algorithm is mainly represented by a YOLO series network model, and the two-stage detection algorithm is represented by an R-CNN series network model. However, most of the classic target detection algorithms lack specific consideration of the detection scene, especially the complex and variable environment scene. The robustness of these mainstream detection algorithms is affected, mainly including the following aspects:

[0004] (1) In a complex low-altitude scene, the contrast and discrimination degree of the UAV target and the environmental background information are low. In this case, the detection model has weak discrimination ability for the UAV target and the background, and thus is prone to cause target mis-detection phenomenon;

[0005] (2) In a complex low-altitude scene, the scale change of the UAV target is large, causing large difference in sensitivity of the feature extraction network to different scale targets, and thus easily causing target missing detection or mis-detection phenomenon. Especially for small scale weak targets, this phenomenon is more obvious;

[0006] (3) In a complex low-altitude scene, the UAV target is severely occluded, which affects the detection effect of the detection algorithm. Through analysis of images in an open source UAV target data set, it can be known that the UAV target occlusion problem mainly includes intra-class occlusion and inter-class occlusion. Among them, the intra-class occlusion refers to mutual occlusion and overlap of UAV targets, and the inter-class occlusion refers to occlusion of the UAV by other targets or occlusion of the UAV by background information. The intra-class occlusion and the inter-class occlusion mainly affect the feature extraction link, and thus cause the information required for reasoning of the algorithm detection module to also decrease, affecting the detection and identification accuracy. SUMMARY

[0007] Therefore, the present application aims to provide a UAV target detection system and method based on multi-scale feature fusion, which introduces an average pooling downsampling module to replace the traditional convolution structure in the backbone network, further optimizes the acquisition of detailed features, constructs a global multi-scale attention enhancement module in the neck network, designs the attention enhancement module from the perspectives of channel attention and spatial attention, improves the positioning ability of the system to key feature information, introduces an adaptive multi-scale feature fusion strategy in the detection head, further improves the sensitivity of the system to multi-scale UAV targets, and thus improves the detection accuracy.

[0008] To achieve the above object, the technical scheme of the present application is as follows:

[0009] A UAV target detection system based on multi-scale feature fusion for identifying UAVs in complex aerial scenes, comprising: a backbone network obtained by replacing the convolution operation in the original backbone network of the YOLOv11 network architecture with an average pooling downsampling operation; a neck network obtained by introducing an average pooling downsampling operation and a global multi-scale attention enhancement operation into the original neck network of the YOLOv11 network architecture; a head network comprising multiple original detection heads of the YOLOv11 network architecture and a feature fusion sub-module matched with each original detection head, each feature fusion sub-module fusing multi-scale features from the neck network, the fused features being input into the corresponding detection head to obtain an initial target detection box; an aggregated non-maximum suppression module selecting a reserved detection box with the highest confidence score from the initial target detection boxes output by the head network and eliminating redundant detection boxes with a high overlap ratio with the reserved detection box; generating a judgment score for measuring the anti-elimination ability of the reserved detection box according to the aggregation characteristics of the redundant detection boxes, and then taking the judgment score as the basis for non-maximum suppression to retain correct detection results with severe occlusion and overlap, thereby obtaining the final target detection box.

[0010] Further, in the average pooling downsampling operation: the input features are subjected to average pooling, the pooled features are subjected to channel splitting to obtain two parts of features; one part of the features is subjected to convolution operation, and the other part of the features is subjected to max pooling and convolution operation; the processed two parts of features are spliced to obtain output features.

[0011] Further, in the global multi-scale attention enhancement operation: the input features are grouped for processing; for each sub-feature map, average pooling and maximum pooling operations are respectively performed along the height and width directions of the sub-feature map, and the obtained four pooled features are nonlinearly mapped; the two pooled features in the height direction are element-wise added, and the two pooled features in the width direction are element-wise added; sigmoid activation operations are respectively performed on the two added features to obtain the weights of the sub-feature map in the height and width directions; the two weights are multiplied with the sub-feature map to obtain an enhanced sub-feature map; the enhanced sub-feature maps are combined to obtain the output features.

[0012] Further, in the feature fusion sub-module: the multi-scale features from the neck network are fused by the following formula:

[0013] ;

[0014] wherein, , and represent the feature vectors at the coordinates (i, j) in the three feature maps of the l-th level, represents the corresponding fusion result, , and represent the corresponding fusion weights, and the sum of the three is 1.

