Improved unmanned aerial vehicle target detection method and system based on YOLO11s

Abstract

The invention relates to the technical field of image processing, in particular to an improved unmanned aerial vehicle target detection method and system based on YOLO11s, and the method comprises the steps: selecting YOLO11s as a basic network, and improving the basic network architecture in a manner that a CMSC module is introduced into a C3K2 module in a backbone network, and the CMSC module and the C3K2 module are combined to serve as a C3K2-CMSC module; when the C3K parameter is True, the C3K-CMSC module is adopted, and when the C3K parameter is False, only the CMSC module is adopted; an original Concat module in a neck network of the YOLO11s is replaced by an SDI (Serial Digital Interface) module; in a YOLO11s detection head module, a P2 detection layer for a small target is added; the original loss function is improved according to the Wise-IoU loss function, the Inner-IoU loss function and the MPDIOU loss function. According to the improved YOLO11s, the detection precision of a high aspect ratio target and a small target can be improved.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and more particularly to an improved UAV target detection method and system based on YOLO11s. Background Art

[0002] Drone aerial photography technology uses drone-mounted cameras to capture images of the ground from the air. With the advancement of image sensors and wireless communications, drone aerial photography has rapidly developed and is widely used in reconnaissance patrols, resource remote sensing, and farmland mapping. Object detection is a critical task in these applications. However, the limited resolution of aerial images, complex backgrounds, clusters of small and high-aspect-ratio objects, and limited drone resources make real-time detection and accurate positioning extremely difficult.

[0003] Traditional methods use manual annotation or feature extraction, but due to the difficulty in processing large amounts of data, they perform poorly in complex drone aerial imagery. With the development of deep learning technology, deep neural networks have been widely used for target detection in drone aerial imagery. Currently, there are two main approaches: two-stage and single-stage. The two-stage approach divides target detection into two consecutive stages: first generating candidate regions, then classifying and regressing these candidate regions. This type of method has high detection accuracy, but due to the large number of model parameters and complexity, it is generally not suitable for drone aerial imagery target detection scenarios. The single-stage approach simplifies the detection process, completing target category prediction and bounding box regression in a single stage. As a single-stage target detection algorithm, the YOLO series is fast, accurate, and efficient, and therefore has been widely used.

[0004] In recent years, many researchers have done a lot of work in this area. Xue et al. designed YOLO-AUV based on Yolov8s. By introducing the C2f-EMBC feature processing module into the backbone, they promoted the fusion of deep and shallow features, added a small target detection head, and used the NWD loss function to improve the accuracy of small target detection. Peng et al. improved YOLOv8s by introducing the RepLKNet framework to obtain a larger receptive field, reduced the number of model parameters by optimizing the SCconv module, and selected the MPDIoU loss function to further improve model accuracy. Yang et al. designed YOLO-ELD based on YOLOv5n. By integrating the CAA attention module into the backbone and replacing some Conv with Adown, they improved feature extraction capabilities and reduced the number of model parameters. They also used Shape-Inner-IoU to improve the accuracy of bounding box regression.

[0005] The above method has achieved good results, but there are still the following problems: low detection accuracy for many small targets in drone aerial images, lack of optimization for high aspect ratio targets, and insufficient performance in cross-level feature fusion. Summary of the Invention

[0006] In view of this, the present invention provides an improved UAV target detection method and system based on YOLO11s, which can improve the detection accuracy of high aspect ratio targets and small targets.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] In a first aspect, the present invention provides an improved drone target detection method based on YOLO11s, comprising the following steps:

[0009] YOLO11s is selected as the basic network, and the basic network architecture is improved by introducing the CMSC module into the C3K2 module in the backbone network. The two are combined into a C3K2-CMSC module. When the C3K parameter is True, the C3K2-CMSC module is used, and when the C3K parameter is False, only the CMSC module is used.

[0010] The C3K-CMSC module consists of two CMSC modules connected in series. The CMSC module is composed of two residual submodules, namely the Strip Block module and the CGLU module. The Strip Block module is used to extract high-aspect-ratio features from UAV aerial images to obtain spatial feature information. The CGLU module is used to fuse the spatial feature information through channel mixing.

