Unmanned aerial vehicle multi-view scene-oriented adaptive dynamic receptive field infrared target detection method and equipment

Abstract

The invention provides a self-adaptive dynamic receptive field infrared target detection method and device for a multi-view scene of an unmanned aerial vehicle. The method comprises the following steps: constructing a detection network, namely replacing a C3K2 module in yolov11 with a multi-expansion-rate convolution module MFDB, adding a multi-scale pooling attention module MPA behind a C2PSA module, and adding a multi-scale feature and channel modulation module MFCM at the tail end of an original neck network; and inputting a to-be-detected infrared image into the detection network to obtain a detection result, the detection result comprising the selected target, the target category and the corresponding probability. According to the method, the problem that scale drift and semantic inconsistency of the target in the feature space are caused by significant change of geometric scale, spatial attitude and thermal radiation distribution of the infrared target under different aerial photography visual angles can be solved.

Description

Technical Field

[0001] This invention relates to the field of infrared target detection technology, and in particular to an adaptive dynamic receptive field infrared target detection method and device for multi-view scenarios of unmanned aerial vehicles. Background Technology

[0002] With the rapid development of artificial intelligence, computer vision, and drone technology, drone-based target detection has shown broad application potential in fields such as intelligent transportation, public safety, disaster monitoring, and military reconnaissance. However, due to the complex and ever-changing imaging environment during drone aerial photography, target detection tasks still face many challenges.

[0003] Traditional visible light imaging relies on ambient light and can acquire relatively clear image information during the day or in good weather conditions. However, at night, in fog, haze, rain, snow, or other adverse weather conditions, visible light images are easily affected by insufficient light, shadow interference, and complex backgrounds, leading to a significant decrease in detection performance. In contrast, infrared imaging detects the thermal radiation distribution of targets, does not rely on external light sources, and has all-weather, all-time imaging capabilities. It can still provide stable observation information under low light, strong occlusion, and complex background conditions, thus becoming an important supplement and alternative for UAV target detection tasks. However, due to differences in flight altitude and shooting angle, targets may appear in various forms such as top-view, side-view, or oblique-view, and their geometric structure and thermal radiation distribution vary significantly from different perspectives. This perspective diversity leads to intra-class differences in target appearance that are much greater than in visible light scenes, and the scale variation of the same target is also large, exacerbating the difficulty of modeling feature alignment and semantic consistency. Therefore, how to construct a feature extraction mechanism with adaptive dynamic receptive field to adapt to scale changes and semantic drift under multi-view conditions of UAVs is a key research issue.

[0004] With the development of deep learning, object detection methods based on convolutional neural networks (CNNs) and visual Transformers have made breakthrough progress on large-scale visible light datasets. However, their designs are generally based on static receptive fields and high-texture visible light data, making it difficult to directly adapt to the low-texture and viewpoint variation characteristics of infrared scenes. Currently, some studies have enhanced the model's ability to detect targets at different scales by constructing multi-scale feature pyramid structures; or by introducing attention mechanisms to improve the model's feature representation ability from spatial and channel dimensions; and by combining infrared and visible light data to achieve cross-modal feature complementarity. However, multi-scale structures and complex attention often increase inference overhead, and cross-modal methods are limited by the availability of visible light data, making them difficult to apply under single infrared imaging conditions. Summary of the Invention

[0005] In view of the significant changes in the geometric scale, spatial attitude, and thermal radiation distribution of infrared targets under different aerial perspectives, which leads to scale drift and semantic inconsistency in the feature space, this invention proposes an adaptive dynamic receptive field infrared target detection method and device for UAV multi-view scenarios. This method enhances the feature representation capability of multi-view targets by multi-receptive field context modeling, thereby improving the detection effect of targets under different perspectives in infrared images.

