Remote sensing small target detection network optimization method based on YOLOv11

Abstract

The invention provides a remote sensing small target detection network optimization method based on YOLOv11, and the method achieves the performance improvement through full-link collaborative optimization: preprocessing a remote sensing image data set, and constructing an improved YOLOv11 target detection network, which comprises the steps: adding a feature enhancement sub-branch, reconstructing a backbone network of an SPPF module, and strengthening the sparse feature extraction of a small target; the cross-scale attention fusion Neck network of the fusion channel and the space attention improves the feature transmission efficiency and inhibits the background noise; a double-branch structure and a Head network matched with an anchor frame are adopted, and the positioning precision is optimized; and a multi-loss mixed loss function is combined, and the small target learning priority is enhanced. And inputting the preprocessed data set into a network for training, and outputting optimized detection data through cooperative processing of each module. According to the method, missing detection and false detection are effectively reduced, high-precision and high-robustness detection of remote sensing small targets is realized, the anti-interference capability of complex scenes is improved, the actual requirements of low-altitude security and protection monitoring, emergency rescue investigation and the like are met, and the practicability and the popularization value are remarkable.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and specifically provides an optimization method for remote sensing small target detection networks based on YOLOv11. Background Technology

[0002] With the widespread application of aerial remote sensing technology in low-altitude security monitoring, emergency rescue reconnaissance, field resource patrol, and urban traffic control, the demand for accurate detection of small targets in remote sensing images is becoming increasingly urgent. Small targets in remote sensing images typically have characteristics such as low pixel ratio, sparse feature information, large scale differences, and susceptibility to interference from complex backgrounds. This places stringent requirements on the feature extraction capabilities, multi-scale adaptability, and background suppression capabilities of target detection algorithms.

[0003] As a mainstream target detection framework, YOLOv11 has been widely used in general target detection scenarios due to its high detection speed and balanced accuracy. However, when directly applied to the detection of small targets in remote sensing images, it still faces the following core technical challenges: First, small targets in aerial remote sensing images generally have a low pixel ratio, resulting in extremely sparse feature information. YOLOv11's original backbone network is primarily designed for medium to large targets in general scenarios, and its feature extraction modules (such as the C3k2 and SPPF modules) lack a targeted mechanism for capturing sparse features. This makes it difficult to effectively extract key features of small targets from complex remote sensing images, thus affecting the detection rate of small targets. Second, the multi-scale feature fusion mechanism used in YOLOv11's Neck network does not fully consider the scale characteristics and feature distribution patterns of small remote sensing targets. When fusing features at different scales, this mechanism easily introduces redundant information from the background region along with the small target features, causing the small target features to be submerged in noise. This not only reduces the effectiveness of feature fusion but also further affects the accuracy of subsequent detection heads in recognizing small targets. Finally, remote sensing images typically contain complex background elements such as terrain, vegetation, and buildings. Small targets and background textures and grayscale features are often highly similar. However, the YOLOv11 algorithm has a weak ability to suppress interference from complex backgrounds and lacks a dedicated background denoising and feature differentiation mechanism. This makes it easy for the model to confuse small targets with background features during training and inference, leading to missed detections and false detections, which makes it difficult to meet the high-precision application requirements for remote sensing small target detection.

[0004] Therefore, in view of the above-mentioned shortcomings of the YOLOv11 algorithm in the scenario of small target detection in remote sensing images, there is an urgent need for an improved scheme that can enhance small target feature extraction, optimize multi-scale feature fusion, and suppress complex background interference, so as to improve the accuracy and reliability of remote sensing small target detection. Summary of the Invention

[0005] In order to address the problems in the prior art, the present invention is proposed to provide solutions or partial solutions to the aforementioned problems.

