Multi-scale target image detection method in unmanned airport scenario

Abstract

A multi-scale target image detection method in an unmanned airport scene, characterized in that it comprises the following steps: step 1: constructing a multi-scale target image detection system in an unmanned airport scene; step 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 airport scene; step 3: a feature extraction module in the multi-scale target image detection model in the unmanned airport scene performs feature extraction operation on the multi-scale target image X to obtain image feature data Y and transmits it to a multi-scale feature fusion module; step 4: the multi-scale feature fusion module performs multi-scale feature fusion operation on the image feature data Y to obtain fusion data Z and transmits it to a detection and recognition module; step 5: the detection and recognition module performs detection and recognition operation on the fusion data Z and outputs the target detection result. Effect: The model's ability to extract multi-scale and directional information is enhanced.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to a multi-scale target image detection method in a drone scenario. Background Technology

[0002] With continuous technological advancements, the flexibility, flight capabilities, and wide coverage of unmanned aerial vehicles (UAVs) have significantly improved, enhancing their effectiveness in aerial reconnaissance missions and making them a prominent focus of artificial intelligence research. This technology primarily covers multiple fields, including wilderness rescue, agricultural plant protection, forest fire early warning, and traffic safety monitoring. However, despite its numerous advantages, UAV technology also faces several challenges. Targets captured in high-altitude operations are typically small, with complex backgrounds and constantly changing aerial environments. Given the high flight speed of UAVs, they are sensitive to external factors such as position and lighting, leading to blurred aerial images and increased noise. Therefore, changes in the shooting angle degrade image quality and increase complexity. To better accomplish the various tasks related to UAV technology, improving the detection capabilities of UAVs in high-altitude environments is crucial.

[0003] Traditional object detection methods primarily rely on object color and texture, or utilize Histogram-Oriented Gradient (HOG) feature extraction techniques to identify objects in images. HOG feature extraction mainly analyzes local gradients in the image and employs Support Vector Machines (SVMs) to detect object shapes. These traditional methods are typically time-consuming and cumbersome, exhibiting suboptimal performance when handling multi-scale targets. With increased computing power, neural networks have emerged as a remedy for these limitations, reducing reliance on manual labor by automatically learning target features. Currently, data augmentation techniques or feature fusion methods are commonly used to address image quality degradation. Furthermore, the feature extraction process has been improved, and various label assignment strategies have been explored. However, these methods have limitations in mitigating the noise specific to UAV aerial photography. One of the most effective solutions is to increase image resolution. However, this approach significantly increases the computational load on GPUs.

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

[0005] This invention provides a multi-scale target image detection method in UAV scenarios, which enhances the model's ability to extract multi-scale and directional information.

[0006] To achieve the above objectives, this invention provides a multi-scale target image detection method in unmanned aerial vehicle (UAV) scenarios, which, as a key feature, includes the following steps:

[0007] Step 1: Construct a multi-scale target image detection system for UAV scenarios. The multi-scale target image detection system for UAV scenarios is equipped with an image acquisition module, which is connected to;

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

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

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

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

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

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

[0014] Preferably, in step 3, the feature extraction module BackBone is provided with the following modules connected in sequence: 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 outputs of the second C3k-SWC module, the third C3k-SWC module, and the enhanced convolutional module C2PSA are connected to the multi-scale feature fusion module Neck.

[0016] The Fast Spatial Pyramid Pooling (SPPF) module is used to extract contextual features through pooling operations of different sizes, and features can be extracted across different receptive fields.

[0017] The enhanced convolutional module C2PSA combines channel-dimensional attention and spatial-dimensional attention to explore the relationships between features from two dimensions, thereby enhancing the network's ability to model both local and global aspects.

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

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

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

[0021] Step 3.3: The first C3k-SWC module performs multi-scale convolution operation on the second convolutional data y2 to obtain the third convolutional data y3, and passes 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 operation on the third convolutional data y3 to obtain the first fused data y4, and passes it to the third basic convolutional module CBS-3;

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

[0024] Step 3.6: The second C3k-SWC module performs multi-scale convolution operation on the fourth convolutional data y5 to obtain the fifth convolutional data y6, and passes 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 performs convolution, batch normalization, and SiLU activation operations on the fifth convolution data y6 in sequence to obtain the sixth convolution data y7, and then passes it to the third C3k-SWC module;

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

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

[0028] Step 3.10: The C3k2-SWC module performs variable convolution and channel separation operations on the eighth convolutional data y9 to obtain the ninth convolutional data y10, and passes 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 operation on the ninth convolutional data y10 to obtain the second fused data y11, and passes it to the fast spatial pyramid pooling module SPPF.