[0011] Improve the original loss function based on Wise-IoU, Inner-IoU and MPDIoU loss functions;

[0012] Target detection is performed on drone aerial images based on the improved YOLO11s.

[0013] Furthermore, the Strip Block module contains a depthwise convolution and two strip convolutions in different directions; the depthwise convolution is used to extract channel feature information, and the two strip convolutions are used to extract spatial feature information in the horizontal and vertical dimensions respectively;

[0014] The CGLU module divides the spatial feature information extracted by the Strip Block module into two parts, one of which is processed by depthwise separable convolution as a gating function and then multiplied element-by-element with the other part, so as to achieve deeper fusion of the spatial feature information through channel mixing.

[0015] Furthermore, the basic network architecture is improved by replacing the original Concat module in the neck network of YOLO11s with an SDI module. The SDI module is used to connect multiple layers of feature maps and perform multi-level feature fusion.

[0016] Furthermore, the SDI module applies the spatial attention mechanism and channel attention mechanism to the feature layer of each level to process the features of each level and integrate local spatial information and global channel information; for feature maps at the same level, they are directly passed downward after identity mapping; for high-level feature maps, their size is reduced by average pooling so that higher-level semantic information can be integrated into the current level; for low-level feature maps, their size is increased by upsampling so that more detailed information can be integrated into the current feature map; for each level feature map, after identity mapping, average pooling or upsampling, the size of each feature map needs to be adjusted through smooth convolution; finally, features of different sizes are fused through the Hadamard product operation.

[0017] Further improvements to the basic network architecture include: adding a P2 detection layer for small targets in the detection head module of YOLO11s, whose predicted feature map size is 160×160.

[0018] Furthermore, the improved loss function is expressed as:

[0019] L Wise-Inner-MPDIoU =r×L MPDIoU +IoU-IoU Inner

[0020] Among them, r is the dynamic focusing coefficient in Wise-IoU loss, L MPDIoU Represents the MPDIoU loss function, IoU represents the intersection-over-union ratio between the predicted bounding box and the true bounding box in the standard IoU loss, and IoU Inner Represents the intersection-over-union ratio between the predicted bounding box and the true bounding box of the auxiliary box in the Inner-IoU loss.

[0021] Furthermore, the calculation formula of the dynamic focusing coefficient r is:

[0022]

[0023] Among them, x, y, x gt 、y gt Represents the horizontal and vertical coordinates of the center points of the predicted box and the real box respectively, W g 、H gRepresents the width and height of the minimum bounding rectangle of the predicted box and the true box, respectively. β represents the sample outlier. A small outlier value means a high-quality anchor box. α and δ are hyperparameters. * indicates that the update will be continuously calculated according to the target detection situation during the training process. R WIoU Indicates distance attention.

[0024] Furthermore, IoU Inner The calculation formula is:

[0025]

[0026] union=(w gt ×h gt )×(ratio) 2 +(w×h)×(ratio) 2 -inner

[0027] in, Represent the right, left, bottom and top boundaries of the real box respectively, b r 、b l 、b b 、b t Represents the right, left, bottom, and top boundaries of the prediction box, w gt 、h gt Represents the width and height of the real box, w and h represent the width and size of the predicted box, ratio represents the scaling ratio of the auxiliary box; inter represents the intersection of the auxiliary boxes; union represents the union of the auxiliary boxes.

[0028] Furthermore, the expression of the MPDIoU loss function is:

[0029] L MPDIoU =1-MPDIoU

[0030]

[0031] in, Refers to the upper left corner and lower right corner of the predicted box and the real box respectively, ρ 2 Represents the distance between corresponding points; MPDIoU represents.

[0032] In a second aspect, the present invention provides an improved drone target detection system based on YOLO11s, comprising:

[0033] Image acquisition module, used to acquire drone aerial images;

[0034] The target detection module is used to perform target detection on drone aerial images using the improved YOLO11s as described above.

[0035] It can be seen from the above technical solutions that compared with the prior art, the present invention has the following beneficial effects:

[0036] 1. By introducing the CMSC module into the YOLO11s backbone network, the present invention enables the network to pay more attention to features along the horizontal or vertical axis, collect directional features on the two spatial axes, enhance the representation of slender or narrow structures in the spatial dimension, and improve the detection accuracy of targets with high aspect ratios.