[0006] In a first aspect, this invention provides an optimization method for a remote sensing small target detection network based on YOLOv11, comprising: preprocessing a remote sensing image dataset; constructing an improved target detection network based on YOLOv11, the network comprising: an improved backbone network: by adding a small target feature enhancement sub-branch to the C3k2 module of the original YOLOv11 backbone network and reconstructing the SPPF module of the original backbone network by pooling kernel combination, so as to achieve enhanced extraction of sparse features of small targets; and an optimized Neck network: by constructing a cross-scale attention fusion module, combining channel attention mechanism and spatial attention mechanism to perform multi-scale attention fusion. Scale features are processed to improve the efficiency of small target feature transfer and suppress background noise; the improved head network adopts a dual-branch structure for classification and localization, and configures anchor boxes adapted to remote sensing small targets to optimize the localization accuracy of small targets; the hybrid loss function for small target detection combines classification loss, localization loss and hard example mining loss to enhance the learning priority of small target samples in the model; the preprocessed dataset is input into the target detection network based on YOLOv11 for training, and the optimized data for remote sensing small target detection is output through the collaborative processing of the various components of the network.

[0007] In one technical solution of the above-mentioned remote sensing small target detection network optimization method based on YOLOv11, the preprocessing includes at least one of image resolution normalization, random cropping, scale transformation, and noise addition.

[0008] In one technical solution of the above-mentioned optimization method for remote sensing small target detection network based on YOLOv11, the small target feature enhancement sub-branch of the C3k2 module includes a convolutional layer, a depthwise separable convolutional layer, a normalization layer, and an activation function; the output features of the enhancement sub-branch are fused with the residual branch output features and ordinary branch output features of the original YOLOv11 C3k2 module through a weighted concatenation method.

[0009] In one technical solution of the above-mentioned optimization method for remote sensing small target detection network based on YOLOv11, the pooling kernel combination reconstruction of the SPPF module is specifically as follows: multi-scale combined pooling kernels are adopted, the output features of each pooling branch are spliced ​​after being uniformly processed by the number of channels, and then the spliced ​​features are filtered by attention gating based on feature variance.

[0010] In one technical solution of the above-mentioned optimization method for remote sensing small target detection network based on YOLOv11, the processing procedure of the cross-scale attention fusion module includes: performing channel alignment processing on the input feature maps of different scales; calculating channel attention weights through the Squeeze-and-Excitation mechanism, and weighting the features based on the channel attention weights; generating a spatial attention mask, and weighting the features based on the spatial attention mask; and fusing the weighted features using a bidirectional feature pyramid structure.

[0011] In one technical solution of the above-mentioned optimization method for remote sensing small target detection network based on YOLOv11, the improved Head network adopts a stacked convolutional layer structure for the classification branch, and a small target confidence calibration layer is set after the convolutional layer; the localization branch connects to a coordinate attention module after the convolutional layer, and then outputs the bounding box coordinates through the convolutional layer; the anchor box adapted to remote sensing small targets is generated by clustering the small target sizes of the publicly available remote sensing small target detection dataset.

[0012] In one technical solution of the above-mentioned optimization method for remote sensing small target detection network based on YOLOv11, the hybrid loss function includes a classification loss of weighted cross-entropy loss, which strengthens the model's learning of small target features by specifically configuring the weights of small target samples; a localization loss of improved CIoU loss, which configures a corresponding penalty coefficient for the size characteristics of small targets; and a hard example mining loss of Focal loss.

[0013] In one technical solution of the above-mentioned optimization method for remote sensing small target detection network based on YOLOv11, the method further includes a step of compressing the target detection network improved based on YOLOv11, wherein the compression includes at least one of structured pruning and knowledge distillation.

[0014] In one technical solution of the above-mentioned remote sensing small target detection network optimization method based on YOLOv11, the structured pruning specifically involves removing redundant convolutional kernels in the improved backbone network and Neck network, while retaining the convolutional kernels of the branches related to small target feature enhancement; the knowledge distillation uses the unpruned optimized network as the teacher model and the pruned network as the student model, and transfers knowledge related to small target detection through feature distillation and output distillation.

