An unmanned aerial vehicle aerial small target detection method, device, equipment and storage medium

Abstract

The application discloses a kind of unmanned aerial vehicle aerial small target detection method, device, equipment and storage medium.The high and low information flow soft grouping is realized by space adaptive information flow modulation, and the fusion of differentiating double-flow processing and space channel adaptive gating is matched, the on-demand allocation of computing resources is realized, the target and background area of aerial image are effectively distinguished, the calculation waste of low-value background is reduced, and the target feature expression is strengthened.Lightweight parallel multi-scale feature extraction is used to eliminate redundancy by sharing the main body, channel segmentation and multi-scale deep convolution parallel calculation are used to solve the serial bottleneck, accelerate model convergence, improve multi-scale feature extraction efficiency and lightweight degree.High and low frequency feature pseudo-separation and global context adaptive fusion are realized by spatial frequency perception feature decomposition, small target edge, texture and other high-frequency details are accurately captured, the image low-frequency semantic structure is retained, background interference is effectively suppressed, and the frequency information demand of different aerial scenes is adapted.

Description

Technical Field

[0001] This invention relates to the field of target detection, and in particular to a method, apparatus, equipment, and storage medium for detecting small targets in drone aerial photography. Background Technology

[0002] Aerial image target detection is widely used in traffic monitoring, disaster assessment, agricultural inspection, and security inspection. This technology often needs to be deployed on resource-constrained embedded devices carried by drones, thus requiring strict real-time performance and accuracy of the detection algorithm.

[0003] Deep learning-based target detection methods provide technical support for target detection in aerial images, and are mainly divided into two categories: two-stage and single-stage detectors. Two-stage detectors require generating candidate regions before precise localization, which suffers from high computational complexity and slow inference speed. Single-stage detectors directly predict the target category and location on the feature map, resulting in a simpler process. Among them, the YOLO series models have become the mainstream choice in the field of drone aerial photography due to their balance between detection speed and accuracy.

[0004] However, aerial images are characterized by drastic changes in target scale, a high proportion of small targets, complex backgrounds, and uneven target density distribution, which means that mainstream single-stage detection methods still have many shortcomings. First, the allocation of computational resources lacks spatial sensitivity. Existing channel attention mechanisms uniformly scale all spatial locations within the same channel, failing to distinguish between target and background regions and resulting in wasted computational resources. Second, feature fusion uses a fixed weight strategy, which cannot be dynamically adjusted according to sample content and spatial location, making it difficult to fully utilize the complementarity of multi-stream features. Third, it is difficult to balance detection efficiency and accuracy; optimizing the detection effect of small targets while simultaneously achieving both performance remains a significant challenge. Summary of the Invention

[0005] This invention provides a method, apparatus, device, and storage medium for small target detection in UAV aerial photography. By designing three core modules—spatial adaptive information flow modulation, lightweight parallel multi-scale feature extraction, and spatial frequency-aware feature decomposition—and fusing them to construct an IFMPF-YOLO model, it solves the technical problems in existing UAV aerial target detection technologies, such as unreasonable allocation of computational resources, low efficiency of multi-scale processing, fixed feature fusion strategies, and difficulty in balancing lightweight design with detection accuracy.

[0006] According to one aspect of the present invention, a method for detecting small targets in aerial photography using a drone is provided, the method comprising: Aerial images captured by drones are acquired, and basic features of the backbone network of the drone aerial images are extracted by the spatial information flow modulation parallel frequency sensing model IFMPF-YOLO to generate input features. Based on the input features, spatial adaptive information flow modulation, lightweight parallel multi-scale feature extraction, and spatial frequency-aware feature decomposition are performed to generate output features; Small target detection is performed based on the output features, and the small target detection results are output.

[0007] Optionally, spatial adaptive information flow modulation is performed based on input features to generate output features, including: performing adaptive global average pooling on the input features to obtain global statistical descriptors for each channel; generating channel importance weights based on the global statistical descriptors through depthwise convolution and activation functions; calculating high-information-flow output features and low-information-flow output features based on the importance weights; concatenating the high-information-flow output features and low-information-flow output features along the channel dimension to obtain concatenated features; generating gating weights based on the concatenated features through unit convolution, batch normalization, and activation functions; and performing spatial channel adaptive weighted fusion of the high-information-flow output features and low-information-flow output features based on the gating weights to obtain output features.

[0008] Optionally, based on importance weights, high information flow output features and low information flow output features are calculated, including: decomposing the input features into high information flow and low information flow based on importance weights; performing multi-scale group convolution processing on the high information flow features to obtain high information flow output features, and performing lightweight linear transformation on the low information flow features to generate low information flow output features.

[0009] Optionally, lightweight parallel multi-scale feature extraction is performed based on the input features to generate output features, including: based on the input features, a shared backbone network is used to eliminate redundancy and obtain shared features; in the channel dimension, the shared features are divided into group features according to a specified ratio; each group feature is subjected to a specified scale of depth convolution in parallel to obtain features at each scale; and the features at each scale are concatenated and fused to obtain output features.

[0010] Optionally, spatial frequency-aware feature decomposition is performed based on the input features to generate output features, including: extracting low-frequency smooth features based on the input features to obtain low-frequency features; approximating high-frequency detail features based on the input features to obtain high-frequency features; extracting channel-level statistical information through global average pooling, processing the statistical information through a two-layer convolutional network to generate adaptive fusion weights; performing weighted fusion of high-frequency and low-frequency features based on the adaptive fusion weights to obtain fused features; performing unit convolution cross-channel information interaction on the fused features, and introducing residual connections for feature enhancement to obtain output features.

[0011] Optionally, low-frequency smoothing feature extraction is performed based on the input features to obtain low-frequency features, including: selecting a depthwise separable convolution with a pre-defined kernel size as a low-pass filter operator; assigning independent learnable convolution weights to each channel of the input features based on the low-pass filter operator; and performing element-wise weighted summation on each channel in the spatial neighborhood of the pre-defined kernel size based on the learnable convolution weights to obtain low-frequency features.