[0015] Further, the three fusion weights , and are obtained by the following formula:

[0016] ;

[0017] ;

[0018] ;

[0019] wherein, , and represent the vector values of the feature vectors at the coordinates (i, j) in the three feature maps of the l-th level.

[0020] Further, in the aggregation non-maximum suppression module: an overlap score between any two initial target detection boxes is calculated; if the overlap score is greater than a preset overlap threshold, the two initial target detection boxes related to the overlap score are defined as an aggregation detection box; the above operation is repeated to calculate the overlap score of each initial target detection box and obtain all aggregation detection boxes related to each initial target detection box, to obtain an aggregation detection box set of each initial target detection box, and define the detection box with the maximum confidence in each aggregation detection box set as a reserved detection box, to construct a reserved detection box set; in each aggregation detection box set, the overlap scores of each aggregation detection box are sorted in descending order, and the serial number of each aggregation detection box is used to perform interval mapping on the corresponding overlap score, to obtain a discrimination suppression score of each aggregation detection box; the discrimination suppression score and the reserved detection box are used to perform non-maximum suppression on each initial target detection box, until all reserved detection boxes in the reserved detection box set are traversed, to obtain a final target detection box.

[0021] Further, the overlap score is:

[0022] ;

[0023] wherein, GScore ij represents the overlap score between the ithinitial target detection box Box i and the jthinitial target detection box Box j , IoG(Box i , Box j ) represents the overlap ratio between the ithinitial target detection box Box i and the jthinitial target detection box Box j , and represent the coordinate values of the upper left corner and the lower right corner of the ithinitial target detection box Box i , and AMap represents a mapping relationship function, which is:

[0024] ;

[0025] wherein, λ represents the input of the mapping relationship, σ t represents a preset parameter threshold.

[0026] Further, the acquisition process of the discrimination inhibition score comprises: defining each aggregation bounding box set, retaining the bounding box outside the bounding box as a redundant bounding box; in each aggregation bounding box set, the overlap score of the retained bounding box and all redundant bounding boxes is sorted in descending order to obtain the serial number coefficient of each redundant bounding box; the overlap score is divided by the serial number coefficient of the corresponding redundant bounding box, and the sum of all divided quotients is obtained; the total value of the quotient is interval mapped by using a mapping function to obtain the corresponding discrimination inhibition score:

[0027] ;

[0028] Wherein, BMap represents a mapping relationship function, λ represents an input of the mapping relationship, and ε t represents a preset parameter threshold.

[0029] A UAV target detection method based on multi-scale feature fusion comprises:

[0030] S1: Obtain a low-altitude image dataset, and pre-process the low-altitude image dataset to obtain a training set;

[0031] S2: Construct a UAV target detection system based on multi-scale feature fusion as provided by the present application;

[0032] S3: Train the target detection system constructed in step S2 using the training set obtained in step S1 to obtain a target detection model;

[0033] S4: Input an image to be detected into the target detection model obtained in step S3 to obtain a detection result.

[0034] Further, the target detection system is trained in step S3 by using a loss function as follows:

[0035] L = η1 x L1 + η2 x L2

[0036] Wherein, L represents a loss function for training the target detection system, η1 and η2 represent loss weights, and the sum of the two is 1, L1 represents a bounding box regression loss function, and L2 represents a Gaussian distribution distance loss function;

[0037] The bounding box regression loss function is:

[0038] L1 = 1 - IoU + α x factor distance + β x factor shape ;

[0039] Wherein, IoU represents an intersection over union value between a predicted bounding box and a real labeled box, α and β represent factor weights, factor distance represents a distance correction factor, and factor shapedenotes a shape correction factor;

[0040] distance correction factor factor distance is:

[0041] ;

[0042] wherein (x Pre ,y Pre ) denotes the center coordinates of the predicted detection frame output by the target detection model, (x GT ,y GT ) denotes the center coordinates of the real labeled frame, d denotes the diagonal line length of the minimum circumscribed frame of the predicted detection frame and the real labeled frame, and ρ h and ρ w respectively denote the weight coefficients in the vertical direction and the horizontal direction;

[0043] The weight coefficient ρ h in the vertical direction is:

[0044] ;

[0045] wherein h GT denotes the height of the real labeled frame, w GT denotes the width of the real labeled frame, and Scale denotes a scale factor;

[0046] The weight coefficient ρ w in the horizontal direction is:

[0047] ;

[0048] The shape correction factor factor shape is:

[0049] ;

[0050] wherein h Pre denotes the height of the predicted detection frame, and w Pre denotes the width of the predicted detection frame;

[0051] The Gaussian distribution distance loss function L2 is:

[0052] ;

[0053] wherein ζ denotes a constant value, D denotes the Gaussian distribution distance between the predicted detection frame and the real labeled frame, and is:

[0054] .