[0006] In a first aspect, the present invention provides an adaptive dynamic receptive field infrared target detection method for multi-view scenarios of unmanned aerial vehicles (UAVs), comprising: The detection network is constructed by replacing the C3K2 module in yolov11 with the multi-dilation rate convolution module MFDB, adding a multi-scale pooling attention module MPA after the C2PSA module, and adding a multi-scale feature and channel modulation module MFCM at the end of the original neck network. The infrared image to be detected is input into the detection network to obtain the detection result, which includes the selected target, the target category, and the corresponding probability.

[0007] Furthermore, the feature map processing procedure of the MDFB module includes: Input feature map First pass To expand the channel dimension, a split operation is then used to divide the expanded feature map into two feature maps along the channel dimension. and ; Feature map First pass To expand the channel dimension, a three-branch dilated convolutional feature extraction module (TDF) is used to achieve information mixing of the token path. Then, the SE module adaptively recalibrates the fused feature map in the channel dimension, and then... Projection back and feature map With the same channel dimensions, the feature map is finally connected via residual connections. Weighted fusion is performed to obtain feature maps ; Feature map The feature map is obtained by first passing it through layer normalization and bottleneck MLP layers. Then, through residual connections and feature maps... Weighted fusion is performed to obtain feature maps ; Feature map , and First, stitch the components along the channel dimension, then... Obtain the input feature map .

[0008] Furthermore, the TDF includes sequentially connected multi-scale dilated convolutional blocks, adaptive weight fusion layers, and... and The function layer; wherein the multi-scale dilated convolutional block is used to process the input feature map using three parallel, depthwise separable convolutions with different dilation rates; the adaptive weight fusion layer is used to weight and fuse the three feature maps output by the multi-scale dilated convolutional block according to learnable weights.

[0009] Furthermore, the feature map processing procedure of the MPA module includes: Applying adaptive average pooling operations at multiple scales to the input feature map yields pooling results at multiple scales. The size is ;in, This indicates the batch size, and C represents the number of channels. Represents the spatial dimension of the feature map. This represents the pooling scale for the i-th adaptive average pooling operation; Pooling results at all scales are spatially compressed to... The channel compression vector is obtained. ; An adaptive weight fusion layer is used to weight and fuse all channel compressed vectors according to learnable weights to obtain the fused channel attention. ; The fused channel attention Channel attention weights are obtained through an attention generation network. ; Input feature map and channel attention weights Perform channel-by-channel multiplication to obtain the output feature map.

[0010] Furthermore, the feature map processing procedure of the MFCM module includes: The input feature map is first normalized by two-dimensional grouping, and then activated by Sigmoid to generate a learnable channel gated weight map. The learnable channel gated weight map is then multiplied element-wise with the input feature map to obtain the reweighted features. The reweighted features are processed by the Multi-Scale Spatial Awareness Module (MSAB) to enhance their ability to represent contextual information, resulting in enhanced features. ; Enhanced features The feature maps are divided into two sets along the channel dimension. One set of feature maps enters the semantic context enhancement branch (CEB) for full-channel feature fusion and enhancement, resulting in the final feature map. Another set of feature maps enters the detail supplementation branch (DRB) to model local details and low-response region features, resulting in feature maps. ; Enhanced features ,feature and The features are added together to obtain the output feature map.

[0011] Furthermore, the MSAB includes five parallel branches; three branches employ 3×3 dilated convolutions with different dilation rates, one branch employs 3×3 convolution and AvgPool operation, and the last branch employs 3×3 convolution and MaxPool operation; correspondingly, the feature map processing procedure of the MSAB includes: processing the input feature map in parallel using the five branches, and adding the outputs of the five branches to obtain the output feature map.