[0012] Optionally, high-frequency features are obtained by approximating high-frequency detail features based on input features, including: selecting an average pooling filter of a preset size as a low-pass filter, performing spatial pooling operation on the input features through the low-pass filter to obtain low-pass approximate features of the input features; and subtracting the low-pass approximate features from the input features to obtain high-frequency features.

[0013] According to another aspect of the present invention, a small target detection device for drone aerial photography is provided, the device comprising: The input feature generation module is used to acquire drone aerial images and extract basic features of the drone aerial images through the spatial information flow modulation parallel frequency sensing model IFMPF-YOLO to generate input features; The input feature optimization module is used to perform spatial adaptive information flow modulation, lightweight parallel multi-scale feature extraction, and spatial frequency-aware feature decomposition based on input features to generate output features. The target detection module is used to perform target detection based on output features and output the target detection results.

[0014] According to another aspect of the present invention, an electronic device is provided, the electronic device comprising: At least one processor; and a memory communicatively connected to the at least one processor; The memory stores a computer program that can be executed by the at least one processor, which is then executed by the at least one processor to enable the at least one processor to perform a method for detecting small targets in aerial photography by a drone, as described in any embodiment of the present invention.

[0015] According to another aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions, the computer instructions being configured to cause a processor to execute and implement a method for detecting small targets in aerial photography by a drone as described in any embodiment of the present invention.

[0016] The technical solution of this invention achieves soft grouping of high and low information streams through spatial adaptive information stream modulation, combined with differentiated dual-stream processing and spatial channel adaptive gating fusion, enabling on-demand allocation of computing resources, effectively distinguishing target and background regions in aerial images, reducing computational waste from low-value backgrounds, and enhancing target feature representation. Lightweight parallel multi-scale feature extraction eliminates redundancy by sharing a backbone, and parallel computation of channel segmentation and multi-scale deep convolution solves the serial bottleneck, accelerates model convergence, and improves the efficiency and lightweight nature of multi-scale feature extraction. Spatial frequency-aware feature decomposition achieves pseudo-separation of high and low frequency features and adaptive fusion of global context, accurately capturing high-frequency details such as small target edges and textures, preserving the low-frequency semantic structure of the image, effectively suppressing background interference, and adapting to the frequency information requirements of different aerial photography scenarios.

[0017] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of a method for detecting small targets by drone aerial photography according to Embodiment 1 of the present invention; Figure 2 This is a spatial adaptive information flow modulation method provided in Embodiment 2 of the present invention; Figure 3 This is a lightweight parallel multi-scale feature extraction method provided in Embodiment 3 of the present invention; Figure 4 This is a spatial frequency sensing feature decomposition method provided in Embodiment 4 of the present invention; Figure 5 This is a schematic diagram of the structure of a small target detection device for drone aerial photography according to Embodiment 5 of the present invention; Figure 6 This is a schematic diagram of the structure of an electronic device that implements a method for detecting small targets in drone aerial photography according to an embodiment of the present invention. Detailed Implementation

[0020] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0021] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0022] Example 1 Figure 1 This is a flowchart of a method for detecting small targets in UAV aerial photography, provided in Embodiment 1 of the present invention. This embodiment is applicable to UAV aerial target detection scenarios. The method can be executed by a UAV small target detection device, which can be implemented in hardware and / or software and can be configured in a computer controller. Figure 1 As shown, the method includes: S110. Acquire aerial images taken by the UAV, and extract basic features of the UAV aerial images by using the spatial information flow modulation parallel frequency sensing model IFMPF-YOLO to generate input features.

[0023] Among them, the Information Flow Modulation and Parallel Frequency-aware YOLO (IFMPF-YOLO) is a lightweight model specifically designed for small target detection in UAV aerial photography. It is constructed by integrating three core innovative modules—Information Flow Modulation (IFM), Parallel Multi-scale Feature (PMF), and Spatial Frequency-aware Feature Decomposition (SFD)—with the YOLO detection framework. Adaptable to various deployment environments from embedded devices to servers, it addresses issues such as unreasonable allocation of computational resources, low efficiency in multi-scale processing, and fixed feature fusion in small target detection of aerial images while achieving model lightweighting, thus balancing detection accuracy and inference speed. The backbone network's basic feature extraction is the fundamental feature extraction step in the target detection algorithm. This involves the IFMPF-YOLO model's backbone network performing basic operations such as convolution, pooling, and activation on the input UAV aerial image, transforming the original image pixel information into input features with preliminary semantic information and spatial characteristics. Input features refer to the multidimensional feature map data generated after the aerial image is extracted using the IFMPF-YOLO backbone network. The input features are represented as follows: ,in, Indicates input features, The input feature is represented as a four-dimensional real tensor, where B is the batch size, C is the number of channels, and H and W are the height and width of the feature map.

[0024] S120: Based on input features, perform spatial adaptive information flow modulation, lightweight parallel multi-scale feature extraction, and spatial frequency-aware feature decomposition to generate output features.

[0025] Spatial adaptive information flow modulation addresses the problem of indiscriminate spatial location processing and wasted computational resources in existing channel attention mechanisms. It employs channel importance assessment and soft grouping, differentiated dual-stream processing, and spatial channel adaptive gating fusion to segment and refine the input features. This includes dividing features into high-information and low-information streams, applying high / low computational complexity processing strategies respectively, and then dynamically fusing them through spatial adaptive gating weights. Ultimately, this achieves on-demand allocation of computational resources, reduces invalid computation in background regions, and enhances the feature representation of target regions. Lightweight parallel multi-scale feature extraction addresses the serial bottleneck and parameter redundancy issues of traditional multi-scale feature extraction. It adopts a strategy of shared feature backbone, parallel grouped convolution, and single-stage fusion. This involves first extracting basic shared features through a shared backbone, then dividing the shared features into multiple groups along the channel dimension, and applying depthwise convolutions of different scales in parallel to extract multi-scale features. The data-independent parallel operation improves computational efficiency, resulting in shorter gradient paths and faster convergence. It also avoids redundant computation of independently extracting low-level features at different scales, achieving lightweight multi-scale feature extraction. Spatial frequency-aware feature decomposition addresses the problem of lost high-frequency details and entanglement of high- and low-frequency information in small targets in aerial images. Inspired by frequency decomposition theory in signal processing, it employs four steps: low-frequency smoothing feature extraction, high-frequency detail feature approximation, global context adaptive fusion weight generation, frequency-aware feature reconstruction, and residual enhancement. This involves pseudo-frequency separation and adaptive fusion of features, including separating low-frequency and high-frequency features from the image. Based on the frequency requirements of different aerial shooting scenarios, adaptive weights are generated to fuse high and low-frequency features, enhancing the edge and texture details of small targets and improving the accuracy of small target detection. The output features refer to the highly discriminative, multi-scale, and frequency-adaptive feature maps generated after the input features have undergone spatial adaptive information flow modulation, lightweight parallel multi-scale feature extraction, and spatial frequency-aware feature decomposition. These maps fully preserve the semantic, spatial, edge, and texture information of small targets in aerial images, effectively suppressing background interference, and serve as the core data foundation for subsequent small target detection.