[0037] 2. The present invention introduces the SDI module into the neck network to replace the Concat module of YOLO11s, fully utilizes the complementarity between spatial and depth information, and extracts more expressive and robust feature information through multi-level feature processing and fusion, which can effectively improve the performance and effect of the target detection network in detecting drone aerial photography tasks.

[0038] 3. The present invention enhances the network's ability to perceive small targets by adding a P2 detection layer for small targets in the detection head.

[0039] 4. The present invention improves the original loss function based on the Wise-IoU, Inner-IoU and MPDIoU loss functions, so that the improved loss function can enhance the bounding box regression effect and reduce the false detection rate of dense target scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0041] Figure 1 The overall architecture diagram of the improved YOLO11s provided by the present invention;

[0042] Figure 2 A schematic structural diagram of the C3K2-CMSC module provided by the present invention;

[0043] Figure 3 A schematic structural diagram of the CMSC module provided by the present invention;

[0044] Figure 4 A schematic structural diagram of the SDI module provided by the present invention;

[0045] Figure 5 Schematic diagram for comparing the detection results of the method of the present invention with other existing methods. DETAILED DESCRIPTION

[0046] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 making creative efforts are within the scope of protection of the present invention.

[0047] like Figure 1 As shown, the embodiment of the present invention discloses an improved drone target detection method based on YOLO11s, comprising the following steps:

[0048] YOLO11s is selected as the basic network and the basic network architecture is improved. YOLO11s consists of three modules: backbone network (Backbone), neck structure (Neck) and detection head (Head). The improvement method is as follows:

[0049] The CMSC module is introduced into the C3K2 module in the backbone network, and the original Bottleneck module in the C3K2 module in the YOLO11s network is replaced by the CMSC module. The two are combined as the C3K2-CMSC module. When the C3K parameter is True, the C3K-CMSC module is used, and when the C3K parameter is False, only the CMSC module is used.

[0050] The original Concat module in the neck network of YOLO11s is replaced with the SDI module. The SDI module is used to connect multiple layers of feature maps and perform multi-level feature fusion.

[0051] In the YOLO11s detection head module, a P2 detection layer for small targets is added, and its predicted feature map size is 160×160;

[0052] Improve the original loss function based on Wise-IoU, Inner-IoU and MPDIoU loss functions;

[0053] Target detection is performed on drone aerial images based on the improved YOLO11s.

[0054] The improvement process of the above parts is further explained below.

[0055] (1) Improvement of backbone network.

[0056] Backbone network such as Figure 1 As shown in part (a), it includes Conv module, C3K2-CMSC module, SPPF module and C2PSA module.

[0057] The Conv module consists of 3×3 convolution, batch normalization, and SiLU activation function for downsampling.

[0058] The present invention replaces the C3K2 module in YOLO11s with the C3K2-CMSC module. The C3K2-CMSC module structure is as follows: Figure 2 The C3K parameter indicates whether to use a module containing a C3K2 structure. When C3K is True, the C3K-CMSC module is used, while when C3K is False, only the CMSC module is used. The C3K2-CMSC module primarily extracts high-aspect ratio features and increases network depth. The C3K-CMSC module consists of three Conv modules, two CMSC modules, and a Concat module. After passing through two parallel Conv modules, the input information is divided into two parts. One part passes through two serially connected CMSC modules, and the other part is concatenated with the Concat module. The concatenated part is then input into a single Conv module.

[0059] The SPPF module expands receptive fields of different sizes and integrates multi-scale features through multiple layers of pooling operations of different sizes.

[0060] The network of the C2PSA module is continued, and multi-scale spatial information is extracted through convolution kernels of different sizes in the PSA module.

[0061] Aerial drone images contain numerous high-aspect-ratio objects, characterized by their elongated bounding boxes. These objects exhibit unique spatial dimensionality, with rich features in one dimension and fewer in another. To better extract these features, the present invention constructs a CMSC module.

