An unmanned aerial vehicle small target detection method and system based on an improved D-FINE-n model
By improving the backbone network and encoder of the D-FINE-n model, and combining it with a cross-scale feature focusing module and windmill convolution, the problem of low detection accuracy of small targets in UAV aerial images was solved, achieving more efficient feature extraction and fusion, and improving detection accuracy.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- JIANGXI NORMAL UNIV
- Filing Date
- 2026-04-16
- Publication Date
- 2026-07-31
AI Technical Summary
Existing drone aerial imagery suffers from low accuracy in small target detection, and mainstream target detection algorithms are insufficient in feature extraction capabilities for drone aerial images, making it difficult to meet practical application needs.
An improved D-FINE-n model is adopted, which obtains feature maps of different scales through backbone network preprocessing. Then, the encoder and cross-scale feature focusing module are used to perform feature transformation, information integration and fusion. Combined with windmill convolution, deep feature extraction is performed to improve the detection accuracy of small targets.
It improves the detection accuracy of small targets in UAV aerial images, enhances the ability to extract and fuse feature information, and improves the detection effect.
Smart Images
Figure CN122067148B_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of computer vision technology, specifically relating to a method and system for small target detection in unmanned aerial vehicles based on an improved D-FINE-n model. Background Technology
[0002] As a flexible and economical remote sensing platform, unmanned aerial vehicles (UAVs) have been widely used in agriculture, inspection, security, and other fields. Deep learning-based target detection technology is a core component for enabling UAVs to achieve autonomous perception and scene understanding. Its goal is to locate and identify specific objects from aerial images, providing crucial information for subsequent task execution.
[0003] However, drone aerial images typically cover vast areas and contain a large number of densely distributed small targets (such as pedestrians and vehicles). Current mainstream target detection algorithms are mainly designed for general scenarios. When dealing with drone aerial images that primarily contain small targets, their feature extraction capabilities are insufficient, leading to decreased detection accuracy and making it difficult to meet the needs of practical applications. Summary of the Invention
[0004] The purpose of this application is to provide a method and system for detecting small targets on unmanned aerial vehicles (UAVs) based on an improved D-FINE-n model, which can improve the detection accuracy of small targets.
[0005] To solve the above-mentioned technical problems, this application is implemented as follows: In a first aspect, embodiments of this application provide a method for small target detection by unmanned aerial vehicles (UAVs) based on an improved D-FINE-n model, the method comprising: In response to receiving the drone aerial image, the drone aerial image is preprocessed based on the backbone network of the improved D-FINE-n model to obtain small-scale feature maps, medium-scale feature maps and large-scale feature maps; The encoder based on the improved D-FINE-n model performs a 1x1 first convolution operation with a stride of 1 on the small-scale feature map, the medium-scale feature map, and the large-scale feature map, respectively. Based on the encoder, feature transformation and information integration are performed on the small-scale feature map after convolution. Based on the cross-scale feature focusing module (CFFM) of the encoder, a fusion operation is performed on the large-scale feature map after convolution, the medium-scale feature map after convolution, and the integrated small-scale feature map to obtain the first fused feature map. The first fused feature map is upsampled, and the upsampled first fused feature map is concatenated with the convolved large-scale feature map through channels, and deep feature extraction is performed to obtain the first deep feature map. The first fused feature map is downsampled based on the windmill convolution PConv, and the downsampled first fused feature map is concatenated with the integrated small-scale feature map through channels, and deep feature extraction is performed to obtain the second deep feature map. Based on the CFFM, the first fused feature map, the first deep feature map, and the second deep feature map are fused to obtain the second fused feature map; Based on the PConv, the second fused feature map is downsampled, and the downsampled second fused feature map is concatenated with the second deep feature map through channels, and deep feature extraction is performed to obtain the third deep feature map. The small target detection result in the UAV aerial image is obtained based on the second fused feature map and the third deep feature map output by the encoder.
[0006] Secondly, embodiments of this application provide a UAV small target detection system based on an improved D-FINE-n model, wherein the UAV small target detection system based on the improved D-FINE-n model includes: The preprocessing module is used to preprocess the drone aerial image based on the backbone network of the improved D-FINE-n model in response to receiving the drone aerial image to obtain small-scale feature maps, medium-scale feature maps and large-scale feature maps. The convolution module is used to perform a 1x1 first convolution operation with a stride of 1 on the small-scale feature map, the medium-scale feature map and the large-scale feature map respectively, based on the encoder of the improved D-FINE-n model; The first fusion module is used to perform feature transformation and information integration on the small-scale feature map after convolution based on the encoder, and to perform fusion operation on the large-scale feature map after convolution, the medium-scale feature map after convolution, and the integrated small-scale feature map based on the cross-scale feature focusing module (CFFM) of the encoder to obtain the first fused feature map. The upsampling module is used to perform upsampling on the first fused feature map, and to concatenate the upsampled first fused feature map with the convolved large-scale feature map through channels, and to perform deep feature extraction to obtain the first deep feature map. The downsampling module is used to perform downsampling operation on the first fused feature map based on windmill convolution PConv, and to perform channel concatenation of the downsampled first fused feature map and the integrated small-scale feature map, as well as to perform deep feature extraction to obtain the second deep feature map. The second fusion module is used to perform a fusion operation on the first fused feature map, the first deep feature map, and the second deep feature map based on the CFFM to obtain the second fused feature map; The channel stitching module is used to perform downsampling operation on the second fused feature map based on the PConv, and to perform channel stitching of the downsampled second fused feature map and the second deep feature map, as well as to perform deep feature extraction to obtain a third deep feature map; The detection module is used to obtain the small target detection result in the UAV aerial image based on the second fused feature map output by the encoder and the third deep feature map.
[0007] Thirdly, embodiments of this application provide a computer device including a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the method described in the first aspect.
[0008] Fourthly, embodiments of this application provide a computer-readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the steps of the method described in the first aspect.
