A method, device and product for detecting small targets in drone aerial photography

Abstract

This invention discloses a method, apparatus, and product for detecting small targets in drone aerial photography, relating to the field of computer vision, to improve the detection performance and parameter efficiency of small targets in complex background scenes. In the visual feature extraction process, the aerial image is downsampled hierarchically using a spatial-to-channel rearrangement approach to obtain multi-layered downsampled feature maps. A multi-branch convolutional structure is used to enhance the receptive field of each downsampled feature map, resulting in multi-scale visual features. In the text semantic enhancement process, at least one cue vector most relevant to the aerial image is selected from a cue vector library. The cue vector is semantically enhanced to obtain enhanced semantic features, which are then fused cross-modally with the multi-scale visual features. A detection head is then used to identify the target based on the cross-modal fused features. This invention improves the accuracy, robustness, and generalization ability of small target detection in complex environments.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a method, apparatus and product for detecting small targets in drone aerial photography. Background Technology

[0002] Traditional UAV aerial small target detection methods primarily rely on computer vision and image processing technologies. They typically use RGB aerial images or video sequences as input, employing image enhancement, candidate region generation, and convolutional neural networks for target detection and classification. In practice, multi-scale feature pyramids (FPN), sliding windows, or anchor boxes are commonly used to locate potential target regions, combined with deep learning models such as YOLO and Faster R-CNN (Region-CNN) for detection. However, due to the high viewing angle and large imaging range of UAVs, targets occupy very few pixels in the image, often becoming lost in complex backgrounds. In complex environments such as rain, snow, nighttime, or backlighting, image contrast is significantly reduced, and noise and blurring are exacerbated, making traditional data-driven methods based on fixed features and supervised learning prone to missed detections and false alarms. Furthermore, existing methods often depend on large-scale, scene-consistent labeled data. When the detection scene changes (e.g., weather, lighting, or terrain), the model's generalization ability is insufficient, making it difficult to quickly adapt to new environments, becoming a key bottleneck in UAV aerial small target detection.

[0003] In recent years, with the development of multimodal perception and large-scale model technology, deep learning-based UAV aerial small target detection methods have continuously evolved in terms of structural design and training strategies. On the one hand, researchers have enhanced the model's understanding of global context and cross-scale features by introducing Transformer structures, attention mechanisms, and temporal modeling methods. On the other hand, some works have attempted to combine infrared, low-light imaging, or event camera data to compensate for the lack of information in single visible light images under harsh environments. Although the above methods have improved detection performance to some extent, most models still adopt an end-to-end, fixed-task detection paradigm, which cannot be flexibly adjusted according to different task objectives or complex scenes. Especially in small target detection, models often struggle to distinguish which targets to focus on and which interferences to ignore, and the utilization of prior knowledge mainly relies on implicit learning, lacking controllability and interpretability. When the scene background is complex or there are dense interference targets, the model's ability to focus on key information remains limited. Summary of the Invention

[0004] The purpose of this invention is to provide a target detection method based on receptive field optimization and semantic enhancement, which addresses all or part of the problems mentioned above, in order to improve the detection performance and parameter efficiency of small targets in complex background scenes captured by UAVs.

[0005] The technical solution adopted in this invention is as follows: In a first aspect, the present invention provides a method for detecting small targets in aerial photography using a drone, comprising: Visual feature extraction process: The aerial image is downsampled in layers using a spatial-to-channel rearrangement method to obtain multi-layer downsampled feature maps; a multi-branch convolutional structure is used to enhance the receptive field of each downsampled feature map to obtain multi-scale visual features; Text semantic enhancement process: Select at least one cue vector most relevant to the aerial image from the cue vector library; perform semantic enhancement on the cue vector to obtain enhanced semantic features; The multi-scale visual features and the enhanced semantic features are fused across modally to obtain cross-modal fused features; The detection head is used to identify the target based on the cross-modal fusion features.

[0006] Furthermore, the aerial images are downsampled using a spatial-to-channel rearrangement approach, including: In the shallow feature extraction layer, the feature maps of the aerial image are spatially rearranged according to row and column parity to obtain multiple sub-features; then, local structural information of each sub-feature is extracted by reparameterized bottleneck structure C3Rep and convolutional branches; the local structural information of each sub-feature is spliced ​​in the channel dimension and fused and compressed to the target number of channels by one-dimensional convolution. In the middle feature extraction layer, the two-by-two neighborhood of the feature map of the aerial image is encoded into the channel dimension using the standard space-to-channel rearrangement method; channel compression is performed by one-dimensional convolution; and feature enhancement processing is performed using C3Rep before outputting the target number of channels. In the deep feature extraction layer, the two-by-two neighborhood of the feature map of the aerial image is encoded into the channel dimension using a standard space-to-channel rearrangement method; channel compression is performed through one-dimensional convolution; and feature enhancement is performed using the feature fusion structure C2f, and then the result is output with the target number of channels.

[0007] Furthermore, the multi-branch convolutional structure includes a 3x3 convolutional branch, a 1x1 convolutional branch, and an identity mapping branch; Methods for enhancing the receptive field of downsampled feature maps using multi-branch convolutional structures include: During the training phase, features are extracted from the downsampled feature map using three branches, and then added element by element. During the inference phase, the convolution kernels and biases of the three branches are fused, and the fused convolution kernels and biases are used to extract features from the downsampled feature map.

[0008] Furthermore, at least one cue vector most relevant to the aerial image is selected from the cue vector library, including: The aerial images are mapped to a semantic space to obtain semantic description vectors; The semantic description vector is compared with the prompt vectors in the prompt vector library to measure their relevance, and at least one prompt vector with the highest relevance to the semantic description vector is selected.

[0009] Furthermore, mapping the aerial images to a semantic space includes: Obtain the multi-scale visual features of the aerial image.

[0010] Furthermore, semantic enhancement is performed on the cue vector, including: The cue vector is fused with the multi-scale visual features to obtain a first fused feature; the key semantic channels of the first fused feature are recalibrated using a channel attention mechanism. Perform multi-scale pooling on the recalibrated first fusion feature; Parallel semantic modeling of pooling features is performed using a local semantic enhancement path and a global semantic increment path. The local semantic enhancement path is used to extract edge texture features, while the global semantic increment path is used to extract global features. The edge texture features and the global features are adaptively fused using a channel-space joint attention mechanism.

[0011] Furthermore, the cross-modal fusion of the multi-scale visual features and the enhanced semantic features includes: Align the multi-scale visual features with the enhanced semantic features at the same scale; Visual features and enhanced semantic features of the same scale are fused separately.

