Efficient detection method and system for unmanned aerial vehicle remote sensing target detection

Abstract

The application relates to the technical field of target detection, in particular to an efficient detection method and system for unmanned aerial vehicle remote sensing target detection. In application, the ESPMStem module, the CM-RepNCSPELAN4 module and the ECA-DyHead layer are cooperatively designed. The ESPMStem module guides the downsampling process through spatial importance modulation to retain shallow detail information related to the target. The CM-RepNCSPELAN4 module integrates global context modeling GC, efficient channel attention ECA and reparameterization convolution RepNCSP. The ECA-DyHead layer integrates efficient channel attention ECA and dynamic detection head DyHead to enhance the representation ability of multi-scale target features.

Description

Technical Field

[0001] This application relates to the field of target detection technology, specifically to an efficient detection method and system for UAV remote sensing target detection. Background Technology

[0002] With the rapid development of UAV technology, its applications in remote sensing, surveying and mapping, agricultural monitoring, and security inspection are becoming increasingly widespread. UAV remote sensing images are characterized by high resolution and wide field of view, but also by large differences in target scale and complex backgrounds, which places extremely high demands on the accuracy and real-time performance of target detection algorithms.

[0003] Traditional target detection methods, such as algorithms based on handcrafted features (e.g., HOG, SIFT) and classifiers (e.g., SVM), have limitations in feature representation capabilities and are difficult to effectively handle complex scenarios in UAV remote sensing images, such as varying target scales, uneven lighting, and dense target occlusion. As a result, the detection accuracy and robustness are insufficient to meet the needs of practical applications.

[0004] In recent years, object detection algorithms based on deep convolutional neural networks have made significant progress. Among them, the YOLO series of algorithms has been widely used in industry due to its good balance between detection speed and accuracy. However, directly applying the YOLO series of algorithms to UAV remote sensing scenarios still faces some challenges.

[0005] First, existing algorithms typically lack effective modeling of global contextual information. In UAV remote sensing images, there are complex semantic relationships between targets and the background. Traditional convolutional networks mainly rely on local receptive fields for feature extraction, making it difficult to fully capture long-range dependency information. This leads to false positives and false negatives in complex backgrounds or scenes with dense targets.

[0006] Secondly, existing methods still have shortcomings in preserving shallow features and fusing multi-scale features. Traditional networks typically reduce the resolution of feature maps rapidly through downsampling in the early stages of feature extraction, which can easily lead to the loss of detailed information such as target edges and textures, especially affecting the feature representation ability of small targets. Meanwhile, how to efficiently and adaptively fuse feature maps of different scales, so as to fully utilize the spatial details in shallow features and the semantic information in deep features, thereby balancing the detection performance of both large and small targets, remains a problem that current technology needs to solve.

[0007] In addition, considering that drone platforms are usually limited by power consumption and computing power, how to improve detection accuracy while maintaining the lightweight nature of the model and efficient inference speed is also a key bottleneck restricting the actual deployment of this technology. Summary of the Invention

[0008] In view of this, this application provides an efficient detection method and system for UAV remote sensing target detection, which can improve detection accuracy while maintaining the lightweight nature and high efficiency of the model.

[0009] In a first aspect, this application provides an efficient target detection method for UAV remote sensing, comprising: extracting a first feature map from the image to be processed, and obtaining a second feature map through spatial gating generation and feature response modulation; performing two pre-modulation operations on the second feature map to obtain a shallow feature map, wherein each pre-modulation operation includes one downsampling and one CM-RepNCSPELAN4 feature enhancement process; performing downsampling and CM-RepNCSPELAN4 feature enhancement on the shallow feature map to obtain a middle feature map; performing downsampling and CM-RepNCSPELAN4 feature enhancement on the middle feature map to obtain a deep feature map; performing SPPF and C2PSA on the deep feature map to obtain a reference deep map; and then performing upsampling on the reference deep map and concatenating it with the middle feature map. The following steps are performed: First, a reference mid-layer image is obtained by merging the reference mid-layer image with the shallow feature image. Then, CM-RepNCSPELAN4 feature enhancement processing is performed to obtain a shallow output image. Next, the shallow output image is downsampled and merged with the reference mid-layer image using CM-RepNCSPELAN4 feature enhancement processing to obtain a mid-layer output image. Finally, the mid-layer output image is downsampled and merged with the reference deep image using CM-RepNCSPELAN4 feature enhancement processing to obtain a deep output image. The shallow, mid, and deep output images are then imported into a dynamic detection head, where scale awareness, spatial awareness, and task awareness are performed to obtain the final decoded classification result, center point regression result, and bounding box regression result.

[0010] In conjunction with the first aspect, in one possible implementation, the step of obtaining a second feature map by spatial gating generation and feature response modulation after extracting the first feature map from the image to be processed includes: calling a first convolutional unit to perform shallow feature extraction on the image to be processed to obtain the first feature map; calling a second convolutional unit to perform cross-channel information aggregation on the first feature map and generate a single-channel spatial response map, which is then normalized by Sigmoid to obtain a spatial gating map; applying a fixed scaling factor 'a' to the spatial gating map for weighted adjustment, and multiplying the weighted spatial gating map element-wise with the first feature map to obtain an intermediate feature map; and using a third convolutional unit to downsample the spatially modulated intermediate feature map to obtain the second feature map.

[0011] In conjunction with the first aspect, in one possible implementation, the first convolutional unit is a 3x3 convolution with a stride of 1, the second convolutional unit is a 1x1 convolution, and the third convolutional unit is a 3x3 convolution with a stride of 2.

[0012] In conjunction with the first aspect, in one possible implementation, the step of importing the shallow output map, the middle output map, and the deep output map into the dynamic detection head, performing scale awareness, spatial awareness, and task awareness to obtain the final decoded classification result, center point regression result, and bounding box regression result includes: importing the shallow output map, the middle output map, and the deep output map into the scale awareness submodule of the dynamic detection head; associating the features of different scale feature maps with each other and then importing them into the spatial awareness submodule of the dynamic detection head; performing attention enhancement on the spatial awareness submodule and then importing it into the task awareness submodule of the dynamic detection head; and decoding and outputting the classification result, center point regression result, and bounding box regression result by the task awareness submodule.

[0013] In conjunction with the first aspect, one possible implementation further includes: embedding ECA modulation into the scale-aware submodule, the spatial-aware submodule, and the task-aware submodule, and performing local cross-channel interaction through the ECA modulation to recalibrate attention features.

[0014] In conjunction with the first aspect, in one possible implementation, the CM-RepNCSPELAN4 feature enhancement processing includes: performing convolution and adaptive average pooling on the input feature map to obtain first branch data and second branch data; performing convolution on the first branch data and importing it into a first Concat pooling pool; performing convolution on the second branch data and importing it into the first Concat pooling pool; performing convolution, RepNCSP modulation, GC modulation, and convolution sequentially on the second branch data to form a third branch data, and importing the third branch data into the first Concat pooling pool; performing RepNCSP modulation, GC modulation, and convolution again on the third branch data to form a fourth branch data, and importing the fourth branch data into the first Concat pooling pool; the first Concat pooling pool performs channel merging on the data and outputs merged data; performing ECA modulation, Dynamic Gate modulation, and convolution sequentially on the merged data and outputting CM-RepNCSPELAN4 feature enhancement processing data; wherein, local cross-channel interaction is performed through ECA modulation to recalibrate attention features.

[0015] In conjunction with the first aspect, in one possible implementation, the RepNCSP modulation includes: performing adaptive average pooling on the input feature map to obtain fifth branch data and sixth branch data; performing convolution on the fifth branch data and then passing it through a repeat stacker before importing it into a second Concat pooling pool; performing convolution on the sixth branch data and then importing it into the second Concat pooling pool; the second Concat pooling pool performs channel merging on the data and then performs convolution before outputting RepNCSP modulated data.