[0009] Fifthly, embodiments of this application also provide a computer program product, including a computer program that, when executed by a processor, implements the steps of the method described in the first aspect.
[0010] In this embodiment, in response to receiving a drone aerial image, the drone aerial image is preprocessed using the backbone network of the improved D-FINE-n model to obtain small-scale feature maps, medium-scale feature maps, and large-scale feature maps. The encoder of the improved D-FINE-n model performs a 1x1 first convolution operation with a stride of 1 on the small-scale feature map, the medium-scale feature map, and the large-scale feature map, respectively. Based on the encoder, feature transformation and information integration are performed on the convolved small-scale feature map, and based on the cross-scale feature focusing module (CFFM) of the encoder, a fusion operation is performed on the convolved large-scale feature map, the convolved medium-scale feature map, and the integrated small-scale feature map to obtain a first fused feature map. The first fused feature map is upsampled, and the upsampled first fused feature map is concatenated with the convolved large-scale feature map, and deep feature extraction is performed to obtain a first deep feature map. PConv performs a downsampling operation on the first fused feature map, and then concatenates the downsampled first fused feature map with the integrated small-scale feature map through channels, and performs deep feature extraction to obtain a second deep feature map. Based on CFFM, the first fused feature map, the first deep feature map, and the second deep feature map are fused to obtain a second fused feature map. Based on PConv, the second fused feature map is downsampled, and then concatenates the downsampled second fused feature map with the second deep feature map through channels, and performs deep feature extraction to obtain a third deep feature map. Based on the second fused feature map and the third deep feature map output by the encoder, the small target detection result in the UAV aerial image is obtained. This improves the ability to extract small target feature information, thereby better performing secondary information extraction and feature fusion on the feature map output by the backbone network, and thus improving the detection accuracy of small targets in UAV aerial images. Attached Figure Description
[0011] Figure 1 This is a flowchart illustrating the UAV small target detection method based on the improved D-FINE-n model provided in this application. Figure 2 This is a schematic diagram of the backbone network of the D-FINE-n model involved in the UAV small target detection method based on the improved D-FINE-n model in this application; Figure 3 This is a schematic diagram of the improved D-FINE-n model provided in the UAV small target detection method based on the improved D-FINE-n model of this application; Figure 4 A schematic diagram of the FCM structure provided by the UAV small target detection method based on the improved D-FINE-n model in this application; Figure 5 This is a schematic diagram of the sigmoid activation function and the SiLU activation function involved in the UAV small target detection method based on the improved D-FINE-n model in this application; Figure 6 This is a schematic diagram of the PConv structure involved in the UAV small target detection method based on the improved D-FINE-n model in this application; Figure 7 A schematic diagram of the CFFM structure provided for the UAV small target detection method based on the improved D-FINE-n model in this application; Figure 8 A comparison of the number of parameters and detection accuracy of the FCM-CFFM-D-FINE model provided by this application for the UAV small target detection method based on the improved D-FINE-n model, with mainstream target detection models and the original D-FINE-n model; Figure 9 A structural block diagram of a UAV small target detection system based on an improved D-FINE-n model is provided for some embodiments of this application; Figure 10 Internal structural diagrams of a computer device provided for some embodiments of this application. Detailed Implementation
[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0013] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.
[0014] It should be noted that the application scenario of the UAV small target detection method based on the improved D-FINE-n model provided in this application embodiment can be UAV small target detection.
[0015] In this application scenario, drones are now widely used in various work scenarios, such as agriculture, urban transportation, and search and rescue. Thanks to their ease of use, flexibility, and relatively moderate cost, drones have become popular instruments in the field of remote sensing imagery. Object detection based on deep learning models is an important perception task when using drones. Object detection is defined as the process of determining the class of an instance of a target and estimating its location by outputting a bounding box or the contours around the target. Information is extracted from the raw images acquired by the drone, enabling it to identify targets in specific work scenarios and thus complete the designated task.
[0016] Small objects make up the majority of drone aerial images, making small object detection a crucial part of drone aerial image target detection tasks. Small objects are defined as those with a ratio less than 0.1 relative to the original image or a resolution less than 32x32 pixels. Current mainstream deep learning object detection models are primarily designed for general object detection datasets, such as the D-FINE-n model. While capable of detecting both small and large targets, its feature extraction capabilities are weaker and detection accuracy is lower for drone aerial image datasets where small objects constitute a large proportion.
[0017] The D-FINE-n model can include a backbone network, an encoder, and a detection head. The backbone network is used to receive aerial images taken by the UAV, and the detection head is used to output the detection results of targets in the UAV aerial images, such as pedestrians.
[0018] The following description, in conjunction with the accompanying drawings, details the UAV small target detection method based on the improved D-FINE-n model provided in this application, through specific embodiments and application scenarios.
[0019] Reference Figure 1 , Figure 1 This is a flowchart illustrating a UAV small target detection method based on an improved D-FINE-n model, provided in some embodiments of this application. The method includes steps 101-107. Wherein: Step 101: In response to receiving the drone aerial image, the drone aerial image is preprocessed based on the backbone network of the improved D-FINE-n model to obtain small-scale feature maps, medium-scale feature maps and large-scale feature maps.
[0020] In some embodiments, drones are typically positioned at a relatively high altitude above the ground when taking images. Therefore, targets such as pedestrians and vehicles appear relatively small in the drone's aerial images. Consequently, the detection of targets within drone aerial images is termed small target detection.
[0021] In some embodiments, the backbone network includes structural blocks for rapid early feature extraction and downsampling of UAV aerial images; such as the Stem Block; and the HG_Stage module for feature extraction and downsampling. That is, the backbone network can perform different levels of preprocessing on UAV aerial images through the Stem Block and the HG_Stage module.
[0022] In some embodiments, after receiving the output of the initial block (StemBlock), the HG_Stage module in the backbone network performs feature extraction and downsampling, and outputs the result to the next HG_Stage module. This process continues until all components of the backbone network have completed processing, thereby completing the preprocessing of the UAV aerial image.