[0012] Furthermore, methods for fusing visual features and enhanced semantic features at the same scale include: Visual features and enhanced semantic features of the same scale are added element-wise, adaptively fused based on gating weights, or channel-wise concatenated and processed by convolution.

[0013] In a second aspect, the present invention also provides a small target detection device for drone aerial photography, which includes a processor and a storage medium, wherein the storage medium stores computer instructions, and the processor executes the computer instructions to perform the above-described small target detection method for drone aerial photography.

[0014] In a third aspect, the present invention also provides a computer program product, including a computer program that, when run by a processor, executes the above-described method for detecting small targets in UAV aerial photography.

[0015] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: This invention employs a spatial-to-channel rearrangement downsampling method, explicitly mapping local pixel structures to the channel dimension. While reducing resolution, it preserves texture and edge information of small targets, significantly alleviating the problem of low pixel ratios for distant targets in aerial photography leading to feature overload. This reduces false negatives and improves localization stability from the source. A hierarchical visual feature extraction strategy makes algorithm deployment more convenient, enabling plug-and-play functionality. Through reparameterized receptive field enhancement, a larger effective receptive field and more stable contextual modeling capability are achieved even in highly interfering backgrounds, thereby improving the separability of small targets from the background and reducing false positives. Compared to traditional deep learning detection methods, this invention introduces a text semantic enhancement process parallel to the visual feature extraction process. By dynamically matching the most relevant cue vector in the current image and performing semantic enhancement, the detection method can stably focus on key target semantics even in scenes with significantly changing conditions such as fog, rain, night / backlight, etc., improving cross-scene generalization ability. Furthermore, this invention adopts a multi-scale visual-semantic feature fusion strategy, balancing high-resolution details with global semantic consistency, reducing localization jitter and category confusion in dense small target scenes, resulting in more stable overall output. Compared to deep learning detection methods, this invention is more robust and has a wider range of applications for typical extreme conditions of UAVs, such as dense small targets, occlusion, low contrast, severe weather, and complex backgrounds. Attached Figure Description

[0016] The present invention will be described by way of example and with reference to the accompanying drawings.

[0017] Figure 1 This is a flowchart of a method for detecting small targets using drone aerial photography.

[0018] Figure 2 This is a schematic diagram of the implementation process of a drone aerial photography small target detection method in one embodiment.

[0019] Figure 3 This is a comparison image of the multi-scale visual features extracted by this application and the multi-scale visual features extracted by YOLOv8 under the same aerial image; where P3, P4, and P5 are the multi-scale visual features extracted by this application, and P3', P4', and P5' are the multi-scale visual features extracted by YOLOv8.

[0020] Figure 4 This is a network structure diagram of the detection module in one embodiment of the UAV aerial photography small target detection method of this application.

[0021] Figure 5 yes Figure 4 Network structure diagram of the PromptSelector module.

[0022] Figure 6 yes Figure 4 Network structure diagram of the TextGate module.

[0023] Figure 7 yes Figure 4 Network structure diagram of the ContextAware module.

[0024] Figure 8 yes Figure 4 Network structure diagram of the DualPath module.

[0025] Figure 9 This is a data flow diagram of one embodiment of the UAV aerial photography small target detection method.

[0026] Figure 10 This is a network structure diagram of the YOLOv8 object detection model. Detailed Implementation

[0027] All features disclosed in this specification, or steps in all methods or processes disclosed herein, may be combined in any way, except for mutually exclusive features and / or steps.

[0028] Any feature disclosed in this specification (including any appended claims and abstract) may be replaced by other equivalent or similar features, unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is merely one example of a series of equivalent or similar features.

[0029] In existing technologies for small target detection in UAV aerial photography, most are based on a single-stage target detection network. Feature downsampling is achieved through stride convolution or pooling operations. While reducing the resolution of the feature map, this can easily lead to the loss of texture and edge information of small targets, resulting in a high false negative rate for distant, low-contrast targets. Furthermore, traditional convolutional structures are limited by a fixed receptive field, making it difficult to balance local detail representation and global semantic modeling in complex backgrounds or scenes with significant scale changes. Deeper network layers or larger convolutional kernels are often needed to improve performance, introducing a large parameter scale and computational complexity. In addition, most existing technologies rely solely on visual features for target discrimination, lacking proactive utilization of scene semantic priors. In foggy, rainy, cluttered, or densely packed small target conditions, the model's ability to distinguish between targets and background is insufficient, limiting its robustness and scene adaptability.

[0030] In existing technologies, deep learning-based drone aerial target detection methods have extremely limited ability to focus on key information. Against this backdrop, prompt-guided drone aerial target detection methods have gradually emerged. The core idea is to guide the model to focus on task-related target features in specific scenarios by introducing explicitly controllable prompts. Known methods mainly include three categories: text-based prompts, visual prompts, and multimodal joint prompts. For example, natural language descriptions (such as "pedestrians on a road at night" or "low-altitude drones in rain or snow") are used as text prompts, inputting them along with visual features into the detection model to strengthen semantically relevant responses during feature extraction; or prior regions, key points, or coarse-grained masks are introduced into the image as visual prompts to guide the model to focus on potential small target regions. These methods exhibit better adaptability and robustness than simple deep learning-based detection methods in complex scenarios, and can quickly adjust the detection focus under different weather and lighting conditions. However, these methods still suffer from limitations such as simple fusion of prompts and visual features, insufficient guidance on fine-grained features of small targets, and limitations in real-time performance and computational efficiency. Therefore, researching a cue-guided detection method that can fully utilize cue information, enhance the feature representation capability of UAV aerial photography of small targets, and adapt to complex scenarios such as rain, snow, and darkness has important engineering application value and research significance for improving the performance of UAV aerial photography of small targets.

[0031] To address the shortcomings of existing technologies and known solutions, this application proposes a UAV aerial photography small target detection method based on receptive field optimization and semantic enhancement, aiming to improve the detection performance of small targets in complex scenes. This method preserves key texture and structural information of small targets during multi-scale feature extraction through spatial rearrangement detail-preserving downsampling, and enhances the effective receptive field and feature representation capability without increasing the inference computation path by utilizing reparameterized receptive field optimization. Simultaneously, a text semantic-driven feature enhancement mechanism is introduced, combining pre-constructed scene cue semantic priors to dynamically guide the feature enhancement process based on the overall semantics of the image. This improves detection robustness and generalization ability under complex weather conditions and high-density target scenes, and enhances overall detection performance and parameter efficiency while maintaining relatively controllable parameter scale.