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

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

[0032] Preferably, the first high-efficiency feature fusion module EFFM-1 and the second high-efficiency feature fusion module EFFM-2 have the same structure, both having an input layer. The output of this input layer is connected to the input of the depthwise separable convolutional layer DWconv and the second addition unit. The output of the depthwise separable convolutional layer DWconv is connected to the input of the average pooling layer Avgpool, the max pooling layer Maxpool, and the second multiplication module. The outputs of the average pooling layer Avgpool and the max pooling layer Maxpool are connected to the input of the first addition unit. The output of the first addition unit is connected to the input of the spatial attention mechanism module SA and the channel attention mechanism CA. The outputs of the spatial attention mechanism module SA and the channel attention mechanism CA are connected to the input of the first multiplication unit. The output of the first multiplication unit is connected to the input of the second multiplication unit via the convolutional layer Covn. The output of the second multiplication unit is connected to the input of the second addition unit. The output of the second addition unit is connected to the output layer.

[0033] The calculation expression for 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] Where a represents the input data of the efficient feature fusion module EFFM, BN represents the batch normalization operation, σ represents the sigmoid function, and σ is used for activation operations in formulas (3) and (4); F output The output data is the efficient feature fusion module EFFM; h and v are the two dimensions of the average pooling layer and the max pooling layer, respectively.

[0041] In the first efficient feature fusion module EFFM-1, a represents the third convolutional data y3, and F... output The first fused data is y4;

[0042] In the second efficient feature fusion module EFFM-2, 'a' represents the ninth convolutional data y10, and F... output The second fused data is y11.

[0043] Traditional small object detection methods typically integrate attention mechanisms into the backbone network, but struggle with multi-scale feature extraction. This invention introduces the efficient feature fusion module EFFM into the backbone network, effectively enhancing the model's multi-scale feature learning and context-aware capabilities.

[0044] In the efficient feature fusion module EFFM, firstly, separable convolutions enhance feature extraction, improve model accuracy, and reduce redundant computation. Next, convolutional 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 preserves initial information and alleviates the gradient vanishing problem.

[0045] The efficient feature fusion module EFFM optimizes the feature extraction process through multi-level feature fusion, attention mechanism and residual connection, which significantly improves model performance in tasks such as small target detection.

[0046] Preferably, the first, second, and third C3k-SWC modules have the same structure, each equipped with an eighth basic convolutional module CBS-8. The output of the eighth basic convolutional module CBS-8 is connected to the input of the fifth channel stitching module Concat-5 via a ninth basic convolutional module CBS-9. There are also n sequentially connected SWC-Bottleneck modules between the output of the eighth basic convolutional module CBS-8 and the input of the fifth channel stitching module Concat-5. The output of the fifth channel stitching module Concat-5 is connected to the tenth basic convolutional module CBS-10.

[0047] The C3k2-SWC module has an eleventh basic convolutional module CBS-11 and a twelfth basic convolutional module CBS-12 arranged in parallel. The output terminals of the eleventh basic convolutional module CBS-11 and the twelfth basic convolutional module CBS-12 have a built-in Split function.

[0048] The output of the eleventh basic convolutional module CBS-11 is connected to the input of the sixth channel stitching module Concat-6. Between the output of the eleventh basic convolutional module CBS-11 and the input of the sixth channel stitching module Concat-6, m sequentially connected SWC-Bottleneck modules are also provided, with the outputs of each of the m SWC-Bottleneck modules connected to the input of the sixth channel stitching module Concat-6. The output of the sixth channel stitching module Concat-6 is connected to the thirteenth basic convolutional module CBS-13.

[0049] The output of the twelfth basic convolutional module CBS-12 is connected to the input of the seventh-channel stitching module Concat-7. Between the output of the twelfth basic convolutional module CBS-12 and the input of the seventh-channel stitching module Concat-7, m C3k-SWC modules C3k-SWC are connected in sequence. The outputs of the m C3k-SWC modules C3k-SWC are all connected to the input of the seventh-channel stitching module Concat-7. The output of the seventh-channel stitching module Concat-7 is connected to the fourteenth basic convolutional module CBS-14.

[0050] In YOLOv11, the C3k2 and C3k modules rely on simple convolutional operations, lack advanced feature fusion, and are prone to misinterpreting important information, especially in small object detection. To address this issue, this invention proposes the SWC-Bottleneck module to replace the bottleneck modules in C3k2 and C3k.

[0051] Preferably, the SWC-Bottleneck module includes an input unit whose output is connected to the input of a first convolutional layer Conv-1. The output of the first convolutional layer Conv-1 is connected to the inputs of a first progressive convolutional block and a second progressive convolutional block. The outputs of the first and second progressive convolutional blocks are connected to the input of an addition unit, and the output of the addition unit is connected to the input of a residual connection layer. The output of the input unit is also connected to the input of the residual connection layer via a second convolutional layer Conv-2.

[0052] The first and second channel-wise convolutional blocks have the same structure, both of which are provided with a mask unit, a split unit, a tensor dimension adjustment unit, a clipping unit, a splicing unit, and a normalization layer (BN) connected in sequence.