[0015] In one technical solution of the above-mentioned remote sensing small target detection network optimization method based on YOLOv11, the optimized data of the remote sensing small target detection includes the category information, bounding box coordinate information and confidence information of the small targets in the remote sensing image; the remote sensing small targets include small targets in low-altitude security monitoring, emergency rescue reconnaissance, field resource patrol and urban traffic control scenarios.

[0016] The beneficial effects of the remote sensing small target detection network optimization method based on YOLOv11 provided by this invention are as follows: This invention conducts a full-link collaborative design by enhancing the extraction of sparse features of small targets in the backbone network, optimizing the cross-scale feature fusion and background noise suppression in the Neck network, improving the positioning accuracy of the Head network, and prioritizing the small target learning in the hybrid loss function. This enables the improved network to accurately capture the key features of remote sensing small targets, efficiently transmit effective features and shield interference information, accurately locate the position of small targets, and focus on small target sample learning. This avoids the low detection rate problem caused by insufficient feature extraction of small targets in traditional YOLOv11, avoids feature confusion and accuracy reduction caused by inappropriate multi-scale fusion, avoids false negatives and false negatives caused by weak background interference suppression, and avoids the disadvantage of model bias towards medium and large targets caused by insufficient learning priority of small targets. It achieves high-precision and high-robustness detection of remote sensing small targets, improves cross-scale adaptability and anti-interference performance in complex scenarios, meets the stringent requirements of remote sensing small target detection in practical application scenarios such as low-altitude security monitoring and emergency rescue reconnaissance, and thus improves the practicality and promotion value of the technology. Attached Figure Description

[0017] The disclosure of this invention will become more readily understood with reference to the accompanying drawings. It will be readily understood by those skilled in the art that these drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. Furthermore, similar numbers in the drawings are used to denote similar components, wherein: Figure 1 This is a schematic diagram of a remote sensing small target detection network optimization method based on YOLOv11 according to an embodiment of the present invention. Detailed Implementation

[0018] Some embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0019] like Figure 1 As shown, this invention provides a method for optimizing remote sensing small target detection networks based on YOLOv11, including: Step S1: Preprocess the remote sensing image dataset; Step S2: Construct an object detection network based on an improved YOLOv11, the network comprising: Improved backbone network: By adding a small target feature enhancement sub-branch to the C3k2 module of the original YOLOv11 backbone network and reconstructing the SPPF module of the original backbone network by pooling kernel combination, the enhanced extraction of sparse features of small targets can be achieved. Optimized Neck Network: By constructing a cross-scale attention fusion module, combining channel attention mechanism and spatial attention mechanism to process multi-scale features, the efficiency of small target feature transfer and background noise suppression are improved. Improved Head Network: By adopting a dual-branch structure for classification and localization, and configuring anchor frames adapted to small remote sensing targets, the localization accuracy of small targets is optimized. A hybrid loss function adapted for small object detection: By combining classification loss, localization loss and hard example mining loss, the learning priority of small object samples in the model is enhanced; Step S3: Input the preprocessed dataset into the target detection network based on YOLOv11 for training. Through the collaborative processing of the various components of the network, the optimized data for remote sensing small target detection is output.

[0020] In this embodiment, the specific implementation of each improved module in step S2 can also be adjusted according to the actual application scenario. For example, the kernel size of the small target feature enhancement sub-branch, the attention weight allocation method of the cross-scale attention fusion module, and the number and size of anchor boxes adapted to remote sensing small targets can be reasonably designed by those skilled in the art based on the characteristics and detection requirements of remote sensing small targets.

[0021] The training process in step S3 can employ adaptive learning rate adjustment strategies, batch normalization, and other techniques to improve training effectiveness. The output optimized remote sensing small target detection data can include information such as target category, location, and confidence level to meet the needs of different remote sensing application scenarios.

[0022] In one embodiment, the preprocessing includes at least one of image resolution normalization, random cropping, scaling, and noise addition.

