Multi-scale target image detection method in unmanned aerial vehicle scene

Abstract

A multi-scale target image detection method in an unmanned aerial vehicle scene is characterized by comprising the following steps: step 1, constructing a multi-scale target image detection system in the unmanned aerial vehicle scene; 2, an image acquisition module acquires a multi-scale target image X and transmits the multi-scale target image to a multi-scale target image detection model in an unmanned aerial vehicle scene; 3, performing feature extraction operation on the multi-scale target image X by a feature extraction module in the multi-scale target image detection model in the unmanned aerial vehicle scene to obtain image feature data Y, and transmitting the image feature data Y to a multi-scale feature fusion module; 4, performing multi-scale feature fusion operation on the image feature data Y by a multi-scale feature fusion module to obtain fusion data Z, and transmitting the fusion data Z to a detection and identification module; and step 5, the detection and identification module carries out detection and identification operation on the fusion data Z and outputs a target detection result. The method has the effect that the capability of extracting multi-scale and direction information by the model is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and particularly to a multi-scale target image detection method in an unmanned aerial vehicle (UAV) scenario. Background Art

[0002] With the continuous progress of technology, the flexibility, flight ability, and wide coverage of UAV technology have been significantly improved, thereby enhancing its effectiveness in aerial detection tasks and making it a prominent focus of artificial intelligence research. This technology mainly covers multiple fields, including wild rescue, agricultural plant protection, forest fire warning, and traffic safety monitoring. However, despite the numerous advantages of UAV technology, it also faces some challenges. The targets captured in high-altitude scenarios are usually small, the background is complex, and the aerial environment is constantly changing. Given the high flight speed of UAVs, they are sensitive to external factors such as position and lighting, resulting in blurred aerial images and increased noise. Therefore, due to the change in the shooting angle, this will reduce the image quality and increase the complexity. To better complete various tasks related to UAV technology, it is crucial to improve the detection ability of UAVs in high-altitude environments.

[0003] Traditional object detection methods mainly rely on object color and texture, or use Histogram of Oriented Gradients (HOG) feature extraction technology to identify objects in images. The HOG feature extraction technology mainly analyzes the local gradients of images and uses Support Vector Machines (SVMs) to detect object shapes. These traditional methods are usually time-consuming and cumbersome, and show suboptimal performance when dealing with multi-scale targets. With the improvement of computing power, neural networks have become a remedy for these limitations by automatically learning object features to reduce the dependence on manual labor. Currently, to solve the problem of degraded image quality, data augmentation techniques or feature fusion methods are usually adopted. In addition, the feature extraction process has been improved, and various label assignment strategies have been explored. However, these methods are limited in reducing the noise specific to UAV aerial photography. One of the most effective solutions is to increase the image resolution. However, this method significantly increases the computational load on the GPU.

[0004] Disadvantages of the prior art: The existing image detection methods perform poorly when dealing with multi-scale targets. Summary of the Invention

[0005] A multi-scale target image detection method in a UAV scenario provided by the present invention enhances the ability of the model to extract multi-scale and directional information.

[0006] To achieve the above object, a multi-scale target image detection method in a UAV scenario provided by the present invention is characterized in that it includes the following steps:

[0007] Step 1: Construct a multi-scale target image detection system for the UAV scenario. The multi-scale target image detection system for the UAV scenario is provided with an image acquisition module, and the image acquisition module is connected;

[0008] The multi-scale target image detection model ESGN-YOLO for the UAV scenario is provided with a feature extraction module BackBone, a multi-scale feature fusion module Neck, and a detection and recognition module Head that are connected in sequence;

[0009] Step 2: The image acquisition module acquires a multi-scale target image X and transfers the multi-scale target image to the multi-scale target image detection model ESGN-YOLO for the UAV scenario;

[0010] Step 3: The feature extraction module BackBone in the multi-scale target image detection model ESGN-YOLO for the UAV scenario performs feature extraction operations on the multi-scale target image X to obtain image feature data Y and transfers it to the multi-scale feature fusion module Neck;

[0011] Step 4: The multi-scale feature fusion module Neck performs multi-scale feature fusion operations on the image feature data Y to obtain fusion data Z and transfers it to the detection and recognition module Head;

[0012] Step 5: The detection and recognition module Head performs detection and recognition operations on the fusion data Z and outputs the target detection result.

[0013] Through the above design, the accurate detection of multi-scale target images in the UAV scenario is realized, the ability of the ESGN-YOLO model to extract multi-scale and direction information is improved, redundant features are effectively suppressed, the model's fine-grainedness is increased, and thus the ability of the model to extract local details is improved.

[0014] Preferably: In the step 3, the feature extraction module BackBone is provided with a first basic convolution module CBS-1, a second basic convolution module CBS-2, a first C3k-SWC module, a first efficient feature fusion module EFFM-1, a third basic convolution module CBS-3, a second C3k-SWC module, a fourth basic convolution module CBS-4, a third C3k-SWC module, a fifth basic convolution module CBS-5, a C3k2-SWC module, a second efficient feature fusion module EFFM-2, a fast spatial pyramid pooling module SPPF, and an enhanced convolution module C2PSA;

[0015] The output ends of the second C3k-SWC module, the third C3k-SWC module, and the enhanced convolution module C2PSA are connected to the multi-scale feature fusion module Neck.

[0016] The Fast Spatial Pyramid Pooling Module SPPF is used to extract context features through pooling operations of different sizes, and can extract features on different receptive fields.

[0017] The Enhanced Convolution Module C2PSA is used to combine channel - dimension attention and spatial - dimension attention, mine the relationship between features from two dimensions, and enhance the network's local and global modeling capabilities.

[0018] Preferably, the feature extraction module BackBone performs feature extraction operations on the multi - scale target image X, including the following steps:

[0019] Step 3.1: The first basic convolution module CBS - 1 obtains the multi - scale target image X, and sequentially performs convolution, batch normalization, and SiLU activation operations on it to obtain the first convolution data y1, and transmits it to the second basic convolution module CBS - 2;

[0020] Step 3.2: The second basic convolution module CBS - 2 sequentially performs convolution, batch normalization, and SiLU activation operations on the first convolution data y1 to obtain the second convolution data y2, and transmits it to the first C3k - SWC module;

[0021] Step 3.3: The first C3k - SWC module performs multi - scale convolution operations on the second convolution data y2 to obtain the third convolution data y3, and transmits it to the first efficient feature fusion module EFFM - 1;

[0022] Step 3.4: The first efficient feature fusion module EFFM - 1 performs efficient feature fusion operations on the third convolution data y3 to obtain the first fusion data y4, and transmits it to the third basic convolution module CBS - 3;

[0023] Step 3.5: The third basic convolution module CBS - 3 sequentially performs convolution, batch normalization, and SiLU activation operations on the first fusion data y4 to obtain the fourth convolution data y5, and transmits it to the second C3k - SWC module;

[0024] Step 3.6: The second C3k - SWC module performs multi - scale convolution operations on the fourth convolution data y5 to obtain the fifth convolution data y6, and transmits it to the fourth basic convolution module CBS - 4 and the multi - scale feature fusion module Neck;

[0025] Step 3.7: The fourth basic convolution module CBS - 4 sequentially performs convolution, batch normalization, and SiLU activation operations on the fifth convolution data y6 to obtain the sixth convolution data y7, and transmits it to the third C3k - SWC module;

[0026] Step 3.8: The third C3k-SWC module performs multi-scale convolution operations on the sixth convolution data y7 to obtain the seventh convolution data y8, and transmits it to the fifth basic convolution module CBS-5 and the multi-scale feature fusion module Neck;

[0027] Step 3.9: The fifth basic convolution module CBS-5 sequentially performs convolution, batch normalization, and SiLU activation operations on the seventh convolution data y8 to obtain the eighth convolution data y9, and transmits it to the C3k2-SWC module;

[0028] Step 3.10: The C3k2-SWC module performs deformable convolution and channel separation operations on the eighth convolution data y9 to obtain the ninth convolution data y10, and transmits it to the second efficient feature fusion module EFFM-2;

[0029] Step 3.11: The second efficient feature fusion module EFFM-2 performs efficient feature fusion operations on the ninth convolution data y10 to obtain the second fusion data y11, and transmits it to the fast spatial pyramid pooling module SPPF;

[0030] Step 3.12: The fast spatial pyramid pooling module SPPF performs multi-scale pooling operations on the second fusion data y11 to obtain the pooled data y12, and transmits it to the enhanced convolution module C2PSA;

[0031] Step 3.13: The enhanced convolution module C2PSA performs multi-scale feature extraction operations on the pooled data y12 to obtain the image feature data Y, and transmits it to the multi-scale feature fusion module Neck.