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

[0054]

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

[0056]

[0057] Where ° represents the positional multiplication of corresponding matrices, ceil(·) represents rounding up, Clip(·) represents the clipping operation, BN(·) represents the normalization operation, Conv represents the convolution operation, and Split(·) represents the splitting function; i represents the i-th channel-wise convolutional block, i = 1, 2; j represents the j-th segmentation block; Q output The output of the SWC-Bottleneck module is represented by Q, the input of the SWC-Bottleneck module is represented by w, and the kernel is represented by kerner. big Represents the maximum kernel size, kernel small σ represents the minimum convolution kernel size; σ represents the sigmoid function, and σ in formula (7) is used for masking operations; Squeeze(·) represents the 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 kernel to preserve key information of the input image. In the second path, after the convolution operation, two feature gating mechanisms are introduced to dynamically adjust the convolution result through learned weights, forming two different branches. As shown in Equation (7), the Mask operation uses channel-wise shift multiplication, utilizing learnable weights w to shift Q... i The constraints are within the range [0,1]. Subsequently, each branch undergoes a tensor reshaping operation, transforming the tensor from Q... i ∈R{[b,c1,h1,w1]} is converted to Q i ∈R{[b,c2,nk,h1,w1]}, where nk is defined in formula (8). Next, the Split operation splits each Q... i The tensor is decomposed into nk blocks, and the Squeeze operation is used to compress the dimensionality of the tensor, as shown in Equation (9). The compressed result is expanded in both the horizontal and vertical directions, and unimportant regions are pruned to retain key information. Subsequently, the features of the two branches are concatenated and batch normalized, and their results are added together to form the output of the second path. Finally, the outputs of the two paths are merged to generate the final output. Equation (10) further provides a formal definition.

[0059] In small object detection, background regions often dominate. The SWC-Bottleneck module suppresses background features through mask operations while enhancing target features. Tensor reshaping and splitting can more effectively capture local features.

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

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

[0062] Preferably, the multi-scale feature fusion module Neck performs a 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 acquires the image feature data Y, performs an upsampling operation on it to obtain the first sampled data z1, and passes it to the first channel stitching 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 convolutional data y8 to obtain the first spliced ​​data z2, and passes it to the first C3k2 module;

[0065] Step 4.3: The first C3k2 module performs variable convolution and channel separation operations on the first spliced ​​data z2 to obtain the first variable convolution data z3, and passes 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 an upsampling operation on the first variable convolution data z3 to obtain the second sampled data z4, and passes it to the second channel stitching module Concat-2;

[0067] Step 4.5: The second channel splicing module Concat-2 performs channel splicing operation on the second sampled data z4 and the fifth convolutional data y6 to obtain the second spliced ​​data z5, and passes it to the second C3k2 module;

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

[0069] Step 4.7: The sixth basic convolution module CBS-6 performs convolution, batch normalization, and SiLU activation operations on the second variable convolution data z6 in sequence to obtain the tenth convolution data z7, and then passes it to the third channel splicing module Concat-3;

[0070] Step 4.8: The third channel splicing module Concat-3 performs channel splicing operation on the tenth convolution data z7 and the first variable convolution data z3 to obtain the third spliced ​​data z8, and passes it to the third C3k2 module;

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

[0072] Step 4.10: The seventh basic convolution module CBS-7 ​​performs convolution, batch normalization, and SiLU activation operations on the third variable convolution data z9 in sequence to obtain the eleventh convolution data z10, and then passes it to the fourth channel stitching module Concat-4.

[0073] Step 4.11: The fourth channel stitching module Concat-4 performs channel stitching operation on the eleventh convolution data z10 and the image feature data Y to obtain the fourth stitched data z11, and passes it to the fourth C3k2 module;

[0074] Step 4.12: The fourth C3k2 module performs variable convolution and channel separation operations on the fourth spliced ​​data z11 to obtain fused data Z, and then passes it to the detection and recognition module Head.

[0075] Preferably, the detection and recognition module Head is equipped 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 terminals 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. The output terminals of the three are connected to the input terminal of the fourth convolutional group normalization block CGS-4. The output terminal of the fourth convolutional group normalization block CGS-4 is connected to the input terminals 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 terminals 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.

[0076] Small / medium / large target detection heads are arranged side by side with a decoupled classification Conv-Cls module and a regression module Conv-Box. The output of the regression module Conv-Box is connected to the input of the model scaling control module Scale.

[0077] The Convolutional Group Normalization Block (CGS) is configured with sequentially connected convolutional layers and group normalization layers, and the output of the group normalization layer has a built-in SiLU activation function.

[0078] In YOLOv11, the detection head employs a multi-task architecture with shared features, which can impact the performance of classification and regression. To address this issue, this invention proposes the Group Normalized Scale Detection Head (GNSH), a lightweight detection head that improves feature utilization while maintaining high efficiency by decoupling the classification module (Conv-Cls) and the regression module (Conv-Box).

[0079] Traditional batch normalized (BN) performs poorly in small-batch scenarios, especially in small object detection. Therefore, this invention employs batch size-independent group normalized (GN) to better preserve details. GNSH extracts high-level features from the Neck module, fuses multi-scale features through convolutional Conv and GN layers, and uses independent branches for classification and regression. Furthermore, 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 fused data Z, with the following steps:

[0081] Step 5.1: The first convolutional group normalization block CGS-1 acquires the fused 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.