[0023] In one embodiment, the small target feature enhancement sub-branch of the C3k2 module includes a convolutional layer, a depthwise separable convolutional layer, a normalization layer, and an activation function; the output features of the enhancement sub-branch are fused with the residual branch output features of the original YOLOv11 C3k2 module (preserving the basic information of the input features and avoiding gradient vanishing) and the ordinary branch output features through a weighted concatenation method.

[0024] In this embodiment, the convolutional layer uses small-sized convolutional kernels (such as 1×1 or 3×3) to perform preliminary feature extraction on the input feature map, filtering out basic features related to small targets (such as minute contours and local textures), avoiding the "dilution" of small target features by large kernel convolution; the depthwise separable convolutional layer splits the standard convolution into "depthwise convolution + pointwise convolution", which significantly reduces the amount of computation while precisely capturing the channel features and spatial details of small targets (such as edge gradients and texture differences of small targets), solving the problem of sparse features and weak signals of small targets; the normalization layer standardizes the convolutional output features, stabilizes the feature distribution, avoids training instability caused by fluctuations in the numerical values ​​of small target features, and helps the model learn small target feature patterns more efficiently; the activation function introduces non-linear activation to enhance the model's ability to express complex features of small targets, such as distinguishing subtle feature differences between small targets and background noise.

[0025] In one embodiment, the pooling kernel combination reconstruction of the SPPF module specifically involves: using multi-scale combined pooling kernels, concatenating the output features of each pooling branch after unified processing of the number of channels, and then filtering the concatenated features through attention gating based on feature variance.

[0026] In this embodiment, the multi-scale combined pooling kernel includes pooling kernels of different sizes such as 1×1, 3×3, and 5×5. Two or more pooling kernel sizes can be selected to form parallel pooling branches based on the size distribution characteristics of remote sensing small targets. Channel number unification is achieved through a 1×1 convolutional layer, adjusting the number of channels in the output features of each pooling branch to a preset unified value (e.g., 64 channels, 128 channels). The feature variance-based attention gating filtering logic includes core steps such as variance calculation, weight allocation, and feature weighting, determining the priority of effective features by quantifying the variance values ​​of each channel's features. Of course, the specific implementation is not limited to the scenarios listed above. The size and number of multi-scale combined pooling kernels, the target value for unified channel number, and the variance threshold setting for attention gating can be flexibly adjusted and set by those skilled in the art according to the accuracy requirements and computational resource limitations of remote sensing small target detection.

[0027] In one embodiment, the processing steps of the cross-scale attention fusion module include: performing channel alignment processing on the input feature maps of different scales; calculating channel attention weights through a Squeeze-and-Excitation mechanism, and weighting the features based on the channel attention weights; generating a spatial attention mask, and weighting the features based on the spatial attention mask; and fusing the weighted features using a bidirectional feature pyramid structure.

[0028] In this embodiment, channel alignment is achieved through a 1×1 convolutional layer, unifying the number of channels in feature maps of different scales to a preset value (such as 64, 128, 256, etc.) to ensure consistent feature dimensions. The Squeeze-and-Excitation mechanism includes compression operations of global average pooling on the feature maps, activation operations implemented through fully connected layers and activation functions, and then outputting the attention weights of each channel. The weighting process involves multiplying the weights by the corresponding channel features one channel at a time. The spatial attention mask is generated by performing global average pooling or max pooling on the feature maps, followed by convolutional layers and a sigmoid activation function. The weighting process involves multiplying the mask by the feature maps pixel by pixel. The fusion of the bidirectional feature pyramid structure includes bottom-up detail feature transfer and top-down semantic feature transfer. The fusion method can be feature addition or channel concatenation. Of course, the specific implementation methods mentioned above are not limited to the situations listed above. The preset number of channels for channel alignment, the number of neurons in the fully connected layer in the Squeeze-and-Excitation mechanism, the generation method of the spatial attention mask, and the fusion strategy of the bidirectional feature pyramid can be flexibly adjusted and set by those skilled in the art according to the characteristics of remote sensing small targets and the requirements for detection accuracy.