[0032] Preferably: The first efficient feature fusion module EFFM-1 and the second efficient feature fusion module EFFM-2 have the same structure, and both are provided with an input layer. The output end of the input layer is connected to the input end of the depthwise separable convolution layer DWconv and the second addition unit; the output end of the depthwise separable convolution layer DWconv is connected to the input ends of the average pooling layer Avgpool, the max pooling layer Maxpool, and the second multiplication module. The output ends of the average pooling layer Avgpool and the max pooling layer Maxpool are connected to the input end of the first addition unit; the output end of the first addition unit is connected to the input ends of the spatial attention mechanism module SA and the channel attention mechanism CA. The output ends of the spatial attention mechanism module SA and the channel attention mechanism CA are connected to the input end of the first multiplication unit; the output end of the first multiplication unit is connected to the input end of the second multiplication unit through the convolution layer Covn. The output end of the second multiplication unit is connected to the input end of the second addition unit, and the output end of the second addition unit is connected to the output layer;

[0033] The calculation expression of the Efficient Feature Fusion Module (EFFM) is as follows:

[0034] F = DWconv(a) (1)

[0035] F global = ∑ d∈{h,v} (Avgpool d (F) + Maxpool d (F)) (2)

[0036] CA = σ(Conv(ReLU(BN.Conv(F global )) / )) (3)

[0037] SA = σ(Conv.Avgpool(F global )) / (4)

[0038] F refined = Conv(SA(F global )) · CA(F global )) (5)

[0039] F output = F · F refined + a (6)

[0040] Among them, a is the input data of the Efficient Feature Fusion Module (EFFM), BN is the batch normalization operation; σ is the sigmoid function, and σ is used for activation operations in formulas (3) and (4); F output is the output data of the Efficient Feature Fusion Module (EFFM); h and v are the two dimensions of the average pooling layer and the maximum pooling layer.

[0041] In the first Efficient Feature Fusion Module (EFFM-1), a is the third convolution data y3, and F output is the first fusion data y4;

[0042] In the second Efficient Feature Fusion Module (EFFM-2), a is the ninth convolution data y10, and F output is the second fusion data y11.

[0043] Traditional small object detection methods usually integrate the attention mechanism into the backbone network, but it is difficult to perform multi-scale feature extraction. The present invention introduces the Efficient Feature Fusion Module (EFFM) into the backbone network, effectively enhancing the model's multi-scale feature learning ability and context awareness ability.

[0044] In the Efficient Feature Fusion Module (EFFM), first, the separable convolution enhances feature extraction, improves model accuracy, and reduces redundant computations. Next, the convolution operations in different horizontal and vertical directions capture multi-scale global information, enhancing the model's ability to extract texture and edge details. Then, various attention mechanisms are applied to the global features obtained in the previous step, further improving the model's ability to identify and emphasize important features. Finally, the residual module retains the initial information and alleviates the problem of vanishing gradients.

[0045] The Efficient Feature Fusion Module (EFFM) optimizes the feature extraction process through multi-level feature fusion, attention mechanisms, and residual connections, significantly improving the model performance in tasks such as small object detection.

[0046] Preferably: the first C3k-SWC module, the second C3k-SWC module, and the third C3k-SWC module have the same structure, and each is provided with an eighth basic convolution module CBS-8. The output end of the eighth basic convolution module CBS-8 is connected to the input end of the fifth channel splicing module Concat-5 through a ninth basic convolution module CBS-9. Between the output end of the eighth basic convolution module CBS-8 and the input end of the fifth channel splicing module Concat-5, there are also n sequentially connected SWC-Bottleneck modules. The output end of the fifth channel splicing module Concat-5 is connected to a tenth basic convolution module CBS-10;

[0047] The C3k2-SWC module is provided with an eleventh basic convolution module CBS-11 and a twelfth basic convolution module CBS-12 in parallel. The output ends of the eleventh basic convolution module CBS-11 and the twelfth basic convolution module CBS-12 are built-in with a Split splitting function;

[0048] The output end of the eleventh basic convolution module CBS-11 is connected to the input end of the sixth channel splicing module Concat-6. Between the output end of the eleventh basic convolution module CBS-11 and the input end of the sixth channel splicing module Concat-6, there are also m sequentially connected SWC-Bottleneck modules. The output ends of the m SWC-Bottleneck modules are all connected to the input end of the sixth channel splicing module Concat-6. The output end of the sixth channel splicing module Concat-6 is connected to a thirteenth basic convolution module CBS-13;

[0049] The output end of the twelfth basic convolution module CBS-12 is connected to the input end of the seventh channel splicing module Concat-7. Between the output end of the twelfth basic convolution module CBS-12 and the input end of the seventh channel splicing module Concat-7, there are also m C3k-SWC modules C3k-SWC connected in sequence. The output ends of the m C3k-SWC modules C3k-SWC are all connected to the input end of the seventh channel splicing module Concat-7; the output end of the seventh channel splicing module Concat-7 is connected to the fourteenth basic convolution module CBS-14.

[0050] In YOLOv11, the C3k2 and C3k modules rely on simple convolution operations, lack advanced feature fusion, and are prone to misinterpreting important information, especially in small object detection. To solve this problem, the present invention proposes an SWC-Bottleneck module for replacing the bottleneck modules in C3k2 and C3k.

[0051] Preferably: The SWC-Bottleneck module is provided with an input unit. The output end of the input unit is connected to the input end of the first convolutional layer Conv-1. The output end of the first convolutional layer Conv-1 is connected to the input ends of the first per-channel convolutional block and the second per-channel convolutional block. The output ends of the first per-channel convolutional block and the second per-channel convolutional block are connected to the input end of the addition unit. The output end of the addition unit is connected to the input end of the residual connection layer; the output end of the input unit is also connected to the input end of the residual connection layer through the second convolutional layer Conv-2;

[0052] The first per-channel convolutional block and the second per-channel convolutional block have the same structure and are both provided with a mask unit Mask, a split unit Split, a tensor dimension adjustment unit Squeeze, a clipping unit Clip, a splicing unit, and a batch normalization layer BN connected in sequence;

[0053] The calculation expression of the SWC-Bottleneck module is as follows:

[0054]

[0055] nk = ceil(kerner big / kerner small ) (8)

[0056]

[0057] Among them, ° represents multiplication of corresponding matrix positions, ceil(·) represents ceiling operation, Clip(·) represents clipping operation, BN(·) represents normalization operation, Conv represents convolution operation, Split(·) represents splitting function; i represents the i-th per-channel convolution block, i = 1, 2; j represents the j-th segmentation block; Q output represents the output of the SWC-Bottleneck module, Q represents the input of the SWC-Bottleneck module, w represents the weight, kerner big represents the largest convolution kernel size, kerner small represents the smallest convolution kernel size; σ represents the sigmoid function, and σ in formula (7) is used for mask operation; Squeeze(·) represents tensor compression operation.

[0058] The SWC-Bottleneck module processes the input image through two parallel paths. In the first path, convolution is performed using a smaller convolution kernel to retain the key information of the input image. In the second path, after the convolution operation, two feature gating mechanisms are introduced, and the convolution results are dynamically adjusted through the learned weights to form two different branches. As shown in formula (7), the Mask operation adopts per-channel shift multiplication, and uses the learnable weight w to constrain Q i within the range of [0, 1]. Subsequently, each branch undergoes a tensor reshaping operation to transform the tensor from Q i ∈R{[b, c1, h1, w1]} to Q i ∈R{[b, c2, nk, h1, w1]}, where nk is defined in formula (8). Then, the Split operation decomposes each Q i into nk blocks, and the Squeeze operation is used to compress the dimensions of the tensor, as shown in formula (9). The compressed result is expanded in the horizontal and vertical directions, and the unimportant regions are clipped to retain the key information. Subsequently, the features of the two branches are concatenated and batch-normalized, and their results are added to form the output of the second path. Finally, the outputs of the two paths are merged to generate the final output. Formula (10) further gives the formal definition.

