Road aerial photography small target detection method fusing direction feature enhancement and scale adaptation
By improving the YOLO11n model and integrating directional feature enhancement and scale adaptation methods, the problems of insufficient accuracy and real-time performance in small target detection in road aerial photography were solved, and efficient detection of vehicles and pedestrians was achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- LIAONING UNIVERSITY
- Filing Date
- 2026-01-15
- Publication Date
- 2026-05-01
AI Technical Summary
Traditional methods for detecting small targets in aerial road photography struggle to balance detection accuracy and real-time performance in real-time UAV monitoring, especially given the challenges of small target features being easily masked by background noise, difficulty in adapting to targets from multiple directions, and low positioning accuracy for targets at multiple scales.
We employ a method that combines directional feature enhancement and scale adaptation. By improving the YOLO11n model, we design the C3K2-DAP module to extract target features of different orientations through asymmetric padding and directional convolution. We combine this with the CGA-SWS module to perform multi-scale feature fusion and introduce the SD dynamic loss mechanism to optimize the weight allocation of the loss function.
It significantly improves the accuracy and real-time performance of small target detection in road aerial photography, and can achieve accurate positioning and classification of vehicles and pedestrians while ensuring the lightweight characteristics of the model, thus meeting the real-time monitoring needs of UAVs.
Smart Images
Figure CN121962984A_ABST
Abstract
Description
A road aerial photography method for small target detection that combines directional feature enhancement and scale adaptation Technical Field
[0001] This invention belongs to the field of computer vision and target detection, and specifically relates to a method for detecting small targets in road aerial photography scenarios. It proposes a method for detecting small targets in road aerial photography that integrates directional feature enhancement and scale adaptation. This method is applicable to intelligent traffic perception scenarios such as road traffic monitoring, dynamic tracking of moving vehicles and pedestrians using UAV aerial photography technology. It can efficiently solve the problems of low pixel ratio of small targets in aerial images, easy obscuring of features by the background, random target orientation, and large scale differences, which lead to insufficient detection accuracy. Background Technology
[0002] With the deep integration of drone aerial photography technology and intelligent transportation, small target detection in road aerial photography has become a core supporting technology for traffic flow statistics and rapid response to emergencies. This technology acquires road images from a high-altitude perspective using drones, enabling the identification and localization of small targets such as vehicles and pedestrians with low pixel ratios. However, traditional detection methods struggle to balance accuracy and real-time performance due to issues such as sparse features, random orientation, large scale differences, and background noise interference associated with small targets.
[0003] In traditional object detection frameworks, two-stage methods (such as Faster R-CNN) offer high localization accuracy, but their reliance on fixed-scale anchor point design makes them unsuitable for the sparse features of small targets in aerial photography, and their detection frame rate is below 15fps, failing to meet the real-time monitoring needs of drones. Single-stage methods (such as YOLOv8 and YOLOv11n) offer significantly improved frame rates, but suffer from three major drawbacks: First, small target features are easily obscured by background noise such as road textures and cloud shadows, resulting in a false negative rate of over 30%. Second, road target orientations (horizontal, vertical, 45°, etc.) are random and exhibit significant scale differences (the scale ratio of trucks in the foreground to pedestrians in the distance reaches 10:1), making it difficult for fixed convolution and single-scale fusion mechanisms to effectively extract features. Third, the fixed-weight loss function lacks differentiation in its constraints on small targets (where the Intersection over Union (IoU) is easily affected by annotation errors) and large targets (where the impact of positional deviations is amplified), leading to low localization accuracy.
[0004] Existing methods for improving small targets in aerial photography mostly focus on optimizing single aspects (such as expanding the receptive field and optimizing scale fusion), failing to construct a full-link optimization mechanism of "directional feature enhancement - scale-adaptive fusion - dynamic loss constraints." Some methods expand the receptive field by stacking convolutional layers but ignore the randomness of target orientation; some methods improve multi-scale fusion strategies but fail to model the local details of small targets with blurred edges; some methods adjust the weights of the loss function but fail to dynamically allocate constraint priorities based on the target's true scale. These shortcomings make it difficult for existing methods to simultaneously achieve accuracy and real-time performance in small target detection in road aerial photography scenarios, failing to meet the practical application needs of intelligent transportation. Summary of the Invention
[0005] To address the problems of existing technologies, this invention proposes a road aerial photography small target detection method that integrates directional feature enhancement and scale adaptation. This method effectively solves the problems of small target features being easily obscured by the background, difficulty in adapting to targets in multiple directions, and low positioning accuracy for targets at multiple scales. While ensuring the lightweight nature of the model, it significantly improves detection accuracy and real-time performance, meeting the real-time small target detection requirements of road aerial photography.