[0023] In this embodiment, the backbone network includes at least three HG_Stage modules, resulting in a feature map output by each HG_Stage module. In some embodiments, the feature maps may be output by the next HG_Stage module after one or more intervals. See, for example, [link to example]. Figure 2 , Figure 2 This is a schematic diagram of the backbone network of the D-FINE-n model involved in the UAV small target detection method based on the improved D-FINE-n model in this application. Figure 2 It includes 4 HG_Stage modules.
[0024] Step 102: The encoder based on the improved D-FINE-n model performs a 1x1 first convolution operation with a stride of 1 on the small-scale feature map, the medium-scale feature map, and the large-scale feature map, respectively.
[0025] Step 103: Based on the encoder, perform feature transformation and information integration on the convolutional small-scale feature map, and based on the cross-scale feature focus module (CFFM) of the encoder, perform a fusion operation on the convolutional large-scale feature map, the convolutional medium-scale feature map, and the integrated small-scale feature map to obtain the first fused feature map.
[0026] In some embodiments, the Transformer module can be used to perform feature transformation and information integration on the large-scale feature map after convolution.
[0027] Step 104: Upsample the first fused feature map, and concatenate the upsampled first fused feature map with the convolved large-scale feature map, and perform deep feature extraction to obtain the first deep feature map.
[0028] Step 105: The first fused feature map is downsampled based on the Pinwheel-shaped Convolution (PConv) module, and the downsampled first fused feature map is concatenated with the integrated small-scale feature map through channels, and deep feature extraction is performed to obtain the second deep feature map.
[0029] Step 106: Based on the CFFM, perform a fusion operation on the first fused feature map, the first deep feature map, and the second deep feature map to obtain the second fused feature map.
[0030] Step 107: Based on the PConv, perform a downsampling operation on the second fused feature map, and then perform channel concatenation between the downsampled second fused feature map and the second deep feature map, and perform deep feature extraction to obtain the third deep feature map.
[0031] Step 108: Based on the second fused feature map and the third deep feature map output by the encoder, the small target detection result in the UAV aerial image is obtained.
[0032] In some embodiments, deep feature extraction can be implemented using the RepNCSPELAN4 module, and downsampling operations can be implemented using pooling layers or convolutional operations.
[0033] In this embodiment, in response to receiving drone aerial images, the backbone network of the improved D-FINE-n model preprocesses the drone aerial images to obtain small-scale, medium-scale, and large-scale feature maps. The encoder of the improved D-FINE-n model performs a 1x1 first convolution operation with a stride of 1 on the small-scale, medium-scale, and large-scale feature maps respectively. Based on the encoder, feature transformation and information integration are performed on the convolved small-scale feature map. Then, based on the encoder's cross-scale feature focusing module (CFFM), a fusion operation is performed on the convolved large-scale feature map, the convolved medium-scale feature map, and the integrated small-scale feature map to obtain a first fused feature map. The first fused feature map is upsampled, and the upsampled first fused feature map is concatenated with the convolved large-scale feature map, and deep feature extraction is performed to obtain a first deep feature map. Based on PC... The onv performs a downsampling operation on the first fused feature map, and then concatenates the downsampled first fused feature map with the integrated small-scale feature map through channels, and performs deep feature extraction to obtain a second deep feature map. Based on the CFFM, the first fused feature map, the first deep feature map, and the second deep feature map are fused to obtain a second fused feature map. Based on the PConv, the second fused feature map is downsampled, and then concatenates the downsampled second fused feature map with the second deep feature map through channels, and performs deep feature extraction to obtain a third deep feature map. Based on the second fused feature map and the third deep feature map output by the encoder, the small target detection result in the UAV aerial image is obtained. This improves the ability to extract small target feature information, thereby better performing secondary information extraction and feature fusion on the feature map output by the backbone network, and thus improving the detection accuracy of small targets in UAV aerial images.
[0034] In some embodiments, the fusion operation performed via CFFM includes: The three input feature maps before fusion are scale aligned; the three input feature maps before fusion are: the large-scale feature map after convolution, the medium-scale feature map after convolution, and the small-scale feature map after integration, or the first fused feature map, the first deep feature map, and the second deep feature map.
[0035] Channel stitching is performed on the scale-aligned feature maps to obtain cross-scale fused feature maps.
[0036] The cross-scale fused feature map is subjected to multi-scale deep convolution processing and post-processing.
[0037] Scale alignment is achieved through the following formulas (1), (2), and (3): (1) (2) (3) in, , , These are the small-scale feature map after feature transformation and integration, the medium-scale feature map after convolution, and the large-scale feature map after convolution, respectively. For upsampling operation, To The branch output obtained after performing one upsampling operation. ; For a 1x1 convolution operation, To The branch output obtained by performing a 1x1 convolution operation, where, ; For downsampling operation, To The branch output obtained after performing a downsampling operation, where, Finally, three outputs of the same size were obtained.
[0038] By using the following formula (4), channel stitching is performed on the scale-aligned feature map. It can be understood that cross-scale information fusion is achieved through scale alignment and channel stitching.
[0039] (4) in, For channel splicing operations; This is a cross-scale fusion feature map.
[0040] In some embodiments, performing multi-scale deep convolution processing on the cross-scale fused feature map includes: The cross-scale fused feature maps are subjected to a second convolution operation using multiple convolution kernels of different sizes. The convolutional cross-scale fused feature maps are then stacked and summed.
[0041] The number of convolution kernels can be 3, 4, etc., and this embodiment does not impose a specific limitation.
[0042] In some embodiments, the number of convolutional kernels is 4. The second convolution operation is performed on the cross-scale fusion feature map by multiple convolutional kernels of different sizes, and the convolutional cross-scale fusion feature maps are stacked and summed. This is achieved by the following formula (5): (5) in, This refers to multi-scale deep convolutional features obtained by performing multi-scale deep convolution processing on cross-scale fused feature maps; For summation operations; For stacking operations; This is the second convolution operation, where the kernel size is... Depth-separable convolution operations, The value can be 3, 5, 7 or 9.