[0012] Furthermore, the process of feature fusion and enhancement across the entire channel range in the CEB includes: The input feature map first passes through Compression involves performing pointwise convolution and segmentation operations on the compressed features. The compressed features are then divided into two parts by a segmentation operation, and these two parts are then fed into independent... Parallel processing is performed, and the results of the two parallel processing paths are spliced ​​along the channel dimension. The compressed features are then added to the concatenated result after pointwise convolution to obtain the final features. ; feature After passing through the AvgPool operation and the Softmax function in sequence, the channel attention distribution is obtained. The input feature map is then weighted using this channel attention distribution to obtain the output feature map.

[0013] Furthermore, the process of modeling local details and low-response region features using DRB includes: The input feature map first passes through Compression is performed on the compressed features, followed by pointwise convolution and depthwise separable convolution. The features after both convolutions are then summed to obtain the final feature value. ; feature The concatenated features are then concatenated with the compressed features. The concatenated results are then processed by the AvgPool operation and the Softmax function to obtain the channel attention distribution. This channel attention distribution is then used to weight the input feature map to obtain the output feature map.

[0014] In a second aspect, the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method described in the first aspect.

[0015] Thirdly, the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method described in the first aspect.

[0016] The beneficial effects of this invention are as follows: (1) The adaptive dynamic receptive field infrared target detection network ADRF-Det proposed in this invention for multi-view scenarios of UAVs effectively improves the feature expression ability and anti-interference ability of targets with scale changes under multiple views. It can solve the problem of scale drift and semantic inconsistency of targets in feature space caused by significant changes in the geometric scale, spatial attitude and thermal radiation distribution of infrared targets under different aerial photography views.

[0017] (2) This invention proposes a multi-scale dilated convolution module (MDFB), which uses multi-scale dilated convolution to replace self-attention, thereby enhancing the perception capability at different scales and contexts. This design can effectively integrate local details and global scene information while maintaining high resolution, thereby alleviating feature instability in cross-view detection and improving detection robustness.

[0018] (3) The present invention designs a multi-scale pooling attention module (MPA). To address the problem that traditional channel attention relies on single-scale global pooling and is difficult to adapt to the semantic differences of targets under different receptive fields, a multi-scale compression and weighted fusion mechanism is constructed to achieve scale-sensitive channel recalibration, thereby enhancing the selectivity of infrared image features and the robustness of cross-view detection.

[0019] (4) The present invention constructs a multi-scale feature and channel modulation module (MFCM), and through the multi-scale spatial perception module (MSAB) and the differentiated modeling of semantic context enhancement branch and detail supplementation branch, multi-dimensional semantic feature fusion is realized, which effectively alleviates the problem of insufficient feature expression of weakly salient targets in UAV multi-view aerial photography scenarios.

[0020] (5) The proposed method was validated on different public infrared image datasets. The experimental results show that the ADRF-Det model outperforms other advanced methods in overall performance and can detect targets in infrared images from different shooting angles more accurately. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the structure of the Adaptive Dynamic Receptive Field Infrared Target Detection Network (ADRF-Det) for multi-view scenarios of unmanned aerial vehicles provided in an embodiment of the present invention. Figure 2 The MDFB structure provided in the embodiments of the present invention; Figure 3 This is a structural diagram of the MPA module provided in an embodiment of the present invention; Figure 4 This is a structural diagram of the MFCM module provided in an embodiment of the present invention; Figure 5 The comparison results of the HIT dataset provided in the embodiments of the present invention; Figure 6 Visualization comparison results of the M3FD dataset provided in the embodiments of the present invention; Figure 7 Comparison results of the DroneVehicle dataset provided in the embodiments of the present invention Figure 8 This is a structural block diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0023] The core objective of this invention is to improve the model's ability to extract features from targets of different scales in infrared images of UAVs in multi-view scenarios, while effectively reducing the loss of feature information, so as to achieve more accurate and efficient detection of small targets in infrared images. To achieve this objective, this invention designs an adaptive dynamic receptive field infrared target detection network ADRF-Det for UAV multi-view scenarios. Figure 1 The overall structure of ADRF-Det is demonstrated. This network uses YOLOv11 as its basic framework and improves upon it. Specifically, the C3K2 module in YOLOv11 is replaced with a multi-dilation rate convolutional module (MFDB), a multi-scale pooling attention module (MPA) is added after the C2PSA module, and a multi-scale feature and channel modulation module (MFCM) is added at the end of the original neck network. The infrared image to be detected is input into the ADRF-Det network to obtain the detection results, which include the selected target, the target category, and the corresponding probability.