[0016] In conjunction with the first aspect, in one possible implementation, the step of performing convolution on the fifth branch data, passing it through a repeated stacker, and then importing it into the second Concat merge pool includes: using the data after performing convolution on the fifth branch data as the first stacker input data; performing RepConvN and convolution on the stacker input data to obtain stacker output data; inputting the stacker output data back into the repeated stacker as the stacker input data; inputting N stacker output data back into the repeated stacker N times as the stacker input data to obtain stacker exported data; and importing the stacker exported data into the second Concat merge pool.

[0017] In conjunction with the first aspect, in one possible implementation, the step of performing local cross-channel interaction through the ECA modulation to recalibrate attention features includes: invoking the ECA module to perform global average pooling on the input feature map; performing one-dimensional convolution on the globally average pooled data; the kernel size of the one-dimensional convolution is related to the number of channels; generating attention weights on the one-dimensional convolution data using the Sigmoid function; and multiplying the attention weights with the input feature map channel by channel to recalibrate attention features.

[0018] Secondly, this application provides an efficient detection system for UAV remote sensing target detection, comprising: an ESPMStem module configured to: extract a first feature map from the image to be processed, and then obtain a second feature map through spatial gating generation and feature response modulation; a feature fusion module communicatively connected to the ESPMStem module, the feature fusion module being configured to: sequentially perform two pre-modulations on the second feature map to obtain a shallow feature map, each pre-modulation including one downsampling and one CM-RepNCSPELAN4 feature enhancement processing; perform downsampling and CM-RepNCSPELAN4 feature enhancement processing on the shallow feature map to obtain a mid-level feature map; perform downsampling and CM-RepNCSPELAN4 feature enhancement processing on the mid-level feature map to obtain a deep feature map; perform SPPF and C2PSA on the deep feature map to obtain a reference deep map, and then perform upsampling on the reference deep map and combine it with the mid-level feature map; The images are concatted to obtain a reference mid-layer image; the reference mid-layer image is upsampled and then concatted with the shallow feature image, followed by CM-RepNCSPELAN4 feature enhancement processing to obtain a shallow output image; the shallow output image is downsampled and then concatted with the reference mid-layer image, followed by CM-RepNCSPELAN4 feature enhancement processing to obtain a mid-layer output image; the mid-layer output image is downsampled and then concatted with the reference deep image, followed by CM-RepNCSPELAN4 feature enhancement processing to obtain a deep output image; a dynamic detection module is communicatively connected to the feature fusion module, and the dynamic detection module is configured to: import the shallow output image, the mid-layer output image, and the deep output image into the dynamic detection head, perform scale awareness, spatial awareness, and task awareness, and obtain the final decoded classification result, center point regression result, and bounding box regression result.

[0019] In its application, this application utilizes the collaborative design of the CM-RepNCSPELAN4 module, which integrates Global Context Modeling (GC), efficient Channel Attention (ECA), and Reparameterized Convolution (RepNCSP). The ECA is then incorporated into the dynamic detection head to form the ECA-DyHead layer. This method systematically resolves the contradiction between accurate multi-scale target detection and efficient deployment on resource-constrained platforms in UAV remote sensing scenarios without significantly increasing computational burden. It achieves a good balance between high accuracy, lightweight design, and strong robustness, and significantly improves the detection capability for small targets. Attached Figure Description

[0020] Figure 1 The diagram shown is a schematic representation of the steps of an efficient target detection method for UAV remote sensing provided in an embodiment of this application.

[0021] Figure 2 A schematic diagram illustrating the steps of a method for preprocessing an image to be processed;

[0022] Figure 3 The diagram shows the processing flow of the ESPMStem module.

[0023] Figure 4 The diagram shown is a schematic representation of the overall process of this application.

[0024] Figure 5 The diagram shows the processing flow of the ECA module.

[0025] Figure 6 The diagram shows the processing flow of the CM-RepNCSPELAN4 module.

[0026] Figure 7 The diagram shows the steps of the CM-RepNCSPELAN4 feature enhancement process.

[0027] Figure 8 The diagram shows the steps of the RepNCSP modulation method.

[0028] Figure 9 The diagram shows the processing steps of the repeating stacker.

[0029] Figure 10 The diagram shows the ECA processing steps.

[0030] Figure 11 The image shows a comparison of the detection performance of each model on the VisDrone2019-DET dataset;

[0031] Figure 12 The image shows a comparison of the visualization detection results of each model on the UAVDT dataset. Detailed Implementation

[0032] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0033] Figure 1 The diagram illustrates the steps of an efficient target detection method for UAV remote sensing, according to an embodiment of this application. This application provides an efficient target detection method for UAV remote sensing; in one embodiment, as shown... Figure 1 As shown, the method includes:

[0034] Step 101: After extracting the first feature map from the image to be processed, a second feature map is obtained through spatial gating generation and feature response modulation. This step constitutes the modulation process of the ESPMStem module (Enhance-Spatial-Preserving ModulationStem).

[0035] Step 102: Perform two pre-modulation operations on the second feature map to obtain a shallow feature map. Each pre-modulation operation includes one downsampling and one CM-RepNCSPELAN4 feature enhancement process. Perform downsampling and CM-RepNCSPELAN4 feature enhancement processes on the shallow feature map to obtain a middle feature map. Perform downsampling and CM-RepNCSPELAN4 feature enhancement processes on the middle feature map to obtain a deep feature map.

[0036] Step 103: After performing SPPF and C2PSA on the deep feature map to obtain the reference deep map, perform upsampling on the reference deep map and then Concat it with the mid-layer feature map to obtain the reference mid-layer map.

[0037] Step 104: After upsampling the reference mid-layer map, perform Concat merging with the shallow feature map, and then perform CM-RepNCSPELAN4 feature enhancement processing to obtain the shallow output map.

[0038] Step 105: After downsampling the shallow layer output image, perform Concat merging with the reference middle layer image, and then perform CM-RepNCSPELAN4 feature enhancement processing to obtain the middle layer output image.

[0039] Step 106: After downsampling the mid-layer output image, perform Concat merging (channel merging) with the reference deep layer image, and then perform CM-RepNCSPELAN4 feature enhancement processing to obtain the deep layer output image.

[0040] Step 107: Import the shallow output map, the middle output map, and the deep output map into the dynamic detection head, perform scale perception, spatial perception, and task perception, and obtain the final decoded classification result, center point regression result, and bounding box regression result.

[0041] In this embodiment, the introduction of the CM-RepNCSPELAN4 module improves the correlation of global contextual information, effectively enhancing the network's ability to extract and express key features, thereby achieving higher target detection accuracy in complex scenarios such as UAV remote sensing images. Furthermore, the SPPF module fuses multi-scale pooling features, and the UCF-Neck structure performs upsampling, stitching, and fusion processing on shallow, medium, and deep feature maps with a unified number of channels, enabling the network to simultaneously and effectively detect targets of different scales (large, medium, and small), especially improving the detection capability for small targets.

[0042] Figure 2 This is a schematic diagram illustrating the steps of a method for preprocessing an image to be processed. In one embodiment, as... Figure 2 As shown, step 101 includes:

[0043] Step 201: Call the first convolutional unit to perform shallow feature extraction on the image to be processed to obtain the first feature map. In this step, the first convolutional unit can obtain the first feature map while maintaining the spatial resolution of the feature map.

[0044] Step 202: Call the second convolutional module to perform cross-channel information aggregation on the first feature map and generate a single-channel spatial response map. After Sigmoid normalization, a spatial gating map is obtained.

[0045] Step 203: The spatial gating map is weighted and adjusted using a fixed scaling factor a, and the weighted spatial gating map is multiplied element by element with the first feature map to achieve spatial importance modulation of the feature response and obtain an intermediate feature map.

[0046] Step 204: The intermediate feature map after spatial modulation is downsampled using a third convolutional unit to obtain the second feature map. In this step, the second feature map can be obtained while preserving key spatial response information using the third convolutional unit.