[0059] In small object detection, the background region often dominates. The SWC-Bottleneck module suppresses background features through the Mask operation while enhancing target features. The reshaping (Reshape) and splitting (Split) of tensors can capture local features more effectively.

[0060] Preferably, in the step 4, the multi-scale feature fusion module Neck is provided with a first upsampling module Upsample-1, a first channel splicing module Concat-1, a first C3k2 module, a second upsampling module Upsample-2, a second channel splicing module Concat-2, a second C3k2 module, a sixth basic convolution module CBS-6, a third channel splicing module Concat-3, a third C3k2 module, a seventh basic convolution module CBS-7, a fourth channel splicing module Concat-4 and a fourth C3k2 module connected in sequence;

[0061] The input ends of the first upsampling module Upsample-1, the first channel splicing module Concat-1, the second channel splicing module Concat-2 and the fourth channel splicing module Concat-4 are connected to the feature extraction module BackBone; the output end of the first C3k2 module is connected to the input end of the third channel splicing module Concat-3; the output ends of the second C3k2 module, the third C3k2 module and the fourth C3k2 module are connected to the detection and recognition module Head.

[0062] Preferably, the multi-scale feature fusion module Neck performs multi-scale feature fusion operation on the image feature data Y, including the following steps:

[0063] Step 4.1: The first upsampling module Upsample-1 obtains the image feature data Y, performs upsampling operation on it to obtain a first sampled data z1, and transmits it to the first channel splicing module Concat-1;

[0064] Step 4.2: The first channel splicing module Concat-1 performs channel splicing operation on the first sampled data z1 and the seventh convolution data y8 to obtain a first spliced data z2, and transmits it to the first C3k2 module;

[0065] Step 4.3: The first C3k2 module performs deformable convolution and channel separation operation on the first spliced data z2 to obtain a first deformable convolution data z3, and transmits it to the second upsampling module Upsample-2 and the third channel splicing module Concat-3;

[0066] Step 4.4: The second upsampling module Upsample-2 performs upsampling operation on the first deformable convolution data z3 to obtain a second sampled data z4, and transmits it to the second channel splicing module Concat-2;

[0067] Step 4.5: The second channel concatenation module Concat-2 performs a channel concatenation operation on the second sampled data z4 and the fifth convolutional data y6 to obtain the second concatenated data z5, and transfers it to the second C3k2 module;

[0068] Step 4.6: The second C3k2 module performs deformable convolution and channel separation operations on the second concatenated data z5 to obtain the second deformable convolutional data z6, and transfers it to the sixth basic convolutional module CBS-6 and the detection and recognition module Head;

[0069] Step 4.7: The sixth basic convolutional module CBS-6 sequentially performs convolution, batch normalization, and SiLU activation operations on the second deformable convolutional data z6 to obtain the tenth convolutional data z7, and transfers it to the third channel concatenation module Concat-3;

[0070] Step 4.8: The third channel concatenation module Concat-3 performs a channel concatenation operation on the tenth convolutional data z7 and the first deformable convolutional data z3 to obtain the third concatenated data z8, and transfers it to the third C3k2 module;

[0071] Step 4.9: The third C3k2 module performs deformable convolution and channel separation operations on the third concatenated data z8 to obtain the third deformable convolutional data z9, and transfers it to the seventh basic convolutional module CBS-7 and the detection and recognition module Head;

[0072] Step 4.10: The seventh basic convolutional module CBS-7 sequentially performs convolution, batch normalization, and SiLU activation operations on the third deformable convolutional data z9 to obtain the eleventh convolutional data z10, and transfers it to the fourth channel concatenation module Concat-4;

[0073] Step 4.11: The fourth channel concatenation module Concat-4 performs a channel concatenation operation on the eleventh convolutional data z10 and the image feature data Y to obtain the fourth concatenated data z11, and transfers it to the fourth C3k2 module;

[0074] Step 4.12: The fourth C3k2 module performs deformable convolution and channel separation operations on the fourth concatenated data z11 to obtain the fusion data Z, and transfers it to the detection and recognition module Head.

[0075] Preferably, the detection and recognition module Head is provided with a first convolutional group normalization block CGS-1, a second convolutional group normalization block CGS-2, and a third convolutional group normalization block CGS-3. The input ends of the first convolutional group normalization block CGS-1, the second convolutional group normalization block CGS-2, and the third convolutional group normalization block CGS-3 are connected to the multi-scale feature fusion module Neck, and the output ends of the three are connected to the input end of the fourth convolutional group normalization block CGS-4. The output end of the fourth convolutional group normalization block CGS-4 is connected to the input ends of the small object detection head BSC-1, the medium object detection head BSC-2, and the large object detection head BSC-3, and the output ends of the small object detection head BSC-1, the medium object detection head BSC-2, and the large object detection head BSC-3 output the object detection results.

[0076] The decoupled classification Conv-Cls module and the regression module Conv-Box are arranged in parallel in the small / medium / large object detection heads, and the output end of the regression module Conv-Box is connected to the input end of the model scaling ratio control module Scale.

[0077] The convolutional group normalization block CGS is provided with a convolutional layer and a group normalization layer connected in sequence, and a SiLU activation function is built in the output end of the group normalization layer.

[0078] In YOLOv11, the detection head adopts a multi-task architecture that shares features, which may affect the performance of classification and regression. To solve this problem, the present invention proposes a group normalization scale detection head GNSH, that is, the detection and recognition module Head, which is a lightweight detection head. By decoupling the classification module Conv-Cls and the regression module Conv-Box, it improves the feature utilization rate while maintaining high efficiency.

[0079] Traditional batch normalization BN performs poorly in small batch scenarios, especially in small object detection. Therefore, the present invention adopts group normalization GN that is independent of the batch size to better retain details. GNSH extracts high-level features from the Neck module, fuses multi-scale features through convolution Conv and GN layers, and uses independent branches for classification and regression. In addition, the Scale module further optimizes the bounding box prediction results.

[0080] Preferably, the detection and recognition module Head performs detection and recognition operations on the fusion data Z, and the steps are as follows:

[0081] Step 5.1: The first convolutional group normalization block CGS-1 obtains the fusion data Z, performs convolution and group normalization operations on it, and then outputs the first group-normalized data c1 to the fourth convolutional group normalization block CGS-4 through the SiLU activation function;

[0082] The second convolutional group normalization block CGS-2 takes the third variable convolutional data z9, performs convolutional and group normalization operations on it, and then outputs the second group normalization data c2 to the fourth convolutional group normalization block CGS-4 through the SiLU activation function;

[0083] The third convolutional group normalization block CGS-3 takes the second variable convolutional data z6, performs convolutional and group normalization operations on it, and then outputs the third group normalization data c3 to the fourth convolutional group normalization block CGS-4 through the SiLU activation function;

[0084] Step 5.2: The fourth convolutional group normalization block CGS-4 performs convolutional and group normalization operations on the first group normalization data c1, the second group normalization data c2, and the third group normalization data c3, and then outputs the fourth group normalization data c4 to the small target detection head BSC-1, the medium target detection head BSC-2, and the large target detection head BSC-3 through the SiLU activation function;

[0085] Step 5.3: The small target detection head BSC-1 performs small target detection on the fourth group normalization data c4 and outputs the small target detection result;

[0086] The medium target detection head BSC-2 performs medium target detection on the fourth group normalization data c4 and outputs the medium target detection result;

[0087] The large target detection head BSC-3 performs large target detection on the fourth group normalization data c4 and outputs the large target detection result.

[0088] The small target detection result, the medium target detection result, and the large target detection result together constitute the multi-scale target detection result.