[0055] Compared with the prior art, the present application can achieve the following beneficial effects:

[0056] (1) In the unmanned aerial vehicle target detection system based on multi-scale feature fusion, an average pooling downsampling module is introduced in the backbone network to replace the traditional convolution structure, thereby avoiding the problem of increased resource consumption caused by large convolution kernel calculation, and further optimizing the acquisition of detail features. Secondly, a global multi-scale attention enhancement module is constructed in the neck network, which is designed from the perspectives of channel attention and spatial attention, aiming to improve the positioning ability of the network to key feature information, and thus more help to improve the accuracy of unmanned aerial vehicle target recognition. Then, from the perspective of data-driven, an adaptive enhanced multi-scale feature fusion strategy is introduced, aiming to further improve the sensitivity of the system to multi-scale unmanned aerial vehicle targets, thereby improving the detection precision;

[0057] (2) In the unmanned aerial vehicle target detection system based on multi-scale feature fusion, an aggregated non-maximum suppression module (Gather-NMS) is constructed after the head network. First, the detection box with the highest confidence score in the set of overlapping detection boxes is selected, and the detection boxes with too high overlap ratio with the box are removed. Then, a judgment score is generated for the detection box with the highest confidence score in the aggregated detection box set according to the aggregation characteristics of the redundant detection boxes, which measures the anti-removal ability of the detection box. Then, the judgment score is used as the basis for non-maximum suppression, so as to retain as many correct detection results as possible that are occluded and severely overlapped. Compared with the traditional non-maximum suppression algorithm, the Gather-NMS algorithm constructed by the present application can better adapt to complex and dense scenes, and is particularly helpful for improving the detection accuracy of small-scale targets.

[0058] (3) In the unmanned aerial vehicle target detection method based on multi-scale feature fusion, the loss function part is optimized and improved. The shape and size of the bounding box are considered, and the similarity of the real box and the predicted box distribution is further measured. The two are weighted and fused to fully utilize their respective advantages and improve the recognition accuracy of unmanned aerial vehicle targets. BRIEF DESCRIPTION OF DRAWINGS

[0059] The accompanying drawings, which form a part of this application, are included to provide a further understanding of the application and are incorporated in and constitute a part of this application. The embodiments of the present application and their description are used to explain the present application and do not constitute an improper limitation of the present application. In the drawings:

[0060] Figure 1 A schematic diagram of the unmanned aerial vehicle target detection system based on multi-scale feature fusion according to the embodiments of the present application;

[0061] Figure 2 A schematic diagram of the average pooling downsampling operation according to the embodiments of the present application;

[0062] Figure 3 Schematic diagram of a global multi-scale attention enhancement operation described in an embodiment of the present invention;

[0063] Figure 4 Flowchart of a UAV target detection method based on multi-scale feature fusion described in an embodiment of the present invention. DETAILED DESCRIPTION

[0064] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, and do not constitute a limitation on the present invention. It should be noted that the embodiments in the present invention and the features in the embodiments can be combined with each other without conflict.

[0065] The present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0066] As shown in Figure 1 , the UAV target detection system based on multi-scale feature fusion described in an embodiment of the present invention is used to identify UAVs in complex aerial scenes, and includes a backbone network, a neck network, a head network and an aggregated non-maximum suppression module. The backbone network is obtained by replacing the convolution operation in the original backbone network of the YOLOv11 network architecture with an average pooling down-sampling operation; the neck network is obtained by introducing an average pooling down-sampling operation and a global multi-scale attention enhancement operation into the original neck network of the YOLOv11 network architecture; the head network includes multiple original detection heads of the YOLOv11 network architecture, and a feature fusion sub-module matched with each original detection head, each feature fusion sub-module fuses multi-scale features from the neck network, and the fused features are input into the corresponding detection head to obtain an initial target detection box; the aggregated non-maximum suppression module selects a reserved detection box with the highest confidence score from the initial target detection boxes output by the head network, and eliminates redundant detection boxes with too high overlap ratio with the reserved detection box; according to the aggregated characteristics of the redundant detection boxes, a judgment score measuring the anti-elimination ability of the reserved detection box is generated, and then the judgment score is used as the basis for non-maximum suppression to retain correct detection results with severe occlusion and overlap, thereby obtaining the final target detection box.