[0026] S130. Perform small target detection based on output features and output the small target detection results.

[0027] Specifically, small target detection refers to the process where the IFMPF-YOLO model's detection head predicts target categories and regresses bounding boxes based on the processed output features. By calculating metrics such as intersection-union ratio, precision, and recall, it identifies small targets in aerial images, such as pedestrians and small vehicles, and determines their spatial locations. The final output includes the target category, location, and confidence score. The small target detection result is the final output of this method, containing relevant information about the small targets detected in the UAV aerial image, including: the target category, such as pedestrians, bicycles, and small cars; the target's spatial location in the image, bounding box coordinates; the confidence score of the target detection, such as the model's credibility with the detection results; and in some scenarios, it may also include detection accuracy metrics. The results can intuitively display the distribution and details of small targets in aerial images, meeting the needs of aerial detection scenarios such as traffic monitoring, disaster assessment, and security inspection.

[0028] The technical solution of this invention achieves soft grouping of high and low information streams through spatial adaptive information stream modulation, combined with differentiated dual-stream processing and spatial channel adaptive gating fusion, enabling on-demand allocation of computing resources, effectively distinguishing target and background regions in aerial images, reducing computational waste from low-value backgrounds, and enhancing target feature representation. Lightweight parallel multi-scale feature extraction eliminates redundancy by sharing a backbone, and parallel computation of channel segmentation and multi-scale deep convolution solves the serial bottleneck, accelerates model convergence, and improves the efficiency and lightweight nature of multi-scale feature extraction. Spatial frequency-aware feature decomposition achieves pseudo-separation of high and low frequency features and adaptive fusion of global context, accurately capturing high-frequency details such as small target edges and textures, preserving the low-frequency semantic structure of the image, effectively suppressing background interference, and adapting to the frequency information requirements of different aerial photography scenarios.

[0029] Example 2 Figure 2 This is a flowchart of a spatial adaptive information flow modulation method provided in Embodiment 2 of the present invention. This embodiment adds a specific process for generating output features based on input features during spatial adaptive information flow modulation, building upon Embodiment 1. For example... Figure 2 As shown, the method includes: S210. Perform adaptive global average pooling on the input features to obtain the global statistical descriptor of the channel.

[0030] The adaptive global average pooling operation calculates the average value of all pixel values ​​for each channel of the input feature in the spatial dimension, as expressed by the following formula (1): (1) in, Represents the global statistics descriptor. H and W represent the input features, and H and W represent the height and width of the feature map, respectively. This indicates adaptive global average pooling. Represents the pixel index along the height direction of the feature map, with values ​​from 1 to H. The pixel index represents the width direction of the feature map, with values ​​from 1 to W. The global statistical descriptor can globally represent the information density and feature representation capability of the corresponding channel, achieving effective compression of the dimension of the input feature space.

[0031] S220: Based on global statistical descriptors, importance weights of channels are generated through depthwise convolution and activation functions.

[0032] In this method, depthwise convolution only performs convolution calculations on each channel of the global statistical descriptor individually, without cross-channel information fusion, thus completing feature transformation while reducing the number of parameters and computational cost. The activation function is the Sigmoid function, which can map the output value of depthwise convolution to the interval [0,1]. The generated value can be directly used as the channel importance weight to quantify the importance of each channel. The closer the value is to 1, the higher the information value of the corresponding channel, and the closer it is to 0, the lower the information value. The importance weight is expressed by the following formula (2): (2) in, Indicates importance weight, This represents the Sigmoid function. express, This represents element-wise multiplication. This represents the global statistics descriptor.

[0033] S230. Based on importance weights, calculate the high information flow output features and low information flow output features.

[0034] Optionally, based on importance weights, high information flow output features and low information flow output features are calculated, including: decomposing the input features into high information flow and low information flow based on importance weights; performing multi-scale group convolution processing on the high information flow features to obtain high information flow output features, and performing lightweight linear transformation on the low information flow features to generate low information flow output features.