[0006] The present invention is achieved through the following technical solution: a method for detecting small targets in road aerial photography that integrates directional feature enhancement and scale adaptation, the steps of which are: Step 1: Divide the dataset: randomly divide the road dataset obtained by aerial photography to obtain a training set, a validation set and a test set, and determine the target category in the dataset.
[0007] The specific method is as follows: The selected dataset should cover images of various road scenes, aerial angles, and different lighting environments, and the images should be labeled with corresponding categories; the dataset should be randomly divided in a ratio of 8:1:1 to obtain training set, validation set and test set.
[0008] Step 2: Build an improved YOLO11 detection model: Select YOLO11n as the baseline model for improvement. Replace the C3K2 module in the backbone network with C3K2-DAP, add the designed CGA-SWS module in front of the detection head, and introduce SD Loss.
[0009] The specific method is as follows: The YOLO11n model is selected as the baseline improvement model, and improvements are made in the following way: First, a DAP module is designed and applied to C3K2 in the YOLO11n model backbone network to obtain C3K2-DAP: This involves extracting different orientation enhancement features F of the target by performing asymmetric padding and directional convolution on the input image. out The C3K2-DAP module first employs four sets of asymmetric-padded directional convolutional kernels, corresponding to the horizontal, vertical, 45°, and 135° directions, respectively. For each directional branch, the padding parameter P is dynamically adjusted. nTo ensure that the output feature maps after convolution in all directions have the same size, the padding operation is as follows:
[0010] Where X is the input feature tensor with dimensions C×H×W (C is the number of channels, H is the height, and W is the width), X n The filled feature in the nth direction (n∈{1,2,3,4}); the filled feature X n Apply the directional convolution kernel K in the corresponding direction n Generate partition features F n (The number of channels is reduced to C'), and its calculation formula is as follows:
[0011] Where Conv represents the convolution operation.
[0012] Calculate the feature F for each partition n Directional gradient energy E n The gradient energy E measures the degree of matching between the branch and the target direction. n The calculation formula is as follows:
[0013] Where C is the number of feature channels, and H and W are the height and width of the feature tensor, respectively. n For the gradient operator in the corresponding direction, F n,c,i,j Let be the eigenvalue of the nth feature tensor at the spatial position (i,j) in the c-th channel.
[0014] E is obtained through the Softmax function n Normalization, generating dynamic weights W n The enhanced feature F is obtained by weighted concatenation of features from the four directions. out The splicing operation is represented as follows:
[0015] Where Concat is the channel concatenation operation; dynamic weights
[0016] Secondly, the CGA-SWS module is designed before being applied to the detection head: The CGA-SWS module consists of two parts: CGA fusion and SWS enhancement. The design idea is to fuse three branches of CGA: channel attention, spatial attention, and pixel attention, and then use SWS enhancement combined with the block-based approach to obtain processed local feature sub-blocks. The mid-level feature x output from layer C5 of the neck FPN network is concatenated with the high-level semantically guided feature y of the detection head after 1×1 convolution dimensionality reduction to obtain the initial feature. The initial feature is then input into the channel attention branch and the spatial attention branch in parallel. The outputs of the two branches are fused to obtain the coarse-grained weights patchn1 that focus on the global feature distribution. Pattn1 is concatenated with the pixel-level detail weights output from the pixel attention branch and activated by Sigmoid to obtain the fine-grained weights patchn2. The initial feature initia is then element-wise weighted with [x;y] (the concatenation of the channels of x and y) using patchn2 to obtain the fused feature F. fusion , fusion feature F fusion The calculation formula is as follows:
[0017] in This is element-wise multiplication. This is a feature fusion operation.
[0018] F fusion feature F fusion Divide the height H and width W into two equal parts each, resulting in four local sub-blocks B. k (k∈{1,2,3,4}). Mean pixel value for each sub-block. With variance Calculate pixel saliency score The calculation formula is as follows:
[0019] Will Enhanced weights are obtained through Sigmoid activation. , for B k Element-level weighting is performed, and then the four enhanced sub-blocks are spliced together according to their original spatial positions to obtain the local feature sub-blocks after block enhancement processing. The calculation formula is as follows:
[0020] Where i is the traversal index of the pixels within the sub-block.