[0029] In one embodiment, the improved Head network employs a stacked convolutional layer structure for the classification branch, with a small target confidence calibration layer set after the convolutional layers; the localization branch connects to a coordinate attention module after the convolutional layers, and then outputs bounding box coordinates through the convolutional layers; the anchor boxes adapted to remote sensing small targets are generated by clustering the small target sizes of publicly available remote sensing small target detection datasets.

[0030] In this embodiment, the convolutional layer stacking structure of the classification branch can adopt 3 or 5 concatenated convolutional layers. Each convolutional layer is followed by a batch normalization layer and an activation function to gradually extract the category features of small targets. The small target confidence calibration layer improves the accuracy of small target classification by introducing an adaptive threshold adjustment mechanism to correct the confidence of the initial prediction. The coordinate attention module of the localization branch strengthens the feature response of the target boundary region by assigning attention weights to the x-axis and y-axis of the feature map, helping the model to more accurately regress the bounding box coordinates. Subsequent convolutional layers are responsible for mapping the attention-weighted features to the final bounding box position parameters (such as center point coordinates, width and height). During the anchor box generation process, public datasets such as DOTA and HRSC2016 can be selected to statistically analyze the width and height distribution of small targets. The K-Means clustering algorithm is used to cluster targets with similar sizes into one class. The cluster center of each class is the width and height of an anchor box. The number of clusters can be set to 3 to 6 depending on the diversity of small target sizes. Of course, the specific implementation methods mentioned above are not limited to the situations listed above. The number of convolutional layers in the classification branch, the threshold adjustment strategy of the small target confidence calibration layer, the specific implementation form of the coordinate attention module, the dataset used for anchor box generation, and the number of clusters can be flexibly adjusted and set by those skilled in the art according to the detection needs of remote sensing small targets.

[0031] In one embodiment, the hybrid loss function includes a classification loss of weighted cross-entropy loss, which enhances the model's learning of small target features by specifically configuring the weights of small target samples; a localization loss of improved CIoU loss, which configures a corresponding penalty coefficient for the size characteristics of small targets; and a hard example mining loss of Focal loss.

[0032] In this embodiment, the weight configuration of the weighted cross-entropy loss focuses on the characteristics of the small target samples themselves. In view of the possible scarcity of categories and differences in clarity of remote sensing small targets, higher weights are assigned to scarce category small targets and blurry small targets (for example, the weight of scarce category small targets is set to 1.5~2.0, and the weight of regular category small targets is set to 1.0). The specific weights can be adaptively determined by the distribution statistics of small target categories and the sample quality classification. The core is to amplify the penalty for the classification error of key small targets, forcing the model to focus on capturing the subtle features of small targets (such as small textures and weak edge information). The improved CIoU loss's penalty coefficient configuration adapts to the size differences within small targets: even among small targets, the bounding box regression error of ultra-small targets (e.g., occupying only 5-10 pixels) has a greater impact on detection accuracy. Therefore, a larger penalty coefficient is configured for ultra-small targets (e.g., for targets with a pixel area of ​​5-10, the penalty coefficient is set to 1.3-1.5; for targets with a pixel area of ​​10-30, the penalty coefficient is set to 1.0-1.2). By strengthening the bounding box regression constraints of small targets of different sizes, the overall localization accuracy is improved (e.g., accurately fitting the edge contours of various small targets). The hard example mining loss adopts Focal loss, the core of which is to amplify the loss proportion for hard example samples among small targets (e.g., overlapping small targets, small targets with blurred edges, small targets occluded by the background). For example, when the model's classification probability for a certain overlapping small target is only 0.3 (hard to classify), the modulation factor (1-0.3) is used. 2 A modulus of 0.49 will significantly amplify the loss of this sample; while for clear, small targets (easy to classify) with a classification probability of 0.9, the modulus factor (1-0.9) is more effective. 2 A value of 0.01 will suppress the loss, thereby guiding the model to focus on feature learning for difficult examples of small targets. Of course, the specific implementation methods mentioned above are not limited to the cases listed above: for example, the weight values ​​of the weighted cross-entropy loss can be dynamically adjusted according to the class distribution of the small target dataset; the penalty coefficient of the improved CIoU loss can be finely tuned on the validation set to determine a more refined size-coefficient mapping relationship; the focusing parameter of the Focal loss can be flexibly set according to the proportion of difficult examples of small targets (e.g., 3 when the proportion of difficult examples is high, and 2 when the proportion is low). Those skilled in the art can flexibly adjust and set the parameters according to the specific detection requirements of remote sensing small targets (e.g., accuracy priority, sample characteristics).