[0089] Advantages of the present invention:

[0090] 1. Integrate the efficient feature fusion module EFFM into the feature extraction module BackBone to enhance the model's ability to extract multi-scale and directional information.

[0091] 2. Adopt the SWC-Bottleneck module to improve the model's granularity and enhance its ability to focus on extracting local image features.

[0092] 3. Design a lightweight detection head called the group normalization scale detection head GNSH to reduce the number of model parameters. Description of the Drawings

[0093] Figure 1 It is the network structure diagram of the multi-scale target image detection model ESGN-YOLO for the UAV scenario;

[0094] Figure 2 It is a schematic diagram of the SWC-Bottleneck module structure;

[0095] Figure 3 It is a comparison diagram of the heat maps of YOLOv11n and the ESGN-YOLO of the present invention;

[0096] Figure 4 It is a comparison diagram of the detection effects of YOLOv11n and the ESGN-YOLO of the present invention. Specific implementation manners

[0097] The present invention will be further described in detail below with reference to the accompanying drawings and specific examples. The following examples or drawings are used to illustrate the present invention, but not to limit the scope of the present invention.

[0098] A multi-scale target image detection method in a drone scenario includes the following steps:

[0099] Step 1: Construct a multi-scale target image detection system in a drone scenario. The multi-scale target image detection system in the drone scenario is provided with an image acquisition module, and the image acquisition module is connected;

[0100] The multi-scale target image detection model ESGN-YOLO in the drone scenario is provided with a feature extraction module BackBone, a multi-scale feature fusion module Neck, and a detection and recognition module Head that are connected in sequence;

[0101] Step 2: The image acquisition module acquires a multi-scale target image X and transfers the multi-scale target image to the multi-scale target image detection model ESGN-YOLO in the drone scenario;

[0102] Step 3: The feature extraction module BackBone in the multi-scale target image detection model ESGN-YOLO in the drone scenario performs feature extraction operations on the multi-scale target image X to obtain image feature data Y, and transfers it to the multi-scale feature fusion module Neck;

[0103] Step 4: The multi-scale feature fusion module Neck performs multi-scale feature fusion operations on the image feature data Y to obtain fusion data Z, and transfers it to the detection and recognition module Head;

[0104] Step 5: The detection and recognition module Head performs detection and recognition operations on the fusion data Z and outputs the target detection result.

[0105] Such as Figure 1As shown in (a), in the step 3, the feature extraction module BackBone is provided with a first basic convolution module CBS-1, a second basic convolution module CBS-2, a first C3k-SWC module, a first efficient feature fusion module EFFM-1, a third basic convolution module CBS-3, a second C3k-SWC module, a fourth basic convolution module CBS-4, a third C3k-SWC module, a fifth basic convolution module CBS-5, a C3k2-SWC module, a second efficient feature fusion module EFFM-2, a fast spatial pyramid pooling module SPPF, and an enhanced convolution module C2PSA, which are connected in sequence.

[0106] The output ends of the second C3k-SWC module, the third C3k-SWC module, and the enhanced convolution module C2PSA are connected to the multi-scale feature fusion module Neck.

[0107] The feature extraction module BackBone performs feature extraction operations on the multi-scale target image X, including the following steps:

[0108] Step 3.1: The first basic convolution module CBS-1 obtains the multi-scale target image X, and sequentially performs convolution, batch normalization, and SiLU activation operations on it to obtain the first convolution data y1, and transmits it to the second basic convolution module CBS-2;

[0109] Step 3.2: The second basic convolution module CBS-2 sequentially performs convolution, batch normalization, and SiLU activation operations on the first convolution data y1 to obtain the second convolution data y2, and transmits it to the first C3k-SWC module;

[0110] Step 3.3: The first C3k-SWC module performs multi-scale convolution operations on the second convolution data y2 to obtain the third convolution data y3, and transmits it to the first efficient feature fusion module EFFM-1;

[0111] Step 3.4: The first efficient feature fusion module EFFM-1 performs efficient feature fusion operations on the third convolution data y3 to obtain the first fusion data y4, and transmits it to the third basic convolution module CBS-3;

[0112] Step 3.5: The third basic convolution module CBS-3 sequentially performs convolution, batch normalization, and SiLU activation operations on the first fusion data y4 to obtain the fourth convolution data y5, and transmits it to the second C3k-SWC module;

[0113] Step 3.6: The second C3k-SWC module performs multi-scale convolution operations on the fourth convolutional data y5 to obtain the fifth convolutional data y6, and transfers it to the fourth basic convolutional module CBS-4 and the multi-scale feature fusion module Neck;

[0114] Step 3.7: The fourth basic convolutional module CBS-4 sequentially performs convolution, batch normalization, and SiLU activation operations on the fifth convolutional data y6 to obtain the sixth convolutional data y7, and transfers it to the third C3k-SWC module;

[0115] Step 3.8: The third C3k-SWC module performs multi-scale convolution operations on the sixth convolutional data y7 to obtain the seventh convolutional data y8, and transfers it to the fifth basic convolutional module CBS-5 and the multi-scale feature fusion module Neck;

[0116] Step 3.9: The fifth basic convolutional module CBS-5 sequentially performs convolution, batch normalization, and SiLU activation operations on the seventh convolutional data y8 to obtain the eighth convolutional data y9, and transfers it to the C3k2-SWC module;

[0117] Step 3.10: The C3k2-SWC module performs deformable convolution and channel separation operations on the eighth convolutional data y9 to obtain the ninth convolutional data y10, and transfers it to the second efficient feature fusion module EFFM-2;

[0118] Step 3.11: The second efficient feature fusion module EFFM-2 performs efficient feature fusion operations on the ninth convolutional data y10 to obtain the second fusion data y11, and transfers it to the fast spatial pyramid pooling module SPPF;

[0119] Step 3.12: The fast spatial pyramid pooling module SPPF performs multi-scale pooling operations on the second fusion data y11 to obtain the pooled data y12, and transfers it to the enhanced convolutional module C2PSA;

[0120] Step 3.13: The enhanced convolutional module C2PSA performs multi-scale feature extraction operations on the pooled data y12 to obtain the image feature data Y, and transfers it to the multi-scale feature fusion module Neck.

[0121] As Figure 1As shown in (f), the first efficient feature fusion module EFFM-1 and the second efficient feature fusion module EFFM-2 have the same structure. Both are provided with an input layer, and the output end of this input layer is connected to the input end of the depthwise separable convolution layer DWconv and the second addition unit; the output end of the depthwise separable convolution layer DWconv is connected to the input ends of the average pooling layer Avgpool, the max pooling layer Maxpool, and the second multiplication module. The output ends of the average pooling layer Avgpool and the max pooling layer Maxpool are connected to the input end of the first addition unit; the output end of the first addition unit is connected to the input ends of the spatial attention mechanism module SA and the channel attention mechanism CA. The output ends of the spatial attention mechanism module SA and the channel attention mechanism CA are connected to the input end of the first multiplication unit; the output end of the first multiplication unit is connected to the input end of the second multiplication unit through the convolution layer Covn. The output end of the second multiplication unit is connected to the input end of the second addition unit, and the output end of the second addition unit is connected to the output layer;

[0122] The calculation expression of the efficient feature fusion module EFFM is as follows:

[0123] F = DWconv(a) (1)

[0124] F global = ∑ d∈{h,v} (Avgpool d (F) + Maxpool d (F)) (2)

[0125] CA = σ(Conv(ReLU(BN.Conv(F global )) / )) (3)

[0126] SA = σ(Conv.Avgpool(F global )) / (4)

[0127] F refined = Conv(SA(F global ) · CA(F global )) (5)

[0128] F output = F · F refined + a (6)

[0129] Among them, a is the input data of the efficient feature fusion module EFFM, BN is the batch normalization operation, σ is the sigmoid function, and F output is the output data of the efficient feature fusion module EFFM; h and v are the two dimensions of the average pooling layer and the max pooling layer.