[0035] Specifically, the partitioning process adopts a soft grouping method with information conservation, which does not result in the loss of feature information. Specifically, it is implemented by element-wise multiplication, which multiplies the input features and channel importance weights element-wise to obtain a high information flow, as shown in the following formula (3): (3) in, Indicates high information flow. Indicates input features, This represents element-wise multiplication. The importance weight is indicated. The high information flow mainly contains information from the high-value channels in the input features, mainly the target area in the aerial image, which accounts for only 5-15% and is rich in discriminative features that can characterize the target; at the same time, the input features are multiplied element-wise by the result of 1 minus the channel importance weight to obtain the low information flow, as shown in the following formula (4): (4) in, Indicates low information flow. Indicates input features, This represents element-wise multiplication. The importance weight is indicated. The low information flow mainly contains information from the low-value channels of the input features, mainly the background region features of the aerial image, with the background region accounting for 85-95%. It is only necessary to maintain the information flow and no complex processing is required. Then, the system will perform multi-scale group convolution processing on the divided high information flow features to obtain the high information flow output features. It can be seen that the high information flow contains the discriminative features of the target, and fine feature extraction is required to enhance the target expression. The processing is carried out in two steps. The first step is to perform cross-channel information interaction on the high information flow features through 1×1 convolution to realize feature fusion and dimension adjustment between different channels, as shown in the following formula (5): (5) in, This represents the intermediate features after the high-information stream has undergone a 1×1 convolution. This represents a convolution operation with a kernel size of 1×1. This indicates high information flow. The second step, based on cross-channel information interaction, involves parallel application of 3×3 and 5×5 grouped convolutions for multi-scale feature extraction. The number of groups in the grouped convolution is set to C / 2, meaning each group has 2 input channels and generates 1 output channel, further refining the high information flow feature representation. The processing results of 3×3 grouped convolution and 5×5 grouped convolution are obtained respectively, as shown in the following formulas (6) and (7): (6) (7) in, This represents the features of the high-information stream after 3×3 grouped convolution. This represents the features of the high-information stream after 5×5 grouped convolution. This represents the intermediate features after the high-information stream has undergone a 1×1 convolution. This indicates a grouped convolution operation with a kernel size of 3×3. This indicates a grouped convolution operation with a kernel size of 5×5. This means the number of groups is half the number of channels, with each group having two input channels and generating one output channel. The results of these two types of multi-scale grouped convolutions are fused to form the final high-information-flow output feature. This fully explores the multi-scale feature information of the target area. At the same time, a lightweight linear transformation is performed on the low information flow features to generate low information flow output features. Since the low information flow is mainly composed of background features, no complex feature extraction is required. The information flow can be maintained simply by lightweight operation. Therefore, a lightweight linear transformation method combining 1×1 convolution and batch normalization is adopted, as shown in the following formula (8): (8) in, This indicates a low information flow output characteristic. Indicates batch normalization, This represents a convolution operation with a kernel size of 1×1. This indicates low information flow. 1×1 convolution achieves simple interaction of channel information while ensuring lightweight computation. Batch normalization standardizes the features after convolution, mapping feature values ​​to a distribution with a mean of 0 and a variance of 1, accelerating model convergence and avoiding feature distribution shift. The resulting low-information-flow output features retain background information while reducing computational overhead. Through differential processing, computational resources are directed towards high-value target region features, reducing invalid computation in background regions. This improves the targeting of feature extraction and achieves model lightweighting.

[0036] S240. The high information flow output features and low information flow output features are concatenated along the channel dimension to obtain the concatenated features.

[0037] Among them, channel dimension concatenation merges two features in dimension C without changing the batch, height and width dimensions of the features, and only superimposes the number of channels. If the number of channels of the high and low information flow output features is C, the number of channels after concatenation is 2C, as shown in the following formula (9): (9) in, Indicates splicing characteristics, This indicates a high information flow output characteristic. This indicates a low information flow output characteristic. This indicates a splicing operation. This indicates concatenation along the channel dimension. The concatenated features integrate high-information-flow target discriminative features and low-information-flow background features, providing more comprehensive feature information for the subsequent generation of gating weights, making the gating weights more closely match the actual feature distribution of the samples.

[0038] S250. Based on the splicing features, gating weights are generated through unit convolution, batch normalization and activation functions. Based on the gating weights, spatial channel adaptive weighted fusion is performed on the high information flow output features and low information flow output features to obtain the output features.

[0039] Among them, unit convolution is 1×1 convolution. By performing channel-dimensional feature transformation and dimensionality reduction on the concatenated features, the 2C-channel concatenated features are restored to C channels, keeping the spatial resolution of the features unchanged. Batch normalization is used to standardize the features after unit convolution, mapping the feature values ​​to a distribution with a mean of 0 and a variance of 1, which speeds up the convergence of the model and alleviates the gradient vanishing problem. Sigmoid is still used as the activation function, mapping the processed feature values ​​to the [0,1] interval, as shown in the following formula (10): (10) in, Indicates the gating weight, This represents the Sigmoid function. Indicates splicing characteristics, Indicates batch normalization, This indicates a convolution operation with a kernel size of 1×1. The gate weights have spatial resolution, with each spatial location of the feature map having an independent weight value. After generating the gate weights, spatial channel adaptive weighted fusion is performed using the following formula (11): (11) in, Indicates output features, Indicates the gating weight, This represents element-wise multiplication. This indicates a high information flow output characteristic. This represents low-information-flow output features. Through spatial channel adaptive weighted fusion, the model can assign higher weights to high-information-flow output features in the target region and higher weights to low-information-flow output features in the background region. This achieves spatial adaptive allocation of computational resources, reduces invalid computation in the background region, strengthens the feature representation of the target region, and ultimately yields output features rich in target discriminative information, providing a high-quality feature foundation for subsequent small target detection.

[0040] The technical solution of this invention addresses the problem of spatial allocation insensitivity in traditional channel attention mechanisms by employing spatial adaptive information flow modulation. Through soft grouping based on channel importance assessment and information conservation, features are accurately divided into high-information streams for the target region and low-information streams for the background. Multi-scale grouped convolutions are used to refine the high-information streams, enhancing discriminative features, while lightweight linear transformations are applied to the low-information streams to reduce unnecessary computation, achieving on-demand allocation of computing resources. By designing a spatial channel adaptive gating fusion strategy, high and low information streams are concatenated to generate spatially resolved gating weights. Position-wise adaptive weighted fusion of the output features from both high and low information streams accurately distinguishes between target and background regions, reduces computational resource waste from low-value backgrounds, strengthens the feature representation of the target region, and makes feature fusion more closely reflect the actual distribution and spatial location differences of the samples, thereby improving feature discriminability and utilization.

[0041] Example 3 Figure 3 This is a flowchart of a lightweight parallel multi-scale feature extraction method provided in Embodiment 3 of the present invention. This embodiment adds a specific process for generating output features based on the input features during lightweight parallel multi-scale feature extraction, building upon Embodiment 1. For example... Figure 3 As shown, the method includes: S310. Based on the input features, a shared backbone network is used to eliminate redundancy and obtain shared features.