[0024] In infrared weak target detection tasks, the feature extraction capability of the backbone network directly determines the detection performance. Most current mainstream detection backbones (such as CSPDarknet53) are constructed using standard convolution and residual structures. While they achieve good results in general target detection, their convolution operators are limited by a fixed receptive field, making it difficult to capture the multi-scale differential features of infrared weak targets in complex backgrounds. On the other hand, Transformer methods achieve powerful global modeling capabilities through self-attention mechanisms, but their computational complexity increases quadratically with the input scale, making them difficult to directly apply to infrared devices with extremely high real-time requirements. Therefore, this invention proposes a multi-dilation rate convolutional module (MDFB) as the backbone unit for infrared target detection. It achieves efficient spatial-channel feature fusion through residual paths and SE Attention, aiming to enhance the network's multi-scale feature extraction capability while maintaining lightweight design.

[0025] Furthermore, in the multi-scale feature fusion stage, shallow features can retain rich spatial details, but their semantic expression is limited; deep features, although containing stronger semantic information, are insufficient in the spatial localization of small targets due to their lower resolution. Traditional convolutional neck structures (such as FPN / PAN) often struggle to balance detail preservation and high-level semantic fusion in infrared small target detection, leading to the easy masking of small target signals. Therefore, this embodiment of the invention also introduces MDFB into the neck structure to fully utilize its multi-receptive field and channel adaptive capabilities, thereby improving the fusion effect of cross-scale information.

[0026] In summary, in this detection network, MDFB can enhance the model's ability to perceive different scales and contexts. By using multi-scale dilated convolution to replace self-attention, it can effectively integrate local details and global scene information while maintaining high resolution, thereby alleviating feature instability in cross-view detection and improving detection robustness.

[0027] Furthermore, in image scenarios with complex backgrounds or frequent multi-scale targets, traditional channel attention mechanisms typically employ single-scale global average pooling to model inter-channel dependencies, making it difficult to adapt to the semantic response differences of targets in different receptive fields. This static, single-scale channel modeling approach easily leads to a decline in the generalization ability of feature representation, especially when facing challenges such as small target blurring and strong background interference in infrared images. Therefore, to address the problem that traditional channel attention relies on single-scale global pooling and cannot adapt to the semantic differences of targets in different receptive fields, this invention proposes a multi-scale pooling attention module (MPA). This module aims to extract scale-sensitive channel attention information in different receptive fields and achieve adaptive recalibration of cross-scale channel representations through a fusion mechanism, thereby enhancing the selectivity of infrared image features and the robustness of cross-view detection. The core idea of ​​MPA is to spatially compress the input feature map at multiple scales, capture the channel response characteristics in multiple receptive fields, model them through a shared channel attention network, and finally use learnable weights for inter-scale fusion, improving the flexibility and expressiveness of channel selection.

[0028] Finally, in order to improve the discrimination ability of infrared image target detection models under complex background interference, and at the same time enhance the modeling effect of multi-scale features, weak semantic information and channel imbalance response, this invention proposes a multi-scale feature and channel modulation module (MFCM). This module achieves multi-dimensional semantic feature fusion through differentiated modeling of multi-scale spatial perception module (MSAB) and semantic context enhancement branch (CEB) and detail supplementation branch (DRB), effectively alleviating the problem of insufficient feature representation of weakly salient targets in UAV multi-view aerial photography scenes.