[0062] Specifically, the structure of the CMSC module is as follows Figure 3 As shown in (a) of the figure, the CMSC module consists of two residual submodules that undergo a normalization-processing-residual process: the Strip Block module and the CGLU module. The first residual submodule is processed by the Strip Block module, and the second residual submodule is processed by the CGLU module. The Strip Block module is used to extract high-aspect ratio features from drone aerial images to obtain spatial feature information; the CGLU module is used to fuse this spatial feature information through channel mixing.

[0063] The Strip Block module consists of a depthwise convolution and two strip convolutions in different directions: H Strip Conv and V Strip Conv. The depthwise convolution is used to extract channel feature information, while the two strip convolutions are used to extract spatial feature information in the horizontal and vertical dimensions, respectively. Through two orthogonal strip convolutions, H Strip Conv (horizontal strip convolution) and V Strip Conv (vertical strip convolution), unlike standard convolutions that extract features from square areas, strip convolutions allow the network to focus more on features along the horizontal or vertical axis. The combined use of horizontal and vertical strip convolutions enables the network to collect directional features along both spatial axes, enhancing the representation of slender or narrow structures in the spatial dimension and improving the extraction of objects with high aspect ratios. By increasing the number of convolutional network layers, the network depth is increased.

[0064] The CGLU module is essentially a channel mixer. It splits the spatial feature information extracted by the Strip Block module into two parts. One part undergoes depthwise separable convolution as a gating function, and then multiplies it element-by-element with the other part. This allows for deeper fusion of the spatial feature information through channel mixing. Compared to the traditional SE (Squeeze Excitation) channel mixer, the CGLU module can dynamically adjust the channel mixing method in a more detailed manner.

[0065] like Figure 3 As shown in (b) and (c), the two strip convolutions in different directions are H strip Conv and V stripConv, both in the Strip Module in the Strip Block.

[0066] The workflow of the Strip Block module is as follows: first, the channel dimension is adjusted through 1*1 convolution, then nonlinearity is introduced through the GELU activation function, then through the Strip Module, and then the number of channels is adjusted through 1*1 convolution, and finally connected with the initial input residual of the Strip Block module.

[0067] The Strip Moudle module first extracts channel feature information through DW convolution, then extracts horizontal and vertical features through H stripConv and V strip Conv, and then adjusts the number of channels through 1*1 convolution, and finally connects with the StripMoudle module input residual.

[0068] like Figure 3 As shown in (d), the input information of the CGLU module is divided into two parts. One part is activated by the DW convolution as the gating function and multiplied element-by-element with the other part. After adjusting the channel dimension through 1*1 convolution, it is connected with the original input information residual.

[0069] (2) Improvement of the neck network.

[0070] The present invention introduces the SDI (Semantics and Detail Infusion) module to replace the Concat module of YOLO11s. The SDI module structure is as follows Figure 4 shown.

[0071] The SDI module applies spatial attention mechanism and channel attention mechanism to the feature layer of each level to process the features of each level and integrate local spatial information and global channel information; for feature maps at the same level, their size remains unchanged and is directly passed downward after identity mapping; for high-level feature maps, their size is reduced by average pooling so that higher-level semantic information can be integrated into the current level; for low-level feature maps, their size is increased by upsampling so that more detailed information can be integrated into the current feature map; since the feature map sizes in each level decoder are different, for each level feature map, after identity mapping, average pooling or upsampling, it is necessary to adjust the size of each feature map through smoothing convolution (3×3 convolution); finally, features of different sizes are fused through the Hadamard product operation to enhance the features of each layer.

[0072] The SDI module utilizes the complementarity between spatial and depth information, and extracts more expressive and robust feature information through multi-level feature processing and fusion, which can effectively improve the performance and effect of the target detection network in detecting drone aerial photography tasks.

[0073] (3) Improvements to the detection head module.

[0074] The initial resolution of the input image is 640×640 pixels. The YOLO11s network downsamples the image size to 1 / 8, 1 / 16, and 1 / 32 respectively through the Conv module. Therefore, the predicted feature map sizes of the P3, P4, and P5 detection layers are 80×80, 40×40, and 20×20 respectively.