[0042] It should be noted that this application addresses the problem of parameter redundancy and computational repetition caused by the independent extraction of low-level features at different scales in traditional multi-scale feature extraction. It constructs a lightweight shared backbone network to perform unified low-level feature extraction on the input features. Specifically, this is accomplished by sequentially performing 3×3 depthwise separable convolution, batch normalization, and SiLU activation function operations on the input features, as shown in the following formula (12): (12) in, Indicates shared feature backbone, This represents the SiLU activation function. express, Indicates batch normalization, This represents the input features. Depthwise separable convolution significantly reduces the number of parameters and computational cost compared to ordinary convolution. Batch normalization standardizes the feature distribution and accelerates model convergence. The SiLU activation function introduces non-linearity into the features to enhance their expressive power, ultimately resulting in shared features. As a unified underlying feature for feature extraction at all scales, it completely eliminates the redundancy problem of repeatedly extracting underlying features at different scales, laying the foundation for subsequent multi-scale feature extraction.

[0043] S320. In the channel dimension, the shared features are divided into group features according to a specified ratio.

[0044] S330. Perform depthwise convolutions of a specified scale on each group of features in parallel to obtain features at each scale.

[0045] Specifically, in the channel dimension, the shared features are divided into group features according to a specified ratio. In order to achieve parallel multi-scale feature extraction, the shared features are precisely segmented in the channel dimension. The specific segmentation ratio is 1 / 4, 1 / 4 and 1 / 2 of the total number of channels C, as shown in the following formulas (13), (14) and (15): (13) (14) (15) in, This indicates that the shared features are in the first group of the channel dimension, accounting for 1 / 4 of the total number of channels. This indicates that the shared features are in the second group of the channel dimension, accounting for 1 / 4 of the total number of channels. This indicates that the shared features are in the third group of the channel dimension, accounting for 1 / 2 of the total number of channels. This indicates a depthwise separable convolution with a kernel size of 7×7. This represents the multi-scale features extracted at a 7×7 scale. This indicates a depthwise separable convolution with a kernel size of 5×5. This represents the multi-scale features extracted at a 5×5 scale. This indicates a depthwise separable convolution with a kernel size of 3×3. This represents multi-scale features extracted at a 3×3 scale, i.e., shared features. Divide into the first group of features, corresponding to the subsequent 7×7 scale feature extraction, and... Divide into the second group of features, corresponding to the subsequent 5×5 scale feature extraction, and... The features are divided into a third group, corresponding to the subsequent 3×3 scale feature extraction. The three groups of depthwise convolution operations have no data dependency, so they can be executed in parallel on the GPU. Moreover, since the three are independent, this module has a shorter gradient path, more balanced magnitude, and faster convergence speed compared to the serial feature fusion module.

[0046] Furthermore, the three groups of features are fed into depthwise separable convolutions of corresponding scales for feature extraction. The three operations have no data dependency and can be executed in parallel on the GPU. The first group of features is processed by a 7×7 depthwise separable convolution to extract large-scale features, resulting in 7×7 scale features. The second set of features is processed by a 5×5 depthwise separable convolution to extract mid-scale features, resulting in 5×5 scale features. The third set of features is processed by a 3×3 depthwise separable convolution to extract small-scale features, resulting in 3×3 scale features. Different scales of depth convolution can capture feature information of targets at different scales in aerial images, adapting to the characteristics of drastic changes in target scale in aerial images. At the same time, the lightweight nature of depth separable convolution and the parallel execution method improve the efficiency of multi-scale feature extraction. Compared with the traditional serial architecture, parallel processing makes the gradient path shorter, the magnitude more balanced, and the model converges faster.

[0047] S340. The features at each scale are spliced ​​and fused to obtain the output features.

[0048] Specifically, the features at each scale are finally spliced ​​and fused to obtain the output features. The 7×7, 5×5, and 3×3 scale features extracted in parallel are spliced ​​along the channel dimension to integrate the target feature information captured at different scales. The fused features contain both the global semantic information captured by large-scale convolution and the local detail information captured by medium and small-scale convolution, which can fully adapt to the detection needs of small targets at different scales in aerial images. The features obtained after splicing and fusion are the output features of the lightweight parallel multi-scale feature extraction module. This feature has both multi-scale expressive power and lightweight characteristics, effectively solving the problems of serial bottleneck and parameter redundancy in traditional multi-scale modules, and providing a high-quality multi-scale feature foundation for subsequent aerial small target detection.

[0049] The technical solution of this invention eliminates computational redundancy caused by independently extracting low-level features at different scales through lightweight parallel multi-scale feature extraction, reducing the number of model parameters and computational cost. Simultaneously, shared features are proportionally divided into three groups along the channel dimension, and different sizes of depthwise convolutions are applied in parallel to extract multi-scale features. These three operations are independent of data and can be executed synchronously on a GPU, overcoming the efficiency bottleneck of serial processing. Furthermore, it results in shorter gradient paths and more balanced magnitudes in the model, accelerating convergence.

[0050] Example 4 Figure 4 This is a flowchart of a spatial frequency sensing feature decomposition method provided in Embodiment 4 of the present invention. This embodiment adds a specific process for performing spatial frequency sensing feature decomposition based on input features to generate output features, based on Embodiment 1 described above. For example... Figure 4 As shown, the method includes: S410. Extract low-frequency smooth features based on input features to obtain low-frequency features.

[0051] Optionally, low-frequency smoothing feature extraction is performed based on the input features to obtain low-frequency features, including: selecting a depthwise separable convolution with a pre-defined kernel size as a low-pass filter operator; assigning independent learnable convolution weights to each channel of the input features based on the low-pass filter operator; and performing element-wise weighted summation on each channel in the spatial neighborhood of the pre-defined kernel size based on the learnable convolution weights to obtain low-frequency features.