[0021] Finally, the SD dynamic loss mechanism is introduced, and the SDB Loss is designed: a scale mapping relationship is established, and the original input image size is W. o ×H o The current feature map size is W c ×H c Define the scaling ratio R between the feature map and the original image. oc The calculation formula is as follows:
[0022] Based on the area B of the target's true bounding box in the original image coordinate system gt Combined with the maximum defined size B of the small target gt-max Calculate the dynamic weighting coefficient β using the balance coefficient δ. B The calculation formula is as follows:
[0023] The min function is used to limit β. B The weight should not exceed δ to avoid weight imbalance.
[0024] Based on β B Obtain the scale loss weight β LBS Location loss weight β LBL Specifically, it is expressed as follows:
[0025] The final SDB Loss representation is as follows:
[0026] Where L BS For scale loss, L BL This is due to position loss.
[0027] Step 3: Preprocess the data in the training set obtained by segmentation: reverse, scale, randomly occlude, add noise appropriately, and perform grayscale transformation on the image data.
[0028] The specific method is as follows: Aerial image quality is easily affected by external interference, such as low contrast between the target and background in nighttime images, and obstruction by birds or trees. However, the number of images containing such interference in the dataset is relatively small. To train the model, the dataset is expanded by inverting, scaling, randomly occluding, appropriately adding noise, and performing grayscale transformations on the image data to simulate nighttime or occluded scenes.
[0029] Step 4 involves inputting the preprocessed training set and the pre-divided validation set into the improved YOLO11 model for training until the loss curve converges.
[0030] The specific method was as follows: Training was conducted on the Linux 5.15 operating system. For hardware configuration, an Intel(R) Xeon(R) Platinum 8362 CPU @ 2.80GHz processor, 45GB of memory, and an RTX4090 GPU (24GB) were used. For software configuration, PyTorch 2.5.1 and Python 3.12 were used, along with CUDA version 12.4. The number of iterations was set to 300, the batch size to 16, and both the initial and final learning rates to 0.01. The SGD optimizer was employed. The improved YOLO11 model showed convergence of its loss curve within 300 iterations.
[0031] Step 5: Model performance verification: Input the partitioned test set into the trained and improved YOLO11 model to obtain the object detection results.
[0032] The specific method is as follows: The test set image data is input into the trained improved YOLO11 model; when loading the test set images, the image size specifications are kept consistent with those of the training set preprocessing to avoid affecting the detection accuracy due to size differences; the model first extracts features from the input test set images through the improved backbone network, then enhances the target orientation features and suppresses background noise through the C3K2-DAP module, and then optimizes the multi-scale feature fusion through the CGA-SWS module of the neck network; finally, the detection head is combined with SDB Loss to output the target category, confidence score and bounding box coordinates, and the detection results with confidence scores higher than a preset threshold (such as 0.5) are selected to obtain the final road aerial small target detection output, including accurate localization and classification information of targets such as pedestrians and vehicles.
[0033] The beneficial effects of this invention are as follows: First, the YOLO11 model is optimized for aerial road scenes by designing a C3K2-DAP module, which extracts features in different directions through asymmetric padding and directional convolution, and combines dynamic directional attention to suppress background interference. Second, to address the issue that local details are easily diluted when fusing multi-scale target features in road scenes, a CGA-SWS scale-adaptive fusion module is designed, introducing a three-level collaborative attention mechanism of "channel-space-pixel" to deepen the coupling of multi-scale information, and combining block processing to optimize the feature representation of blurred targets at the edges. Finally, to address the issue of large scale differences in small targets in road aerial photography and poor adaptability of traditional loss constraints, an SD dynamic loss mechanism is introduced, and a loss function SDB Loss is designed that dynamically allocates scale and position loss weights based on the true scale of the target to ensure positioning accuracy. This invention maintains high detection accuracy for vehicle and pedestrian targets in road scenes while keeping the model size relatively low, making it suitable for real-time small target detection tasks in road aerial photography. Attached Figure Description
[0034] Figure 1 shows the overall flow of the present invention; Figure 2 shows the structure of the improved YOLO11 model in the present invention; Figure 3 shows the structure of the C3K2-DAP module in the present invention; Figure 4 shows the structure of the DAP module in the C3K2-DAP module in the present invention; Figure 5 shows the CGA fusion structure in the CGA-SWS module in the present invention; Figure 6 shows the SWS enhancement structure in the CGA-SWS module in the present invention. Detailed Implementation
[0035] A method for detecting small targets in road aerial photography that integrates directional feature enhancement and scale adaptation includes the following steps: The invention will be described in detail with reference to the accompanying drawings.