[0067] In the backbone network provided in the embodiment of the present application, the input unmanned aerial vehicle image of a low-altitude complex background is convolved twice in succession, the processed features are extracted by the first C3k2 module with the class False, and then the average pooling downsampling is used to further optimize the acquisition of the detail features; the output features of the average pooling downsampling are processed in turn by the second C3k2 module with the class False and the average pooling downsampling; the processed features are processed in turn by the third C3k2 module with the class True and the average pooling downsampling; the processed features are processed in turn by the fourth C3k2 module with the class True, the SPFF module and the C2PSA module, and at this time, the output features of the second C3k2 module, the third C3k2 module and the C2PSA module are the multi-scale features output by the backbone network. In the neck network provided in the embodiment of the present application, the output features of the C2PSA module are upsampled, and then are concatenated with the output features of the third C3k2 module; the concatenated features are processed by the fifth C3k2 module with the class False and the upsampling operation, and then are concatenated with the output features of the second C3k2 module; the concatenated features are processed by the sixth C3k2 module with the class False and the average pooling downsampling operation, and then are concatenated with the output features of the fifth C3k2 module, and the concatenated features are input into the seventh C3k2 module with the class False for processing, and then the output features are processed in turn by the global multi-scale attention enhancement operation and the average pooling downsampling operation; the processed features are concatenated with the output features of the C2PSA module, and then are input into the eighth C3k2 module with the class True for feature extraction, and then the extracted features are processed by the global multi-scale attention enhancement operation, and at this time, the output features of the sixth C3k2 module, the seventh C3k2 module and the last global multi-scale attention enhancement operation are the multi-scale features output by the neck network. In the head network provided in the embodiment of the present application, three detection heads and three feature fusion sub-modules corresponding to the detection heads are included, the sixth C3k2 module is input into the respective corresponding feature fusion sub-modules, the output features after the feature fusion performed by the feature fusion sub-modules are input into the respective corresponding detection heads, and initial target detection boxes are output. After all the initial target detection boxes are processed by the Gather-NMS module, the detection result of the unmanned aerial vehicle target in the image to be detected is output.

[0068] In some embodiments, the processing process of the average pooling downsampling operation includes: performing average pooling on the input features, performing channel splitting on the pooled features to obtain two parts of features; performing convolution operation on one part of the features, and performing maximum pooling and convolution operation on the other part of the features; and concatenating the processed two parts of features to obtain output features.

[0069] The processing process of the average pooling downsampling operation provided in the embodiment of the present application is as followsFigure 2 As shown, the input features are subjected to average pooling operation to obtain overall features and reduce the amount of calculation; the features after average pooling are split into two parts along the channel dimension, wherein the first part is subjected to maximum pooling operation to highlight the key features of the target, and then a 1x1 convolution kernel is used to further enhance the expression ability of the features; the second part uses a 3x3 convolution kernel to enhance the learning and acquisition of detailed features. The outputs of the two processed parts are spliced along the channel dimension to obtain the output features.

[0070] In some embodiments, the processing process of the global multi-scale attention enhancement operation includes: grouping processing on the input features; for each sub-feature map, respectively performing average pooling and maximum pooling operations along the height and width directions of the sub-feature map, and performing nonlinear mapping on the obtained four pooled features; element-wise adding the two pooled features in the height direction and element-wise adding the two pooled features in the width direction; respectively performing sigmoid activation operation on the two added features to obtain the weights of the sub-feature map in the height and width directions; multiplying the two weights with the sub-feature map to obtain the enhanced sub-feature map; combining the enhanced sub-feature maps to obtain the output features. The present application improves the detection sensitivity of the system to different scale unmanned aerial vehicle targets by performing multi-scale feature fusion on the features from the backbone network, and improves the detection and recognition accuracy; and the global multi-scale attention enhancement operation is designed from the aspects of channel attention and spatial attention to improve the positioning ability of the network to key feature information, thereby further improving the accuracy of unmanned aerial vehicle target recognition.