[0032] The small target detection method for UAV aerial photography proposed in this invention achieves the identification of small targets in UAV aerial photography based on a processing flow of "detail-preserving downsampling—re-parameterized receptive field enhancement; cue vector selection—semantic enhancement; cross-modal multi-scale fusion—detection output". For example... Figure 1 , Figure 2 As shown, the UAV aerial photography small target detection method proposed in this invention includes the following stages: S1, Visual Feature Extraction Process.

[0033] In stage S1, aerial images are downsampled hierarchically using a spatial-to-channel rearrangement approach to obtain multi-layer downsampled feature maps. A multi-branch convolutional structure is then used to enhance the receptive field of each downsampled feature map to obtain multi-scale visual features.

[0034] As an optional implementation, in the downsampling step of stage S1, visual features at different resolutions are extracted in three layers. Three downsampling structures based on spatial-to-channel rearrangement (SPD_C3Rep, STD_C3Rep, and STD_Light) are used to map adjacent pixel blocks of the aerial image feature map (image pixels) to the channel dimension according to their spatial positions, thus reducing the spatial resolution while explicitly preserving the correspondence of local structures.

[0035] In the shallow feature extraction layer, the SPD_C3Rep downsampling structure is used to extract the third-layer downsampled feature map. The SPD_C3Rep downsampling structure first spatially rearranges the feature maps of the aerial image according to row and column parity, obtaining multiple sub-features. Then, it extracts local structural information for each sub-feature through the reparameterized bottleneck structure C3Rep (also known as RepC3 or C3 module) and convolutional branches. Finally, the local structural information of each sub-feature is concatenated along the channel dimension and fused and compressed to the target number of channels for output using one-dimensional convolution. This achieves integrated downsampling processing that preserves detail, extracts lightweight features, and integrates channels.

[0036] In the middle feature extraction layer, the STD_C3Rep downsampling structure is used to extract the fourth-layer downsampled feature map. The STD_C3Rep downsampling structure uses a standard spatial-to-channel rearrangement method to encode the 2x2 neighborhood of the feature map of the aerial image into the channel dimension. Channel compression is performed through one-dimensional convolution. After feature enhancement processing using C3Rep, the output is based on the target number of channels to balance detection performance and computational efficiency.

[0037] In the deep feature extraction layer, the STD_Light downsampling structure is used to extract the fifth-layer downsampled feature map. The feature extraction method of the STD_Light downsampling structure is similar to that of the STD_C3Rep downsampling structure, but to address stricter edge computing constraints, a lighter feature fusion structure, C2f, is used instead of C3Rep. This achieves semantic modeling processing after downsampling with lower computational cost, adapting to various deployment requirements. Specifically, the STD_Light downsampling structure uses a standard spatial-to-channel rearrangement method to encode the two-by-two neighborhood of the aerial image's feature map into the channel dimension. Channel compression is performed through one-dimensional convolution. After feature enhancement processing using C2f, the output is the target number of channels.

[0038] By using the downsampling structure design described above, compared with the traditional stride convolution method, the risk of losing texture information of small targets can be significantly reduced, and the recall rate and localization stability of dense small targets can be improved.

[0039] The core idea behind the downsampling structure design in stage S1 is to avoid directly weighting and fusing spatial neighborhoods. Instead, it explicitly maps local pixel structures to the channel dimension and preserves them through spatial rearrangement. The purpose of the SPD_C3Rep downsampling structure is not to reduce spatial size, but to more fully model local sub-pixel structures and fine-grained texture information while maintaining the feature map resolution. This downsampling structure explicitly separates features of different sub-pixel phases in the input feature map through spatial rearrangement and uses a multi-branch, multi-parameter residual structure to independently model and semantically enhance each sub-feature, thus avoiding premature mixing or weakening of different spatial phase information in the early stages. Compared to directly stacking convolutional layers, the SPD_C3Rep downsampling structure significantly improves the discriminative power and local structural sensitivity of features at high resolution, providing more complete and structurally clear input features for subsequent downsampling modules, effectively enhancing the network's ability to perceive small targets and fine-grained structures.

[0040] The specific configuration of the SPD_C3Rep downsampling structure is as follows: The aerial image feature map is denoted as... ;in, For batch size, For the number of channels, These represent the height and width of the feature map, respectively.

[0041] First, the input feature map is spatially rearranged according to its row and column parity indices. Assume the input feature map is divided into four sub-feature maps. (The same applies to the remaining quantities): ; in, This represents four sub-feature maps obtained by sampling from the original feature map at different row and column starting positions.

[0042] Subsequently, the four sub-feature maps are fed into four C3Rep reparameter residual modules with independent parameters and convolutional branches, respectively. The calculation process can be represented as follows: ; in: Indicates the first The C3Rep module of the path is used to extract the local structure and semantic features of the sub-region; Indicates the first The subsequent convolutional layers are preferably convolutions with a kernel size of 3×3. This is used to further perform local fusion and smoothing on the C3Rep module output; For the first The output characteristics of the path, that is, for the th path i Local structural information extracted from individual feature maps.

[0043] The four output features are concatenated along the channel dimension, and then channel compression and information integration are performed through a fusion convolutional layer to obtain the final output features of the SPD_C3Rep downsampling structure. : ; in This represents a concatenation operation on the channel dimension, resulting in a concatenated feature dimension of . ; To merge convolutional layers, a 1×1 convolution is preferred, used to reduce the number of channels from... Compress to target number of channels Output: .

[0044] The SPD_C3Rep downsampling structure has a high computational cost. In order to approximate the expressive power of SPD_C3Rep with a lower computational cost and achieve a balance between information preservation and computational power, this application proposes the STD_C3Rep downsampling structure for mid-level features and the STD_Light downsampling structure for high semantic layers that require stable aggregation and low latency.

[0045] The STD_C3Rep downsampling structure is configured as follows: After STD rearrangement (i.e., encoding the 2x2 neighborhood of the feature map of the aerial image to the channel dimension using the standard space-to-channel rearrangement method), a 1×1 convolution is first used ( Channel compression is performed to obtain intermediate features. : ; in, This represents the features after STD rearrangement. This convolution reduces the number of channels from... Compressed to intermediate channel number Its function is to reduce the amount of computation and improve the efficiency of subsequent feature extraction.

[0046] The compressed features are then fed into the C3Rep module to obtain intermediate features. : ; The C3Rep module learns local and cross-channel features through a multi-branch reparameterized structure.