[0036] The flowchart of the road aerial photography small target detection method that integrates directional feature enhancement and scale adaptation is shown in Figure 1. The specific method flow includes: 1. Selecting and dividing the dataset.
[0037] The selected dataset should cover images of various road scenes, aerial angles, and different lighting environments, and the images should be labeled with corresponding categories. The dataset should be randomly divided into training, validation, and test sets in an 8:1:1 ratio.
[0038] 2. Construct an improved YOLO11 detection model.
[0039] The YOLO11n model was chosen as the baseline for improvement, and the improvement was carried out in the following way: A DAP module was designed and applied to C3K2 in the backbone network of the YOLO11n model to obtain C3K2-DAP. The C3K2-DAP module first uses four sets of asymmetric-padded directional convolutional kernels, corresponding to the horizontal, vertical, 45°, and 135° directions, respectively. For each directional branch, the padding parameter P is dynamically adjusted. n To ensure that the output feature maps after convolution in all directions have the same size, the padding operation is as follows:
[0040] Where X is the input feature tensor with dimensions C×H×W (C is the number of channels, H is the height, and W is the width), X n The filled feature in the nth direction (n∈{1,2,3,4}); the filled feature X n Apply the directional convolution kernel K in the corresponding direction n Generate partition features F n (The number of channels is reduced to C'), and its calculation formula is as follows:
[0041] Where Conv represents the convolution operation.
[0042] Calculate the feature F for each partitionn Directional gradient energy E n The gradient energy E measures the degree of matching between the branch and the target direction. n The calculation formula is as follows:
[0043] Where C is the number of feature channels, and H and W are the height and width of the feature tensor, respectively. n For the gradient operator in the corresponding direction, F n,c,i,j Let be the eigenvalue of the nth feature tensor at the spatial position (i,j) in the c-th channel.
[0044] E is obtained through the Softmax function n Normalization, generating dynamic weights W n The enhanced feature F is obtained by weighted concatenation of features from the four directions. out The splicing operation is represented as follows:
[0045] Where Concat is the channel concatenation operation; dynamic weights
[0046] Before applying the CGA-SWS module to the detection head: The CGA-SWS module consists of two parts: CGA fusion and SWS enhancement. The design idea is to fuse three branches of CGA—channel attention, spatial attention, and pixel attention—and then use SWS enhancement combined with the block-based approach to obtain processed local feature sub-blocks. The mid-level feature x output from layer C5 of the neck FPN network is concatenated with the high-level semantically guided feature y of the detection head after 1×1 convolution dimensionality reduction to obtain the initial feature. The initial feature is then input into the channel attention branch and the spatial attention branch in parallel. The outputs of the two branches are fused to obtain the coarse-grained weights patchn1 that focus on the global feature distribution. Pattn1 is concatenated with the pixel-level detail weights output from the pixel attention branch and activated by Sigmoid to obtain the fine-grained weights patchn2. The initial feature initia is then element-wise weighted with [x;y] (the concatenation of the channels of x and y) using patchn2 to obtain the fused feature F. fusion , fusion feature F fusion The calculation formula is as follows:
[0047] in This is element-wise multiplication. This is a feature fusion operation.
[0048] F fusion feature F fusionDivide the height H and width W into two equal parts each, resulting in four local sub-blocks B. k (k∈{1,2,3,4}). Mean pixel value for each sub-block. With variance Calculate pixel saliency score The calculation formula is as follows:
[0049] Will Enhanced weights are obtained through Sigmoid activation. , for B k Element-level weighting is performed, and then the four enhanced sub-blocks are spliced together according to their original spatial positions to obtain the local feature sub-blocks after block enhancement processing. The calculation formula is as follows:
[0050] Where i is the traversal index of the pixels within the sub-block.