[0043] The above post-processing can be achieved using the following formula (6): (6) in, This is a pointwise convolution operation; For the number of channels 1x1 convolution operation; , which are the output features of post-processing.
[0044] It is understandable that, due to this output feature At the same time, it preserves the original multi-scale features (i.e. ) and multi-scale deep convolutional features (i.e. ), making the output features It contains more comprehensive information and avoids the problem of overfitting.
[0045] In some embodiments, the downsampling operation can be implemented using the PConv module. Unlike typical convolution operations, PConv uses asymmetric padding to create horizontal and vertical convolution kernels for different regions of the image, with a convolution stride set to 1. First, the windmill convolution module processes the input tensor... Z (Such as the tensor of the first fused feature map) The four branches are padded and convolved using the following formulas (7), (8), (9) and (10): (7) (8) (9) (10) in, C1 , H and W They represent the input tensors respectively. Z The number of channels, height, and width; fill parameters These represent the number of pixels filled in the left, right, top, and bottom directions, respectively: 1, 0, 0, and 3, and so on. , and ; For a 1x3 horizontal convolution operation, This is a 3x1 vertical convolution operation; C 2 represents the number of output channels of the windmill convolution module; For batch normalization operation; For activation functions; , , and These are the output features of the four branches of the windmill convolution module.
[0046] It should be noted that, to improve training stability and speed, after convolution processing, each branch further processes the convolution results using batch normalization (BN) and the SiLU activation function; the final four branches have the following output channel counts. C 2 / 4, height is H +1, width is W Tensors of the same scale with +1.
[0047] The expression for the activation function SiLU is as follows: (11) in, This is the input to the activation function SiLU; This is the output of the activation function SiLU.
[0048] Secondly, feature fusion and output are performed using formulas (12) and (13): (12) (13) in, This is the channel concatenation result of the output features from the four branches of the windmill convolution module; Indicates the number of channels is C A convolution operation with a kernel size of 2x2; To The output features obtained by deep feature extraction.
[0049] In some embodiments, the method further includes: The HG_Stage module in the backbone network is replaced with a Feature Complementary Mapping Module (FCM) and a downsampling module, with the output of the FCM serving as the input of the downsampling module.
[0050] In some embodiments, a windmill convolution module is introduced as a downsampling module to connect four FCM modules, thereby forming a lightweight backbone network with low parameter and computational cost. This network transforms UAV aerial images into feature maps of multiple sizes with hierarchical semantic information and spatial structure, providing high-quality basic features for subsequent encoders to further extract and fuse feature map information.
[0051] Based on the aforementioned backbone network, assuming the drone aerial image is P1 with a size of 640x640, the drone aerial image, after passing through the Stem Block, yields a feature map with a 4x downsampling size of 160x160. Then, an FCM module extracts shallow information to obtain feature map P2. Next, a PConv convolution with a stride of 2 and another FCM module are used for downsampling to obtain a feature map P3 with an 8x downsampling size of 80x80. This process is repeated to obtain a feature map P4 with a 16x downsampling size of 40x40. Finally, a PConv convolution with a stride of 1 and another FCM module are used to obtain a deep information feature map P5 with a 16x downsampling size of 40x40.
[0052] For FCM, channel segmentation is first performed using the following formula (14): (14) in, This represents channel segmentation, which is the input feature map. (e.g., P1) Divided by channel and ; , , , The parameter matrix representing the feature map, The ratio for channel segmentation, ranging from 0. 1, C For the number of channels, H and W The height and width of the feature map.
[0053] Secondly, branching is performed using the following formulas (15) and (16): (15) (16) in, Indicates the number of channels is C A convolution operation with a kernel size of 3x3. Indicates the number of channels is C A convolution operation with a kernel size of 1x1. That is... The output features, containing rich channel information, are obtained through two 3x3 convolutions and one pointwise convolution. , After a pointwise convolution, output features that retain more original spatial information are obtained. .
[0054] Afterwards, based on and The complementary mapping is performed, and the specific process is as follows: First, the features are determined using the following formula (17). Perform depthwise separable convolution operations.
[0055] (17) in, This represents the mapping of each feature layer channel. For the first i One convolutional kernel, For the first i One input channel, For the corresponding single output channel; the output result after depthwise separable convolution is .
[0056] Secondly, global average pooling is performed to obtain global information for each channel, and key information weights are obtained through a sigmoid layer, resulting in unique weights for each channel. The process is as follows: (18) (18) in, The output result is The i line, number The values in the column; To Sum and average.
[0057] The expression for the sigmoid activation function is shown in formula (19): (19) in, For the input of the sigmoid activation function, This is the output of the sigmoid activation function.
[0058] For the feature, use the following formula (20). Extract spatial information and generate spatial information weights. N .
[0059] (20) Among them, features Dimensionality reduction convolution is performed using a 1x1 one-dimensional convolution kernel, followed by Batch Normalization, and finally the spatial information weights are output through a sigmoid activation function. .
[0060] Obtain channel information weights M Spatial information weights N Then, they are mapped to the contents respectively. and The features are then obtained by concatenating the two branches by channels and performing a 1x1 convolution to adjust the number of channels. , This feature represents the number of output channels. The feature, which includes a dual mapping of spatial and semantic relations, is calculated as follows (21).
[0061] (twenty one) in, This involves multiplying each element individually.
[0062] The above completes the process of the FCM module processing the input shallow feature information feature map. The process involves channel segmentation and information complementation, ultimately outputting a feature map that contains a more balanced mix of channel and spatial information. .
[0063] For ease of understanding, a specific embodiment will be used as an example: Step 1: Classify and label the drone aerial images to construct a drone target detection dataset.