[0071] The global multi-scale attention enhancement operation provided by the embodiment of the present application includes Figure 2 as shown, comprising:

[0072] performing nonlinear mapping on the pooled features to obtain the first part and the second part of the sub-feature map. dividing the input features into N groups along the channel to obtain a plurality of sub-feature maps , wherein F M represents the Mth sub-feature map, M=1, 2, 3,..., N, H, W and C represent the height, width and channel number of the feature respectively;

[0073] for each sub-feature map F M , respectively performing average pooling and maximum pooling operations along the height and width directions of the sub-feature map as follows:

[0074] ;

[0075] ;

[0076] ;

[0077] ;

[0078] in, and Representing the sub-feature maps F respectively M The features obtained after performing average pooling and max pooling operations along the height direction, respectively. and Representing the sub-feature maps F respectively M The features obtained by performing average pooling and max pooling operations along the width direction, F M (H,i) represents a feature map with width i at height H, F M (W,j) represents a feature map with a width of W and a height of j;

[0079] The four pooling features obtained , , and Perform nonlinear mapping, specifically including mapping the four pooling features respectively. , , and Channel compression is performed using a 1×1 convolutional kernel, followed by normalization and ReLU nonlinear activation. The number of channels is then restored using a 1×1 convolutional kernel to complete the pooling of four features. , , and Nonlinear mapping;

[0080] The attention weights g in the height direction are obtained by element-wise summing of two pooling features in the same direction and normalizing them using the Sigmoid function. H Attention weights g in the width direction W ;

[0081] The two weights g H and g W With sub-feature map F M After multiplying the following formulas, we obtain the enhancer feature map. :

[0082] ;

[0083] Strengthening sub-feature maps Following the inverse operation of the initial channel partitioning, the enhanced sub-feature map is processed. By combining the features, we obtain the output feature F. out .

[0084] When the feature map fused with the global multi-scale attention enhancement module is obtained, considering the target detection algorithm based on the feature pyramid, if the features of different levels are directly fused, part of important feature information will be lost. In view of this, the application considers from the data-driven point of view, introduces a feature fusion sub-module based on an adaptive enhanced multi-scale feature fusion strategy in the head network of the traditional YOLOv11 architecture, adjusts the multi-level features from the neck network to the same resolution as the current level features, and adaptively adjusts the weight information through training, so as to obtain the best fusion result. Specifically, in some embodiments, the processing process of the feature fusion sub-module includes: fusing the multi-scale features from the neck network by the following formula:

[0085] ;

[0086] wherein, 、 and represent the feature vectors at the coordinates (i, j) in the three feature maps of the lth level, represent the corresponding fusion result, 、 and represent the corresponding fusion weights, and the sum of the three is 1. The fusion weights 、 and are obtained by the following formula:

[0087] ;

[0088] ;

[0089] ;

[0090] wherein, 、 and represent the vector values of the feature vectors at the coordinates (i, j) in the three feature maps of the lth level. The fusion weights 、 and can be trained and learned by propagating along the error direction, so as to obtain the optimal value and realize the optimal feature fusion.

[0091] In some embodiments, the processing process of the aggregation non-maximum suppression module includes:

[0092] (1) Calculate the overlap score between any two initial target detection boxes. It can be known through analyzing the preliminary detection results output by the detection head that for a certain real unmanned aerial vehicle target, the system model will generate multiple highly overlapping aggregated detection boxes for prediction; therefore, it can be determined that for a detection region, if multiple overlapping detection boxes are generated in the region, there is a real target in the region. In the aggregation non-maximum suppression module (Gather-NMS) provided in the application, first, the overlap score between any two initial target detection boxes is calculated, which not only considers the overlap ratio between the detection boxes, but also considers and designs the proximity of each edge of the detection box. The specific design idea is that if two initial target detection boxes have more than one edge that is very similar and close, and their overlap ratio is also relatively large, it can be determined that they correspond to the detection result of the same unmanned aerial vehicle target. Specifically, in some embodiments, the overlap score between the ith initial target detection box Box i and the jth initial target detection box Box j is calculated by the following formula:

[0093] ;

[0094] wherein GScore ij represents the overlap score between the ith initial target detection box Box i and the jth initial target detection box Box j , IoG(Box i , Box j ) represents the overlap ratio between the ith initial target detection box Box i and the jth initial target detection box Box j , and represent the coordinate values of the upper left corner and the lower right corner of the ith initial target detection box Box i , k=1 represents the upper left corner of the initial target detection box, k=1 represents the lower right corner of the initial target detection box, and AMap represents a mapping relationship function, which is:

[0095] ;

[0096] wherein λ represents the input of the mapping relationship, and σ t represents a preset parameter threshold value, which is specifically determined according to the data set;

[0097] (2) If the overlap score is greater than a preset overlap threshold value, define two initial target detection boxes related to the overlap score as aggregated detection boxes, as follows:

[0098] ;

[0099] wherein, denotes the set of aggregated bounding boxes related to the ith initial target bounding box Box i φ denotes the overlap threshold, T box denotes the set of all initial target bounding boxes;