[0082] 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 normalized 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 convolution and group normalization operations on it, and then outputs the third group normalized 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 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;

[0085] Step 5.3: The small target detection head BSC-1 performs small target detection on the fourth set of normalized data c4 and outputs the small target detection results;

[0086] The BSC-2 target detection head performs target detection on the fourth set of normalized data c4 and outputs the target detection result.

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

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

[0089] The beneficial effects of this invention are:

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

[0091] 2. The SWC-Bottleneck module is used to improve the granularity of the model and enhance its ability to focus on extracting local image features.

[0092] 3. A lightweight detection head, called the Group Normalized Scale Detection Head (GNSH), was designed to reduce the number of model parameters. Attached Figure Description

[0093] Figure 1 The network structure diagram of ESGN-YOLO, a multi-scale target image detection model for UAV scenarios;

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

[0095] Figure 3 A heatmap comparison diagram of YOLOv11n and the ESGN-YOLO of this invention;

[0096] Figure 4 This is a comparison chart of the detection performance of YOLOv11n and the ESGN-YOLO of this invention. Detailed Implementation

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

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

[0099] Step 1: Construct a multi-scale target image detection system for UAV scenarios. The multi-scale target image detection system for UAV scenarios is equipped with an image acquisition module, which is connected to;

[0100] The multi-scale target image detection model ESGN-YOLO in the UAV scenario is configured with a feature extraction module BackBone, a multi-scale feature fusion module Neck, and a detection and recognition module Head connected in sequence.

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

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

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

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

[0105] like Figure 1As shown in (a), in step 3, the feature extraction module BackBone is configured with the following modules connected in sequence: 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.

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

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

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

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

[0110] Step 3.3: The first C3k-SWC module performs multi-scale convolution operation on the second convolutional data y2 to obtain the third convolutional data y3, and passes 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 operation on the third convolutional data y3 to obtain the first fused data y4, and passes it to the third basic convolutional module CBS-3;

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

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

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

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

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

[0117] Step 3.10: The C3k2-SWC module performs variable convolution and channel separation operations on the eighth convolutional data y9 to obtain the ninth convolutional data y10, and passes 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 operation on the ninth convolutional data y10 to obtain the second fused data y11, and passes it to the fast spatial pyramid pooling module SPPF.

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

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

[0121] like 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 having an input layer. The output of this input layer is connected to the input of the depthwise separable convolutional layer DWconv and the second addition unit. The output of the depthwise separable convolutional layer DWconv is connected to the input of the average pooling layer Avgpool, the max pooling layer Maxpool, and the second multiplication module. The outputs of the average pooling layer Avgpool and the max pooling layer Maxpool are connected to the input of the first addition unit. The output of the first addition unit is connected to the input of the spatial attention mechanism module SA and the channel attention mechanism CA. The outputs of the spatial attention mechanism module SA and the channel attention mechanism CA are connected to the input of the first multiplication unit. The output of the first multiplication unit is connected to the input of the second multiplication unit via the convolutional layer Covn. The output of the second multiplication unit is connected to the input of the second addition unit. The output of the second addition unit is connected to the output layer.

[0122] The calculation expression for 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] Where 'a' represents the input data of the efficient feature fusion module EFFM, BN is the batch normalization operation, σ is the sigmoid function, and F... output The output data is the efficient feature fusion module EFFM; h and v are the two dimensions of the average pooling layer and the max pooling layer, respectively.

[0130] like 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 equipped with an eighth basic convolutional module CBS-8. The output of the eighth basic convolutional module CBS-8 is connected to the input of the fifth channel splicing module Concat-5 via the ninth basic convolutional module CBS-9. There are also n SWC-Bottleneck modules connected in sequence between the output of the eighth basic convolutional module CBS-8 and the input of the fifth channel splicing module Concat-5. The output of the fifth channel splicing module Concat-5 is connected to the tenth basic convolutional module CBS-10.

[0131] like 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 arranged in parallel. The output terminals of the eleventh basic convolution module CBS-11 and the twelfth basic convolution module CBS-12 have built-in Split splitting functions.

[0132] The output of the eleventh basic convolutional module CBS-11 is connected to the input of the sixth channel stitching module Concat-6. Between the output of the eleventh basic convolutional module CBS-11 and the input of the sixth channel stitching module Concat-6, m sequentially connected SWC-Bottleneck modules are also provided, with the outputs of each of the m SWC-Bottleneck modules connected to the input of the sixth channel stitching module Concat-6. The output of the sixth channel stitching module Concat-6 is connected to the thirteenth basic convolutional module CBS-13.

[0133] The output of the twelfth basic convolutional module CBS-12 is connected to the input of the seventh-channel stitching module Concat-7. Between the output of the twelfth basic convolutional module CBS-12 and the input of the seventh-channel stitching module Concat-7, m C3k-SWC modules C3k-SWC are connected in sequence. The outputs of the m C3k-SWC modules C3k-SWC are all connected to the input of the seventh-channel stitching module Concat-7. The output of the seventh-channel stitching module Concat-7 is connected to the fourteenth basic convolutional module CBS-14.