[0047] Ultimately, features The target channel dimension is restored using a 1×1 fusion convolution. The output characteristics of the STD_C3Rep downsampling structure are obtained. : ; The output feature dimension is: .

[0048] The STD_Light downsampling structure is configured as follows: For lighter-weight scenarios, in this embodiment, the STD_Light downsampling structure uses the C2f lightweight bottleneck instead of C3Rep, and its calculation process is as follows: ; C2f is a cross-layer aggregation module, characterized by low computational cost and good gradient fluidity. The intermediate features processed by this module are represented as follows: ; This represents the output characteristics of the STD_Light downsampling structure.

[0049] The output dimension is also: .

[0050] The three downsampling structures designed by the above optional implementation methods can downsample the feature maps of aerial images to obtain three-layer downsampling feature maps. Following the principle of decreasing feature resolution, increasing semantic strength, and decreasing computational budget, the downsampling structure strength is progressively selected layer by layer. The shallow layer uses SPD_C3Rep to preserve local texture and small target information to the maximum extent; the middle layer uses STD_C3Rep, which replaces the multi-branch SPD with low-information-loss downsampling for spatial-to-depth conversion when completing key scale transformations (such as entering feature layers P3 / P4), achieving a balance between expressive power and computational cost through a single-path C3Rep that first compresses and then expresses; the deep layer uses STD_Light, which, at a stage where semantics are sufficiently abstracted, only retains the information rearrangement advantages of STD and uses lightweight C2f for necessary fusion, thereby significantly reducing latency and memory usage, leaving computational power for multi-scale fusion in the neck network and the detection head. This spatial arrangement, from strong expression to balanced expression to light expression, makes network downsampling smoother, gradients more stable, and more friendly to small targets and dense scenes, while also achieving a better speed-accuracy trade-off in engineering.

[0051] Receptive field enhancement was performed on the three downsampled feature maps using a multi-branch convolutional structure, resulting in visual features P3, P4, and P5 at three scales, as shown below. Figure 3 As shown.

[0052] Drone aerial footage suffers from strong background interference (road textures, building edges, and repetitive vegetation textures) and scale variations, making it difficult to achieve both a larger receptive field and stronger expressive power under lightweight constraints using only conventional convolution. As an alternative implementation, this application designs a multi-branch convolutional structure (referred to as the RR module) comprising a 3×3 convolutional branch, a 1×1 convolutional branch, and an identity mapping (Batch Normalization, BN) branch. During training, features are extracted from the downsampled feature map using the three branches, and then added element-wise to fuse local and cross-channel information at different scales, enhancing feature diversity and expanding the effective receptive field. During inference, the convolutional kernels and biases of the three branches are fused, and the fused kernels and biases are used to extract features from the downsampled feature map, maintaining a single-branch inference structure. This achieves strong feature expressive power during training, without introducing additional computational overhead during inference, and ensuring numerical consistency between the training and inference outputs. This structure is suitable for embedding into different depth levels of the backbone network, enabling the third-scale detail feature P3, the fourth-scale mesoscale structural feature P4, and the fifth-scale high semantic feature P5 to all acquire stronger ability to discriminate complex backgrounds.

[0053] The training data used in the training phase was the Visdrone dataset, which contains 10,209 still images taken by drones. This dataset is divided into three parts: a training set (6,471 images), a validation set (548 images), and a test set (1,610 images). The dataset is divided into 10 categories: pedestrians, people, bicycles, cars, vans, trucks, tricycles, awning tricycles, buses, and motorcycles. "Pedestrians" refers to standing pedestrians, and "people" refers to seated people.

[0054] The outputs of the three branches of the multi-branch convolutional structure are as follows: This represents a 3×3 convolution branch with learnable parameters.

[0055] This represents a 1×1 convolution branch with learnable parameters.

[0056] This represents the Batch Normalization identity mapping branch, which learns the mean and variance during the training phase.

[0057] The multi-branch convolutional structure output during the training phase Defined as: ; in For the activation function, ReLU or SiLU is preferably used in this embodiment.

[0058] The pixel-wise summation of the three branches enables multi-scale information fusion, thereby improving the coverage of the receptive field.

[0059] Structural reparameterization during the inference phase: To reduce inference cost, the two convolutional branches and the BN branch are merged into a single 3×3 convolutional kernel during the inference phase. Let the convolutional kernels and biases of the three branches be: The kernel and bias of the convolution branch; Convolutional kernels and biases in convolutional branches; The equivalent convolution kernel and bias of the identity mapping branch.

[0060] A 1×1 convolution kernel is expanded into a 3×3 convolution kernel by zero padding, denoted as: ; This indicates zero-value padding. The equivalent convolutional kernel after fusion during the inference phase. With bias for: ; .

[0061] After fusion, the multi-branch convolutional structure is completely and equivalently replaced by a single 3×3 convolution, thereby obtaining a larger effective receptive field, lower inference latency, and output that is completely consistent with the values ​​in the training phase.

[0062] Output of multi-branch convolutional structure during inference phase Defined as: .

[0063] Indicates that the convolution kernel is , bias is The 3×3 convolution is represented by Act, where Act represents the activation function.

[0064] Multi-scale output explanation: Multi-branch convolutional structures are embedded in feature extraction layers of different depths within the backbone network, ultimately yielding multi-scale visual features P3, P4, and P5. The backbone network uses multi-branch convolutional structures to extract multi-scale visual features. Since visual features are high-level features, they also contain semantic information.

[0065] In stage S1, while maintaining a minimal inference structure, the multi-branch convolutional structure significantly improves the effective receptive field, channel fusion capability, and representational stability of features through multi-branch reparameterization during training. This makes it particularly suitable as a receptive field compensation unit for high-resolution or multi-scale fusion nodes. This embodiment constructs a feature representation path that evolves gradually from high resolution to high semantics by hierarchically and collaboratively designing the reparameterized receptive field enhancement module of the multi-branch convolutional structure with downsampling structures such as SPD_C3Rep, STD_C3Rep, and STD_Light. Specifically, SPD_C3Rep maximizes the preservation of sub-pixel-level details and small target structural information through spatial-to-channel rearrangement in the shallow layer stage; STD_C3Rep completes key scale transformation with low information loss and balances expressive power and computational cost in the mid-layer stage; and STD_Light provides stable and efficient feature aggregation in the high semantic stage to reduce overall inference latency. Building upon this foundation, the multi-branch convolutional structure does not perform downsampling or information preservation. Instead, it is embedded in key feature nodes to perform low-cost receptive field compensation and structural reorganization on downsampled or cross-scale fused features. Through a reparameterization mechanism that uses multi-branch representation during training and is equivalent to a single convolution during inference, the contextual modeling ability and spatial consistency of features are enhanced without increasing inference complexity. This collaborative strategy achieves a feature evolution process that minimizes information loss, gradually abstracts the structure, and continuously enhances the receptive field. It significantly improves the network's ability to perceive and discriminate small targets, dense targets, and complex scenes, while also meeting the engineering requirements of speed and accuracy.