[0047] Specifically, the first convolutional unit is a 3x3 convolution with a stride of 1, the second convolutional unit is a 1x1 convolution, and the third convolutional unit is a 3x3 convolution with a stride of 2.

[0048] In this embodiment, the first feature map is represented as:

[0049] ;

[0050] The first feature map is extracted using a 3x3 convolution with a stride of 1, resulting in a second feature map with high-resolution features.

[0051] ;

[0052] Xh represents the second feature map of the high-resolution feature;

[0053] This represents a 3x3 convolution with a stride of 1;

[0054] This operation effectively enhances local contextual information and edge structure representation while maintaining the original spatial resolution, providing a more comprehensive fine-grained feature foundation for subsequent spatial importance modeling. This design better preserves structural information from earlier stages, providing a more stable feature representation for subsequent feature selection and compression.

[0055] In acquiring the second feature of high-resolution features Figure X After h, a lightweight spatial gating mechanism is introduced to explicitly model the importance of different spatial locations. Specifically, a 1*1 convolution is used to weight and fuse the multi-channel feature responses, and a spatial gating map is generated using a Sigmoid activation function:

[0056] ;

[0057] The spatial gating graph g represents the relative importance of each spatial location to the target detection task, σ(·) is the sigmoid activation function, and Conv1×1 represents a 1×1 convolution. Compared to complex spatial attention mechanisms, this gating design does not introduce additional multi-branch structures, has extremely low computational overhead, and is suitable for use in the early stages of the network.

[0058] To avoid over-suppressing features in the early stages of the network, this application also employs spatial modulation with a fixed ratio. Specifically, a fixed ratio coefficient α is introduced on top of spatial gating to softly modulate the high-resolution features.

[0059] ;

[0060] Where a∈(0,1] are hyperparameters used to control the strength of overall spatial information preservation. Xm represents the intermediate feature map after gating and fixed-ratio spatial modulation. This design ensures that spatial gating is only responsible for characterizing the distribution of relative importance, without directly causing the complete erasure of features in low-response regions, thereby effectively improving training stability and preventing early feature degradation.

[0061] To further investigate the impact of spatial modulation intensity on detection performance in ESPMStem, experiments were conducted on the VisDrone2019 dataset with different spatial modulation ratios 'a'. The experimental results are shown in Table I. As the modulation ratio increases, the model performance initially improves and then decreases, indicating that stronger spatial enhancement does not necessarily guarantee sustained performance gains. Specifically, the model achieves the best detection performance when 'a' = 0.75, with an mAP50 of 33.4%. Compared to the baseline model YOLO11n, the detection accuracy for pedestrians improved from 33.7% to 35.8%, and for bicycles from 8.4% to 9.4%, demonstrating that ESPMStem has a more significant enhancement effect on dense small targets and fine-grained structural targets. This is mainly because moderate spatial modulation can effectively enhance the spatial representation ability of shallow features, strengthen local edge responses, and retain more fine-grained structural information, thereby improving the discriminability of small targets. However, as the modulation intensity increases further, background textures and invalid spatial responses in complex UAV remote sensing scenes are also amplified, introducing additional noise interference and weakening the ability to discriminate small target features, leading to a decline in detection performance. In contrast, the performance of large targets such as cars varies relatively little under different modulation ratios because they already possess sufficient semantic and structural information and have a lower dependence on shallow spatial detail enhancement. Therefore, when a=0.75, ESPMStem achieves a good balance between feature enhancement and background noise suppression, thus obtaining optimal detection performance.

[0062]

[0063] Table I

[0064] After spatial modulation, a 3x3 convolution with a stride of 2 is used to downsample the intermediate feature map to obtain the second feature map:

[0065] ;

[0066] This indicates a 3x3 convolution with a stride of 2, and X0 represents the second feature map output after downsampling.

[0067] Figure 3 The diagram shows the processing flow of the ESPMStem module. (Refer to...) Figure 3 Steps 201 to 204 constitute an ESPMStem module. Figure 3The solid box in the image represents the ESPMStem module. Since the downsampling operation is performed after spatial importance modulation, regions with higher response values ​​are preferentially preserved during feature compression, thus maximizing the preservation of target-related information while reducing resolution. This "importance-guided downsampling" mechanism significantly improves the discriminative power of early features. Although ESPMStem introduces an additional high-resolution convolution operation in the Stem stage, leading to an increase in FLOPs, its computational overhead is mainly concentrated in the early stages with low channel counts, having a limited impact on overall inference efficiency. Experimental results show that this design significantly improves detection accuracy, especially for small targets, at an acceptable computational cost, demonstrating a good trade-off between accuracy and efficiency.

[0068] In summary, ESPMStem effectively alleviates the problem of spatial information loss caused by premature downsampling in the Stem layer of traditional detectors by introducing an importance-aware spatial modulation mechanism before downsampling. This improves the expressive power of early features and provides a more reliable feature basis for the detection of dense targets and small-scale targets in complex UAV remote sensing scenarios.

[0069] Figure 4 The diagram shown is a schematic representation of the overall process of this application. (Combined with...) Figure 4 In the Backbone layer, a 640*640*3 image to be processed is input into the ESPMStem module from left to right. The ESPMStem module performs feature extraction, spatial gating generation, spatial importance modulation, and downsampling to output a 320*320*16 second feature map, which is then processed by the first 3*3 Conv and the first CM-RepNCSPELAN4 feature enhancement process. The output size of the first 3*3 Conv and the first CM-RepNCSPELAN4 feature enhancement process is 160*160*32. The second 3*3 Conv and the second CM-RepNCSPELAN4 feature enhancement process perform downsampling and CM-RepNCSPELAN4 feature enhancement to obtain a shallow feature map of 80*80*64. The third 3*3 Conv and the third CM-RepNCSPELAN4 feature enhancement process, i.e., the downsampling and CM-RepNCSPELAN4 feature enhancement process performed on the shallow feature map in step 102, yields a middle feature map of 40*40*128. The fourth 3*3Conv and the fourth CM-RepNCSPELAN4 feature enhancement process, namely step 102, perform downsampling and CM-RepNCSPELAN4 feature enhancement on the middle layer feature map to obtain a 20*20*256 deep feature map.

[0070] Combination Figure 4The deep feature map (size 20*20*256) is processed with SPPF and C2PSA to obtain a reference deep map (size 20*20*256), which is then input into the UCF-Neck layer. In the UCF-Neck layer, the following process is performed: the reference deep map is upsampled and then concatted (channel merging) with the mid-layer feature map (size 40*40*128) to obtain a reference mid-layer map (size 40*40*128). The reference mid-layer map (size 40*40*128) is then upsampled and concatted (channel merging) with the shallow feature map (size 80*80*64), and finally CM-RepNCSPELAN4 feature enhancement processing is performed to obtain a shallow output map (size 80*80*64). The shallow layer output image (80*80*64) is downsampled and then concat-merged (channel merging) with the reference intermediate layer image (40*40*128). Then, CM-RepNCSPELAN4 feature enhancement processing is performed to obtain the intermediate layer output image (40*40*128). The intermediate layer output image (40*40*128) is downsampled and then concat-merged (channel merging) with the reference deep layer image (20*20*256). Then, CM-RepNCSPELAN4 feature enhancement processing is performed to obtain the deep layer output image (20*20*256).

[0071] Combination Figure 4 The shallow, medium, and deep output maps are imported into the dynamic detection head (ECA-DyHead). Scale perception, spatial perception, and task perception are performed in the dynamic detection head to obtain the final decoded classification results, center point regression results, and bounding box regression results. The classification results include the category label for each target, the center point regression results include the center point coordinates, width, and height of the target box, and the bounding box regression results include the predicted width and height (w, h) or offset of the object's bounding box, thereby determining the size of the object.