[0075] After the input image is processed by the backbone network, the resulting low-level feature maps have higher resolution and often contain more positional and detail information, but contain less abstract semantic information and more noise. High-level feature maps have lower resolution and contain less positional and detail information, but have stronger semantic information. The neck PAN-FPN structure and SDI module design can better utilize the features extracted by the backbone, reusing the positional and detail information of the shallow feature maps of the backbone network and fusing them with the semantic information of the deep feature maps. Small targets occupy fewer pixels and contain insufficient feature information after passing through the multi-layer network, making them difficult to detect with the original three-layer detection head. Therefore, to enhance the network's perception of small targets, a P2 detection layer for small targets was added to the improved YOLO network, with a predicted feature map size of 160×160.

[0076] (4) Improvement of loss function.

[0077] YOLO11 uses the CIoU loss function of the YOLOv8 series. CIoU takes into account the overlapping area, center point distance and aspect ratio consistency of the bounding box, but the detection frame regression effect is relatively general. Therefore, the present invention constructs the Wise-Inner-MPDIoU loss function based on the Wise-IoU, Inner-IoU and MPDIoU loss functions. Its core formula is as follows:

[0078] L Wise-Inner-MPDIoU =r×L MPDIoU +IoU-IoU Inner

[0079] Among them, r is the dynamic focusing coefficient in Wise-IoU loss, L MPDIoU Represents the MPDIoU loss function, IoU represents the intersection-over-union ratio between the predicted bounding box and the true bounding box in the standard IoU loss, and IoU Inner Represents the intersection-over-union ratio between the predicted bounding box and the true bounding box of the auxiliary box in the Inner-IoU loss.

[0080] The Wise-IoU loss function is a bounding box regression loss function based on dynamic weight adjustment. It addresses the issue of the CIoU loss function not taking into account the presence of low-quality annotations in the dataset itself. It introduces a dynamic focusing mechanism to reduce the interference of low-quality samples on network training. The core formula of Wise-IoU is as follows:

[0081] L WIoU =rL IoU

[0082] The calculation formula of dynamic focusing coefficient r is:

[0083]

[0084] Among them, x, y, x gt 、y gt Represents the horizontal and vertical coordinates of the center points of the predicted box and the real box respectively, W g 、H g Represents the width and height of the minimum bounding rectangle of the predicted box and the true box, respectively. β represents the sample outlier. A small outlier value means a high-quality anchor box. α and δ are hyperparameters. * indicates that the update will be continuously calculated according to the target detection situation during the training process. R WIoU Indicates distance attention. L IoU Represents IoU Loss, that is, IoU is used as the loss function of the target detection network.

[0085] The Inner-IoU loss function is a loss function based on auxiliary boxes. It optimizes different detection tasks by controlling the size of the auxiliary boxes. It can reduce background interference by shrinking the box and enhance the positioning of the core area of small targets. Inner The calculation formula is:

[0086]

[0087] union=(w gt ×h gt )×(ratio) 2 +(w×h)×(ratio) 2 -inner

[0088] L Inner-IoU =1-IoU Inner

[0089] in, Represent the right, left, bottom and top boundaries of the real box respectively, b r 、b l 、b b 、b t Represents the right, left, bottom, and top boundaries of the prediction box, w gt 、h gt Represents the width and height of the real box, w and h represent the width and size of the predicted box, ratio represents the scaling ratio of the auxiliary box; inter represents the intersection of the auxiliary boxes; union represents the union of the auxiliary boxes.

[0090] The MPDIoU loss function is an improved loss function based on the minimum point distance of the bounding box. It directly minimizes the difference in the coordinates of the upper left and lower right corners of the predicted box and the ground truth box, which can effectively handle the situation of bounding box overlap. It also solves the problem that the aspect ratio penalty term of the CIoU loss function fails to affect network accuracy when the aspect ratio of the predicted box and the ground truth box is the same. Therefore, it can effectively deal with the situation of bounding box overlap in scenes with dense objects. The expression of the MPDIoU loss function is:

[0091] L MPDIoU =1-MPDIoU

[0092]

[0093] in, Refers to the upper left corner and lower right corner of the predicted box and the real box respectively, ρ 2 represents the distance between corresponding points; MPDIoU represents the MPD intersection-over-union ratio; w and h represent the width and height of the prediction box, respectively.

[0094] In another embodiment, the present invention further provides an improved drone target detection system based on YOLO11s, comprising:

[0095] Image acquisition module, used to acquire drone aerial images;

[0096] The target detection module is used to perform target detection on drone aerial images using the improved YOLO11s as described above.