[0066] S2, Text Semantic Enhancement Process.

[0067] The S2 stage works in parallel with the S1 stage to improve the semantic reasoning ability and robustness of small object detection methods in challenging scenarios such as foggy weather, rainy weather, occlusion, dense small objects, and complex backgrounds.

[0068] In stage S2, at least one cue vector most relevant to the aerial image is selected from the cue vector library. The selected cue vector is then semantically augmented to obtain enhanced semantic features.

[0069] As an alternative implementation, a method for selecting at least one cue vector most relevant to the aerial image from a cue vector library includes: Aerial images are mapped to a semantic space to obtain semantic description vectors. The semantic description vectors are then compared with cue vectors (pre-encoded text description vectors, such as scene cue vectors, conditional cue vectors, etc.) in a cue vector library to measure their relevance, and at least one cue vector with the highest relevance to the semantic description vector is selected.

[0070] Mapping aerial images to a semantic space involves acquiring the semantic features corresponding to the aerial images. In one optional implementation, the mapping from aerial images to the semantic space can be achieved using image-level semantic description vectors extracted by the backbone network, i.e., the visual feature vectors containing semantic information described earlier. Then, by calculating cosine similarity, the extracted semantic description vectors are compared with prompt vectors in a prompt vector library to measure their relevance. One or more prompt vectors with the highest relevance to the current aerial image are selected, achieving dynamic prompt vector selection and avoiding the misleading and redundant nature of static prompts (such as fixed prompt words). Specifically, if the cosine similarity is 1 or close to 1 (e.g., reaching 95%), the prompt vector is selected; otherwise, multiple candidate prompt vectors with the highest cosine similarity (e.g., 5 or other numbers) are averaged to obtain the selected prompt vector. The selection of at least one prompt vector most relevant to the aerial image from the prompt vector library can be accomplished by the PromptSelector module.

[0071] In one optional implementation, the semantic enhancement processing of the cue vector includes the following steps: (1) The cue vector is fused with multi-scale visual features to obtain the first fused feature. The key semantic channels of the first fused feature are relabeled using a channel attention mechanism. This step can be completed by the TextGate module.

[0072] The selected cue vector is concatenated and fused with multi-scale visual features to obtain the first fused feature. Channel modulation weights are generated using a lightweight multilayer perceptron (MLP) (i.e., with a few hidden layers) to adaptively weight the key semantic channels, thus completing the recalibration of the key semantic channels.

[0073] As a gated channel subnetwork, the multilayer perceptron must satisfy "structural constraints + output constraints + lightweight constraints"; otherwise, it cannot guarantee "channel modulation" and stable training. The structural constraint requires that the input must contain visual features and cue vectors, with B×C representing the visual feature channel size and B×D representing the cue vector channel size. The output constraint strictly controls the number of output channels to B×C, and the output range is controllable, for example, by compressing the output using a sigmoid function. The controllable range; Lightweight constraint defines the bottleneck of MLP: (C+D)→C / 4→C.

[0074] For key semantic channels, the definition is: Channels with a weight close to 1 are those that are more important and relevant under the current semantic prompt. Channels with low weight (close to 0) are redundant or potentially interfering channels under the current semantic cues.

[0075] Key semantic channels refer to the channels with higher modulation weights (which can be defined by a threshold) in the output channels of the MLP gated network, which are strengthened under the current semantic prior. They are data-driven and dynamically changing.

[0076] Recalibration: The channel modulation weights learned are used to rescale the response of the feature channels, so that the network emphasizes relevant channels and suppresses irrelevant channels under the semantic cues of the current scene.

[0077] Given the current visual feature x_global and the currently selected cue vector text_feat, the MLP gated network outputs channel_gate. The recalibration is implemented by generating a channel_gate using sigmoid(MLP), and then multiplying it channel by channel with the feature map.

[0078] Given the feature map before enhancement: R represents the feature map. Indicates the length and width dimensions of a single channel; MLP gated network output channel modulation weights: g = sigmoid(MLP([x_global, text_feat])) ; The recalibration operation is: x'= x g;x' represents the visual features recalibrated via the key semantic channel.

[0079] (2) Perform multi-scale pooling on the recalibrated first fusion feature. Combine multi-scale context aggregation to supplement long-distance or large-scale scene information, so as to supplement the overall semantic expression of long-distance targets and complex environments. This step can be completed by the ContextAware module.

[0080] Multi-scale pooling operations (or multi-scale context aggregation) employ a parallel, multi-branch structure rather than a sequential pyramid stacking: context sampling is performed on the same input feature map at multiple scales. The specific pooling method is Adaptive AvgPool2d, which compresses the feature map to a fixed s×s (e.g., s=2, 4) to obtain long-distance / large-scale contextual information. After pooling, a lightweight transformation is performed using a 1×1 convolution, and the image is upsampled back to its original resolution. Finally, the feature map is concatenated with the original features and fused using another 1×1 convolution to obtain the output feature. This approach can introduce a large range of scene contextual information while maintaining low computational cost, improving the stability of feature discrimination in complex scenes.

[0081] (3) Parallel semantic modeling of pooling features is performed using local semantic enhancement path and global semantic increment path, where local semantic enhancement path is used to extract edge texture features and global semantic increment path is used to extract global features.

[0082] The local semantic enhancement path adopts a structure of "1×1 dimensionality reduction convolution → 3×3 depthwise separable convolution → 1×1 dimensionality increase convolution", where the 3×3 depthwise separable convolution is used to enhance the edges and texture details of small objects. The global semantic enhancement path adopts a structure of "1×1 dimensionality reduction convolution → 5×5 depthwise separable convolution → 1×1 dimensionality increase convolution", where the 5×5 depthwise separable convolution provides a larger receptive field to model large-scale scene consistency and contextual semantics. Both paths have C input and output channels, initially outputting B×C×H×W dimensions, which are then concatenated to a B×2C×H×W dimension for subsequent attention modulation and fusion. Both paths are implemented in a lightweight manner, with the core convolutions using depthwise + pointwise convolutions to reduce the number of parameters and computational cost.