[0072] Figure 5 The diagram shows the processing flow of the ECA module. (Refer to...) Figure 5 Step 107 includes:

[0073] Step 1071: After processing the shallow, medium and deep output maps with the feature pyramid, convert them into General View data and then import them into the scale-aware attention submodule of the dynamic detection head. After associating the features of different scale feature maps with each other, import them into the spatial-aware attention submodule of the dynamic detection head.

[0074] Step 1072: After the spatial awareness submodule performs attention enhancement, it is imported into the task-aware attention submodule of the dynamic detection head (DyHead).

[0075] Step 1073: The task-aware submodule decodes and outputs the classification results (Object Classification), center regression results (Center Regression), and bounding box regression results (Box Regression).

[0076] In one embodiment, reference is made to Figure 5 This efficient detection method for UAV remote sensing target detection also includes:

[0077] Step 1070: Embed ECA (Efficient Channel Attention) modulation into the scale perception submodule, spatial perception submodule, and task perception submodule, and perform local cross-channel interaction through ECA modulation to recalibrate attention features.

[0078] Figure 6 The diagram shows the processing flow of the CM-RepNCSPELAN4 module. Figure 7 The diagram illustrates the steps of the CM-RepNCSPELAN4 feature enhancement process. For details, please refer to... Figure 6 ,like Figure 7 As shown, the CM-RepNCSPELAN4 feature enhancement process includes:

[0079] Step 701: Perform convolution and adaptive average pooling on the input feature map to obtain the first branch data and the second branch data.

[0080] Step 702: After performing convolution on the data from the first branch, import it into the first Concat pool. The first Concat pool is used to perform Concat merging, i.e., channel merging, on the data input to this module.

[0081] Step 703: Perform convolution on the data in the second branch and then import it into the first Concat merge pool.

[0082] Step 704: Perform convolution, RepNCSP modulation, GC modulation, and convolution on the second branch data in sequence to form the third branch data, and import the third branch data into the first Concat merging pool.

[0083] Step 705: Perform RepNCSP modulation, GC modulation, and convolution on the third branch data again to form the fourth branch data, and import the fourth branch data into the first Concat merging pool.

[0084] Step 706: The first Concat merging pool merges the data through the execution channel and outputs the merged data.

[0085] Step 707: After sequentially performing ECA modulation, Dynamic Gate modulation, and convolution on the merged data, output CM-RepNCSPELAN4 feature enhancement processing data; wherein, local cross-channel interaction is performed through ECA modulation to recalibrate attention features.

[0086] Figure 8 The diagram illustrates the steps of RepNCSP modulation. (Refer to...) Figure 6 ,like Figure 8 As shown, RepNCSP modulation includes:

[0087] Step 801: Perform adaptive average pooling on the input feature map to obtain the fifth branch data and the sixth branch data.

[0088] Step 802: After performing convolution on the data from the fifth branch, the data is passed through a repeat stacker and then imported into the second Concat pool. The second Concat pool is used to perform concat merging (channel merging) on ​​the data input to this module. In this step, the repeat stacker is used to perform reparameterized convolution RepNCSP.

[0089] Step 803: Perform convolution on the data from the sixth branch and import it into the second Concat merge pool.

[0090] Step 804: The second Concat pool merges the data channels, performs convolution, and outputs RepNCSP modulated data.

[0091] Figure 9 The diagram illustrates the processing steps of a repeating stacker. (Refer to...) Figure 6 ,like Figure 9 The diagram shows the steps of the RepNCSP method for reparameterized convolution. Step 802 includes:

[0092] Step 901: Use the data from the fifth branch after convolution as the input data for the first stacker.

[0093] Step 902: Perform RepConvN and convolution on the input data of the stacker to obtain the output data of the stacker.

[0094] Step 903: Recycle the stacker output data into the repeating stacker as the stacker input data.

[0095] Step 904: Take the output data of N stackers, and input it into the stacker N times as the input data of the stacker to obtain the output data of the stacker.

[0096] Step 905: Import the data exported from the stacker into the second Concat merge pool.

[0097] Figure 10 The diagram illustrates the ECA processing steps. In one embodiment, as shown... Figure 10 As shown, performing local cross-channel interaction via ECA modulation to recalibrate attention features includes:

[0098] Step 1001: Call the ECA module to perform global average pooling on the input feature map.

[0099] Step 1002: Perform a one-dimensional convolution on the data after global average pooling. In this step, the kernel size of the one-dimensional convolution is related to the number of channels and is adaptively determined by the number of channels.

[0100] Step 1003: Generate attention weights for the data after one-dimensional convolution using the Sigmoid function.

[0101] Step 1004: Multiply the attention weights with the input feature map channel by channel to perform attention feature recalibration.

[0102] Reference Figure 5 In the object detection framework, the detection head is responsible for decoding multi-scale features into object category and location predictions. Its structural design directly affects the model's ability to perceive complex scenes and multi-scale objects. The dynamic detection head, by introducing dynamic convolution and multi-dimensional attention mechanisms, enables the detection head to adaptively adjust the feature fusion strategy based on input features, thereby effectively improving the model's expressive power. The dynamic detection head can construct a collaborative attention mechanism in three dimensions: hierarchical, spatial, and channel-based, giving it scale-aware, spatial, and task-aware capabilities. In the hierarchical dimension, the dynamic detection head learns key feature relationships between different semantic levels through a scale-aware attention module; in the spatial dimension, it enhances the detection head's response to key regions through a spatial-aware attention mechanism; and in the channel-based dimension, it uses task-aware attention to focus on classification and regression tasks respectively, thereby improving detection performance.

[0103] To enhance the feature modeling capability of the dynamic detection head in the channel dimension, this application introduces the ECA channel attention mechanism on the basis of the dynamic detection head, forming the ECA-DyHead detection head. ECA achieves local cross-channel dependency modeling through lightweight one-dimensional convolution, and completes the efficient calculation of attention weights without channel dimensionality reduction. ECA does not rely on fully connected mapping for dimensionality compression, which can avoid information loss and the introduction of additional parameters, and more effectively model the correlation between channels while maintaining low computational overhead. The model structure of ECA-DyHead is shown in Figure 5. The ECA mechanism is added after scale awareness, spatial awareness, and task awareness, which enhances the feature modeling capability of the dynamic detection head in the channel dimension.

[0104] To compare the performance of different detection heads, this application conducted comparative experiments on the VisDrone dataset. Yolo11n (UCF-Neck), DyHead, and ECA-DyHead were all built based on UCF-Neck, ensuring consistent channel dimensions for the input features of each detection head. The performance comparison of different detection head structures on the VisDrone dataset is shown in Table II. As can be seen from Table II, after introducing UCF-Neck, the number of parameters in Yolo11n (UCF-Neck) decreased from 2.58M to 1.89M, achieving significant model lightweighting, but with a slight decrease in detection accuracy. When DyHead was further introduced, while maintaining low computational complexity, mAP50 and mAP50–95 increased to 0.34 and 0.20, respectively, indicating that dynamic detection heads can effectively enhance the feature modeling capabilities of detection heads. Building upon this foundation, by introducing ECA channel attention, ECA-DyHead further improves the performance of two metrics to 0.352 and 0.208, achieving optimal detection performance under the same parameter scale and computational complexity. Experimental results demonstrate that the proposed ECA-DyHead effectively enhances the channel feature selection capability of the detection head, further improving detection accuracy while maintaining the model's lightweight characteristics, thus validating the effectiveness of the proposed detection head structure.

[0105]

[0106] Table II

[0107] This application also provides an efficient detection system for UAV remote sensing target detection, including: an ESPMStem module, a feature fusion module, and a dynamic detection module.

[0108] The ESPMStem module is configured to extract a first feature map from the image to be processed, and then obtain a second feature map through spatial gating generation and feature response modulation.