[0052] Specifically, firstly, a depthwise separable convolution with a preset kernel size is selected as the low-pass filter operator. According to signal processing theory, the frequency response of the convolution kernel determines the filtering characteristics. The weighted average operation of the depthwise separable convolution in the spatial domain approximates a low-pass filter, and its frequency response is shown in the following formula (16): (16) in, This represents the frequency response of a depthwise separable convolution. Represents the frequency variable in the frequency domain. The standard deviation of the Gaussian kernel is represented by its frequency response, which is close to 1 in the low-frequency region and decays rapidly in the high-frequency region. It can effectively preserve the smooth structural information of the image and suppress high-frequency noise, which is perfectly suited to the needs of low-frequency smooth feature extraction of aerial images. Therefore, this convolution operation is selected as the low-pass filter operator, and the convolution kernel adopts a preset k×k size. The convolution kernel of this size can cover a reasonable spatial neighborhood of the feature map, ensuring the integrity of low-frequency feature extraction.

[0053] Then, based on the low-pass filter operator, independent learnable convolution weights are assigned to each channel of the input feature. The characteristic of depthwise separable convolution is that each channel of the feature is convolved separately without cross-channel information fusion. The weights are parameters that are continuously optimized during model training. They can adaptively adjust the weighting degree of each channel at different spatial locations according to the feature distribution of the aerial image and the requirements of the detection task, so that each channel can extract low-frequency features in a targeted manner, avoiding the indiscriminate processing of features of different channels by uniform weights, and improving the accuracy of low-frequency feature extraction. Finally, based on the learnable convolution weights, each channel is weighted element-wise and summed in the spatial neighborhood of the convolution kernel of the preset size to obtain the low-frequency features, as shown in the following formula (17): (17) in, Indicates low-frequency characteristics. This represents a depthwise separable convolution operation, approximating a low-pass filter. Indicates input features, This represents the neighborhood of a k×k convolution kernel. This represents the learnable weights of a k×k depth convolution. This represents element-wise multiplication. This represents the channel feature value at position (h+i, w+j) of the input feature. After performing an element-wise weighted summation operation within the spatial neighborhood of all channels of the input feature, the processing results of each channel are integrated to finally obtain the low-frequency feature. The low-frequency feature preserves the main structure and semantic information of the aerial image, effectively filters high-frequency noise in the image, provides a stable low-frequency semantic foundation for the subsequent adaptive fusion of high- and low-frequency features, and adapts to the needs of small target detection in aerial images, avoiding interference from high-frequency noise on the discriminative power of target features.

[0054] S420. Based on the input features, high-frequency detail features are approximated to obtain high-frequency features.

[0055] Optionally, high-frequency features are obtained by approximating high-frequency detail features based on input features, including: selecting an average pooling filter of a preset size as a low-pass filter, performing spatial pooling operation on the input features through the low-pass filter to obtain low-pass approximate features of the input features; and subtracting the low-pass approximate features from the input features to obtain high-frequency features.

[0056] Specifically, firstly, a preset-size average pooling operation is selected as the low-pass filter. This low-pass filter is used to perform spatial pooling on the input features, yielding a low-pass approximation of the input features. Based on the principles of frequency domain decomposition, a high-pass filter can be implemented by the difference between an all-pass filter and a low-pass filter. To balance computational efficiency and lightweight requirements, this implementation does not use a complex ideal low-pass filter, but instead selects a preset-size average pooling operation as an approximation of the low-pass filter. This operation can perform a weighted average processing on the spatial neighborhood of the input features, effectively filtering out high-frequency details and retaining smooth low-frequency information. The low-pass approximation features are calculated using the following formula (18): (18) in, This represents the low-pass approximation feature. This represents the channel feature value of the input feature at position (h+i, w+j). This indicates the neighborhood range of the pooling kernel. This indicates the kernel size for average pooling. This indicates the average pooling operation. This represents the input features.

[0057] Furthermore, in the spatial domain, the difference between the full-pass and low-pass filters is transformed into the residual connection form of the features. Therefore, the original input features are directly subtracted element-wise from the low-pass approximation features obtained above, as shown in the following formula (19): (19) in, Indicates high-frequency characteristics, Indicates input features, Representing the low-pass approximation features, the above subtraction operation is mathematically equivalent to the discrete approximation of the Laplacian operator. The frequency response of the Laplacian operator increases with the square of the frequency, naturally enhancing high-frequency components. In the spatial domain, this is reflected as "the center pixel minus the average of its neighborhood," effectively highlighting rapidly changing local details such as edges and textures in the features. The resulting high-frequency features focus on discriminative details such as edges and textures of small targets in aerial images, effectively removing interference from low-frequency background information. This provides a high-frequency information foundation rich in small target details for subsequent adaptive fusion of high and low-frequency features, improving the model's ability to identify and locate small targets in aerial images.

[0058] S430: Channel-level statistical information is extracted through global average pooling, and the statistical information is processed through a two-layer convolutional network to generate adaptive fusion weights.

[0059] Specifically, the system extracts channel-level statistical information through global average pooling, and then processes the statistical information through two convolutional networks to generate adaptive fusion weights. Different aerial photography scenarios have significantly different requirements for high and low frequency information. Dense small target scenarios require more high-frequency details, while large-scale target scenarios require more low-frequency semantics. Therefore, global average pooling is first performed on the input features to calculate channel-level global statistical information, as shown in the following formula (20): (20) in, This represents global statistical information. H and W represent the input features, and H and W represent the height and width of the feature map, respectively. This indicates adaptive global average pooling. The input feature represents the channel feature value at position (h, w), where h and w are the height and width indices of the feature map spatial location. Global statistics can characterize the frequency feature distribution of each channel. This statistical information is then fed into a two-layer convolutional network for processing, as shown in the following formula (21): in, ;(twenty one) in, Indicates adaptive fusion weights, This represents the Sigmoid function. ReLU activation function introduces non-linearity into the features. This represents the learnable parameters of the first convolutional layer. The channel compression ratio is 4. This represents the learnable parameters of the second convolutional layer. , It represents global statistical information. Based on the information bottleneck theory, the channel compression ratio is designed to be 4, which can achieve lightweighting while ensuring the coding frequency preference information.

[0060] S440. Based on adaptive fusion weights, high-frequency features and low-frequency features are weighted and fused to obtain fused features.