[0083] A dual-path feature enhancement method is designed, constructing a local semantic enhancement path and a global semantic increment path that operate in parallel. The local semantic enhancement path is used to extract edge texture features, focusing on the edge texture and fine-grained structural features of small objects; the global semantic increment path is used to extract global features, focusing on scene consistency and contextual semantic modeling.

[0084] (4) Adaptive fusion of edge texture features and global features is performed through a channel-space joint attention mechanism, thereby maintaining the stability of the detected features under conditions of occlusion, low contrast, and dense distribution. This step is the same as step (3) and can be completed by the DualPath module.

[0085] This embodiment employs a channel-space joint attention mechanism to adaptively modulate dual-path features. First, in the channel attention branch, the outputs of the local semantic enhancement path and the global semantic increment path are concatenated in the channel dimension to obtain the features. The channel attention branch performs global average pooling on F to obtain... The dimensionality is reduced by 1×1 convolution (2C→C / r), ReLU is applied, and then increased by 1×1 convolution (C / r→2C), where r is an intermediate variable; the channel weights are output via Sigmoid. F is weighted channel by channel. Then the spatial attention branch is applied. The weighted feature maps are then summed by calculating the mean and maximum values ​​in the channel dimension and concatenating them. Spatial weights are obtained by k×k convolution (k=3 for lightweighting, otherwise k=5) and Sigmoid. The F function is weighted position by position. Finally, a 1×1 convolution is used to compress the number of channels from 2C to C, and a residual connection is formed with the input to the output. The channel attention branch and the spatial attention branch are concatenated in two stages (channel first, then spatial), and finally fused by a 1×1 convolution and residual connection is formed.

[0086] Specific implementation of input features: ; Dual-path output: The local semantic enhancement path outputs `local_feat`, and the global semantic enhancement path outputs `global_feat`. ; Concatenation: F = Cat(local_feat, global_feat), where Cat represents the concatenation operation; Then, channel attention → spatial attention → fusion processing is performed:

[0087] in These represent the outputs of the channel attention module and the spatial attention module, respectively. These represent the weighting processing of channel weights and the weighting processing of spatial weights, respectively. This represents element-wise multiplication; This indicates an enhanced semantic feature that has undergone semantic enhancement.

[0088] As an example, the structural design of the channel-space joint attention module is as follows: Channel attention module network structure: Input / output definitions: enter: ; Output weights: This indicates that the 2C channels are weighted one by one; Corresponding code: self.channel_attention = nn.Sequential( nn.AdaptiveAvgPool2d(1),# ->[B, 2C, 1, 1] nn.Conv2d(2C, C / / r, 1),# r=reduction=16 ->[B, C / r, 1, 1] nn.ReLU(inplace=True), nn.Conv2d(C / / r, 2C, 1),# ->[B, 2C, 1, 1] nn.Sigmoid() ) Spatial attention module network structure: Input / output definitions: Input: Channel weighted ; Output space weights This indicates a weighted average for each spatial location; First, perform channel statistics, then use convolution to generate a spatial mask: Spatial statistical characteristics: Avg = mean(F, dim=1, keepdim=True) →

[0089] Max = max(F, dim=1, keepdim=True) →

[0090] Concatenation: S = Cat(Avg, Max) →

[0091] Convolution generates spatial weights (k=3 for lightweight implementation, otherwise k=5): kernel_size = 3 if lightweight else 5 self.spatial_attention = nn.Sequential( nn.Conv2d(2, 1, kernel_size=kernel_size, padding=kernel_size / / 2), nn.Sigmoid() )get: W_s = spatial_attention(S)

[0092] effect: . Fusion and residual connections (completing "adaptive modulation after joint attention"): The attention-modulated features remain B×2C×H×W, and are finally compressed back into the C channel using a 1×1 convolution, with residual connections to the input. self.fusion_conv = Conv(2C, C, k=1) # 1×1 conv + BN + SiLU output =self.fusion_conv(F) + x ). Based on the three scales of visual features P3, P4, and P5 extracted from aerial images as examples above, in the text semantic enhancement process, a corresponding enhanced semantic feature, represented as T2, T3, and T4, will be generated for each scale of visual feature. The same principle applies to visual features at more scales.

[0093] like Figure 4 The diagram shows the network structure of the detection model designed with the YOLOv8 architecture. Figure 4 In the example, the network structures of the PromptSelector module, TextGate module, ContextAware module, and DualPath module are as follows: Figure 5 , Figure 6 , Figure 7 , Figure 8 As shown in the diagram, the PromptSelector module selects the most relevant task prompt vectors, the TextGate module performs feature map recalibration based on visual features and prompt vectors, the ContextAware module acquires multi-scale scene context, and the DualPath module performs local and global dual-path enhancement and adaptive modulation of visual features through channel and spatial joint attention. This endows the detection model with cross-modal semantic understanding capabilities, enabling it to maintain high robustness under conditions such as fog, rain, occlusion, and dense scenes.

[0094] For the text semantic enhancement branch, we have: The PromptSelector module selects a prompt vector: This module is used to select the text semantic vector that is most relevant to the current visual content from a pre-encoded library of text hints.

[0095] Let the visual features be: ; First, the global visual description vector is obtained through global average pooling: ; Project it onto the text alignment space: ; in, For visual projection matrix, This refers to the alignment dimension between text and visual features.

[0096] The pre-stored text prompt vector is represented as follows: ; Select the text hint vector that is most semantically relevant: ; in, This is the cosine similarity function.

[0097] The TextGuided module performs feature map recalibration: This module uses textual semantic cues to guide visual channel feature enhancement.

[0098] Perform global average pooling on visual feature x: ; Concatenate visual features and cue vectors: ; Channel weight generation: ; in, This represents the Sigmoid activation function. This represents the channel weight.

[0099] Finally, the channel recalibration was completed: ; in, Multiplication for channel broadcast.

[0100] The ContextAware module is responsible for multi-scale context aggregation: To improve the model's ability to understand large objects, distant targets, and complex backgrounds, a multi-scale pooling mechanism is introduced. Three pooling operations at different scales are defined: ; Piecing together features at different scales: ; Then, context-enhanced features are obtained through convolutional fusion: .

[0101] The DualPath module enhances dual-path functionality: This module includes local semantic enhancement paths and global semantic enhancement paths.

[0102] Local semantic enhancement path: ; Global semantic enhancement path: .

[0103] (1) Channel attention generation.

[0104] Global pooling is performed after concatenation: ; Generate channel weights: ; (2) Spatial attention generation.