[0109] The feature fusion module communicates with the ESPMStem module. The feature fusion module is configured as follows: It performs two pre-modulation operations on the second feature map to obtain a shallow feature map. Each pre-modulation operation includes one downsampling and one CM-RepNCSPELAN4 feature enhancement process. It then performs downsampling and CM-RepNCSPELAN4 feature enhancement on the shallow feature map to obtain a mid-level feature map. Finally, it performs downsampling and CM-RepNCSPELAN4 feature enhancement on the mid-level feature map to obtain a deep feature map. SPPF and C2PSA are then applied to the deep feature map to obtain a reference deep map. Finally, the reference deep map is further processed... After sampling, the image is concatted with the intermediate feature map to obtain a reference intermediate image. The reference intermediate image is then upsampled and concatted with the shallow feature map, and then CM-RepNCSPELAN4 feature enhancement processing is performed to obtain a shallow output image. The shallow output image is then downsampled and concatted with the reference intermediate image, and then CM-RepNCSPELAN4 feature enhancement processing is performed to obtain a intermediate output image. The intermediate output image is then downsampled and concatted with the reference deep image, and then CM-RepNCSPELAN4 feature enhancement processing is performed to obtain a deep output image.

[0110] The dynamic detection module communicates with the feature fusion module. The dynamic detection module is configured to import the shallow output map, the middle output map, and the deep output map into the dynamic detection head, perform scale perception, spatial perception, and task perception, and obtain the final decoded classification result, center point regression result, and bounding box regression result.

[0111] This application provides an efficient target detection method for UAV remote sensing, which improves detection accuracy, lightweight design, and multi-scale adaptability. By using a sigmoid spatial gating map and soft modulation with a fixed scaling factor, it achieves adaptive preservation of spatial information and noise suppression in shallow feature maps, ensuring that spatial details of small targets are not lost in subsequent processing. Through the CM-RepNCSPELAN4 module, global context modeling (GC) information is integrated into deep feature maps, and this global semantic information is progressively passed to mid-level and shallow feature maps through upsampling, channel concatenation, and re-modulation. This gives all feature maps used for detection rich context awareness capabilities, effectively distinguishing targets from background and significantly improving the detection accuracy of small and occluded targets. By embedding efficient channel attention (ECA) into the scale awareness, spatial awareness, and task awareness sub-modules of the dynamic detection head (DyHead), it achieves fine-grained recalibration of feature map channels, enabling the network to more accurately focus on key features, thereby improving the accuracy of classification and regression.

[0112] The ECA module used achieves local cross-channel interaction through one-dimensional convolution, avoiding channel dimensionality reduction. It has very few parameters and negligible computational overhead, making it an efficient means of achieving channel attention recalibration. RepNCSP modulation, through the RepConvN process of multi-branch training and single-branch inference, significantly reduces the parameters and computational load in the inference stage while maintaining high accuracy, thus accelerating the model's running speed on UAV edge devices.

[0113] In the UCF-Neck layer, channel compression and unified processing of feature maps at different scales effectively control the number of channels in the feature maps, avoiding the surge in parameters caused by feature splicing, and maintaining the lightweight and real-time performance of the network.

[0114] In the ECA-DyHead layer, the scale-aware submodule of the dynamic detection head adaptively correlates and fuses feature maps at different scales, enabling the model to flexibly cope with scale differences caused by variations in flight altitude and target size in UAV images. By capturing global dependencies in deep features through GC (Global Context Modeling), the model can effectively resist common interferences in UAV remote sensing images, such as uneven illumination, background clutter, and dense target occlusion, thus improving the stability and robustness of detection.

[0115] In summary, the technical solution of this application, through the collaborative design of the ESPMStem module, the CM-RepNCSPELAN4 module, and the ECA-DyHead layer, systematically solves the contradiction between accurate multi-scale target detection and efficient deployment of resource-constrained platforms in UAV remote sensing scenarios without significantly increasing the computational burden, and achieves a good balance between detection accuracy, model lightweighting, and environmental robustness.

[0116] To verify the technical effectiveness of this application, data verification is presented below.

[0117] VisDrone2019-DET is a widely used benchmark dataset for object detection in the field of UAV vision research. This dataset contains 10,209 real aerial images, of which 6,471 were used for training, 548 for validation, and 3,190 for testing. All images were acquired by different types of UAV platforms under various environmental and lighting conditions, covering typical scenes such as urban blocks, rural roads, and highways. The dataset labels 10 object classes, including pedestrians (PT), people (PP), bicycles (BC), cars (CA), vans (VA), trucks (TR), tricycles (TC), canopy tricycles (AT), buses (BU), and motorcycles (MC). In addition, the dataset provides supplementary attribute information such as object occlusion degree, truncation ratio, and object scale. Because objects occupy a small proportion of the image from a UAV's top-down view, and are densely distributed with significant scale variations, this dataset provides a challenging experimental platform for evaluating the performance of object detection algorithms in complex aerial photography scenarios.

[0118] To further verify the model's generalization ability, this application also conducted supplementary experiments on the UAVDT dataset. UAVDT primarily targets target detection and tracking tasks in UAV traffic monitoring scenarios. Its data comes from UAV video sequences in real urban traffic environments, covering different weather conditions, lighting variations, and flight altitudes. Compared to VisDrone2019-DET, UAVDT focuses more on vehicle targets, mainly including cars, buses, and trucks. Due to the high flight altitude of UAVs, targets are typically small in scale within the images, accompanied by complex background interference and significant scale variations, making this dataset also quite challenging to detect. Since the raw UAVDT data is provided in video sequence form, consecutive frames within the same video often contain a large amount of similar content, easily leading to data redundancy and affecting fair comparisons between different detection models. Therefore, this application selected and constructed 9,651 images from the original video sequences for target detection experiments. To avoid data leakage, this application adopts a strict video-level partitioning strategy, that is, using video sequences as the smallest partitioning unit, dividing images from the same video into the same subset, and dividing them into training set, validation set and test set according to the ratio of 70% / 20% / 10%, thereby ensuring the fairness and reliability of experimental evaluation.

[0119] Overall, VisDrone2019-DET focuses on UAV target detection in multi-class dense scenes, while UAVDT highlights the challenge of small-scale vehicle detection in traffic scenes. The two datasets are highly complementary in terms of scene type, class distribution, and target density, which helps to comprehensively evaluate the model's generalization ability and robustness in complex UAV remote sensing environments.

[0120] All experiments were conducted on the same computer configuration using Python 3.9 and the PyTorch 2.1.2 framework. The hardware environment included a 12th Gen Intel® Core™ i7-12700KF @3.61 GHz processor and an NVIDIA GeForce RTX 4070, 12282MiB graphics card. Experimental results from YOLOv11n were used as a baseline for comparative analysis. During the training phase, no pre-trained weights were used to initialize the model. The initial learning rate was set to 0.01, and the momentum coefficient was 0.937. During training, the batch size was set to 16, the SGD optimizer was selected, and the core parameters were lr=0.01 and momentum=0.9.

[0121] To improve the scientific rigor and stability of model selection, an Early Stopping mechanism is introduced: training automatically terminates when the mAP50 and mAP50-95 metrics on the validation set fail to improve for 10 consecutive rounds, reducing the risk of overfitting; the maximum number of training rounds is fixed at 400. The model weights that best perform in terms of mAP50 and mAP50-95 on the validation set are retained for testing and evaluation. To evaluate the complexity and performance of the proposed algorithm, multiple evaluation metrics are used for quantification. The number of model parameters and the number of floating-point operations are used as measures of complexity. Detection performance is evaluated using precision (P), recall (R), and mean average precision (mAP).