[0134] like Figure 2As shown, the SWC-Bottleneck module is equipped with an input unit. The output of the input unit is connected to the input of the first convolutional layer Conv-1. The output of the first convolutional layer Conv-1 is connected to the input of the first channel-wise convolutional block and the second channel-wise convolutional block. The outputs of the first channel-wise convolutional block and the second channel-wise convolutional block are connected to the input of the addition unit. The output of the addition unit is connected to the input of the residual connection layer via the second convolutional layer Conv-2.

[0135] The first and second channel-wise convolutional blocks have the same structure, both of which are provided with a mask unit, a split unit, a tensor dimension adjustment unit, a clipping unit, a splicing unit, and a normalization layer (BN) connected in sequence.

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

[0137]

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

[0139]

[0140] Where ° represents the positional multiplication of corresponding matrices, ceil(·) represents rounding up, Clip(·) represents the clipping operation, BN(·) represents the normalization operation, Conv represents the convolution operation, and Split(·) represents the splitting function; i represents the i-th channel-wise convolutional block, i = 1, 2; j represents the j-th segmentation block; Q output The output of the SWC-Bottleneck module is represented by Q, the input of the SWC-Bottleneck module is represented by w, and the kernel is represented by kerner. big Represents the maximum kernel size, kernel small σ represents the minimum convolution kernel size; σ represents the sigmoid function, and σ in formula (7) is used for masking operations; Squeeze(·) represents the tensor compression operation.

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

[0142] The input terminals of the first upsampling module Upsample-1, the first channel stitching module Concat-1, the second channel stitching module Concat-2, and the fourth channel stitching module Concat-4 are connected to the feature extraction module BackBone; the output terminal of the first C3k2 module is connected to the input terminal of the third channel stitching module Concat-3; and the output terminals of the second, third, and fourth C3k2 modules 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 acquires the image feature data Y, performs an upsampling operation on it to obtain the first sampled data z1, and passes it to the first channel stitching module Concat-1;

[0145] Step 4.2: The first channel splicing module Concat-1 performs channel splicing operation on the first sampled data z1 and the seventh convolutional data y8 to obtain the first spliced ​​data z2, and passes it to the first C3k2 module;

[0146] Step 4.3: The first C3k2 module performs variable convolution and channel separation operations on the first spliced ​​data z2 to obtain the first variable convolution data z3, and passes 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 variable convolution data z3 to obtain the second sampled data z4, and passes it to the second channel stitching module Concat-2;

[0148] Step 4.5: The second channel splicing module Concat-2 performs channel splicing operation on the second sampled data z4 and the fifth convolutional data y6 to obtain the second spliced ​​data z5, and passes it to the second C3k2 module;

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

[0150] Step 4.7: The sixth basic convolution module CBS-6 performs convolution, batch normalization, and SiLU activation operations on the second variable convolution data z6 in sequence to obtain the tenth convolution data z7, and then passes it to the third channel splicing module Concat-3;

[0151] Step 4.8: The third channel splicing module Concat-3 performs channel splicing operation on the tenth convolution data z7 and the first variable convolution data z3 to obtain the third spliced ​​data z8, and passes it to the third C3k2 module;

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

[0153] Step 4.10: The seventh basic convolution module CBS-7 ​​performs convolution, batch normalization, and SiLU activation operations on the third variable convolution data z9 in sequence to obtain the eleventh convolution data z10, and then passes it to the fourth channel stitching module Concat-4.

[0154] Step 4.11: The fourth channel stitching module Concat-4 performs channel stitching operation on the eleventh convolution data z10 and the image feature data Y to obtain the fourth stitched data z11, and passes it to the fourth C3k2 module;

[0155] Step 4.12: The fourth C3k2 module performs variable convolution and channel separation operations on the fourth spliced ​​data z11 to obtain fused data Z, and then passes it to the detection and recognition module Head.

[0156] like Figure 1As shown in (c), the detection and recognition module Head is equipped 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 terminals 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. The output terminals of the three are connected to the input terminal of the fourth convolutional group normalization block CGS-4. The output terminal of the fourth convolutional group normalization block CGS-4 is connected to the input terminals 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 terminals 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 fused data Z, and the steps are as follows:

[0158] Step 5.1: The first convolutional group normalization block CGS-1 acquires the fused 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 convolutional data z9, performs convolution and group normalization operations on it, and then outputs the second group 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 convolutional data z6, performs convolution and group normalization operations on it, and then outputs the third group 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 set of normalized data c4 and outputs the small target detection results;

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

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

[0165] like Figure 1 As shown in (d), the small / medium / large target detection heads are equipped with a decoupled classification Conv-Cls module and a regression module Conv-Box arranged in parallel. The output of the regression module Conv-Box is connected to the input of the model scaling control module Scale.

[0166] like Figure 1 As shown in (e), the Convolutional Group Normalization Block (CGS) is provided with convolutional layers and group normalization layers connected in sequence, and the output of the group normalization layer has a built-in SiLU activation function.

[0167] To evaluate the performance of this invention, experiments were conducted on two representative datasets, VisDrone2019 and TinyPerson.