[0130] Such as Figure 1As shown in (h), the first C3k-SWC module, the second C3k-SWC module, and the third C3k-SWC module have the same structure, and each is provided with an eighth basic convolution module CBS-8. The output end of the eighth basic convolution module CBS-8 is connected to the input end of a fifth-channel splicing module Concat-5 through a ninth basic convolution module CBS-9. Between the output end of the eighth basic convolution module CBS-8 and the input end of the fifth-channel splicing module Concat-5, n sequentially connected SWC-Bottleneck modules are also provided. The output end of the fifth-channel splicing module Concat-5 is connected to a tenth basic convolution module CBS-10;

[0131] As Figure 1 As shown in (g), the C3k2-SWC module is provided with an eleventh basic convolution module CBS-11 and a twelfth basic convolution module CBS-12 in parallel. The output ends of the eleventh basic convolution module CBS-11 and the twelfth basic convolution module CBS-12 are built-in with a Split splitting function;

[0132] The output end of the eleventh basic convolution module CBS-11 is connected to the input end of a sixth-channel splicing module Concat-6. Between the output end of the eleventh basic convolution module CBS-11 and the input end of the sixth-channel splicing module Concat-6, m sequentially connected SWC-Bottleneck modules are also provided. The output ends of the m SWC-Bottleneck modules are all connected to the input end of the sixth-channel splicing module Concat-6. The output end of the sixth-channel splicing module Concat-6 is connected to a thirteenth basic convolution module CBS-13;

[0133] The output end of the twelfth basic convolution module CBS-12 is connected to the input end of a seventh-channel splicing module Concat-7. Between the output end of the twelfth basic convolution module CBS-12 and the input end of the seventh-channel splicing module Concat-7, m sequentially connected C3k-SWC modules C3k-SWC are also provided. The output ends of the m C3k-SWC modules C3k-SWC are all connected to the input end of the seventh-channel splicing module Concat-7. The output end of the seventh-channel splicing module Concat-7 is connected to a fourteenth basic convolution module CBS-14.

[0134] As Figure 2As shown, the SWC-Bottleneck module is provided with an input unit, the output end of which is connected to the input end of the first convolutional layer Conv-1. The output end of the first convolutional layer Conv-1 is connected to the input ends of the first per-channel convolutional block and the second per-channel convolutional block. The output ends of the first per-channel convolutional block and the second per-channel convolutional block are connected to the input end of the addition unit, and the output end of the addition unit is connected to the input end of the residual connection layer. The output end of the input unit is also connected to the input end of the residual connection layer through the second convolutional layer Conv-2;

[0135] The first per-channel convolutional block and the second per-channel convolutional block have the same structure, and are both provided with a mask unit Mask, a split unit Split, a tensor dimension adjustment unit Squeeze, a clipping unit Clip, a splicing unit, and a batch normalization layer BN connected in sequence;

[0136] The calculation expression of the SWC-Bottleneck module is as follows:

[0137]

[0138] nk = ceil(kerner big / kerner small ) (8)

[0139]

[0140] Among them, ° represents element-wise multiplication at the corresponding matrix positions, ceil(·) represents ceiling operation, Clip(·) represents clipping operation, BN(·) represents batch normalization operation, Conv represents convolution operation, Split(·) represents splitting function; i represents the i-th per-channel convolutional block, i = 1, 2; j represents the j-th split block; Q output represents the output of the SWC-Bottleneck module, Q represents the input of the SWC-Bottleneck module, w represents the weight, kerner big represents the largest convolutional kernel size, kerner small represents the smallest convolutional kernel size; σ represents the sigmoid function, and σ in formula (7) is used for mask operation; Squeeze(·) represents tensor compression operation.

[0141] As Figure 1As shown in (b), in the step 4, the multi-scale feature fusion module Neck is provided with a first upsampling module Upsample-1, a first channel splicing module Concat-1, a first C3k2 module, a second upsampling module Upsample-2, a second channel splicing module Concat-2, a second C3k2 module, a sixth basic convolution module CBS-6, a third channel splicing module Concat-3, a third C3k2 module, a seventh basic convolution module CBS-7, a fourth channel splicing module Concat-4, and a fourth C3k2 module, which are connected in sequence;

[0142] The input ends of the first upsampling module Upsample-1, the first channel splicing module Concat-1, the second channel splicing module Concat-2, and the fourth channel splicing module Concat-4 are connected to the feature extraction module BackBone; the output end of the first C3k2 module is connected to the input end of the third channel splicing module Concat-3; the output ends of the second C3k2 module, the third C3k2 module, and the fourth C3k2 module are connected to the detection and recognition module Head.

[0143] The multi-scale feature fusion module Neck performs multi-scale feature fusion operations on the image feature data Y, including the following steps:

[0144] Step 4.1: The first upsampling module Upsample-1 obtains the image feature data Y, performs an upsampling operation on it to obtain a first sampling data z1, and transmits it to the first channel splicing module Concat-1;

[0145] Step 4.2: The first channel splicing module Concat-1 performs a channel splicing operation on the first sampling data z1 and the seventh convolution data y8 to obtain a first splicing data z2, and transmits it to the first C3k2 module;

[0146] Step 4.3: The first C3k2 module performs deformable convolution and channel separation operations on the first splicing data z2 to obtain a first deformable convolution data z3, and transmits it to the second upsampling module Upsample-2 and the third channel splicing module Concat-3;

[0147] Step 4.4: The second upsampling module Upsample-2 performs an upsampling operation on the first deformable convolution data z3 to obtain a second sampling data z4, and transmits it to the second channel splicing module Concat-2;

[0148] Step 4.5: The second channel concatenation module Concat-2 performs a channel concatenation operation on the second sampled data z4 and the fifth convolutional data y6 to obtain the second concatenated data z5, and transfers it to the second C3k2 module;

[0149] Step 4.6: The second C3k2 module performs deformable convolution and channel separation operations on the second concatenated data z5 to obtain the second deformable convolutional data z6, and transfers it to the sixth basic convolutional module CBS-6 and the detection and recognition module Head;

[0150] Step 4.7: The sixth basic convolutional module CBS-6 sequentially performs convolution, batch normalization, and SiLU activation operations on the second deformable convolutional data z6 to obtain the tenth convolutional data z7, and transfers it to the third channel concatenation module Concat-3;

[0151] Step 4.8: The third channel concatenation module Concat-3 performs a channel concatenation operation on the tenth convolutional data z7 and the first deformable convolutional data z3 to obtain the third concatenated data z8, and transfers it to the third C3k2 module;

[0152] Step 4.9: The third C3k2 module performs deformable convolution and channel separation operations on the third concatenated data z8 to obtain the third deformable convolutional data z9, and transfers it to the seventh basic convolutional module CBS-7 and the detection and recognition module Head;

[0153] Step 4.10: The seventh basic convolutional module CBS-7 sequentially performs convolution, batch normalization, and SiLU activation operations on the third deformable convolutional data z9 to obtain the eleventh convolutional data z10, and transfers it to the fourth channel concatenation module Concat-4;

[0154] Step 4.11: The fourth channel concatenation module Concat-4 performs a channel concatenation operation on the eleventh convolutional data z10 and the image feature data Y to obtain the fourth concatenated data z11, and transfers it to the fourth C3k2 module;

[0155] Step 4.12: The fourth C3k2 module performs deformable convolution and channel separation operations on the fourth concatenated data z11 to obtain the fusion data Z, and transfers it to the detection and recognition module Head.

[0156] As Figure 1As shown in (c), the detection and recognition module Head is provided with a first convolutional group normalization block CGS-1, a second convolutional group normalization block CGS-2, and a third convolutional group normalization block CGS-3. The input ends of the first convolutional group normalization block CGS-1, the second convolutional group normalization block CGS-2, and the third convolutional group normalization block CGS-3 are connected to the multi-scale feature fusion module Neck, and the output ends of the three are connected to the input end of the fourth convolutional group normalization block CGS-4. The output end of the fourth convolutional group normalization block CGS-4 is connected to the input ends of the small target detection head BSC-1, the medium target detection head BSC-2, and the large target detection head BSC-3, and the output ends of the small target detection head BSC-1, the medium target detection head BSC-2, and the large target detection head BSC-3 output the target detection results.