[0064] In this step, drones can be used to collect aerial images or existing drone aerial image databases. Labeling tools such as AnyLabeling can be used to label the images to obtain a drone target detection dataset.
[0065] Step 2: By improving the backbone network and encoder of the D-FINE-n model, an improved D-FINE-n target detection model is obtained. The main improvements to this model are the use of complementary feature mapping technology and cross-scale feature focusing technology. Therefore, the improved model can be simply referred to as FCM-CFFM-D-FINE. In this step, specific improvements to the backbone network and encoder include: (2.1) FCM was introduced into the backbone network to replace the original HG_Stage module, which made the model lighter and alleviated the problem of imbalance between spatial and channel information in feature extraction.
[0066] (2.2) CFFM was introduced into the encoder, the structure of the encoder was redesigned, and the model’s ability to understand feature map details and global information was enhanced.
[0067] (2.3) At the same time, the windmill convolution module PConv was introduced into the backbone network and encoder, which improved the model's ability to extract small target features and reduced the number of model parameters and computational cost.
[0068] like Figure 3 As shown, Figure 3 This diagram illustrates the structure of the improved D-FINE-n model provided in this application for the UAV small target detection method. In the backbone network, the original HG_Stage module is replaced with an FCM for feature extraction. Simultaneously, PConv is introduced as a downsampling convolutional module connecting the four FCMs, forming a lightweight backbone network with low parameter and computational complexity. This network transforms UAV aerial images into feature maps of multiple sizes with hierarchical semantic information and spatial structure, providing high-quality basic features for the subsequent encoder (neck) to further extract and fuse feature map information.
[0069] Specifically, in the backbone, assuming the drone aerial image is P1 with a size of 640x640, the image is processed by Stem Block to obtain a feature map with a size of 160x160 after being downsampled by 4 times. Then, an FCM module is used to extract shallow information from the image to obtain feature map P2. After being downsampled by a PConv convolution with a stride of 2 and an FCM module, a feature map P3 with a size of 80x80 after being downsampled by 8 times is obtained. The operation is repeated to obtain a feature map P4 with a size of 40x40 after being downsampled by 16 times. Finally, a PConv convolution with a stride of 1 and an FCM module are used to obtain a deep information feature map P5 with a size of 40x40 after being downsampled by 16 times.
[0070] like Figure 4 As shown, Figure 4 This is a schematic diagram of the FCM provided in the UAV small target detection method based on the improved D-FINE-n model of this application. The specific workflow of the FCM in the backbone in this embodiment is as follows: Channel splitting is performed using the above formula (14); branching is performed using the above formulas (15) and (16); complementary mapping is performed using the above formulas (17), (18), (19) and (20); and feature aggregation is performed using formula (21).
[0071] like Figure 5 As shown, Figure 5 This diagram illustrates the sigmoid and SiLU activation functions involved in the UAV small target detection method based on the improved D-FINE-n model in this application.
[0072] like Figure 6 As shown, Figure 6 This is a schematic diagram of the PConv structure involved in the UAV small target detection method based on the improved D-FINE-n model in this application. The specific workflow of Pconv in this embodiment is as follows: Branch filling is performed using formulas (7), (8), (9), (10), and (11). Feature fusion and output are performed using formulas (12) and (13). First, the outputs of each branch are concatenated by channel to obtain the concatenated features. Its channel count is C2, and its size remains unchanged; then... A convolutional operation is performed using a convolutional layer with C2 channels and a 2x2 kernel size. Padding is removed to adjust the height and width of the feature map after convolution, so that its height and width become H and W. Finally, normalized Batch Normalization and the SiLU activation function are applied to obtain the final output of the module. Among them, for normalized BN and activation function SiLU, Figure 6 Not shown in the image.
[0073] like Figure 3 As shown, a CFFM module is introduced into the encoder (neck). The CFFM module enhances the detection model's ability to understand global information of the feature map while retaining the model's ability to understand detailed information. The PConv module is used as a downsampling convolution module, and the structure of the encoder is redesigned, which makes the encoder more capable of extracting feature information of small targets. It can better perform secondary information extraction and feature fusion on the feature maps of multiple scales output by the backbone network, and provide the detection head with a higher quality fused feature map that is more conducive to the detection of small targets.
[0074] Specifically, the encoder receives feature maps P2, P3, and P5 output from the backbone network. Each input feature map undergoes a 1x1 convolution operation with a stride of 1 to obtain P2', P3', and P5'. P5' then passes through a Transformer module to obtain P5'', which is then input into the CFFM module to extract information from the three scale feature maps, completing cross-scale feature information fusion and outputting a multi-scale fused feature map P6 containing both detailed and global information. P6 undergoes upsampling and channel concatenation with P2', and then passes through the RepNCSPELAN4 module for deep feature information processing. The information is extracted to obtain P7. P6 is downsampled and concatenated with P5'', and then processed by the RepNCSPELAN4 module for deep feature information extraction to obtain P8. Then, P6, P7 and P8 are input into the CFFM module for deep feature information extraction and fusion to output P9. P9 is downsampled and concatenated with P8, and then processed by the RepNCSPELAN4 module for deep feature information extraction to obtain P10. Finally, feature maps P9 and P10 are input into the detection head to complete target category prediction and bounding box regression, and determine the category and location of each target in the image.
[0075] like Figure 7 As shown, Figure 7 This is a schematic diagram of the CFFM structure provided in the UAV small target detection method based on the improved D-FINE-n model of this application. The specific workflow of CFFM in this embodiment is as follows: Branching is performed using formulas (1), (2), and (3) above; feature fusion is performed using formula (4) above; multi-scale deep convolution is performed using formula (5) above; post-processing is performed using formula (6) above, and feature output is performed.
[0076] Step 3: Based on the aforementioned UAV target detection dataset, refine the improved D... The FINE-n model is trained to obtain a small target detection model for drones.