[0168] VisDrone2019: The VisDrone2019 dataset was collected by the AISKYEYE team at Tianjin University, covering 14 cities, including urban and rural areas, with scenes containing targets of varying densities, such as pedestrians and vehicles. Data was collected under different drone platforms, weather conditions, and lighting environments. This paper uses the officially defined training and validation sets for experiments.

[0169] TinyPerson: The TinyPerson dataset is sourced from the internet, and the images primarily depict people near the beach. This dataset contains 1,610 labeled images, with 794 used for training and 816 for testing. Most images have a resolution of 1920×1080. Each image contains an average of 25 small objects, representing no more than 0.87% of the total pixels. Furthermore, all small objects are no larger than 20×20 pixels and are sparsely distributed.

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

[0171] This experiment used VisDrone2019 and TinyPerson datasets for evaluation. Experimental results and comparisons with YOLO series models and other enhanced models demonstrate that the model proposed in this invention outperforms YOLOv11n in small target detection.

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

[0173]

[0174] VisDrone 2019: As shown in Table 1, this invention outperforms the second-ranked model PP-PicoDet-L by 0.3% in mAP@0.5, with mAP@0.5 for the "bus" and "car" categories being 2.4% and 1.8% higher than the second-ranked model, respectively. Compared to YOLOv11n, the overall mAP@0.5 is improved by 1.9%. Except for a slight decrease in mAP@0.5 for the "awning-tricycle" and "bicycle" categories, the detection performance of other categories is improved. The results indicate that ESGN-YOLOv11n has good applicability in small target detection tasks in UAV aerial imagery.

[0175] Table 2. Comparison of mAP@.5 for different networks on the TinyPerson dataset.

[0176]

[0177] TinyPerson: As shown in Table 2, this invention outperforms state-of-the-art (SOTA) methods on the validation set. GNSH-YOLOv11n achieves an mAP@0.5 of 20.3%, 0.3% higher than the second-ranked model, Sf-YOLOv5. Furthermore, the mAP@0.5 for the "earth-person" and "sea-person" categories reaches 15.5% and 25% respectively, still higher than other models. The overall mAP@0.5, as well as the mAP@0.5 for "earth-person" and "sea-person," are improved by 1%, 1.6%, and 0.3% respectively. Experimental results demonstrate that ESGN-YOLO performs better on small object detection tasks.

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

[0179]

[0180] To evaluate the impact of this invention, ablation experiments were conducted on the VisDrone 2019 validation set, focusing on three key improvements: the introduction of the EFFM module, modification of the C3k2 module, and replacement of the detection head with GNSH. Table 3 summarizes the experimental results, showing that the introduction of the EFFM module significantly improved mAP@0.5 and mAP@[0.5:0.95] by 1.3% and 1.1%, respectively. Modifying the SWC module alone had a smaller impact, improving mAP@0.5 by only 0.3%, while mAP@[0.5:0.95] remained unchanged. Combining SWC and EFFM(S+E) resulted in a significant performance improvement compared to using SWC(+SWC(S)) alone. Specifically, adding E+S improved mAP@0.5 by 1.8 percentage points from baseline, while improving mAP@[0.5:0.95] by 1.4 percentage points. Using GNSH alone improved recall and reduced the number of model parameters from 2.6M to 2.4M, but failed to improve mAP. Finally, combining GNSH with EFFM and SWC further improved recall and mAP[0.5:0.95], while significantly reducing the number of parameters (from 3.9M to 3.7M), but precision decreased slightly.

[0181] Next, the thermal feature maps of the final backbone network layers were analyzed to evaluate the feature extraction and fusion performance of the improved model. For example... Figure 3 As shown, Figure 3 The first row shows the detection results of ESGN-YOLO, and the second row shows the detection results of YOLOv11n. Compared with YOLOv11n(b), ESGN-YOLOv11n(c) shows greater attention to small targets and stronger brightness in the region of interest, indicating that the model's detection capability is improved in complex backgrounds and occluded scenes. Furthermore, experiments were conducted on YOLOv11 and ESGN-YOLOv11n to further verify the effectiveness of the 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, such as... Figure 4 As shown, Figure 4 The first line shows the detection results of ESGN-YOLO, and the second line shows the detection results of YOLOv11n. The model of this invention can detect more small and occluded targets, and its detection performance is superior to the original version.

[0182] This invention proposes ESGN-YOLOv11n, an enhanced version of YOLOv11n, designed to improve the performance of small object detection in images. The main contributions of this invention are as follows, and have been verified through extensive experiments. First, this invention integrates the EFFM module into the backbone network to enhance the extraction 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-normalized scale detection head GNSH to reduce resource consumption. Experimental results verify the effectiveness of the proposed method.