[0100] (3) Repeat the operations of (1)~(2) to calculate the overlap scores of each initial target bounding box and obtain all aggregated bounding boxes related to each initial target bounding box, to obtain the set of aggregated bounding boxes of each initial target bounding box, and define the bounding box with the highest confidence in each set of aggregated bounding boxes as the reserved bounding box, and construct the set of reserved bounding boxes, as follows:

[0101] ;

[0102] wherein, Save box denotes the set of reserved bounding boxes, score i and score j denote the confidence scores of the ith and jth initial target bounding boxes Box i ;

[0103] (4) In each set of aggregated bounding boxes, sort the overlap scores of each aggregated bounding box in descending order, and use the serial number of each aggregated bounding box to map the corresponding overlap score to an interval to obtain the discrimination suppression score of each aggregated bounding box. In some embodiments, the process of obtaining the discrimination suppression score includes:

[0104] Define the bounding boxes other than the reserved bounding boxes in each set of aggregated bounding boxes as redundant bounding boxes, and the role of the redundant bounding boxes is to measure the anti-rejection ability of the reserved bounding boxes, and to remove these redundant boxes in the set of detection results to reduce their interference with the final detection results;

[0105] In each set of aggregated bounding boxes, sort the overlap scores of the reserved bounding boxes and all redundant bounding boxes from high to low in descending order to obtain the serial number coefficient of each redundant bounding box;

[0106] Divide the overlap score by the serial number coefficient of the corresponding redundant bounding box, and then sum all the quotients, and use the mapping function of the following formula to map the total value of the quotients to an interval to obtain the corresponding discrimination suppression score, as follows:

[0107] ;

[0108] wherein, JScore iDei represents the discriminative suppression score of the i-th redundant detection frame, DeOrder represents a function of the corresponding serial number coefficient after the overlap score is sorted in descending order, BMap represents a mapping relationship function, and specifically:

[0109] ;

[0110] Wherein, λ represents the input of the mapping relationship, and ε t represents a preset parameter threshold, which is specifically set to 0.05 in the embodiment of the application.

[0111] (5) Using the discriminative suppression score and the retained detection frame, non-maximum suppression is performed on each initial target detection frame until all retained detection frames in the retained detection frame set are traversed, and the final target detection frame is obtained. This process specifically includes:

[0112] When the final discriminative suppression score is higher than or equal to the preset threshold of non-maximum suppression, the redundant detection frame is removed; when the final discriminative suppression score is lower than the preset threshold of non-maximum suppression, the redundant detection frame is retained and sent to the retained detection frame set of the next round of non-maximum suppression; until all rounds of non-maximum suppression are completed, that is, the retained detection frame set is empty, the post-processing process of the algorithm ends, and the result obtained is the final target detection result.

[0113] The application also provides a UAV target detection method based on multi-scale feature fusion, which combines Figure 1 and Figure 4 , and includes:

[0114] S1: Obtain a low-altitude image dataset, pre-process the low-altitude image dataset, and obtain a training set. In the embodiment of the application, the low-altitude image dataset of Anti-UAV and Anti-UAV410 is selected, and the UAV targets in the images are relabeled to obtain the training set.

[0115] S2: Construct a UAV target detection system based on multi-scale feature fusion as provided by the application.

[0116] S3: Train the target detection system constructed in step S2 using the training set obtained in step S1 to obtain a target detection model.

[0117] Since the scale of the UAV target is not uniform in a complex low-altitude scene, the boundary box regression accuracy is affected. To this end, the application comprehensively considers the shape and scale of the boundary box and further measures the similarity of the distribution of the real frame and the predicted frame, and proposes an improved loss function. Specifically, in some embodiments, the target detection system is trained by the loss function as follows:

[0118] L = η1 × L1 + η2 × L2;

[0119] wherein, L represents a loss function of training the target detection system, η1 and η2 represent loss weights, the sum of which is 1, and both are learned and updated to the optimal value through the training of the model, L1 represents a bounding box regression loss function, and L2 represents a Gaussian distribution distance loss function;

[0120] The bounding box regression loss function is:

[0121] L1=1-IoU+α×factor distance +β×factor shape ;

[0122] wherein, IoU represents the intersection over union value between the predicted detection box and the real labeled box, that is, , Box Pre represents the predicted detection box, Box GT represents the real labeled box, and α and β represent factor weights, factor distance represents a distance correction factor, and factor shape represents a shape correction factor;

[0123] The distance correction factor factor distance is:

[0124] ;