[0033] In one embodiment, the method further includes a step of compressing the YOLOv11-based improved object detection network, the compression including at least one of structured pruning and knowledge distillation.

[0034] In one embodiment, the structured pruning specifically involves removing redundant convolutional kernels from the improved backbone network and Neck network, while retaining the convolutional kernels of branches related to small target feature enhancement; the knowledge distillation uses the unpruned optimized network as the teacher model and the pruned network as the student model, transferring knowledge related to small target detection through feature distillation and output distillation.

[0035] In this embodiment, redundant convolutional kernel identification in structured pruning can be achieved through methods such as L1 regularization and batch normalization layer weight analysis. The pruning ratio can be flexibly set according to the model compression target (e.g., compression rate of 50% or 70%). The core is to ensure that the convolutional kernels of key structures such as the small target feature enhancement sub-branch and the cross-scale attention fusion module are not mistakenly deleted. Feature distillation for knowledge distillation can use L2 loss or cosine similarity loss of intermediate layer feature mapping, and output distillation can be adjusted using a Softmax temperature coefficient (usually with a temperature value of 1 to 10). The soft label of the teacher model guides the student model to learn the fine feature patterns of small targets. Of course, the specific implementation methods are not limited to the situations listed above. The regularization strength of pruning, the combination of loss functions for knowledge distillation, and the value of the temperature coefficient can be flexibly adjusted and set by those skilled in the art according to the accuracy requirements and inference speed requirements of the compressed model.

[0036] In one embodiment, the optimized data for remote sensing small target detection includes category information, bounding box coordinate information, and confidence information of small targets in remote sensing images; the remote sensing small targets include small targets in scenarios such as low-altitude security monitoring, emergency rescue reconnaissance, field resource patrol, and urban traffic control.

[0037] In this embodiment, the category information refers to the specific type of small target identified by the model, such as small drones, wild animals, small road vehicles, trapped personnel, low-altitude aircraft, etc.; the bounding box coordinate information can be pixel coordinates (such as the x / y coordinates of the upper left corner and the x / y coordinates of the lower right corner) or geographic coordinates (coordinates converted from latitude and longitude) to accurately locate the position of the small target in the remote sensing image; the confidence information is the prediction reliability output by the model (value range 0~1), and the higher the confidence, the more reliable the target recognition result. In the scenarios corresponding to remote sensing small targets, small targets in low-altitude security monitoring scenarios also include suspicious floating objects, small illegal equipment, etc.; small targets in emergency rescue and reconnaissance scenarios also include rescue boats, temporary shelters, etc.; small targets in wild resource patrol scenarios also include rare plants and animals, illegal mining equipment, etc.; and small targets in urban traffic control scenarios also include electric bicycles, pedestrian gathering points, etc. Of course, the specific content mentioned above is not limited to the situations listed above. The optimized data may also include the size information and attitude information of the small target. The application scenarios of remote sensing small targets can also be extended to agricultural pest and disease monitoring, ecological environment observation, etc. Those skilled in the art can flexibly adjust and set according to actual application needs.