[0157] The detection and recognition module Head performs detection and recognition operations on the fusion data Z, and the steps are as follows:

[0158] Step 5.1: The first convolutional group normalization block CGS-1 obtains the fusion data Z, performs convolution and group normalization operations on it, and then outputs the first group of normalized data c1 to the fourth convolutional group normalization block CGS-4 through the SiLU activation function;

[0159] The second convolutional group normalization block CGS-2 takes the third variable convolution data z9, performs convolution and group normalization operations on it, and then outputs the second group of normalized data c2 to the fourth convolutional group normalization block CGS-4 through the SiLU activation function;

[0160] The third convolutional group normalization block CGS-3 takes the second variable convolution data z6, performs convolution and group normalization operations on it, and then outputs the third group of normalized data c3 to the fourth convolutional group normalization block CGS-4 through the SiLU activation function;

[0161] Step 5.2: The fourth convolutional group normalization block CGS-4 performs convolution and group normalization operations on the first group of normalized data c1, the second group of normalized data c2, and the third group of normalized data c3, and then outputs the fourth group of normalized data c4 to the small target detection head BSC-1, the medium target detection head BSC-2, and the large target detection head BSC-3 through the SiLU activation function;

[0162] Step 5.3: The small target detection head BSC-1 performs small target detection on the fourth group of normalized data c4 and outputs the small target detection result;

[0163] The medium target detection head BSC-2 performs medium target detection on the fourth group of normalized data c4 and outputs the medium target detection result;

[0164] The large object detection head BSC-3 performs large object detection on the fourth set of normalized data c4 and outputs the large object detection result.

[0165] As Figure 1 (d) shows, the small / medium / large object detection heads are provided with a decoupled classification Conv-Cls module and a regression module Conv-Box in parallel, and the output end of the regression module Conv-Box is connected to the input end of the model scaling ratio control module Scale.

[0166] As Figure 1 (e) shows, the convolutional group normalization block CGS is provided with a convolutional layer and a group normalization layer connected in sequence, and a SiLU activation function is built in at the output end of the group normalization layer.

[0167] To evaluate the performance of the present invention, experiments are carried out in this embodiment on two representative datasets, VisDrone2019 and TinyPerson.

[0168] VisDrone2019: The VisDrone2019 dataset is collected by the AISKYEYE team of Tianjin University, covering 14 cities, including urban and rural areas. The scenes contain targets with different densities, such as pedestrians and vehicles. The data is collected under different drone platforms, weather conditions and lighting environments. In this paper, the officially divided training set and validation set are used for experiments.

[0169] TinyPerson: The TinyPerson dataset comes from the Internet, and the targets in the images are mainly people near the seaside. This dataset contains 1,610 annotated images, of which 794 are used for training and 816 are used for testing. The resolution of most images is 1920×1080. Each image contains an average of 25 small targets, accounting for no more than 0.87% of the total pixels. In addition, the size of all small targets does not exceed 20×20, and the target distribution is relatively sparse.

[0170] This experiment is based on the open-source deep learning framework Pytorch for machine learning and deep learning, and runs on an NVIDIA GeForce RTX 4070 SUPER GPU. The input image size of the VisDrone and TinyPerson datasets is set to 640×640 pixels, the batch size is 32, and the number of training epochs is 200 to optimize the training accuracy of the model. The optimizer uses Stochastic Gradient Descent (SGD), and the initial learning rate is set to 0.01.

[0171] This experiment uses the VisDrone2019 and TinyPerson datasets for evaluation. The experimental results and their comparison with YOLO series models and other enhanced models show that the model proposed in the present invention is superior to YOLOv11n in small object detection.

[0172] Table 1. Comparison of mAP@0.5 of different networks on the VisDrone dataset

[0173]

[0174] VisDrone2019: As shown in Table 1, the present invention is 0.3% higher than the second-ranked model PP-PicoDet-L in terms of the mAP@0.5 metric. The mAP@0.5 of the "bus" and "car" categories is 2.4% and 1.8% higher than that of the second-ranked model respectively. Compared with YOLOv11n, the overall mAP@0.5 has increased by 1.9%. Except for a slight decrease in the mAP@0.5 of the "awning-tricycle" and "bicycle" categories, the detection performance of other categories has improved. The results show that ESGN-YOLOv11n has good applicability in the small object detection task of drone aerial images.

[0175] Table 2. Comparison of mAP@0.5 of different networks on the TinyPerson dataset

[0176]

[0177] TinyPerson: As shown in Table 2, the present invention is superior to the SOTA method on the validation set. The mAP@0.5 of GNSH-YOLOv11n reaches 20.3%, which is 0.3% higher than the second-ranked model Sf-YOLOv5. In addition, the mAP@0.5 of the "earth-person" and "sea-person" categories reaches 15.5% and 25% respectively, still higher than other models. The overall mAP@0.5 and the mAP@0.5 of "earth-person" and "sea-person" have increased by 1%, 1.6% and 0.3% respectively. The experimental results prove that ESGN-YOLO performs better in the small object detection task.

[0178] Table 3. Ablation experiment on the Visdrone dataset

[0179]

[0180] To evaluate the impact of the present invention, ablation experiments were conducted on the VisDrone2019 validation set, focusing on analyzing three key improvements: introducing the EFFM module, modifying the C3k2 module, and replacing the detection head with GNSH. Table 3 summarizes the experimental results, showing that after introducing the EFFM module, mAP@0.5 and mAP@[0.5:0.95] increased by 1.3% and 1.1% respectively, with significant effects. The impact of modifying the SWC module alone was relatively small, with mAP@0.5 only increasing by 0.3% while mAP@[0.5:0.95] remaining unchanged. Combining SWC and EFFM (S+E) brought a significant performance improvement compared to using SWC alone (+SWC(S)). Specifically, after adding E+S, mAP@0.5 increased by 1.8 percentage points from the baseline, and mAP@[0.5:0.95] increased by 1.4 percentage points. After adopting GNSH alone, the recall rate increased, and the number of model parameters decreased from 2.6M to 2.4M, but mAP was not improved. Finally, combining GNSH with EFFM and SWC further improved the recall rate and mAP@[0.5:0.95], while significantly reducing the number of parameters (from 3.9M to 3.7M), but the precision decreased slightly.

[0181] Next, the heatmap of the final backbone network layer was analyzed to evaluate the feature extraction and fusion performance of the improved model. As Figure 3 shown, Figure 3 the first row in it shows the detection results of ESGN-YOLO, and the second row shows the detection results of YOLOv11n. Compared with YOLOv11n(b), ESGN-YOLOv11n(c) pays more attention to small targets and has stronger brightness in the region of interest, indicating that the detection ability of this model has been improved in complex background and occlusion scenarios. In addition, experiments were conducted on YOLOv11 and ESGN-YOLOv11n to further verify the effectiveness of this method in four scenarios: urban parking lots, intersections, traffic roads, and environments with poor lighting conditions. These scenarios contain a large number of small targets, as Figure 4 shown, Figure 4 the first row in it shows the detection results of ESGN-YOLO, and the second row shows the detection results of YOLOv11n. The model of the present invention can detect more small and occluded targets, and its detection performance is better than the original version.

[0182] The present invention proposes ESGN-YOLOv11n, which is an enhanced version of YOLOv11n, aiming to improve the detection performance of small targets in images. The main contributions of the present invention include the following three points and are verified through a large number of experiments. First, the present invention integrates the EFFM module in the backbone network to enhance the extraction ability of directional and multi-scale information. Second, the Bottleneck module in the C3k2 and C3k structures is replaced with SWC-Bottleneck to improve the feature granularity of the model and enhance the ability to capture local features. Finally, the original detection head is replaced with the group normalization scale detection head GNSH to reduce resource consumption. The experimental results verify the effectiveness of the proposed method.