[0061] Specifically, based on the adaptive fusion weights, high-frequency features and low-frequency features are then weighted and fused to obtain fused features, and the adaptive fusion weights are then used to... As weights for high-frequency features, As the weights for low-frequency features, the weighted fusion is completed using the following formula (22): ;(twenty two) in, Indicates fusion features, Indicates adaptive fusion weights, Indicates high-frequency characteristics, This indicates low-frequency features. When detecting a scene of densely packed small targets with rich high-frequency textures... The fused features will approach 1, and will more closely resemble high-frequency features, capturing more detailed information about small targets; however, when detecting large-scale target scenes dominated by low-frequency smoothing... The fusion value will approach 0, and the fused features will be closer to the low-frequency features, retaining more semantic structural information of the image, thus achieving scene adaptive fusion of high and low frequency features.

[0062] S450. Perform unit convolution on the fused features to exchange information across channels, and introduce residual connections to enhance the features, thus obtaining the output features.

[0063] Specifically, the fused features are then subjected to unit convolution for cross-channel information interaction, and residual connections are introduced for feature enhancement to obtain the output features. First, a 1×1 unit convolution is used to process the fused features. The system processes features to achieve information interaction and fusion between different channels, further enriching the expressive power of features without changing their spatial resolution. Then, residual connections are introduced to combine the features processed by unit convolution with the original input features of the module, effectively alleviating the gradient vanishing problem during deep model training and preserving the original feature information. The features enhanced by unit convolution cross-channel interaction and residual connections are the output features of the spatial frequency-aware feature decomposition module. These features not only preserve the low-frequency semantic structure of aerial images but also accurately capture the high-frequency details of small targets. They can also adapt to different aerial detection scenarios, effectively improving the accuracy of subsequent small target detection.

[0064] The technical solution of this invention achieves low-frequency smooth feature extraction through spatial frequency-aware feature decomposition, preserving the main structure and semantic information of the image. It then uses average pooling to approximate low-pass filtering, accurately extracting high-frequency detail features such as small target edges and textures through residuals, compensating for the loss of high-frequency discriminative information during downsampling. Through a global context-adaptive fusion weight generation mechanism, the fusion ratio of high and low-frequency features can be dynamically adjusted according to the frequency information requirements of different aerial photography scenarios. In dense small target scenes, high-frequency details are emphasized, while in large-scale target scenes, low-frequency semantics are emphasized, effectively separating the entanglement between low-frequency background and high-frequency target features and reducing background interference. Finally, unit convolution enables cross-channel information interaction of fused features, and residual connections are introduced to alleviate gradient vanishing and preserve original features, enhancing the uniformity and discriminativeness of feature expression. This makes the extracted features more suitable for the small target detection requirements of aerial images, improving the model's accuracy in recognizing and locating small targets.

[0065] Example 5 Figure 5 This is a schematic diagram of a small target detection device for drone aerial photography provided in Embodiment 5 of the present invention. Figure 5 As shown, the device includes: an input feature generation module 510, which is used to acquire aerial images taken by UAVs and extract basic features of the UAV aerial images through the spatial information flow modulation parallel frequency sensing model IFMPF-YOLO to generate input features; The input feature optimization module 520 is used to perform spatial adaptive information flow modulation, lightweight parallel multi-scale feature extraction, and spatial frequency-aware feature decomposition based on input features to generate output features. The target detection module 530 is used to perform target detection based on output features and output the target detection results.

[0066] Optionally, the input feature optimization module 520 specifically includes: a spatial adaptive information flow modulation unit, used to perform adaptive global average pooling on the input features to obtain global statistical descriptors for the channels; based on the global statistical descriptors, generating channel importance weights through depthwise convolution and activation functions; based on the importance weights, calculating high information flow output features and low information flow output features; concatenating the high information flow output features and low information flow output features along the channel dimension to obtain concatenated features; based on the concatenated features, generating gating weights through unit convolution, batch normalization, and activation functions; and performing spatial channel adaptive weighted fusion of the high information flow output features and low information flow output features based on the gating weights to obtain output features.

[0067] Optionally, the spatial adaptive information flow modulation unit specifically includes: a high- and low-information-flow feature extraction subunit, used to: decompose the input features into high-information-flow and low-information-flow based on importance weights; perform multi-scale group convolution processing on the high-information-flow features to obtain high-information-flow output features, and perform lightweight linear transformation on the low-information-flow features to generate low-information-flow output features.

[0068] Optionally, the input feature optimization module 520 specifically includes: a lightweight parallel multi-scale feature extraction unit, used to: eliminate redundancy based on the input features using a shared backbone network to obtain shared features; divide the shared features into group features according to a specified ratio in the channel dimension; perform depth convolution of a specified scale on each group feature in parallel to obtain features at each scale; and concatenate and fuse the features at each scale to obtain output features.

[0069] Optionally, the input feature optimization module 520 specifically includes: a spatial frequency-aware feature decomposition unit, used for: extracting low-frequency smooth features based on input features to obtain low-frequency features; approximating high-frequency detail features based on input features to obtain high-frequency features; extracting channel-level statistical information through global average pooling, processing the statistical information through a two-layer convolutional network to generate adaptive fusion weights; performing weighted fusion of high-frequency and low-frequency features based on the adaptive fusion weights to obtain fused features; performing unit convolution cross-channel information interaction on the fused features, and introducing residual connections for feature enhancement to obtain output features.

[0070] Optionally, the spatial frequency-aware feature decomposition unit specifically includes: a low-frequency feature extraction subunit, used to: select a depthwise separable convolution with a pre-defined size convolution kernel as a low-pass filter operator; assign independent learnable convolution weights to each channel of the input feature based on the low-pass filter operator; and, based on the learnable convolution weights, perform element-wise weighted summation on each channel within the spatial neighborhood of the pre-defined size convolution kernel to obtain the low-frequency features.

[0071] Optionally, the spatial frequency-aware feature decomposition unit specifically includes: a high-frequency feature extraction subunit, used to: select an average pooling of a preset size as a low-pass filter, perform spatial pooling operation on the input features through the low-pass filter to obtain low-pass approximate features of the input features; and subtract the low-pass approximate features from the input features to obtain high-frequency features.