[0122] As shown in Table III, this application conducts a layer-by-layer comparative analysis of the parameter distribution of the baseline model YOLO11n and the improved model ECE-YOLO at different network stages. The results show that although ECE-YOLO introduces more expressive structures in some local modules, the model ultimately achieves effective parameter compression through the redesign of the overall network structure. In the Backbone stage, ECE-YOLO replaces the original ConvStem with ESPMStem, increasing the number of parameters in Backbone-1. This increase mainly stems from the introduction of high-resolution feature extraction and spatial modulation mechanisms, aiming to more fully preserve key spatial information in the early stages of the network. It should be noted that this parameter increase is mainly concentrated in the low-channel layers of the network front-end, thus having a relatively limited impact on the overall model size. As the network depth increases, ECE-YOLO optimizes the parameter distribution of the middle and later layers by replacing multiple C3K2 modules with CM-RepNCSPELAN4. In the Backbone-7 and Backbone-9 stages, the number of parameters decreased by 29,219 and 116,811 respectively, indicating that the reparameterization structure combined with the efficient feature aggregation mechanism can reduce redundant parameters while maintaining feature expressive power. This structural design of "moderate front-end enhancement and effective back-end compression" makes the model more compact in terms of overall parameter scale. In the Neck part, ECE-YOLO uses UCF-Neck to replace the original feature fusion structure, which significantly reduces the number of parameters from 522,272 to 59,161, becoming one of the main sources of the model parameter reduction. This shows that by reasonably designing the feature fusion structure, the model complexity can be significantly reduced while maintaining the ability to transmit multi-scale information. In the DetectionHead stage, although the introduction of ECA-DyHead brings a small increase in parameters, this increase is offset by the parameter compression in the Backbone and Neck stages. Finally, the total number of parameters in ECE-YOLO decreased from 2,591,790 to 2,101,258, an overall reduction of approximately 18.9%. Overall, this model does not achieve lightweighting by simply reducing the network size, but rather achieves reasonable allocation and efficient utilization of parameters through structural redesign.

[0123] To further verify the performance of the proposed ECE-YOLO in target detection tasks of UAV remote sensing images, this application compared it with several mainstream detection models, including YOLOv5n, YOLOv8n, and YOLO11n, under the same experimental settings. The comparative experiments were conducted on two public datasets: VisDrone2019-DET and UAVDT. In terms of overall detection performance (Table IV), ECE-YOLO achieved 36.6% mAP@50 on the VisDrone dataset, a 3.8 percentage point improvement compared to the baseline model YOLO11n's 32.8%. Furthermore, compared to YOLOv8n (34.1%) and YOLOv5n (31.2%), the model improved by 2.5 and 5.4 percentage points, respectively. In terms of category detection results, the improved model achieved superior detection accuracy in categories such as PT, PP, BC, CA, VA, TR, TC, AT, BU, and MC. For example, the accuracy reaches 0.386 for the PT category, 0.509 for the BU category, and 0.415 for the MC category, indicating that the proposed method can effectively improve the detection capability of different target categories in complex UAV scenarios. Under the more stringent evaluation metric mAP@50–95 (Table V), ECE-YOLO also achieves good performance, with an overall accuracy of 21.9%, which is 2.6 percentage points higher than YOLO11n's 19.3%. This shows that the proposed method not only has advantages under the more lenient IoU threshold condition, but also maintains stable improvement under the more stringent localization evaluation criteria, indicating that the model has good robustness in terms of target localization accuracy and bounding box regression. In addition, as shown in Table VI, on the VisDrone dataset, the overall performance of the proposed ECE-YOLO model is better than YOLO11n. Specifically, the mean precision improved from 0.421 to 0.476, while the recall improved from 0.331 to 0.354, indicating a consistent improvement in detection capability. At the category level, precision improved for most object categories, particularly for PT (from 0.437 to 0.546), VA (from 0.439 to 0.516), and BU (from 0.490 to 0.555), indicating stronger discriminative power under complex background conditions. Recall also improved for most categories, suggesting better detection completeness. For categories with relatively low baseline performance, such as BC (from 0.251 to 0.306) and AT (from 0.240 to 0.308), the proposed method still achieved significant improvements, further validating its robustness. Although precision slightly decreased for the TR (truck) category (from 0.451 to 0.443), recall improved from 0.319 to 0.329, indicating that more truck instances were successfully detected.This observation is consistent with the visualization results, where the proposed model exhibits stronger sensitivity to truck targets, indicating an improved ability to capture such objects despite a slight increase in false positive rate. To further validate the model's generalization ability, additional experiments were conducted on the UAVDT dataset (Table VII). Experimental results show that ECE-YOLO also achieves good detection performance on this dataset, with a precision of 0.482, a recall of 0.407, mAP@50 of 36.4%, and mAP@50–95 of 20.3%, all outperforming models such as YOLOv5n, YOLOv8n, and YOLO11n. This demonstrates that the proposed method not only performs well on the VisDrone dataset but also maintains stable performance under different data distributions, exhibiting good cross-dataset generalization ability. Furthermore, to verify the effectiveness of the proposed ECE-YOLO, comparative experiments were conducted on the VisDrone dataset with several advanced UAV target detection methods, and the experimental results are shown in Table VIII. The results show that ECE-YOLO achieves the best detection performance while maintaining low model complexity, demonstrating a good balance between accuracy and efficiency. Specifically, ECE-YOLO achieves 36.6% mAP@50 with only 2.1M parameters and 7.0 GFLOPs, significantly outperforming most existing lightweight detection models. Compared to RMVAD-YOLO, the model in this application improves mAP@50 by 7.8% with fewer parameters; compared to Gold-YOLO-n, it improves detection accuracy by 3.6% while reducing parameters by approximately 62.6% and GFLOPs by approximately 42.1%. Furthermore, compared to YOLO-DMF, ECE-YOLO improves mAP@50 by 1.3% while reducing parameters and computation by approximately 43.2% and 76.1%, respectively. It is worth noting that although MFFSODNet has a low number of parameters, its GFLOPs are as high as 21.5, indicating that it still has high computational overhead in feature extraction and multi-scale modeling. In contrast, ECE-YOLO achieves higher detection accuracy with lower computational complexity. This result shows that the network structure proposed in this application can improve detection performance while ensuring computational efficiency, achieving a good balance between accuracy and model complexity.

[0124] In summary, system experiments on the VisDrone and UAVDT datasets demonstrate that ECE-YOLO exhibits stable advantages in detection accuracy, localization performance, and model lightweighting. Experimental results show that the proposed structural improvements can enhance the model's ability to represent target features in complex UAV remote sensing scenarios while maintaining low computational complexity, thereby improving overall detection performance.

[0125]

[0126] Table III

[0127]

[0128] Table IV

[0129]

[0130] Table V

[0131]

[0132] Table VI

[0133]

[0134] Table VII

[0135]

[0136] Table VIII

[0137] like Figure 11 and Figure 12 . Figure 11 In the image, pedestrians are represented in red, people in orange, bicycles in yellow, cars in green, vans in blue, trucks in purple, tricycles in cyan, awning-tricycles in pinkish-purple, buses in dark orange, and motors in yellowish-green. (a), (b), (c), and (d) represent four sets of images. Figure 12 In the diagram, blue boxes represent car-type targets, orange boxes represent truck-type targets, and green boxes represent bus-type targets. In the visualization and analysis of target detection results, such as... Figure 11 As shown, the comparison of detection performance of each model on the VisDrone2019-DET dataset is presented. Overall, the proposed model performs better in terms of multi-class discrimination ability and false detection control. Specifically, in the recognition of two similar-looking targets, "car" and "van," the proposed model can distinguish between the two better, while the comparative models YOLO11n and YOLOv8n are more prone to class confusion. Furthermore, in the pedestrian category detection, the comparative models are more prone to false detections. For example, in... Figure 11 In the scenario shown in (d), both YOLO11n and YOLOv8n misidentify streetlights as pedestrians and generate false detection results for cars in some areas without vehicles.