[0077] In this embodiment, the step specifically includes: (3.1) The UAV target detection dataset used in this embodiment is the VisDrone dataset. The VisDrone dataset is a large-scale UAV-view dataset open-sourced by teams such as Tianjin University. The official labeled dataset for target detection tasks contains 6471 images for training, 548 images for validation, and 1610 images for testing, totaling 8629 images. Ten categories of objects are defined, including pedestrians, crowds, bicycles, cars, vans, trucks, tricycles, covered tricycles, buses, and motorcycles. From the UAV perspective, most of these are small objects. The training set allows the model to learn the feature patterns of the data during training. The validation set monitors the training process, tunes hyperparameters, and prevents overfitting. The test set evaluates the model's final generalization ability and provides reliable performance indicators of the model in real-world scenarios.
[0078] (3.2) The training environment and parameters in this embodiment are set as follows: The experimental environment was based on PyTorch 2.4.1 and Python 3.11. Training used two NVIDIA GeForce RTX 3080 GPUs and the AdamW optimizer with default momentum parameters. The batch size was set to 4, and the number of epochs was 200. The global learning rate was set to a default value of 0.0004, the weight decay coefficient was set to 0.0001 to prevent overfitting, the backbone network learning rate was set to 0.0002, weight decay was disabled in both the encoder and decoder, and pre-trained weights were not used during training.
[0079] (3.3) The improved D-FINE-n model is trained using the above dataset and training settings. After reaching the maximum number of training iterations of 200, the weight file of the best iteration is obtained. The improved D-FINE-n model based on this weight is the required UAV small target detection model.
[0080] Step 4: Based on the UAV small target detection model obtained above, detect the UAV aerial images, determine the position and category of different objects in the UAV aerial images, and complete the test of model performance.
[0081] In this embodiment, to verify the performance of the improved model, this step uses the test set in the VisDrone dataset to test the trained model, and compares the test results with the current mainstream object detection models and the original model. Specifically: (4.1) Evaluation Indicators In this embodiment, the evaluation criterion used is the COCO evaluation index, with the main focus on the average precision (AP), calculated as follows: (twenty two) (twenty three) (twenty four) in, TP This indicates the number of detection boxes whose Intersection over Union (IoU) ratio with the ground truth bounding boxes is greater than a preset threshold. FP This indicates the number of detection boxes with an IoU less than the threshold. FN P represents the number of undetected true bounding boxes, P is the accuracy, and R is the recall. The area under the PR curve is the average precision (AP). AP balances both accuracy and recall, providing a more comprehensive evaluation of the model.
[0082] AP metrics can be categorized by IoU threshold. AP@IoU represents the average precision when the threshold is set to IoU. The IoU threshold ranges from 0.5 to 0.95, with a step size of 0.05 for a total of 10 values. Therefore, the mean AP is AP@[0.5:0.95], or mAP. (25) To verify the model's performance in detecting small targets, three metrics were used to decompose mAP by scale: AP_small, AP_medium, and AP_large. AP_small represents the mAP for small targets (area < 32² pixels), AP_medium represents the mAP for medium targets (32² ≤ area < 96² pixels), and AP_large represents the mAP for large targets (area ≥ 96² pixels). Since the model considers both accuracy and parameter count, computational cost (FLOPs) and parameter count (Params) are also important evaluation metrics.
[0083] (4.2) Comparison of experimental parameters for the model In this embodiment, the training environment used for all comparison models is the same as that used for the FCM-CFFM-D-FINE model. That is, the experimental environment is based on PyTorch 2.4.1 and Python 3.11, the GPU used for training is two NVIDIA GeForce RTX3080s, the batch size is 4, the maximum number of iterations is 200 epochs, no pre-trained weights are used, and the other parameters are the default values.
[0084] When testing all models using the VisDrone dataset's test set, the batch size was set to 4, with the rest set to default.
[0085] (4.3) Comparative test results The comparison models selected in this embodiment include the current mainstream object detection models YOLO series, specifically YOLOv8-n, YOLOv9-t, YOLOv10-n, YOLOv11-n, YOLOv12-n, YOLOv13-n, as well as the original model D-FINE-n and the DEIM-n model, which is also based on transformer. The experimental results are all test results of the models on the VisDrone dataset test set.
[0086] The comparative experimental results of this embodiment are shown in Table 1 below: Table 1: Results of the comparative experiment.
[0087] As shown in Table 1, compared with the YOLO series models, the FCM-CFFM-D-FINE model exhibits superior performance with fewer parameters and similar computational cost (FLOPs). In terms of accuracy, compared to the best-performing YOLOv12-n model's 15.6%, FCM-CFFM-D-FINE achieves 19.1% mAP, a 3.5% improvement. Further analysis reveals that while FCM-CFFM-D-FINE's accuracy (AP1) for large targets is 3.8% lower than YOLOv12-n, its accuracy (APs) and (APm) for small and medium targets are improved by 4.4% and 3.3% respectively, ultimately leading to an overall improvement in mAP. This demonstrates the advantage of FCM-CFFM-D-FINE in small target feature recognition. In drone application scenarios, small targets usually occupy a larger proportion. Compared with the YOLO series models, the drone small target detection model based on FCM-CFFM-D-FINE has greater advantages. At the same time, it has fewer parameters and lower computational cost, lower deployment storage and computing requirements, and an FPS of 156 ensures inference speed and real-time performance during operation.
[0088] Compared to the original models D-FINE-n and DEIM-n, FCM-CFFM-D-FINE reduces the number of parameters from 3.73M to 1.62M, while maintaining a relatively constant computational cost. In terms of accuracy (mAP), it improves by 3.6% compared to D-FINE-n and 3.9% compared to DEIM-n. Further analysis shows that compared to the original model D-FINE-n, FCM-CFFM-D-FINE maintains the same accuracy for large targets (34.7%), while improving accuracy for small and medium targets by 2.9% and 4.4%, respectively. This demonstrates that FCM-CFFM-D-FINE retains its ability to detect large targets while enhancing its ability to identify small and medium-sized targets.