[0183] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-scale target image detection method in a drone scenario, characterized in that, Includes the following steps: Step 1: Construct a multi-scale target image detection system for UAV scenarios. The multi-scale target image detection system for UAV scenarios is equipped with an image acquisition module, which is connected to the ESGN-YOLO multi-scale target image detection model for UAV scenarios. The multi-scale target image detection model ESGN-YOLO in the UAV scenario is configured with a feature extraction module BackBone, a multi-scale feature fusion module Neck, and a detection and recognition module Head 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 ESGN-YOLO multi-scale target image detection model in the UAV scene; Step 3: The feature extraction module BackBone in the ESGN-YOLO multi-scale target image detection model in the UAV scenario performs feature extraction on the multi-scale target image X to obtain image feature data Y, and then passes it to the multi-scale feature fusion module Neck. Step 4: The multi-scale feature fusion module Neck performs multi-scale feature fusion on the image feature data Y to obtain fused data Z, and then passes it to the detection and recognition module Head; Step 5: The detection and recognition module Head performs detection and recognition operations on the fused data Z and outputs the target detection result; In step 3, the feature extraction module BackBone is configured with the following modules connected in sequence: a first basic convolutional module CBS-1, a second basic convolutional module CBS-2, a first C3k-SWC module, a first efficient feature fusion module EFFM-1, a third basic convolutional module CBS-3, a second C3k-SWC module, a fourth basic convolutional module CBS-4, a third C3k-SWC module, a fifth basic convolutional 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 convolutional module C2PSA. The outputs of the second C3k-SWC module, the third C3k-SWC module, and the enhanced convolutional module C2PSA are connected to the multi-scale feature fusion module Neck; The first high-efficiency feature fusion module EFFM-1 and the second high-efficiency feature fusion module EFFM-2 have the same structure, both of which have an input layer, and the output of the input layer is connected to a depth-separable convolutional layer. The input terminal of the second addition unit; the depth-separable convolutional layer The output of the first multiplication unit is connected to the input of the average pooling layer (Avgpool), the max pooling layer (Maxpool), and the second multiplication module. The outputs of the average pooling layer (Avgpool) and the max pooling layer (Maxpool) are connected to the input of the first addition unit. The output of the first addition unit is connected to the input of the spatial attention mechanism module (SA) and the channel attention mechanism (CA). The outputs of the spatial attention mechanism module (SA) and the channel attention mechanism (CA) are connected to the input of the first multiplication unit. The output of the first multiplication unit is connected to the input of the second multiplication unit via the convolutional layer (Conv). The output of the second multiplication unit is connected to the input of the second addition unit. The output of the second addition unit is connected to the output layer. The calculation expression for the efficient feature fusion module EFFM is as follows: （1）； （2）； （3）； （4）； （5）； （6）； in, For the input data of the efficient feature fusion module EFFM, For batch normalization operations, For the sigmoid function, The output data is the efficient feature fusion module EFFM; h and v are the two dimensions of the average pooling layer and the max pooling layer, respectively.

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

3. The multi-scale target image detection method in a drone scenario according to claim 1 or 2, characterized in that: The first, second, and third C3k-SWC modules have the same structure, each equipped with an eighth basic convolutional module CBS-8. The output of the eighth basic convolutional module CBS-8 is connected to the input of the fifth channel stitching module Concat-5 via a ninth basic convolutional module CBS-9. There are also n sequentially connected SWC-Bottleneck modules between the output of the eighth basic convolutional module CBS-8 and the input of the fifth channel stitching module Concat-5. The output of the fifth channel stitching module Concat-5 is connected to the tenth basic convolutional module CBS-10. The C3k2-SWC module has an eleventh basic convolutional module CBS-11 and a twelfth basic convolutional module CBS-12 arranged in parallel. The output terminals of the eleventh basic convolutional module CBS-11 and the twelfth basic convolutional module CBS-12 have a built-in Split function. The output of the eleventh basic convolutional module CBS-11 is connected to the input of the sixth channel stitching module Concat-6. Between the output of the eleventh basic convolutional module CBS-11 and the input of the sixth channel stitching module Concat-6, m sequentially connected SWC-Bottleneck modules are also provided, with the outputs of each of the m SWC-Bottleneck modules connected to the input of the sixth channel stitching module Concat-6. The output of the sixth channel stitching module Concat-6 is connected to the thirteenth basic convolutional module CBS-13. The output of the twelfth basic convolutional module CBS-12 is connected to the input of the seventh-channel stitching module Concat-7. Between the output of the twelfth basic convolutional module CBS-12 and the input of the seventh-channel stitching module Concat-7, m C3k-SWC modules C3k-SWC are connected in sequence. The outputs of the m C3k-SWC modules C3k-SWC are all connected to the input of the seventh-channel stitching module Concat-7. The output of the seventh-channel stitching module Concat-7 is connected to the fourteenth basic convolutional module CBS-14.