[0029] In one embodiment, the structure of the MDFB provided by this invention is as follows: Figure 2 As shown. Specifically, the feature map processing of MDFB mainly includes three fusion stages: Phase 1: First, the input feature map First pass Extended channel dimension is , The number of channels output by MDFB ensures that each half has sufficient representational capacity during subsequent splitting. The feature map of one half after splitting is directly output via residual connection, while the feature map of the other half is input into subsequent modules for processing. First, it is processed through... Extended channel dimension is And introduce multiple separable convolutional branches with different dilation rates and hole depths (such as...) Figure 2 The three-branch dilated convolution feature extraction module (TDF) in the model is shown in the figure, thereby realizing the information mixing of the token path.

[0030] The contributions of different branches may vary significantly across different task scenarios. To improve robustness, a learnable weighted fusion mechanism is adopted to dynamically adjust the output contribution of each branch. All branch outputs are weighted and fused according to learnable weights, and then... Compressed to the original channel dimension This mechanism allows the network to flexibly adjust branch contributions under different tasks and inputs, making it more adaptive than a fixed mixer.

[0031] in, This represents the learnable weights corresponding to the output of the i-th hole-depth separable convolution branch.

[0032] Subsequently, the fused features are adaptively recalibrated in the channel dimension using an activation function and the SE Attention module to further enhance the spatial and channel feature representation, suppress redundant background features, highlight potential target responses, and further improve detection sensitivity.

[0033] Then through Projecting back to the original channel dimension and then weighting and fusing the input feature maps through residual connections yields the final result. in, These are learnable weights.

[0034] The second stage includes a residual path and a feedforward structure (including LayerNorm and bottleneck MLP transformation); the residual path passes through... Scaling is used for feedforward output integration.

[0035] In the second stage, a bottleneck approach is adopted. This invention replaces the standard MLP, significantly reducing the number of parameters and computational overhead; it also introduces GeLU activation to enhance the nonlinear representation capability between channels. To prevent training instability, this embodiment of the invention introduces residual paths and learnable scaling factors. This ensures the gradient remains stable during deep propagation, guaranteeing the network's convergence against complex infrared backgrounds. First, LayerNorm is applied to the output after the first stage. A simplified MLP is constructed, with GeLU activation introduced in the middle to enhance the nonlinear expression between channels, and finally output through residual connection.

[0036] The third stage: Concatenate the features output from the second stage with the two parts initially split along the channel dimension to obtain... And then through a Obtain output features .

[0037] In one embodiment, the structure of the MPA module is as follows: Figure 3 As shown. The process of feature map processing in MPA is as follows: Let the input feature map be ,in For batch size, For the number of channels, and These represent the spatial dimensions of the feature map. First, for the input feature map... Multiple adaptive average pooling operations at different scales are applied to construct feature compression paths under multiple receptive fields. Let the set of pooling scales be... Each of them This represents the spatial size of the pooling kernel. At each scale, MPA first processes the input... Adaptive average pooling is performed to obtain the downsampled feature map. Its size is .

[0038] It should be noted that, in Figure 3 In this example, n=3 is used as the only example.

[0039] Subsequently, considering that channel attention requires a unified dimensional representation for fusion, all scale pooling results are spatially compressed again. This yields the channel compression vector: By employing a multi-scale pooling followed by unified compression, the global channel description at different spatial scales can be effectively captured. First, multi-scale pooling provides context awareness across different spatial receptive fields, helping to capture the salience of a target within a large background or small region. Second, unified compression is only used to standardize the output dimensions, facilitating subsequent fusion operations. The advantage of this design is that each... This actually represents a global description of the channels captured at different spatial scales, and their information distribution varies significantly. Stacking the channel descriptions from all scales yields: To integrate this scale information, MPA introduces a learnable fusion vector. Each scale is assigned an attention weight. This fusion vector is softmax normalized during the forward pass to ensure the sum of the weights for each scale is 1, thus achieving weighted fusion of attention. The weighted fusion operation is performed through a broadcast mechanism, and the output is the fused channel attention representation. Its size is This representation integrates channel information from different receptive fields, exhibiting stronger scale adaptability.