[0072] The technical solution of this invention achieves soft grouping of high and low information streams through spatial adaptive information stream modulation, combined with differentiated dual-stream processing and spatial channel adaptive gating fusion, enabling on-demand allocation of computing resources, effectively distinguishing target and background regions in aerial images, reducing computational waste from low-value backgrounds, and enhancing target feature representation. Lightweight parallel multi-scale feature extraction eliminates redundancy by sharing a backbone, and parallel computation of channel segmentation and multi-scale deep convolution solves the serial bottleneck, accelerates model convergence, and improves the efficiency and lightweight nature of multi-scale feature extraction. Spatial frequency-aware feature decomposition achieves pseudo-separation of high and low frequency features and adaptive fusion of global context, accurately capturing high-frequency details such as small target edges and textures, preserving the low-frequency semantic structure of the image, effectively suppressing background interference, and adapting to the frequency information requirements of different aerial photography scenarios.

[0073] The drone aerial photography small target detection device provided in this embodiment of the invention can execute the drone aerial photography small target detection method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0074] Example 6 Figure 6 A schematic diagram of an electronic device 10 that can be used to implement embodiments of the present invention is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices (e.g., helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0075] like Figure 6 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded from storage unit 18 into the RAM 13. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0076] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0077] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as a method for detecting small targets in drone aerial photography.

[0078] In some embodiments, a method for detecting small targets in drone aerial photography can be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program can be loaded and / or installed on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the drone aerial target detection method described above can be performed. Alternatively, in other embodiments, processor 11 can be configured to perform a drone aerial target detection method by any other suitable means (e.g., by means of firmware).

[0079] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0080] Computer programs used to implement the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0081] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0082] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0083] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0084] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system. It addresses the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0085] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0086] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for detecting small targets in drone aerial photography, characterized in that, include: Aerial images captured by a drone are acquired, and basic features of the backbone network are extracted from the drone aerial images using the spatial information flow modulation parallel frequency sensing model IFMPF-YOLO to generate input features. Based on the input features, spatial adaptive information flow modulation, lightweight parallel multi-scale feature extraction, and spatial frequency-aware feature decomposition are performed to generate output features; Small target detection is performed based on the output features, and the small target detection result is output.

2. The method according to claim 1, characterized in that, The step of performing spatial adaptive information flow modulation based on the input features to generate output features includes: Adaptive global average pooling is performed on the input features to obtain the global statistical descriptors of the channels; Based on the global statistical descriptor, channel importance weights are generated through depthwise convolution and activation functions; Based on the aforementioned importance weights, high information flow output features and low information flow output features are calculated; The high information flow output feature and the low information flow output feature are concatenated along the channel dimension to obtain the concatenated feature. Based on the splicing features, gate weights are generated through unit convolution, batch normalization, and activation functions. Based on the gate weights, spatial channel adaptive weighted fusion is performed on the high information flow output features and the low information flow output features to obtain the output features.

3. The method according to claim 2, characterized in that, The calculation of high information flow output features and low information flow output features based on the importance weights includes: Based on the importance weights, the input features are decomposed into high information streams and low information streams; The high information flow features are processed by multi-scale group convolution to obtain high information flow output features, and the low information flow features are processed by lightweight linear transformation to generate low information flow output features.

4. The method according to claim 1, characterized in that, The step of performing lightweight parallel multi-scale feature extraction based on the input features to generate output features includes: Based on the input features, a shared backbone network is used to eliminate redundancy and obtain shared features; At the channel level, the shared features are divided into grouped features according to a specified ratio; Perform depthwise convolutions of a specified scale on each group of features in parallel to obtain features at each scale; The features at each scale are spliced ​​and fused to obtain the output features.

5. The method according to claim 1, characterized in that, The step of performing spatial frequency-aware feature decomposition based on the input features to generate output features includes: Low-frequency smoothing features are extracted based on the input features to obtain low-frequency features; High-frequency features are obtained by approximating high-frequency detail features based on the input features. Channel-level statistical information is extracted by global average pooling, and the statistical information is processed by a two-layer convolutional network to generate adaptive fusion weights. Based on the adaptive fusion weights, the high-frequency features and the low-frequency features are weighted and fused to obtain fused features; The fused features are subjected to unit convolution for cross-channel information interaction, and residual connections are introduced for feature enhancement to obtain the output features.

6. The method according to claim 5, characterized in that, The step of extracting low-frequency smoothing features based on the input features to obtain low-frequency features includes: Selecting a depth-separable convolution kernel of a preset size as a low-pass filter operator; Based on the low-pass filter operator, independent learnable convolution weights are assigned to each channel of the input feature; Based on the learnable convolution weights, each channel is weighted element-wise and summed within the spatial neighborhood of the convolution kernel of a preset size to obtain low-frequency features.

7. The method according to claim 5, characterized in that, The process of approximating high-frequency detail features based on the input features to obtain high-frequency features includes: A preset-size average pooling is selected as a low-pass filter, and the input features are subjected to spatial pooling operation through the low-pass filter to obtain low-pass approximate features of the input features. Subtracting the low-pass approximation feature from the input feature yields the high-frequency feature.

8. A device for detecting small targets in drone aerial photography, characterized in that, include: The input feature generation module is used to acquire aerial images taken by UAVs and extract basic features of the UAV aerial images through the spatial information flow modulation parallel frequency sensing model IFMPF-YOLO to generate input features. The input feature optimization module is used to perform spatial adaptive information flow modulation, lightweight parallel multi-scale feature extraction, and spatial frequency-aware feature decomposition based on the input features to generate output features. The target detection module is used to perform target detection based on the output features and output the target detection result.

9. An electronic device, characterized in that, The electronic device includes: At least one processor; and a memory communicatively connected to the at least one processor; The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-7.

10. A computer storage medium, characterized in that, The computer storage medium stores computer instructions that are used to cause a processor to execute the method of any one of claims 1-7.

Images