[0105] Average pooling: ; Max pooling: ; Spatial weights: .

[0106] (3) Enhance output The final enhanced features are calculated as follows: x ; The corresponding output includes enhanced semantic features at three scales: .

[0107] S3, Multi-scale Feature Fusion Stage.

[0108] In the S3 stage, cross-modal fusion (referred to as FusionModule) of multi-scale visual features and enhanced semantic features is performed to obtain cross-modal fused features.

[0109] The enhanced semantic features correspond one-to-one with the multi-scale visual features. As an optional implementation method, the cross-modal fusion of multi-scale visual features and enhanced semantic features includes: The multi-scale visual features are aligned with the enhanced semantic features at the same scale. For example, P3 is aligned with T2, P4 with T3, and P5 with T4. The visual features and enhanced semantic features at the same scale are then fused. Cross-modal fusion of P3 and T2 yields cross-modal fused feature F3; cross-modal fusion of P4 and T3 yields cross-modal fused feature F4; and cross-modal fusion of P5 and T4 yields cross-modal fused feature F5, and so on. Figure 9 As shown.

[0110] For cross-modal fusion methods of visual features and enhanced semantic features, one optional implementation can be achieved by using any of the following methods: element-wise addition, adaptive fusion based on gated weights, or convolutional processing after channel concatenation. This ensures the effectiveness of semantic guidance while avoiding the negative impact of over-enhancement. This allows small target details (high resolution) to complement global semantics (contextual prior information), improving localization stability and discriminative power in complex backgrounds.

[0111] S4, Small Target Identification Stage.

[0112] In the S4 stage, the detection head is used to identify the target based on cross-modal fusion features.

[0113] Depending on the purpose of the detection method (e.g., classification), the cross-modal fusion features at each scale are input into the model's detection head (Detect) to perform small target detection for the corresponding purpose. For example, identifying the category of people / objects in aerial images. The output format of the detection head can include classification information, confidence scores, and bounding box regression parameters.

[0114] As an optional implementation, the detection head also processes the results identified by cross-modal fusion features using a non-maximum suppression strategy to output the final detection result.

[0115] Through the synergistic effect of the above stages, this application achieves robust detection in typical challenging drone aerial photography scenarios such as foggy weather, rainy weather, complex backgrounds, and dense small targets, without significantly increasing the number of model parameters or inference computational complexity. The backbone visual feature extraction process provides stable multi-scale visual representation, the text semantic enhancement process introduces controllable scene prior information, and the cross-modal fusion mechanism coordinates the complementary relationship between the two types of features. Thus, under edge device conditions, it simultaneously improves detection accuracy, positioning stability, and cross-scene adaptability, demonstrating good engineering application value.

[0116] Compared with deep learning detection methods, the detection method proposed in this application has the following significant advantages: Small targets have higher recall: This application replaces traditional strided convolution / pooling downsampling with spatial rearrangement detail-preserving downsampling (SPD_C3Rep of this application / STD_C3Rep of this application / STD_Light of this application), which explicitly maps the local pixel structure to the channel dimension. While reducing the resolution, it retains the texture and edge information of small targets, which significantly alleviates the problem that the pixel ratio of distant targets in aerial photography is so low that the features are overwhelmed. This reduces the false negative rate and improves the stability of localization from the source.

[0117] Enhanced discrimination capability in complex backgrounds: Through a multi-branch convolutional structure, the model employs 3×3, 1×1, and identity mapping to model local and cross-channel information at different scales in parallel during the training phase. During the deployment phase, it folds into a single 3×3 convolution to achieve equivalent inference. This enables the model to obtain a larger effective receptive field and a more stable context modeling capability in strongly interfering backgrounds such as road textures, building edges, and repetitive vegetation textures, thereby improving the separability of small targets from the background and reducing false detections.

[0118] More comprehensive use of semantic priors and better scene adaptability: Compared with traditional pure visual detection methods that rely solely on implicit learning, this application introduces a text semantic enhancement process. It uses a prompt selection processing unit to dynamically select the scene prompt words most relevant to the current image, and generates channel modulation weights through a text-guided enhancer. Combined with multi-scale context aggregation and local-global dual-path enhancement, it achieves controllable semantic prior-guided feature enhancement, enabling the model to stably focus on key target semantics in scenes with significant changes in conditions such as fog, rain, night / backlight, etc., and improves cross-scene generalization ability.

[0119] More effective multi-scale fusion and more stable detection output: This application performs cross-modal alignment fusion in the P3 / P4 / P5 multi-scale feature layers, and fuses the enhanced semantic features with the backbone multi-scale visual features through addition / gating / convolution, etc., taking into account high-resolution details and global semantic consistency, reducing localization jitter and category confusion in dense small target scenes, and making the overall output more stable.

[0120] Superior computational efficiency and deployment friendliness: The inference stage of the multi-branch convolutional structure can be completely folded into a single-branch convolutional structure, avoiding the introduction of additional inference overhead; at the same time, a layered strategy of shallow strong expression (SPD-C3Rep) - mid-layer balance (STD_C3Rep) - deep light expression (STD_Light) is adopted to control the parameter scale and latency while ensuring the preservation of key details, making the solution more adaptable to edge / airborne computing power constraints, and has the advantages of plug-and-play and engineering implementation.

[0121] More robust to extreme conditions and wider applicability: Relying on the collaborative mechanism of detail-preserving downsampling, receptive field compensation, textual semantic prior guidance, and cross-modal multi-scale fusion, this application has stronger robustness in typical and difficult UAV scenarios such as dense small targets, occlusion, low contrast, severe weather, and complex backgrounds. It can be applied to various aerial photography tasks such as UAV inspection, traffic monitoring, security search and rescue, and disaster emergency response, solving the bottlenecks of existing technologies that are prone to missed detections, false detections, and insufficient generalization in complex environments.

[0122] In summary, compared with deep learning detection methods, this application shows more significant advantages in terms of small target recall and localization stability, anti-interference in complex backgrounds, cross-scene adaptation and controllable semantic utilization, as well as inference efficiency and deployment friendliness. It can provide a more reliable engineering solution for accurate perception of small targets in complex scenes captured by UAVs.