[0089] like Figure 8 As shown, Figure 8 This chart compares the parameter count and detection accuracy of the FCM-CFFM-D-FINE model provided in this application for UAV small target detection based on the improved D-FINE-n model, with mainstream target detection models and the original D-FINE-n model. (Combined with...) Figure 8 It can be seen that FCM-CFFM-D-FINE achieves the highest detection accuracy with minimal storage requirements (fewest parameters), demonstrating the promising prospects for deployment and application of this model.
[0090] In summary, this application addresses the challenges of extracting feature information from small targets and achieving low recognition accuracy in UAV aerial image recognition. It improves the backbone network and encoder of the D-FINE-n target detection model. Feature complementarity mapping alleviates the imbalance between channel and spatial information extraction; cross-scale feature focusing enhances the model's understanding of feature map details and global information; and windmill convolution improves the improved model's ability to extract and recognize small target features. Finally, while maintaining real-time performance and low computational cost, the application achieves higher target detection accuracy in UAV scenarios.
[0091] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.
[0092] Based on the same inventive concept, this application also provides a UAV small target detection system based on an improved D-FINE-n model for implementing the aforementioned UAV small target detection method based on the improved D-FINE-n model. The solution provided by this device is similar to the implementation scheme described in the above method. Therefore, the specific limitations of one or more embodiments of the UAV small target detection system based on the improved D-FINE-n model provided below can be found in the limitations of the UAV small target detection method based on the improved D-FINE-n model described above, and will not be repeated here.
[0093] In one exemplary embodiment, such as Figure 9As shown, a small target detection system for UAVs based on an improved D-FINE-n model is provided, including: a preprocessing module 100, a convolution module 200, a first fusion module 300, an upsampling module 400, a downsampling module 500, a second fusion module 600, a channel stitching module 700, and a detection module 800, wherein: The preprocessing module 100 is used to preprocess the drone aerial image based on the backbone network of the improved D-FINE-n model to obtain small-scale feature maps, medium-scale feature maps and large-scale feature maps in response to receiving the drone aerial image. The convolution module 200 is used to perform a 1x1 first convolution operation with a stride of 1 on the small-scale feature map, the medium-scale feature map and the large-scale feature map respectively by the encoder based on the improved D-FINE-n model; The first fusion module 300 is used to perform feature transformation and information integration on the small-scale feature map after convolution based on the encoder, and to perform fusion operation on the large-scale feature map after convolution, the medium-scale feature map after convolution, and the integrated small-scale feature map based on the cross-scale feature focusing module CFFM of the encoder to obtain the first fused feature map. The upsampling module 400 is used to perform upsampling operation on the first fused feature map, and to perform channel concatenation between the upsampled first fused feature map and the convolved large-scale feature map, and to perform deep feature extraction to obtain the first deep feature map. The downsampling module 500 is used to perform downsampling operation on the first fused feature map based on windmill convolution PConv, and to perform channel concatenation of the downsampled first fused feature map and the integrated small-scale feature map, as well as to perform deep feature extraction to obtain the second deep feature map. The second fusion module 600 is used to perform a fusion operation on the first fused feature map, the first deep feature map, and the second deep feature map based on the CFFM to obtain the second fused feature map. The channel splicing module 700 is used to perform downsampling operation on the second fused feature map based on the PConv, and to perform channel splicing of the downsampled second fused feature map and the second deep feature map, and to perform deep feature extraction to obtain a third deep feature map; The detection module 800 is used to obtain the small target detection result in the UAV aerial image based on the second fused feature map output by the encoder and the third deep feature map.
[0094] In some embodiments, the fusion operation involving the first fusion module 300 and the second fusion module 600 includes: The three input feature maps before fusion are scale aligned; the three input feature maps before fusion are: the large-scale feature map after convolution, the medium-scale feature map after convolution, and the small-scale feature map after integration, or the first fused feature map, the first deep feature map, and the second deep feature map. Channel concatenation is performed on the scale-aligned feature maps to obtain cross-scale fused feature maps. The cross-scale fused feature map is subjected to multi-scale deep convolution processing and post-processing.
[0095] In some embodiments, the fusion operation involved in the first fusion module 300 and the second fusion module 600 specifically includes: The cross-scale fused feature maps are subjected to a second convolution operation using multiple convolution kernels of different sizes. The convolutional cross-scale fused feature maps are then stacked and summed.
[0096] In some embodiments, the fusion operation involving the first fusion module 300 and the second fusion module 600 is implemented by the following formula: Among them, the The multi-scale deep convolutional features are obtained by performing multi-scale deep convolution processing on the cross-scale fused feature map; For summation operations; For stacking operations; This refers to the cross-scale fusion feature map; This is the second convolution operation, where the kernel size is... Depth-separable convolution operations, The value can be 3, 5, 7 or 9.
[0097] In some embodiments, the downsampling operation is implemented via PConv.
[0098] In some embodiments, the apparatus further includes: The replacement module is used to replace the HG_Stage module in the backbone network with a Feature Complementary Mapping Module (FCM) and a downsampling module, wherein the output of the FCM is the input of the downsampling module.
[0099] In some embodiments, the downsampling module is a PConv module.
[0100] The modules in the aforementioned UAV small target detection system based on the improved D-FINE-n model can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.
[0101] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 10 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a small target detection method for unmanned aerial vehicles (UAVs) based on an improved D-FINE-n model.
[0102] Those skilled in the art will understand that Figure 10 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.
[0103] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.
[0104] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.
[0105] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.