[0051] Finally, the SD dynamic loss mechanism is introduced, and the SDB Loss is designed: a scale mapping relationship is established, and the original input image size is W. o ×H o The current feature map size is W c ×H c Define the scaling ratio R between the feature map and the original image. oc The calculation formula is as follows:
[0052] Based on the area B of the target's true bounding box in the original image coordinate system gt Combined with the maximum defined size B of the small target gt-max Calculate the dynamic weighting coefficient β using the balance coefficient δ. B The calculation formula is as follows:
[0053] The min function is used to limit β. B The weight should not exceed δ to avoid weight imbalance.
[0054] Based on β B Obtain the scale loss weight β LBS Location loss weight β LBL Specifically, it is expressed as follows:
[0055] The final SDB Loss representation is as follows:
[0056] Where LBS For scale loss, L BL This is due to position loss.
[0057] 3. Preprocess the training data.
[0058] Training was conducted on the Linux 5.15 operating system. Hardware configuration included an Intel(R) Xeon(R) Platinum 8362 CPU @ 2.80GHz processor, 45GB of memory, and an RTX4090 GPU (24GB). Software configuration used PyTorch 2.5.1 and Python 3.12, with CUDA version 12.4. The number of iterations was set to 300, the batch size to 16, and both the initial and final learning rates to 0.01. The SGD optimizer was used. The improved YOLO11 model showed convergence in its loss curve within 300 iterations.
[0059] 4. Train the improved YOLO11 model using the input training and validation sets.
[0060] The test set image data is input into the trained improved YOLO11 model. When loading the test set images, the image size specifications are kept consistent with those of the training set preprocessing to avoid affecting detection accuracy due to size differences. The model first extracts features from the input test set images through the improved backbone network, then enhances the target orientation features and suppresses background noise through the C3K2-DAP module, and then optimizes multi-scale feature fusion through the CGA-SWS module of the neck network. Finally, the detection head, combined with SDB Loss, outputs the target category, confidence score, and bounding box coordinates. Detection results with confidence scores higher than a preset threshold are selected to obtain the final road aerial small target detection output, including accurate localization and classification information of targets such as pedestrians and vehicles.
Claims
1. A method for detecting small targets in road aerial photography that integrates directional feature enhancement and scale adaptation, characterized in that, The steps are as follows: Step 1: Dataset partitioning: The road dataset obtained from aerial photography is randomly partitioned into training, validation, and test sets to determine the target categories in the dataset; Step 2: Improved YOLO11 detection model: YOLO11n is selected as the baseline model for improvement. The C3K2 module in the backbone network is replaced with C3K2-DAP, and a designed CGA-SWS module is added before the detection head to introduce SD Loss; Step 3: Preprocessing of the data in the partitioned training set: The image data is inverted, scaled, randomly occluded, and appropriately noiseed and grayscale transformed; Step 4: Model training: The preprocessed training set and the pre-partitioned validation set are input into the improved YOLO11 model for training until the loss curve converges; Step 5: Model performance verification: The partitioned test set is input into the trained improved YOLO11 model to obtain the target detection results.
2. The method for detecting small road targets in aerial photography that integrates directional feature enhancement and scale adaptation as described in claim 1, characterized in that, In step 1, the specific method is as follows: the selected dataset should cover images of various road scenes, aerial angles, and different lighting environments, and the images should be labeled with corresponding categories; the dataset should be randomly divided in a ratio of 8:1:1 to obtain a training set, a validation set, and a test set.