[0123] Furthermore, when deploying the detection method of this application, modifications can be made to one-stage object detection networks such as YOLOv8 to reduce the workload of model construction. For example... Figure 10As shown, YOLOv8 models object detection as a single-stage end-to-end regression problem. The input image, after size alignment and normalization preprocessing, is fed into the backbone network. Multi-level convolution and downsampling are used to progressively extract multi-scale feature representations, from low-level geometric texture to high-level semantic information. The C2f structure improves feature reuse and enhances gradient propagation stability through splitting and cascading fusion mechanisms. Subsequently, the Neck part employs a top-down Feature Pyramid Network (FPN) and a bottom-up Path Aggregation Network (PAN) to bidirectionally fuse features at different scales, allowing high-level semantic information and low-level precise localization information to interact fully, thus constructing multi-scale fused features that accommodate both large and small targets. The detection stage uses a decoupled detection head structure, mapping the classification task and bounding box regression task to independent feature subspaces to reduce task conflicts and improve convergence efficiency. Simultaneously, an Anchor-Free prediction mechanism is introduced, using pixels on the feature map as regression reference centers to directly predict the relative distance from the target center to the boundary, avoiding dependence on predefined anchor boxes and enhancing the model's cross-scene generalization ability. To further improve positioning accuracy, YOLOv8 employs a Distributed Focal Loss strategy. This strategy models continuous boundary distances as discrete probability distributions and obtains the final regression value through expectation calculation, thus providing more stable and refined gradient constraints near the boundaries. During training, a positive sample allocation strategy with central region constraints establishes the correspondence between predicted points and real targets. CIoU loss, classification loss, and target confidence loss are used in conjunction to perform end-to-end optimization of the network. In the inference phase, the model generates candidate detection boxes from feature maps at various scales and calculates class confidence. Finally, non-maximum suppression removes redundant predictions, outputting the target's class label, spatial location, and corresponding confidence.

[0124] Based on the concept of this application, this application embodiment also provides a drone aerial photography small target detection device, which includes a processor and a storage medium. The storage medium stores computer instructions, and the processor runs the computer instructions to execute the drone aerial photography small target detection method of the above embodiments and optional implementations.

[0125] In addition, this application also provides a computer program product, including a computer program, which, when run by a processor, executes the UAV aerial photography small target detection method described in the above embodiments and optional implementations.

[0126] This invention is not limited to the specific embodiments described above. The invention extends to any new feature or combination disclosed in this specification, as well as any new method or process step or combination disclosed herein.

Claims

1. A method for detecting small targets in drone aerial photography, characterized in that, include: Visual feature extraction process: The aerial images are downsampled in layers using a spatial-to-channel rearrangement method to obtain multi-layer downsampled feature maps; A multi-branch convolutional structure is used to enhance the receptive field of each downsampled feature map to obtain multi-scale visual features. Text semantic enhancement process: Select at least one cue vector most relevant to the aerial image from the cue vector library; perform semantic enhancement on the cue vector to obtain enhanced semantic features; The multi-scale visual features and the enhanced semantic features are fused across modally to obtain cross-modal fused features; The detection head is used to identify the target based on the cross-modal fusion features.

2. The method for detecting small targets in UAV aerial photography as described in claim 1, characterized in that, The aerial images are downsampled using a spatial-to-channel rearrangement approach, including: In the shallow feature extraction layer, the feature maps of the aerial image are spatially rearranged according to row and column parity to obtain multiple sub-features; then, local structural information of each sub-feature is extracted by reparameterized bottleneck structure C3Rep and convolutional branches; the local structural information of each sub-feature is spliced ​​in the channel dimension and fused and compressed to the target number of channels by one-dimensional convolution. In the middle feature extraction layer, the two-by-two neighborhood of the feature map of the aerial image is encoded into the channel dimension using the standard space-to-channel rearrangement method; channel compression is performed by one-dimensional convolution; and feature enhancement processing is performed using C3Rep before outputting the target number of channels. In the deep feature extraction layer, the two-by-two neighborhood of the feature map of the aerial image is encoded into the channel dimension using a standard space-to-channel rearrangement method; channel compression is performed through one-dimensional convolution; and feature enhancement is performed using the feature fusion structure C2f, and then the result is output with the target number of channels.

3. The method for detecting small targets in UAV aerial photography as described in claim 1, characterized in that, The multi-branch convolutional structure includes a 3x3 convolutional branch, a 1x1 convolutional branch, and an identity mapping branch; Methods for enhancing the receptive field of downsampled feature maps using multi-branch convolutional structures include: During the training phase, features are extracted from the downsampled feature map using three branches, and then added element by element. During the inference phase, the convolution kernels and biases of the three branches are fused, and the fused convolution kernels and biases are used to extract features from the downsampled feature map.

4. The method for detecting small targets in UAV aerial photography as described in claim 1, characterized in that, Select at least one cue vector from the cue vector library that is most relevant to the aerial image, including: The aerial images are mapped to a semantic space to obtain semantic description vectors; The semantic description vector is compared with the prompt vectors in the prompt vector library to measure their relevance, and at least one prompt vector with the highest relevance to the semantic description vector is selected.

5. The method for detecting small targets in UAV aerial photography as described in claim 4, characterized in that, Mapping the aerial images to a semantic space includes: Obtain the multi-scale visual features of the aerial image.

6. The method for detecting small targets in UAV aerial photography as described in claim 1, characterized in that, Semantic enhancement of the cue vector includes: The cue vector is fused with the multi-scale visual features to obtain a first fused feature; the key semantic channels of the first fused feature are recalibrated using a channel attention mechanism. Perform multi-scale pooling on the recalibrated first fusion feature; Parallel semantic modeling of pooling features is performed using a local semantic enhancement path and a global semantic increment path. The local semantic enhancement path is used to extract edge texture features, while the global semantic increment path is used to extract global features. The edge texture features and the global features are adaptively fused using a channel-space joint attention mechanism.

7. The method for detecting small targets in UAV aerial photography as described in claim 1, characterized in that, Cross-modal fusion of the multi-scale visual features and the enhanced semantic features includes: Align the multi-scale visual features with the enhanced semantic features at the same scale; Visual features and enhanced semantic features of the same scale are fused separately.

8. The method for detecting small targets in UAV aerial photography as described in claim 7, characterized in that, Methods for fusing visual features and enhanced semantic features at the same scale include: Visual features and enhanced semantic features of the same scale are added element-wise, adaptively fused based on gating weights, or channel-wise concatenated and processed by convolution.

9. A device for detecting small targets in drone aerial photography, characterized in that, It includes a processor and a storage medium, the storage medium storing computer instructions, the processor executing the computer instructions to perform the UAV aerial photography small target detection method as described in any one of claims 1-8.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is run by the processor, it executes the UAV aerial photography small target detection method as described in any one of claims 1-8.

Images