Claims
1. An unmanned aerial vehicle small target detection method based on an improved D-FINE-n model, characterized in that, The improvement of the D-FINE-n model lies in the introduction of a cross-scale feature focusing module (CFFM) and a windmill convolution (PConv) into the encoder of the D-FINE-n model. The D-FINE-n model includes a backbone network, an encoder, and a detection head. The backbone network is used to receive aerial images captured by the UAV. The detection head is used to output the detection results of targets in the UAV aerial images. The UAV small target detection method based on the improved D-FINE-n model includes: In response to receiving the drone aerial image, the drone aerial image is preprocessed based on the backbone network of the improved D-FINE-n model to obtain small-scale feature maps, medium-scale feature maps and large-scale feature maps; The encoder based on the improved D-FINE-n model performs a 1x1 first convolution operation with a stride of 1 on the small-scale feature map, the medium-scale feature map, and the large-scale feature map, respectively. Based on the encoder, feature transformation and information integration are performed on the small-scale feature map after convolution. Based on the CFFM, a fusion operation is performed on the large-scale feature map after convolution, the medium-scale feature map after convolution, and the integrated small-scale feature map to obtain the first fused feature map. The first fused feature map is upsampled, and the upsampled first fused feature map is concatenated with the convolved large-scale feature map through channels, and deep feature extraction is performed to obtain the first deep feature map. Based on the PConv, the first fused feature map is downsampled, and the downsampled first fused feature map is concatenated with the integrated small-scale feature map through channels, and deep feature extraction is performed to obtain the second deep feature map. Based on the CFFM, the first fused feature map, the first deep feature map, and the second deep feature map are fused to obtain the second fused feature map; Based on the PConv, the second fused feature map is downsampled, and the downsampled second fused feature map is concatenated with the second deep feature map through channels, and deep feature extraction is performed to obtain the third deep feature map. The small target detection result in the UAV aerial image is obtained based on the second fused feature map and the third deep feature map output by the encoder.
2. The UAV small target detection method based on the improved D-FINE-n model according to claim 1, characterized in that, The fusion operation includes: The three input feature maps before fusion are scale aligned; the three input feature maps before fusion are: the large-scale feature map after convolution, the medium-scale feature map after convolution, and the small-scale feature map after integration, or the first fused feature map, the first deep feature map, and the second deep feature map. Channel concatenation is performed on the scale-aligned feature maps to obtain cross-scale fused feature maps. The cross-scale fused feature map is subjected to multi-scale deep convolution processing and post-processing.
3. The UAV small target detection method based on the improved D-FINE-n model according to claim 2, characterized in that, The multi-scale deep convolution processing of the cross-scale fused feature map includes: The cross-scale fused feature maps are subjected to a second convolution operation using multiple convolution kernels of different sizes. The convolutional cross-scale fused feature maps are then stacked and summed.
4. The UAV small target detection method based on the improved D-FINE-n model according to claim 3, characterized in that, The second convolution operation is performed on the cross-scale fused feature map using multiple convolution kernels of different sizes. The convolutional cross-scale fused feature maps are then stacked and summed, as implemented using the following formula: Among them, the The multi-scale deep convolutional features are obtained by performing multi-scale deep convolution processing on the cross-scale fused feature map; For summation operations; For stacking operations; This refers to the cross-scale fusion feature map; This is the second convolution operation, where the kernel size is... Depth-separable convolution operations, The value can be 3, 5, 7 or 9.
5. The UAV small target detection method based on the improved D-FINE-n model according to claim 1, characterized in that, The method further includes: The HG_Stage module in the backbone network is replaced with a Feature Complementary Mapping Module (FCM) and a downsampling module, with the output of the FCM serving as the input of the downsampling module.
6. The UAV small target detection method based on the improved D-FINE-n model according to claim 5, characterized in that, The downsampling module is a PConv module.
7. A small target detection system for unmanned aerial vehicles (UAVs) based on an improved D-FINE-n model, characterized in that, The improvement of the D-FINE-n model lies in the introduction of a cross-scale feature focusing module (CFFM) and a windmill convolution (PConv) into the encoder of the D-FINE-n model; the D-FINE-n model includes a backbone network, an encoder, and a detection head, wherein the backbone network is used to receive aerial images taken by UAVs. The detection head is used to output the detection results of targets in the drone aerial images; The UAV small target detection system based on the improved D-FINE-n model includes: The preprocessing module is used to preprocess the drone aerial image based on the backbone network of the improved D-FINE-n model in response to receiving the drone aerial image to obtain small-scale feature maps, medium-scale feature maps and large-scale feature maps. The convolution module is used to perform a 1x1 first convolution operation with a stride of 1 on the small-scale feature map, the medium-scale feature map and the large-scale feature map respectively, based on the encoder of the improved D-FINE-n model; The first fusion module is used to perform feature transformation and information integration on the small-scale feature map after convolution based on the encoder, and to perform fusion operation on the large-scale feature map after convolution, the medium-scale feature map after convolution, and the integrated small-scale feature map based on the CFFM to obtain the first fused feature map. The upsampling module is used to perform upsampling on the first fused feature map, and to concatenate the upsampled first fused feature map with the convolved large-scale feature map through channels, and to perform deep feature extraction to obtain the first deep feature map. The downsampling module is used to perform downsampling operation on the first fused feature map based on the PConv, and to perform channel concatenation of the downsampled first fused feature map and the integrated small-scale feature map, and to perform deep feature extraction to obtain a second deep feature map. The second fusion module is used to perform a fusion operation on the first fused feature map, the first deep feature map, and the second deep feature map based on the CFFM to obtain the second fused feature map; The channel stitching module is used to perform downsampling operation on the second fused feature map based on the PConv, and to perform channel stitching of the downsampled second fused feature map and the second deep feature map, as well as to perform deep feature extraction to obtain a third deep feature map; The detection module is used to obtain the small target detection result in the UAV aerial image based on the second fused feature map output by the encoder and the third deep feature map.
8. A computer device, characterized in that, It includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein when the program or instructions are executed by the processor, they implement the steps of the UAV small target detection method based on the improved D-FINE-n model as described in any one of claims 1-6.
9. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the steps of the UAV small target detection method based on the improved D-FINE-n model as described in any one of claims 1-6.