3. The method for detecting small road targets in aerial photography that integrates directional feature enhancement and scale adaptation as described in claim 1, characterized in that, In step 2, the specific method is as follows: The YOLO11n model is selected as the baseline improvement model, and improvements are made in the following way: Step 2.1 Design the DAP module, apply it to C3K2 in the YOLO11n model backbone network, to obtain C3K2-DAP: This involves extracting different orientation enhancement features F of the target by performing asymmetric padding and directional convolution on the input image. out The DAP module first employs four sets of asymmetric-padded directional convolutional kernels, corresponding to the horizontal, vertical, 45°, and 135° directions, respectively. For each directional branch, the padding parameter P is dynamically adjusted. n To ensure that the output feature maps after convolution in all directions have the same size, the padding operation is as follows: Where X is the input feature tensor with dimensions C×H×W, C is the number of channels, H is the height, W is the width, and X... n Let X be the filled feature in the nth direction, where n∈{1,2,3,4}; n Apply the directional convolution kernel K in the corresponding direction n Generate partition features F n The number of channels is reduced to C', calculated using the following formula: Where Conv is the convolution operation; calculate the feature F for each partition. n Directional gradient energy E n The gradient energy E measures the degree of matching between the branch and the target direction. n The calculation formula is as follows: Where C is the number of feature channels, and H and W are the height and width of the feature tensor, respectively. n For the gradient operator in the corresponding direction, F n,c,i,j Let E be the eigenvalue of the nth feature tensor in the cth channel, with i,j being the spatial location; and E is obtained by applying the Softmax function. n Normalization, generating dynamic weights W n The enhanced feature F is obtained by weighted concatenation of features from the four directions. out The splicing operation is represented as follows: Where Concat is the channel concatenation operation; dynamic weights Step 2.2 Designing the CGA-SWS module before applying it to the detection head: The CGA-SWS module consists of two parts: CGA fusion and SWS enhancement. It obtains processed local feature sub-blocks by fusing three branches of CGA: channel attention, spatial attention, and pixel attention, and then combining SWS enhancement with the block-based approach. ； The mid-level feature x output from layer C5 of the neck FPN network is concatenated with the high-level semantic guidance feature y of the detection head after 1×1 convolution dimensionality reduction to obtain the initial feature. The initial feature is then input into the channel attention branch and the spatial attention branch in parallel. The outputs of the two branches are fused to obtain the coarse-grained weights patchn1 that focus on the global feature distribution. Pattn1 is concatenated with the pixel-level detail weights output from the pixel attention branch and activated by Sigmoid to obtain the fine-grained weights patchn2. The initial feature initia and [x;y] are then element-wise weighted using patchn2 to obtain the fused feature F. fusion , fusion feature F fusion The calculation formula is as follows: in This is element-wise multiplication. This is a feature fusion operation. The fused features F... fusion Divide the height H and width W into two equal parts each, resulting in four local sub-blocks B. k For k∈{1,2,3,4}, the average pixel value for each sub-block With variance Calculate pixel saliency score The calculation formula is as follows: Will Enhanced weights are obtained through Sigmoid activation. , for B k Element-level weighting is performed, and then the four enhanced sub-blocks are spliced together according to their original spatial positions to obtain the local feature sub-blocks after block enhancement processing. The calculation formula is as follows: Where i is the traversal index of the pixels within the sub-block; Step 2.3 introduces the SD dynamic loss mechanism and designs SDB Loss: establishes a scale mapping relationship, with the original input image size being W. o ×H o The current feature map size is W c ×H c Define the scaling ratio R between the feature map and the original image. oc The calculation formula is as follows: Based on the area B of the target's true bounding box in the original image coordinate system gt Combined with the maximum defined size B of the small target gt-max Calculate the dynamic weighting coefficient β using the balance coefficient δ. B The calculation formula is as follows: The min function is used to limit β. B The weight should not exceed δ to avoid weight imbalance. Based on β B Obtain the scale loss weight β LBS Location loss weight β LBL Specifically, it is expressed as follows: The final SDB Loss representation is as follows: Where L BS For scale loss, L BL This is due to position loss.
4. The method for detecting small road targets in aerial photography that integrates directional feature enhancement and scale adaptation as described in claim 1, characterized in that, In step 4, the specific method is as follows: training is performed under the Linux 5.15 operating system. In terms of hardware configuration, an Intel(R) Xeon(R) Platinum 8362 CPU @ 2.80GHz processor, 45GB of memory, and an RTX4090 (24GB) GPU are used. In terms of software configuration, PyTorch version 2.5.1 and Python version 3.12 are used, and CUDA version 12.4 is used. The number of iterations is set to 300, the batch size is set to 16, and both the initial learning rate and the final learning rate are set to 0.
01. The SGD optimizer is used, and the loss curve of the improved YOLO11 model tends to converge within 300 iterations.
5. The method for detecting small road targets in aerial photography that integrates directional feature enhancement and scale adaptation as described in claim 1, characterized in that, In step 5, the specific method is as follows: the test set image data is input into the trained improved YOLO11 model; when loading the test set images, the image size specifications are kept consistent with those of the training set preprocessing; the model first extracts features from the input test set images through the improved backbone network, then enhances the target direction features and suppresses background noise through the C3K2-DAP module, and then optimizes the multi-scale feature fusion through the CGA-SWS module of the neck network; finally, the target category, confidence score, and bounding box coordinates are output by the detection head combined with SDB Loss, and the detection results with confidence scores higher than the preset threshold are selected to obtain the final road aerial small target detection output localization and classification information.