[0038] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the original technical features, and the technical solutions resulting from these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for optimizing a remote sensing small target detection network based on YOLOv11, characterized in that, include: Preprocess the remote sensing image dataset; Construct an object detection network based on an improved version of YOLOv11, the network comprising: Improved backbone network: By adding a small target feature enhancement sub-branch to the C3k2 module of the original YOLOv11 backbone network and reconstructing the SPPF module of the original backbone network by pooling kernel combination, the enhanced extraction of sparse features of small targets can be achieved. Optimized Neck Network: By constructing a cross-scale attention fusion module, combining channel attention mechanism and spatial attention mechanism to process multi-scale features, the efficiency of small target feature transfer and background noise suppression are improved. Improved Head Network: By adopting a dual-branch structure for classification and localization, and configuring anchor frames adapted to small remote sensing targets, the localization accuracy of small targets is optimized; A hybrid loss function adapted for small object detection: By combining classification loss, localization loss and hard example mining loss, the learning priority of small object samples in the model is enhanced; The preprocessed dataset is input into the target detection network based on YOLOv11 for training. Through the collaborative processing of the various components of the network, optimized data for remote sensing small target detection is output.

2. The method according to claim 1, characterized in that, The preprocessing includes at least one of image resolution normalization, random cropping, scale transformation, and noise addition.

3. The method according to claim 1, characterized in that, The small target feature enhancement sub-branch of the C3k2 module includes a convolutional layer, a depthwise separable convolutional layer, a normalization layer, and an activation function; the output features of the enhancement sub-branch are fused with the residual branch output features and ordinary branch output features of the original YOLOv11 C3k2 module through a weighted concatenation method.

4. The method according to claim 1, characterized in that, The specific reconstruction of the pooling kernel combination in the SPPF module is as follows: a multi-scale combined pooling kernel is adopted, the output features of each pooling branch are spliced ​​after being uniformly processed by the number of channels, and then the spliced ​​features are filtered by attention gating based on feature variance.

5. The method according to claim 1, characterized in that, The processing steps of the cross-scale attention fusion module include: channel alignment of the input feature maps at different scales; calculation of channel attention weights using a Squeeze-and-Excitation mechanism, and weighting of features based on the channel attention weights; generation of a spatial attention mask, and weighting of features based on the spatial attention mask; and fusion of the weighted features using a bidirectional feature pyramid structure.

6. The method according to claim 1, characterized in that, In the improved Head network, the classification branch adopts a stacked convolutional layer structure, and a small target confidence calibration layer is set after the convolutional layer; the localization branch connects to the coordinate attention module after the convolutional layer, and then outputs the bounding box coordinates through the convolutional layer; the anchor box adapted to remote sensing small targets is generated by clustering the small target sizes of the publicly available remote sensing small target detection dataset.

7. The method according to claim 1, characterized in that, In the hybrid loss function, the classification loss is a weighted cross-entropy loss, which strengthens the model's learning of small target features by specifically configuring the weights of small target samples; the localization loss is an improved CIoU loss, which configures a corresponding penalty coefficient for the size characteristics of small targets; and the hard example mining loss is a Focal loss.

8. The method according to claim 1, characterized in that, It also includes a step of compressing the YOLOv11-based improved object detection network, wherein the compression includes at least one of structured pruning and knowledge distillation.

9. The method according to claim 8, characterized in that, The structured pruning specifically involves removing redundant convolutional kernels from the improved backbone network and Neck network, while retaining the convolutional kernels of branches related to small target feature enhancement. The knowledge distillation uses the unpruned optimized network as the teacher model and the pruned network as the student model, transferring knowledge related to small target detection through feature distillation and output distillation.

10. The method according to claim 1, characterized in that, The optimized data for remote sensing small target detection includes category information, bounding box coordinate information, and confidence information of small targets in remote sensing images; the remote sensing small targets include small targets in low-altitude security monitoring, emergency rescue reconnaissance, field resource patrol, and urban traffic control scenarios.

Images