[0040] Merged channel representation Input a two-layer channel attention generation network to predict the final channel-dimensional attention weights. This attention generation network uses a compression-activation-expansion structure consisting of two 1×1 convolutional layers. The first layer is used for dimensionality reduction, and the second layer is used to restore the number of channels. A ReLU activation function is inserted in between, and a Sigmoid activation function is appended at the end to generate the final channel attention weights. .

[0041] in, ReLU represents the activation function. This represents the Sigmoid function.

[0042] Finally, the attention weight It is expanded to the same dimension as the input feature map and with Multiplication completes channel-level feature recalibration: Where ⊙ represents multiplication by channel.

[0043] In one embodiment, the overall structure of the MFCM module is as follows: Figure 4 As shown, it mainly includes a multi-scale space-aware block (MSAB) and a semantic dual-branch enhancement module (CEB and Detail-RefineBranch, DRB).

[0044] The algorithm flow of this module is shown in Algorithm 1. The input features are first normalized by two-dimensional group batch normalization (GroupBatchNorm2d). These normalized weights are further normalized to the [0,1] interval after Sigmoid activation, and their sum is uniformly adjusted to 1, forming a learnable channel-gated weight map. This attention map is then multiplied element-wise with the original input features to achieve explicit channel reweighting, thereby suppressing redundant responses and highlighting key semantics. Subsequently, the reweighted features are fed into the Multi-Scale Spatial Awareness Module (MSAB) to enhance its contextual information representation ability. The MSAB module integrates multiple features with different hole rates ( The convolutional branch of the module is used to capture multi-scale spatial semantics from local textures to mid-to-long-range structures. Simultaneously, two pooling branches, AvgPool2d and MaxPool2d, are introduced to supplement global context awareness and local edge detail modeling, thereby achieving joint enhancement of spatial awareness and channel representation. This module significantly enhances the network's ability to perceive infrared target boundaries, background interference regions, and weak-response targets in the early stages of feature representation. To further refine the distribution representation of channel semantics, the output features of MSAB are divided into two groups along the channel dimension, which are respectively sent to the Semantic Context Enhancement Branch (CEB) and the Detail Enhancement Branch (DRB). The division ratio is adjusted by a hyperparameter. The invention is set up as follows. .

[0045] The Multi-Scale Spatial Awareness Module (MSAB), as the core perceptual structure of this module, is designed to mine spatial semantics under different receptive fields through multi-path convolutional combinations. Firstly, for the input features... The network is processed through five sub-branches: three 3×3 dilated convolutional branches with different dilation rates (dilation = 1, 3, 5); a global perception path containing AvgPool; and an edge enhancement path containing MaxPool. The outputs of the five branches are fused along the channel dimension and uniformly adjusted using a non-linear activation function, ultimately outputting a fused feature that combines spatial awareness and semantic complementarity. This module significantly improves the network's response to targets at different scales, exhibiting stronger robustness, especially in infrared-weak targets and complex background scenes. Finally, the features from all branches are summed and fused to obtain the output feature map. The above processing procedure can be represented as follows: in The convolution kernel corresponding to the a-th branch, with the superscript b indicating a kernel size of b×b and the superscript d indicating the kernel dilation rate; for example This indicates that the first branch corresponds to a convolution kernel with a kernel size of 3×3 and an inflation rate of d=1. (·) indicates ReLU activation. For global average pooling, For global max pooling, For batch normalization.

[0046] The fused features output by the Multi-Scale Space-Aware Block (MSAB) are split into two parts, which are fed into the Semantic Context Enhancement Branch (CEB) and the Detail Supplement Branch (DRB) respectively for differential modeling. The design motivation is to address the problem of insufficient fusion of semantic features across different channel dimensions.