[0125] wherein, (x Pre ,y Pre ) represents the center coordinates of the predicted detection box output by the target detection model, (x GT ,y GT ) represents the center coordinates of the real labeled box, d represents the diagonal length of the minimum circumscribed box of the predicted detection box and the real labeled box, ρ h and ρ w respectively represent the weight coefficients in the vertical direction and the horizontal direction, and depend on the shape of the real box. It should be noted that the minimum circumscribed box here is the minimum circumscribed box of the union of the predicted detection box and the real labeled box;

[0126] The weight coefficient ρ h in the vertical direction is:

[0127] ;

[0128] wherein, h GT represents the height of the real labeled box, w GT represents the width of the real labeled box, and Scale represents a scale factor, which can be specifically set according to the scale of the target in the data set;

[0129] The weight coefficient ρw is:

[0130] ;

[0131] shape correction factor factor shape is:

[0132] ;

[0133] wherein h Pre represents the height of the predicted detection frame, w Pre represents the width of the predicted detection frame;

[0134] The present application reconstructs the regression detection frame into a two-dimensional Gaussian distribution function, and calculates the similarity between the predicted frame and the real frame, thereby strengthening the detection and recognition stability of small-scale unmanned aerial vehicle targets. Specifically, the Gaussian distribution distance loss function L2 is:

[0135] ;

[0136] wherein ζ represents a constant value, the specific value depends on the data set, D represents the Gaussian distribution distance between the predicted detection frame and the real predicted detection frame, and is:

[0137] .

[0138] In the training process of the embodiment of the present application, the random gradient descent optimizer is used for training for 50 epochs, the momentum factor momentum=0.937, the learning rate adopts the warm up strategy, the initial learning rate is set to a lower value, which is set to 1.25x10 -4 In the first iteration to the 800th iteration of each epoch, the learning rate is linearly increased to 1.25x10 -3 Then it remains unchanged. To the 35th epoch, the learning rate is reduced to 10% of the original, and to the 40th epoch, the learning rate becomes 1% of the original.

[0139] S4: input the low-altitude image of the target to be detected into the target detection model obtained in step S3 to obtain a detection result.

[0140] It should be understood that the various forms of the flow shown above can be used to reorder, add or delete steps. For example, the steps described in the present disclosure can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solutions of the present disclosure can be achieved, which is not limited herein.

[0141] The above detailed description does not limit the scope of the application. Various modifications, combinations, sub-combinations and alternatives can be made to the detailed embodiment within the scope of the application. Any modification, equivalent replacement and improvement made without departing from the spirit and principle of the application shall fall within the scope of the application.

Claims

1. A UAV target detection method based on multi-scale feature fusion, characterized in that, Comprise: S1: obtaining a low-altitude image dataset, preprocessing the low-altitude image dataset to obtain a training set; S2: constructing a UAV target detection system based on multi-scale feature fusion; The system is used for identifying UAVs in complex aerial scenes, comprising: A backbone network obtained by replacing convolution operations in the original backbone network of the YOLOv11 network architecture with average pooling down-sampling operations; A neck network obtained by introducing average pooling down-sampling operations and global multi-scale attention enhancement operations into the original neck network of the YOLOv11 network architecture; A head network comprising multiple original detection heads of the YOLOv11 network architecture and a feature fusion sub-module matched with each original detection head, each feature fusion sub-module fusing multi-scale features from the neck network, and the fused features being input into the corresponding detection head to obtain an initial target detection box; An aggregated non-maximum suppression module selecting a reserved detection box with the highest confidence score from the initial target detection boxes output by the head network and eliminating redundant detection boxes with a high overlap ratio with the reserved detection box; generating a judgment score for the reserved detection box to measure its anti-elimination ability according to the aggregated characteristics of the redundant detection boxes, and then using the judgment score as the basis for non-maximum suppression to retain correct detection results that are severely overlapped and occluded, thereby obtaining final target detection boxes; In the aggregated non-maximum suppression module: Calculate the overlap score between any two initial target detection boxes: ； wherein GScore ij represents an overlap score between the ith initial target detection box Box i and the jth initial target detection box Box j , IoG(Box i , Box j ) represents an overlap ratio between the ith initial target detection box Box i and the jth initial target detection box Box j , and represent coordinate values of a top-left corner and a bottom-right corner of the ith initial target detection box Box i , and AMap represents a mapping relationship function, which is: ； wherein λ represents an input of a mapping relationship, σ t represents a preset parameter threshold; If the overlap score is greater than a preset overlap threshold, define the two initial target detection boxes related to the overlap score as aggregated detection boxes; Repeat the above operation to calculate the overlap score of each initial target detection box and obtain all aggregated detection boxes related to each initial target detection box, thereby obtaining an aggregated detection box set for each initial target detection box, and defining the detection box with the highest confidence in each aggregated detection box set as a reserved detection box to construct a reserved detection box set; In each aggregated detection box set, sort the overlap scores of each aggregated detection box in descending order to obtain the serial number of each aggregated detection box; Map the corresponding overlap score using the serial number of each aggregated detection box to obtain the discrimination suppression score of each aggregated detection box; Use the discrimination suppression score and the reserved detection box to perform non-maximum suppression on each initial target detection box until all reserved detection boxes in the reserved detection box set are traversed, thereby obtaining final target detection boxes; S3: training the target detection system constructed in step S2 using the training set obtained in step S1 to obtain a target detection model; S4: inputting the image to be detected into the target detection model obtained in step S3 to obtain a detection result. 2.The method of claim 1, wherein, In the average pooling down-sampling operation: Perform average pooling on the input features, and split the pooled features into two parts; Perform convolution on one part of the features and maximum pooling and convolution on the other part of the features; Concatenate the processed two parts of features to obtain the output features. 3.The method of claim 1, wherein, In the global multi-scale attention enhancement operation: Group the input features for processing; For each sub-feature map, average pooling and maximum pooling operations are respectively performed along the height and width directions of the sub-feature map, and the obtained four pooled features are nonlinearly mapped; The two pooled features in the height direction are element-wise added, and the two pooled features in the width direction are element-wise added; Sigmoid activation operations are respectively performed on the two added features to obtain the weights of the sub-feature map in the height and width directions; After multiplying the two weights with the sub-feature map, a reinforced sub-feature map is obtained; The reinforced sub-feature maps are combined to obtain the output feature. 4.The method of claim 1, wherein, In the feature fusion sub-module: The multi-scale features from the neck network are fused by the following formula: ； wherein, , and respectively represent the feature vector at coordinate (i,j) in the 3 feature maps of the l-th level, represents the result of the corresponding fusion, , and respectively represent the corresponding fusion weights, the sum of which is 1.