[0183] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A multi-scale target image detection method in the scenario of an unmanned aerial vehicle, characterized in that, It includes the following steps: Step 1: Construct a multi-scale target image detection system in the UAV scenario. The multi-scale target image detection system in the UAV scenario is provided with an image acquisition module, and the image acquisition module is connected; The multi-scale target image detection model ESGN-YOLO in the UAV scenario is provided with a feature extraction module BackBone, a multi-scale feature fusion module Neck, and a detection and recognition module Head that are connected in sequence; Step 2: The image acquisition module acquires a multi-scale target image X and transmits the multi-scale target image to the multi-scale target image detection model ESGN-YOLO in the UAV scenario; Step 3: The feature extraction module BackBone in the multi-scale target image detection model ESGN-YOLO in the UAV scenario performs feature extraction operations on the multi-scale target image X to obtain image feature data Y and transmits it to the multi-scale feature fusion module Neck; Step 4: The multi-scale feature fusion module Neck performs multi-scale feature fusion operations on the image feature data Y to obtain fusion data Z and transmits it to the detection and recognition module Head; Step 5: The detection and recognition module Head performs detection and recognition operations on the fusion data Z and outputs the target detection result.

2. The multi-scale target image detection method in the UAV scenario according to claim 1, characterized in that: In Step 3, the feature extraction module BackBone is provided with a first basic convolution module CBS-1, a second basic convolution module CBS-2, a first C3k-SWC module, a first efficient feature fusion module EFFM-1, a third basic convolution module CBS-3, a second C3k-SWC module, a fourth basic convolution module CBS-4, a third C3k-SWC module, a fifth basic convolution module CBS-5, a C3k2-SWC module, a second efficient feature fusion module EFFM-2, a fast spatial pyramid pooling module SPPF, and an enhanced convolution module C2PSA that are connected in sequence; The output ends of the second C3k-SWC module, the third C3k-SWC module, and the enhanced convolution module C2PSA are connected to the multi-scale feature fusion module Neck.

3. The multi-scale target image detection method in the UAV scenario according to claim 2, wherein: The feature extraction module BackBone performs feature extraction operations on the multi-scale target image X, including the following steps: Step 3.1: The first basic convolution module CBS-1 acquires the multi-scale target image X and sequentially performs convolution, batch normalization, and SiLU activation operations on it to obtain first convolution data y1 and transmits it to the second basic convolution module CBS-2; Step 3.2: The second basic convolution module CBS-2 sequentially performs convolution, batch normalization, and SiLU activation operations on the first convolution data y1 to obtain second convolution data y2 and transmits it to the first C3k-SWC module; Step 3.3: The first C3k-SWC module performs multi-scale convolution operations on the second convolution data y2 to obtain third convolution data y3 and transmits it to the first efficient feature fusion module EFFM-1; Step 3.4: The first efficient feature fusion module EFFM-1 performs an efficient feature fusion operation on the third convolutional data y3 to obtain the first fusion data y4, and transfers it to the third basic convolutional module CBS-3; Step 3.5: The third basic convolutional module CBS-3 sequentially performs convolution, batch normalization, and SiLU activation operations on the first fusion data y4 to obtain the fourth convolutional data y5, and transfers it to the second C3k-SWC module; Step 3.6: The second C3k-SWC module performs a multi-scale convolution operation on the fourth convolutional data y5 to obtain the fifth convolutional data y6, and transfers it to the fourth basic convolutional module CBS-4 and the multi-scale feature fusion module Neck; Step 3.7: The fourth basic convolutional module CBS-4 sequentially performs convolution, batch normalization, and SiLU activation operations on the fifth convolutional data y6 to obtain the sixth convolutional data y7, and transfers it to the third C3k-SWC module; Step 3.8: The third C3k-SWC module performs a multi-scale convolution operation on the sixth convolutional data y7 to obtain the seventh convolutional data y8, and transfers it to the fifth basic convolutional module CBS-5 and the multi-scale feature fusion module Neck; Step 3.9: The fifth basic convolutional module CBS-5 sequentially performs convolution, batch normalization, and SiLU activation operations on the seventh convolutional data y8 to obtain the eighth convolutional data y9, and transfers it to the C3k2-SWC module; Step 3.10: The C3k2-SWC module performs deformable convolution and channel separation operations on the eighth convolutional data y9 to obtain the ninth convolutional data y10, and transfers it to the second efficient feature fusion module EFFM-2; Step 3.11: The second efficient feature fusion module EFFM-2 performs an efficient feature fusion operation on the ninth convolutional data y10 to obtain the second fusion data y11, and transfers it to the fast spatial pyramid pooling module SPPF; Step 3.12: The fast spatial pyramid pooling module SPPF performs a multi-scale pooling operation on the second fusion data y11 to obtain the pooled data y12, and transfers it to the enhanced convolutional module C2PSA; Step 3.13: The enhanced convolutional module C2PSA performs a multi-scale feature extraction operation on the pooled data y12 to obtain the image feature data Y, and transfers it to the multi-scale feature fusion module Neck.

4. The multi-scale target image detection method in the UAV scenario according to claim 2 or 3, characterized in that: The first efficient feature fusion module EFFM-1 and the second efficient feature fusion module EFFM-2 have the same structure, both of which are provided with an input layer, and the output end of the input layer is connected to the input end of the depthwise separable convolution layer DWconv and the second addition unit; the output end of the depthwise separable convolution layer DWconv is connected to the input ends of the average pooling layer Avgpool, the max pooling layer Maxpool and the second multiplication module, and the output ends of the average pooling layer Avgpool and the max pooling layer Maxpool are connected to the input end of the first addition unit; the output end of the first addition unit is connected to the input ends of the spatial attention mechanism module SA and the channel attention mechanism CA, and the output ends of the spatial attention mechanism module SA and the channel attention mechanism CA are connected to the input end of the first multiplication unit; the output end of the first multiplication unit is connected to the input end of the second multiplication unit through the convolution layer Covn, the output end of the second multiplication unit is connected to the input end of the second addition unit, and the output end of the second addition unit is connected to the output layer; The calculation expression of the efficient feature fusion module EFFM is as follows: F = DWconv(a) (1) F global = ∑ d∈{h,v} (Avgpool d (F) + Maxpool d (F)) (2) CA = σ(Conv(ReLU(BN.Conv(F global ) / ))) (3) SA = σ(Conv.Avgpool(F global ) / ) (4) F refined = Conv(SA(F global )·CA(F global )) (5) F output = F·F refined + a(6) Among them, a is the input data of the efficient feature fusion module EFFM, BN is the batch normalization operation, σ is the sigmoid function, and F output is the output data of the efficient feature fusion module EFFM; h and v are the two dimensions of the average pooling layer and the max pooling layer.

5. The multi-scale target image detection method in the UAV scenario according to claim 2 or 3, characterized in that: The first C3k-SWC module, the second C3k-SWC module and the third C3k-SWC module have the same structure, and are all provided with the eighth basic convolution module CBS-8. The output end of the eighth basic convolution module CBS-8 is connected to the input end of the fifth channel concatenation module Concat-5 through the ninth basic convolution module CBS-9. Between the output end of the eighth basic convolution module CBS-8 and the input end of the fifth channel concatenation module Concat-5, n sequentially connected SWC-Bottleneck modules are also provided, and the output end of the fifth channel concatenation module Concat-5 is connected to the tenth basic convolution module CBS-10; The C3k2-SWC module is provided with the eleventh basic convolution module CBS-11 and the twelfth basic convolution module CBS-12 in parallel, and the Split splitting function is built in the output ends of the eleventh basic convolution module CBS-11 and the twelfth basic convolution module CBS-12; The output end of the eleventh basic convolution module CBS-11 is connected to the input end of the sixth channel concatenation module Concat-6. Between the output end of the eleventh basic convolution module CBS-11 and the input end of the sixth channel concatenation module Concat-6, m sequentially connected SWC-Bottleneck modules are also provided, and the output ends of the m SWC-Bottleneck modules are all connected to the input end of the sixth channel concatenation module Concat-6; the output end of the sixth channel concatenation module Concat-6 is connected to the thirteenth basic convolution module CBS-13; The output end of the twelfth basic convolution module CBS-12 is connected to the input end of the seventh channel concatenation module Concat-7. Between the output end of the twelfth basic convolution module CBS-12 and the input end of the seventh channel concatenation module Concat-7, there are also m sequentially connected C3k-SWC modules C3k-SWC. The output ends of the m C3k-SWC modules C3k-SWC are all connected to the input end of the seventh channel concatenation module Concat-7; the output end of the seventh channel concatenation module Concat-7 is connected to the fourteenth basic convolution module CBS-14.