[0138] For small-scale targets like bicycles, detection is challenging due to their limited pixel count in images. The YOLO11n model frequently misidentifies motors as bicycles, resulting in significant misclassification. In contrast, the method presented in this application reduces false detections for this type of target. Regarding truck category detection, Figure 11 In the scenario shown in (a), the number of missed detections and false detections of the model in this application is significantly less than that of other models; while Figure 11 In (b), only the model in this application successfully detected all three truck targets. For the awning-tricycle category, such as Figure 11 As shown in (c), the comparison models all misclassified it as a truck, while the model of this application was able to correctly identify the category, demonstrating a stronger fine-grained category discrimination ability.

[0139] Furthermore, such as Figure 12 The image shows the visualized detection results of each model on the UAVDT dataset. From... Figure 12 (a) It can be seen that YOLOv5n misdetects a car target as a truck, while YOLO11n incorrectly identifies houses in the background as trucks; at the same time, the model in this application detects more car targets than YOLOv8n. Figure 12 In (b), only the model in this application successfully detected the truck target, while the other models misidentified the target as a car. Figure 12 In the complex scene shown in (c), the ground truth (GT) contains 9 bounding boxes. YOLOv5n detects only 4 targets, YOLO11n also detects only 4 correctly, while YOLOv8n produces 5 false positives. In contrast, the model in this application successfully detects 7 targets, demonstrating better detection completeness and a lower false positive rate.

[0140] The visualization results above show that the method in this application can effectively reduce class confusion and false detection in complex UAV remote sensing scenarios, and has better robustness and detection capability in dense targets and small targets, thus further verifying the effectiveness of the proposed model.

[0141] To explore the importance of the proposed fused network and the three modules in ECE-YOLO, ablation experiments were conducted on the VisDrone2019-DET dataset. Starting from the baseline, this application progressively integrated the ESPMStem, CM-RepNCSPELAN4, and ECA-DyHead modules to evaluate their respective contributions. Table IX shows how adding or removing each component affects performance metrics. √ indicates that the improved module or network was integrated into the baseline model, and - indicates that it was not integrated.

[0142]

[0143] Table IX

[0144] 1)ESPMStem

[0145] As shown in Table IX, after introducing ESPMStem into the baseline YOLOv11n model, the number of model parameters remained unchanged (both were 2.58M), indicating that this module improved detection performance without introducing any additional parameter overhead, demonstrating good computational efficiency. While maintaining the same model complexity, detection performance was steadily improved, with mAP@50 increasing from 0.328 to 0.334 and mAP@50–95 increasing from 0.193 to 0.198. This result shows that ESPMStem can effectively enhance feature representation capabilities while maintaining lightweight characteristics. By introducing an importance-aware spatial modulation mechanism in the initial stage of the network, it makes the network focus more on discriminative spatial regions before feature compression, thereby reducing the propagation of invalid information and improving the utilization rate of effective features. Therefore, ESPMStem achieves a stable improvement in detection performance without increasing model complexity, demonstrating good efficiency and accuracy improvement capabilities.

[0146] 2)CM-RepNCSPELAN4

[0147] When only the CM-RepNCSPELAN4 module is introduced, the number of model parameters increases slightly from 2.58M to 2.59M, with minimal change in overall complexity, but the detection performance is still further improved. Specifically, mAP@50 increases to 0.337, and mAP@50–95 increases from 0.193 to 0.196. This result shows that this module can effectively enhance feature representation capabilities while maintaining the lightweight structure of the model. CM-RepNCSPELAN4, through its reparameterized structure and multi-branch feature aggregation strategy, enables the network to more fully integrate information from different scales and semantic levels, thereby improving the discriminative power of feature representation. Notably, when CM-RepNCSPELAN4 is used in conjunction with ECA-DyHead, the number of model parameters decreases from 2.58M in the baseline model to 2.10M, while mAP@50 increases to 0.352, indicating a good synergistic optimization effect between feature representation and detection head modeling. Specifically, this is because the combination of the CM-RepNCSPELAN4 module and UCF-Neck significantly reduces structural redundancy in the multi-scale feature fusion process. Compared to simply increasing the network depth or width, this design achieves a more efficient feature utilization method through structural optimization, improving accuracy with less computational overhead, demonstrating a good efficiency-performance balance.

[0148] 3) ECA-DyHead

[0149] As shown in Table IX, after introducing ECA-DyHead, the number of model parameters decreased from 2.58M to 2.27M, significantly reducing the model size, while the detection performance was still significantly improved. Specifically, mAP@50 increased to 0.352, and mAP@50-95 increased to 0.208. This indicates that this module not only improves detection accuracy but also further optimizes the model's lightweight nature. ECA-DyHead combines a lightweight channel attention mechanism with a dynamic feature fusion strategy, enabling the network to adaptively strengthen key channel information and suppress redundant features, thereby improving the effectiveness of multi-scale feature fusion. This mechanism improves feature representation quality while reducing model parameters, allowing the model to achieve better detection performance while maintaining high computational efficiency.

[0150] In summary, this application addresses the problems of small target scale, complex background, and easy loss of early features in UAV remote sensing target detection. By introducing ESPMStem, CM-RepNCSPELAN4, and ECA-DyHead modules at different stages of the network, this application systematically optimizes the feature extraction and fusion process, significantly improving detection performance while maintaining low model complexity.

[0151] Specifically, ESPMStem enhances the spatial information preservation capability in the early stages of the network through an importance-aware spatial modulation mechanism, thereby improving the expression quality of key features; CM-RepNCSPELAN4 strengthens the global feature modeling capability through reparameterization and multi-branch feature aggregation strategies, improving feature representation capability while maintaining computational efficiency; and ECA-DyHead combines lightweight channel attention and dynamic feature fusion mechanisms to effectively enhance the interaction capability between high-level features and improve feature discriminability. These modules complement each other at different network layers, enabling the model to more effectively utilize multi-scale feature information in complex UAV remote sensing scenarios.

[0152] Experimental results on the VisDrone2019-DET and UAVDT drone datasets demonstrate that the proposed ECE-YOLO outperforms the baseline model YOLO11n in both detection accuracy and model efficiency. Specifically, on the VisDrone dataset, mAP@50 is improved from 32.8% to 36.6%, and mAP@50–95 from 19.3% to 21.9%, while the number of model parameters is reduced to 2.1M. Experimental results on the UAVDT dataset also validate the effectiveness of the proposed method. Compared to YOLO11n, ECE-YOLO improves mAP@50 from 33.1% to 36.4%, and mAP@50–95 from 18.1% to 20.3%. These results show that the proposed method effectively improves detection accuracy while maintaining a lightweight model, achieving a good trade-off between accuracy and complexity.

[0153] In summary, the ECE-YOLO proposed in this application provides an effective design approach for constructing efficient UAV remote sensing target detection models, demonstrating good application potential in resource-constrained real-world scenarios. Future work will further explore more adaptive spatial modulation strategies and investigate their combination with global modeling methods (such as Transformer) to further enhance the model's generalization ability and robustness in complex remote sensing scenarios.

[0154] The basic principles of this application have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this application are merely examples and not limitations, and should not be considered as essential features of each embodiment of this application. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the application to the necessity of employing the aforementioned specific details for implementation.

[0155] The block diagrams of devices, apparatuses, devices, and systems involved in this application are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as “comprising,” “including,” “having,” etc., are open-ended terms meaning “including but not limited to,” and are used interchangeably with them. The terms “or” and “and” as used herein refer to the terms “and / or,” and are used interchangeably with them unless the context clearly indicates otherwise. The term “such as” as used herein refers to the phrase “such as but not limited to,” and is used interchangeably with it.

[0156] It should also be noted that in the apparatus, equipment, and methods of this application, the components or steps can be disassembled and / or recombined. These disassemblies and / or recombinations should be considered as equivalent solutions of this application.

[0157] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this application. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein can be applied to other aspects without departing from the scope of this application. Therefore, this application is not intended to be limited to the aspects shown herein, but rather to be carried out within the widest scope consistent with the principles and novel features of the invention herein.