[0097] Next, an experiment is conducted to verify the detection performance of the improved network of the present invention.

[0098] (1) Dataset:

[0099] The experiment primarily used the VisDrone2019 dataset, released by the Machine Learning and Data Mining Laboratory of Tianjin University. It is one of the mainstream public datasets of drone aerial images. This dataset was captured by drones in 14 different cities and contains images with different weather conditions, brightness conditions, and target density. It covers 10 categories of aerial detection targets: pedestrians, people, bicycles, cars, vans, trucks, tricycles, sunshade tricycles, buses, and motorcycles. The training set, validation set, and test set images are 6471, 548, and 1610, respectively. Because this dataset contains a large number of detection targets, small targets, and targets with high aspect ratios, it is very suitable for the research of this invention.

[0100] (2) Evaluation indicators:

[0101] The experiment of the present invention evaluates the detection performance and compares it with the baseline model YOLO11s. The present invention uses precision (P), recall (R), mAP50 and mAP50:95, which are commonly used in target detection, as evaluation indicators. P represents the ratio of the number of correct detections to the total number of detections, which is used to evaluate the accuracy of the network. R represents the ratio of the number of correct detections to the total number of detections required, which is used to evaluate the detection rate of the network. mAP50 and mAP50:95 represent the mAP values when the IoU threshold is 0.5 and, respectively, and the average mAP value when 10 IoU thresholds are selected at equal intervals between 0.5 and 0.95, which are used to evaluate the average detection accuracy of all categories.

[0102] (3) Experimental environment:

[0103] The experimental environment settings are shown in Table 1.

[0104] Table 1 Experimental environment

[0105] Configure the environment Configuration Name (Version) operating system Windows 10 GPU NVIDIA GTX4090 Compiler Python 3.10 Deep Learning Framework Pytorch 2.6 Acceleration module CUDA 12.4

[0106] The experimental hyperparameter settings are shown in Table 2, and the rest keep the default parameters of the baseline model.

[0107] Table 2 Experimental hyperparameters

[0108] parameter Parameter value batchsize 16 epochs 640 images_size Python 3.10 Optimizer SGD

[0109] (4) Experimental results:

[0110] To evaluate the effectiveness of the improved YOLO proposed in this paper, a series of experiments were conducted. The effectiveness of each improved module was verified through ablation experiments. The experimental results are shown in Table 3. It can be seen that these four modules, as well as their integration into the baseline network, all contribute to improving the accuracy of drone aerial object detection.

[0111] Table 3 Ablation experiments

[0112]

[0113] Table 4 shows the comparative experiments. Compared with other methods, the comprehensive method of the method of the present invention is the best.

[0114] Table 4 Comparative experiment

[0115] Network P R mAP50 mAP50:90 RT-DETR-R34 0.561 0.413 0.402 0.235 YOLOv8s 0.508 0.379 0.389 0.234 YOLOv10s 0.513 0.387 0.39 0.238 YOLO11n 0.436 0.331 0.33 0.192 Ours 0.559 0.426 0.443 0.269

[0116] The present invention also carries out visual analysis on the experimental results, such as Figure 5The three figures on the left are the prediction results of the baseline network, and the three figures on the right are the inference results of the improved YOLO network. By observing the prediction boxes within the red circles, we can see that the network has improved the detection accuracy of small objects and objects with high aspect ratios.

[0117] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0118] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A drone target detection method based on YOLO11s improvement, characterized in that: The following steps are involved: YOLO11s is selected as the basic network, and the basic network architecture is improved in the following ways: the CMSC module is introduced into the C3K2 module in the backbone network, and the two are combined into a C3K2-CMSC module; when the C3K parameter is True, the C3K-CMSC module is used, and when the C3K parameter is False, only the CMSC module is used; the C3K-CMSC module consists of two CMSC modules connected in series; The CMSC module is composed of two residual submodules: the Strip Block module and the CGLU module. The StripBlock module is used to extract high-aspect-ratio features from drone aerial images and obtain spatial feature information. The CGLU module is used to fuse the spatial feature information through channel mixing. Improve the original loss function based on Wise-IoU, Inner-IoU and MPDIoU loss functions; Target detection is performed on drone aerial images based on the improved YOLO11s.