4. The multi-scale target image detection method in a drone scenario according to claim 3, characterized in that: The SWC-Bottleneck module is equipped with an input unit. The output of the input unit is connected to the input of the first convolutional layer Conv-1. The output of the first convolutional layer Conv-1 is connected to the input of the first channel-wise convolutional block and the second channel-wise convolutional block. The outputs of the first channel-wise convolutional block and the second channel-wise convolutional block are connected to the input of an addition unit. The output of the addition unit is connected to the input of the residual connection layer via the second convolutional layer Conv-2. The first and second channel-wise convolutional blocks have the same structure, both of which are provided with a mask unit, a split unit, a tensor dimension adjustment unit, a clipping unit, a splicing unit, and a normalization layer (BN) connected in sequence. The calculation expression for the SWC-Bottleneck module is as follows: （7）； （8）； （9）； （10）； in, This indicates positional multiplication of corresponding matrices, ceil(·) represents rounding up, Clip(·) represents clipping, and BN(·) represents normalization. This represents the convolution operation. The splitting function is represented by ; i represents the i-th channel-wise convolutional block, i=1,2; j represents the j-th segmentation block; This indicates the output of the SWC-Bottleneck module, and Q indicates the input of the SWC-Bottleneck module. Indicates weight, Indicates the maximum convolution kernel size. This represents the minimum convolution kernel size; Represents the sigmoid function, in formula (7) Used for masking operations; This indicates a tensor compression operation.

5. The multi-scale target image detection method in a drone 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 stitching module Concat-1, a first C3k2 module, a second upsampling module Upsample-2, a second channel stitching module Concat-2, a second C3k2 module, a sixth basic convolution module CBS-6, a third channel stitching module Concat-3, a third C3k2 module, a seventh basic convolution module CBS-7, a fourth channel stitching module Concat-4, and a fourth C3k2 module connected in sequence. The input terminals of the first upsampling module Upsample-1, the first channel stitching module Concat-1, the second channel stitching module Concat-2, and the fourth channel stitching module Concat-4 are connected to the feature extraction module BackBone; the output terminal of the first C3k2 module is connected to the input terminal of the third channel stitching module Concat-3; and the output terminals of the second, third, and fourth C3k2 modules are connected to the detection and recognition module Head.

6. The multi-scale target image detection method in a drone scenario according to claim 5, 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 acquires the image feature data Y, performs an upsampling operation on it to obtain the first sampled data z1, and passes it to the first channel stitching module Concat-1; Step 4.2: The first channel splicing module Concat-1 performs channel splicing operation on the first sampled data z1 and the seventh convolutional data y8 to obtain the first spliced ​​data z2, and passes it to the first C3k2 module; Step 4.3: The first C3k2 module performs variable convolution and channel separation operations on the first spliced ​​data z2 to obtain the first variable convolution data z3, and passes it to the second upsampling module Upsample-2 and the third channel splicing module Concat-3; Step 4.4: The second upsampling module Upsample-2 performs an upsampling operation on the first variable convolution data z3 to obtain the second sampled data z4, and passes it to the second channel stitching module Concat-2; Step 4.5: The second channel splicing module Concat-2 performs channel splicing operation on the second sampled data z4 and the fifth convolutional data y6 to obtain the second spliced ​​data z5, and passes it to the second C3k2 module; Step 4.6: The second C3k2 module performs variable convolution and channel separation operations on the second spliced ​​data z5 to obtain the second variable convolution data z6, and passes it to the sixth basic convolution module CBS-6 and the detection and recognition module Head; Step 4.7: The sixth basic convolution module CBS-6 performs convolution, batch normalization, and SiLU activation operations on the second variable convolution data z6 in sequence to obtain the tenth convolution data z7, and then passes it to the third channel splicing module Concat-3; Step 4.8: The third channel splicing module Concat-3 performs channel splicing operation on the tenth convolution data z7 and the first variable convolution data z3 to obtain the third spliced ​​data z8, and passes it to the third C3k2 module; Step 4.9: The third C3k2 module performs variable convolution and channel separation operations on the third spliced ​​data z8 to obtain the third variable convolution data z9, and passes it to the seventh basic convolution module CBS-7 ​​and the detection and recognition module Head; Step 4.10: The seventh basic convolution module CBS-7 ​​performs convolution, batch normalization, and SiLU activation operations on the third variable convolution data z9 in sequence to obtain the eleventh convolution data z10, and then passes it to the fourth channel stitching module Concat-4. Step 4.11: The fourth channel stitching module Concat-4 performs channel stitching operation on the eleventh convolution data z10 and the image feature data Y to obtain the fourth stitched data z11, and passes it to the fourth C3k2 module; Step 4.12: The fourth C3k2 module performs variable convolution and channel separation operations on the fourth spliced ​​data z11 to obtain fused data Z, and then passes it to the detection and recognition module Head.

7. The multi-scale target image detection method in a drone scenario according to claim 1, characterized in that: The detection and recognition module Head is equipped 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 terminals 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. The output terminals of the three are connected to the input terminal of the fourth convolutional group normalization block CGS-4. The output terminal of the fourth convolutional group normalization block CGS-4 is connected to the input terminals 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 terminals 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.

8. The multi-scale target image detection method in a drone scenario according to claim 7, characterized in that: The detection and recognition module Head performs detection and recognition operations on the fused data Z, and the steps are as follows: Step 5.1: The first convolutional group normalization block CGS-1 acquires the fused 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 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 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 set of normalized data c4 and outputs the small target detection results; The BSC-2 target detection head performs target detection on the fourth set of normalized data c4 and outputs the target detection result. The large target detection head BSC-3 performs large target detection on the fourth set of normalized data c4 and outputs the large target detection result.

Images