6. The multi-scale target image detection method in the UAV scenario according to claim 1, characterized in that: The SWC-Bottleneck module is provided with an input unit. The output end of the input unit is connected to the input end of the first convolutional layer Conv-1. The output end of the first convolutional layer Conv-1 is connected to the input ends of the first per-channel convolutional block and the second per-channel convolutional block. The output ends of the first per-channel convolutional block and the second per-channel convolutional block are connected to the input end of the addition unit. The output end of the addition unit is connected to the input end of the residual connection layer; the output end of the input unit is also connected to the input end of the residual connection layer through the second convolutional layer Conv-2; The first per-channel convolutional block and the second per-channel convolutional block have the same structure, and both are provided with a mask unit Mask, a splitting unit Split, a tensor dimension adjustment unit Squeeze, a clipping unit Clip, a concatenation unit, and a batch normalization layer BN connected in sequence; The calculation expression of the SWC-Bottleneck module is as follows: nk = ceil(kerner bi / kerner small ) (8) Among them, ° represents multiplication of corresponding matrix positions, ceil(·) represents ceiling operation, Clip(·) represents clipping operation, BN(·) represents normalization operation, Conv represents convolution operation, Split(·) represents splitting function; i represents the i-th per-channel convolution block, i = 1, 2; j represents the j-th segmentation block; Q output represents the output of the SWC-Bottleneck module, Q represents the input of the SWC-Bottleneck module, w represents the weight, kerner big represents the largest convolution kernel size, kerner small represents the smallest convolution kernel size; σ represents the sigmoid function, and σ in formula (7) is used for mask operation; Squeeze(·) represents tensor compression operation.

7. The multi-scale target image detection method in the UAV scenario according to claim 1, characterized in that: In step 4, the multi-scale feature fusion module Neck is provided with a first upsampling module Upsample-1, a first channel concatenation module Concat-1, a first C3k2 module, a second upsampling module Upsample-2, a second channel concatenation module Concat-2, a second C3k2 module, a sixth basic convolution module CBS-6, a third channel concatenation module Concat-3, a third C3k2 module, a seventh basic convolution module CBS-7, a fourth channel concatenation module Concat-4, and a fourth C3k2 module connected in sequence; The input ends of the first upsampling module Upsample-1, the first channel concatenation module Concat-1, the second channel concatenation module Concat-2, and the fourth channel concatenation module Concat-4 are connected to the feature extraction module BackBone; the output end of the first C3k2 module is connected to the input end of the third channel concatenation module Concat-3; the output ends of the second C3k2 module, the third C3k2 module, and the fourth C3k2 module are connected to the detection and recognition module Head.

8. The multi-scale target image detection method in the UAV scenario according to claim 7, characterized in that: The multi-scale feature fusion module Neck performs multi-scale feature fusion operations on the image feature data Y, including the following steps: Step 4.1: The first upsampling module Upsample-1 obtains the image feature data Y, performs upsampling operations on it to obtain the first sampled data z1, and transmits it to the first channel concatenation module Concat-1; Step 4.2: The first channel concatenation module Concat-1 performs a channel concatenation operation on the first sampled data z1 and the seventh convolutional data y8 to obtain the first concatenated data z2, and transfers it to the first C3k2 module; Step 4.3: The first C3k2 module performs deformable convolution and channel separation operations on the first concatenated data z2 to obtain the first deformable convolutional data z3, and transfers it to the second upsampling module Upsample-2 and the third channel concatenation module Concat-3; Step 4.4: The second upsampling module Upsample-2 performs an upsampling operation on the first deformable convolutional data z3 to obtain the second sampled data z4, and transfers it to the second channel concatenation module Concat-2; Step 4.5: The second channel concatenation module Concat-2 performs a channel concatenation operation on the second sampled data z4 and the fifth convolutional data y6 to obtain the second concatenated data z5, and transfers it to the second C3k2 module; Step 4.6: The second C3k2 module performs deformable convolution and channel separation operations on the second concatenated data z5 to obtain the second deformable convolutional data z6, and transfers it to the sixth basic convolutional module CBS-6 and the detection and recognition module Head; Step 4.7: The sixth basic convolutional module CBS-6 sequentially performs convolution, batch normalization, and SiLU activation operations on the second deformable convolutional data z6 to obtain the tenth convolutional data z7, and transfers it to the third channel concatenation module Concat-3; Step 4.8: The third channel concatenation module Concat-3 performs a channel concatenation operation on the tenth convolutional data z7 and the first deformable convolutional data z3 to obtain the third concatenated data z8, and transfers it to the third C3k2 module; Step 4.9: The third C3k2 module performs deformable convolution and channel separation operations on the third concatenated data z8 to obtain the third deformable convolutional data z9, and transfers it to the seventh basic convolutional module CBS-7 and the detection and recognition module Head; Step 4.10: The seventh basic convolutional module CBS-7 sequentially performs convolution, batch normalization, and SiLU activation operations on the third deformable convolutional data z9 to obtain the eleventh convolutional data z10, and transfers it to the fourth channel concatenation module Concat-4; Step 4.11: The fourth channel concatenation module Concat-4 performs a channel concatenation operation on the eleventh convolutional data z10 and the image feature data Y to obtain the fourth concatenated data z11, and transfers it to the fourth C3k2 module; Step 4.12: The fourth C3k2 module performs deformable convolution and channel separation operations on the fourth concatenated data z11 to obtain the fusion data Z, and transfers it to the detection and recognition module Head.

9. The multi-scale target image detection method in the UAV scenario according to claim 1, characterized in that: The detection and recognition module Head is provided with a first convolutional group normalization block CGS-1, a second convolutional group normalization block CGS-2, and a third convolutional group normalization block CGS-3. The input ends of the first convolutional group normalization block CGS-1, the second convolutional group normalization block CGS-2, and the third convolutional group normalization block CGS-3 are connected to the multi-scale feature fusion module Neck, and the output ends of the three are connected to the input end of the fourth convolutional group normalization block CGS-4. The output end of the fourth convolutional group normalization block CGS-4 is connected to the input ends of the small target detection head BSC-1, the medium target detection head BSC-2, and the large target detection head BSC-3. The output ends of the small target detection head BSC-1, the medium target detection head BSC-2, and the large target detection head BSC-3 output the target detection results.

10. The multi-scale target image detection method in the UAV scenario according to claim 9, characterized in that: The detection and recognition module Head performs detection and recognition operations on the fusion data Z, and the steps are as follows: Step 5.1: The first convolutional group normalization block CGS-1 obtains the fusion data Z, performs convolution and group normalization operations on it, and then outputs the first group of normalized data c1 to the fourth convolutional group normalization block CGS-4 through the SiLU activation function; The second convolutional group normalization block CGS-2 takes the third variable convolutional data z9, performs convolution and group normalization operations on it, and then outputs the second group of normalized data c2 to the fourth convolutional group normalization block CGS-4 through the SiLU activation function; The third convolutional group normalization block CGS-3 takes the second variable convolutional data z6, performs convolution and group normalization operations on it, and then outputs the third group of normalized data c3 to the fourth convolutional group normalization block CGS-4 through the SiLU activation function; Step 5.2: The fourth convolutional group normalization block CGS-4 performs convolution and group normalization operations on the first group of normalized data c1, the second group of normalized data c2, and the third group of normalized data c3, and then outputs the fourth group of normalized data c4 to the small target detection head BSC-1, the medium target detection head BSC-2, and the large target detection head BSC-3 through the SiLU activation function; Step 5.3: The small target detection head BSC-1 performs small target detection on the fourth group of normalized data c4 and outputs the small target detection result; The medium target detection head BSC-2 performs medium target detection on the fourth group of normalized data c4 and outputs the medium target detection result; The large target detection head BSC-3 performs large target detection on the fourth group of normalized data c4 and outputs the large target detection result.

Images