2. The improved UAV target detection method based on YOLO11s according to claim 1 is characterized in that: The StripBlock module contains a depthwise convolution and two strip convolutions in different directions; the depthwise convolution is used to extract channel feature information, and the two strip convolutions are used to extract spatial feature information in the horizontal and vertical dimensions respectively; The CGLU module divides the spatial feature information extracted by the Strip Block module into two parts, one of which is processed by depthwise separable convolution as a gating function and then multiplied element-by-element with the other part, so as to achieve deeper fusion of the spatial feature information through channel mixing.

3. The improved UAV target detection method based on YOLO11s according to claim 1 is characterized in that: Ways to improve the basic network architecture also include: replacing the original Concat module in the neck network of YOLO11s with the SDI module. The SDI module is used to connect multiple layers of feature maps and perform multi-level feature fusion.

4. The improved UAV target detection method based on YOLO11s according to claim 3 is characterized in that: The SDI module uses the spatial attention mechanism and channel attention mechanism for each level of feature layer to process the features of each level and integrate local spatial information and global channel information; For feature maps at the same level, they are directly passed down after identity mapping; for high-level feature maps, their size is reduced through average pooling so that higher-level semantic information can be integrated into the current level; for low-level feature maps, their size is increased through upsampling so that more detailed information can be integrated into the current feature map; for feature maps at each level, after identity mapping, average pooling or upsampling, the size of each feature map needs to be adjusted through smooth convolution; finally, features of different sizes are fused through the Hadamard product operation.

5. The improved UAV target detection method based on YOLO11s according to claim 1 is characterized in that, Improvements to the basic network architecture also include adding a P2 detection layer for small targets in the YOLO11s detection head module, with a predicted feature map size of 160×160.

6. The improved UAV target detection method based on YOLO11s according to claim 1 is characterized in that: The improved loss function is expressed as: L Wise-Inner-MPDIoU =r×L MPDIoU +IoU-IoU Inner Among them, r is the dynamic focusing coefficient in Wise-IoU loss, I MPDIoU Represents the MPDIoU loss function, IoU represents the intersection-over-union ratio between the predicted bounding box and the true bounding box in the standard IoU loss, and IoU Inner Represents the intersection-over-union ratio between the predicted bounding box and the true bounding box of the auxiliary box in the Inner-IoU loss.

7. The improved UAV target detection method based on YOLO11s according to claim 6 is characterized in that: The calculation formula of dynamic focusing coefficient r is: Among them, x, y, x gt 、y gt Represents the horizontal and vertical coordinates of the center points of the predicted box and the real box respectively, W g 、H g Represents the width and height of the minimum bounding rectangle of the predicted box and the true box respectively, β represents the sample outlier degree, and a small outlier value means a high-quality anchor box. α and δ are hyperparameters, and * represents that they will be continuously calculated and updated according to the target detection situation during the training process; R WIoU Indicates distance attention.

8. The improved UAV target detection method based on YOLO11s according to claim 6 is characterized in that: IoU Inner The calculation formula is: union=(w gt ×h gt )×(ratio) 2 +(w×h)×(ratio) 2 -inner in, Represent the right, left, bottom and top boundaries of the real box respectively, b r 、b l 、b b 、b t Represents the right, left, bottom, and top boundaries of the prediction box, w gt 、h gt Represents the width and height of the real box, w and h represent the width and size of the predicted box, ratio represents the scaling ratio of the auxiliary box; inter represents the intersection of the auxiliary boxes; union represents the union of the auxiliary boxes.

9. The improved UAV target detection method based on YOLO11s according to claim 7 is characterized in that: The expression of MPDIoU loss function is: L MPDIoU =1-MPDIoU in, Refers to the upper left corner and lower right corner of the predicted box and the real box respectively, ρ 2 represents the distance between corresponding points; w and h represent the width and height of the prediction box respectively.

10. An improved UAV target detection system based on YOLO11s, characterized in that: include: Image acquisition module, used to acquire drone aerial images; The target detection module is used to perform target detection on drone aerial images using the improved YOLO11s according to any one of claims 1 to 9.

Images