5. The unmanned aerial vehicle target detection method based on multi-scale feature fusion according to claim 4, characterized in that, three fusion weights , and are obtained from the following equations: ； ； ； wherein, , and denotes the vector value of the feature vector at coordinates (i,j) in the 3 feature maps of the l-th level. 6.The method of claim 1, wherein, The acquisition process of the discrimination suppression score includes: Define the detection boxes other than the retained detection boxes in each aggregated detection box set as redundant detection boxes; In each aggregated detection box set, the overlap scores of the retained detection box and all redundant detection boxes are sorted in descending order to obtain the serial number coefficients of each redundant detection box; Divide the overlap score by the serial number coefficient of the corresponding redundant detection box, and then sum all the quotients; The total value of the quotients is interval mapped using the mapping function to obtain the corresponding discrimination suppression score: ； Wherein, BMap represents a mapping relationship function, λ represents an input of the mapping relationship, and ε t represents a preset parameter threshold. 7.The method of claim 1, wherein, The target detection system is trained in step S3 by using the loss function of the following formula: L = η1 × L1 + η2 × L2; Wherein, L represents the loss function of the training target detection system, η1 and η2 represent loss weights, and the sum of the two is 1, L1 represents a bounding box regression loss function, and L2 represents a Gaussian distribution distance loss function; The bounding box regression loss function is: L1 = 1 - IoU + a x factor distance + b x factor shape ; wherein, IoU represents the intersection over union value between the predicted detection box and the real labeled box, and a and β represent factor weights, factor distance represents a distance correction factor, factor shape represents a shape correction factor; Distance correction factor factor distance is: ； wherein (x Pre ,y Pre ) represents the center coordinates of the predicted detection box output by the target detection model, (x GT ,y GT ) represents the center coordinates of the real labeled box, d represents the diagonal line length of the minimum circumscribed box of the predicted detection box and the real labeled box, and ρ h and ρ w represent the weight coefficients in the vertical direction and the horizontal direction, respectively. Weight coefficient in vertical direction ρ h is: ； wherein h GT represents the height of the real annotation box, w GT represents the width of the real annotation box, and Scale represents a scale factor; Horizontal weight coefficient p w is: ； Shape correction factor factor shape is: ； where h Pre represents the height of the predicted bounding box, w Pre represents the width of the predicted bounding box; The Gaussian distribution distance loss function L2 is: ； Wherein, ζ represents a constant value, D represents the Gaussian distribution distance between the predicted detection box and the real labeled box, and is: 。

Images