[0158] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications or equivalent substitutions made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A highly efficient target detection method for UAV remote sensing, characterized in that, include: After extracting the first feature map from the image to be processed, the second feature map is obtained through spatial gating generation and feature response modulation. The second feature map is subjected to two pre-modulation operations to obtain a shallow feature map. Each pre-modulation operation includes one downsampling and one CM-RepNCSPELAN4 feature enhancement process. The shallow feature map is subjected to downsampling and CM-RepNCSPELAN4 feature enhancement processes to obtain a middle feature map. The middle feature map is subjected to downsampling and CM-RepNCSPELAN4 feature enhancement processes to obtain a deep feature map. After performing SPPF and C2PSA on the deep feature map, a reference deep map is obtained. Then, the reference deep map is upsampled and merged with the mid-layer feature map using Concat to obtain a reference mid-layer map. The reference mid-layer map is upsampled and then concatted with the shallow feature map. Finally, CM-RepNCSPELAN4 feature enhancement processing is performed to obtain the shallow output map. The shallow output image is downsampled and then concatted with the reference mid-layer image. Then, CM-RepNCSPELAN4 feature enhancement processing is performed to obtain the mid-layer output image. The intermediate output image is downsampled and then concatted with the reference deep image. Finally, CM-RepNCSPELAN4 feature enhancement processing is performed to obtain the deep output image. The shallow output map, the middle output map, and the deep output map are imported into the dynamic detection head to perform scale perception, spatial perception, and task perception, and finally decoded classification results, center point regression results, and bounding box regression results are obtained.

2. The efficient detection method for UAV remote sensing target detection according to claim 1, characterized in that, The step of obtaining a second feature map by spatial gating generation and feature response modulation after extracting the first feature map from the image to be processed includes: The first convolutional unit is invoked to perform shallow feature extraction on the image to be processed to obtain the first feature map; The second convolutional module is invoked to aggregate cross-channel information of the first feature map and generate a single-channel spatial response map, which is then normalized by Sigmoid to obtain a spatial gating map. The spatial gating map is weighted and adjusted using a fixed scaling factor a, and the weighted spatial gating map is multiplied element-wise with the first feature map to obtain an intermediate feature map; The second feature map is obtained by downsampling the spatially modulated intermediate feature map using a third convolutional unit.

3. The efficient detection method for UAV remote sensing target detection according to claim 2, characterized in that: The first convolutional unit is a 3x3 convolution with a stride of 1, the second convolutional unit is a 1x1 convolutional unit, and the third convolutional unit is a 3x3 convolutional unit with a stride of 2.

4. The efficient detection method for UAV remote sensing target detection according to claim 1, characterized in that, The process of importing the shallow output map, the middle output map, and the deep output map into the dynamic detection head, performing scale perception, spatial perception, and task perception, and obtaining the final decoded classification result, center point regression result, and bounding box regression result includes: The shallow output map, the middle output map, and the deep output map are all imported into the scale perception submodule of the dynamic detection head, and the feature maps of different scales are correlated with each other and then imported into the spatial perception submodule of the dynamic detection head. After the spatial perception submodule performs attention enhancement, it is imported into the task perception submodule of the dynamic detection head. The task perception submodule decodes and outputs classification results, center point regression results, and bounding box regression results.

5. The efficient detection method for UAV remote sensing target detection according to claim 4, characterized in that, Also includes: ECA modulation is embedded in the scale-aware submodule, the spatial-aware submodule, and the task-aware submodule, and local cross-channel interaction is performed through the ECA modulation to recalibrate attention features.

6. The efficient detection method for UAV remote sensing target detection according to claim 1, characterized in that, The CM-RepNCSPELAN4 feature enhancement process includes: The input feature map is subjected to convolution and adaptive average pooling to obtain the first branch data and the second branch data. After performing convolution on the data in the first branch, import it into the first Concat merge pool; After performing convolution on the data in the second branch, import it into the first Concat merge pool; The second branch data is sequentially subjected to convolution, RepNCSP modulation, GC modulation, and convolution to form the third branch data, and the third branch data is imported into the first Concat merging pool; The third branch data is subjected to RepNCSP modulation, GC modulation, and convolution again to form the fourth branch data, and the fourth branch data is imported into the first Concat merging pool. The first Concat merging pool will merge the data through the execution channel and then output the merged data. The merged data is sequentially subjected to ECA modulation, Dynamic Gate modulation, and convolution to output CM-RepNCSPELAN4 feature enhancement processing data; wherein, local cross-channel interaction is performed through ECA modulation to recalibrate attention features.

7. The efficient detection method for UAV remote sensing target detection according to claim 6, characterized in that, The RepNCSP modulation includes: The input feature map is subjected to adaptive average pooling to obtain the fifth and sixth branch data. After performing convolution on the data in the fifth branch, it is passed through a repeat stacker and then imported into the second Concat merge pool. The data from the sixth branch is convolved and then imported into the second Concat merging pool; The second Concat pool merges the data channels, performs convolution, and then outputs RepNCSP modulated data.

8. The efficient detection method for UAV remote sensing target detection according to claim 7, characterized in that, The step of performing convolution on the data from the fifth branch, passing it through a repeat stacker, and then importing it into the second Concat merge pool includes: The data obtained by convolving the data from the fifth branch is used as the input data for the first stacker. The stacker output data is obtained by performing RepConvN and convolution on the input data of the stacker; The output data of the stacker is fed back into the repeating stacker as the input data of the stacker; After taking the output data of N stackers and inputting it into the repeating stacker N times as the input data of the stacker, the stacker output data is obtained. Import the data exported by the stacker into the second Concat merge pool.

9. The efficient detection method for UAV remote sensing target detection according to claim 5 or 6, characterized in that, The step of performing local cross-channel interaction via the ECA modulation to recalibrate attention features includes: The ECA module is invoked to perform global average pooling on the input feature map; Perform a one-dimensional convolution on the data after global average pooling; the kernel size of the one-dimensional convolution is related to the number of channels; Attention weights are generated from the data after one-dimensional convolution using the Sigmoid function; The attention weights are multiplied channel by channel of the input feature map to perform attention feature recalibration.

10. A high-efficiency detection system for remote sensing target detection using unmanned aerial vehicles (UAVs), characterized in that, include: The ESPMStem module is configured to: extract a first feature map from the image to be processed, and then obtain a second feature map through spatial gating generation and feature response modulation; A feature fusion module, communicatively connected to the ESPMStem module, is configured to: sequentially perform two pre-modulation operations on the second feature map to obtain a shallow feature map, each pre-modulation operation including one downsampling and one CM-RepNCSPELAN4 feature enhancement process; perform downsampling and CM-RepNCSPELAN4 feature enhancement on the shallow feature map to obtain a middle-layer feature map; perform downsampling and CM-RepNCSPELAN4 feature enhancement on the middle-layer feature map to obtain a deep feature map; perform SPPF and C2PSA on the deep feature map to obtain a reference deep map; and then perform SPPF and C2PSA on the reference deep map. After upsampling, the image is concatted with the intermediate feature map to obtain a reference intermediate image; after upsampling the reference intermediate image, it is concatted with the shallow feature map, and then CM-RepNCSPELAN4 feature enhancement processing is performed to obtain a shallow output image; after downsampling the shallow output image, it is concatted with the reference intermediate image, and then CM-RepNCSPELAN4 feature enhancement processing is performed to obtain a intermediate output image; after downsampling the intermediate output image, it is concatted with the reference deep image, and then CM-RepNCSPELAN4 feature enhancement processing is performed to obtain a deep output image. The dynamic detection module is communicatively connected to the feature fusion module. The dynamic detection module is configured to import the shallow output map, the middle output map, and the deep output map into the dynamic detection head, perform scale perception, spatial perception, and task perception, and obtain the final decoded classification result, center point regression result, and bounding box regression result.

Images