[0047] Semantic Context Enhancement Branch (CEB): This branch aims to perform full-channel feature fusion and enhancement on the input features. First, for the input features... Perform 1×1 convolution ( Compression is performed to reduce computational cost while preserving key semantic components. The compressed features are then divided into two parts and fed into two independent 3×3 convolutional branches for parallel processing, capturing spatial patterns and contextual features from different subspaces. The two results are concatenated along the channel dimension and added to a Point-wise Convolution (PWC) output to obtain the output features. This enables full-channel interaction and multi-scale context fusion.

[0048] Then, AvgPool is used to aggregate global channel information, and Softmax is used to calculate the channel attention distribution, which is finally compared with the input features. Weighted average .

[0049] Detail Supplement Branch (DRB): This branch focuses more on modeling local details and low-response region features. It is structurally symmetrical to CEB and provides support for input features. 1×1 convolution compression is performed, and Depthwise Convolution is used to highlight intra-channel spatial details. Point-wise Convolution is used to restore inter-channel interactions, thereby enhancing the expressive power of low-response regions and finally outputting features. .

[0050] Then, AvgPool is used to aggregate global channel information, and Softmax is used to calculate the channel attention distribution, which is finally compared with the input features. Weighting: The final output of the MFCM module: To verify the effectiveness of the present invention, the present invention also provides the following experimental data.

[0051] In this experiment, the method of this invention was mainly implemented based on PyTorch 2.8.0 and CUDA 12.8 frameworks. The SGD optimizer was used for network training, with a momentum of 0.937. The input size of the training images was the original size of the dataset. The network was trained on a single NVIDIA RTX 5060ti (16GB) GPU with a batch size of 16, an initial learning rate of 0.01, a weight_decay of 0.0004, and 200 epochs.

[0052] To verify the effectiveness of this invention, we conducted systematic comparative experiments on the HIT-UAV dataset against several mainstream detection models, including the classic Faster-RCNN, the YOLO series models (YOLOv9m, YOLOv10n, YOLOv11n, YOLOv12n), and the Transformer-based RT-DETR series. We also compared it with other proposed models such as PHSI-RTDETR, AIMED-Net, and YOLO-TSL. The comparison results are shown in Table 1.

[0053] Table 1 Comparison results on the HIT dataset As shown in Table 1, in terms of overall performance, this invention achieves an mAP@0.5 score of 0.887, which is significantly better than the lightweight YOLOv10n (0.832) and YOLOv12n (0.842), and also surpasses YOLOv11n (0.846). It also demonstrates significant competitiveness compared to the larger-scale YOLOv9m (0.865) and RT-DETR-l (0.849). Notably, at the more stringent mAP@0.5:0.95 metric, this invention achieves a score of 0.584, higher than YOLOv10n (0.546), YOLOv11n (0.561), and YOLOv12n (0.543), indicating that the proposed method maintains stable detection performance at different IoU thresholds, demonstrating strong generalization ability. In terms of precision and recall (P, R), this invention achieves a precision of 0.911 and a recall of 0.812, demonstrating a relatively balanced detection performance. Compared to YOLOv9m (P=0.897, R=0.826), this invention further improves precision while maintaining a similar recall; compared to PHSI-RTDETR (R=0.761), the recall of this invention is significantly higher, avoiding a large number of missed detections. Regarding model complexity, this invention requires only 7.61M parameters, far lower than large-scale models such as RT-DETR-L (31.99M), RT-DETR-R50 (41.94M), and AIMED-Net (34.2M), and also only one-third the complexity of YOLOv9m. However, while maintaining a lightweight design, this invention still achieves comprehensive superiority in detection accuracy, verifying its excellent balance between precision and efficiency. This advantage indicates that the invention is not only applicable to high-performance computing platforms, but also has the potential to be deployed on resource-constrained devices such as drone payloads.

[0054] To more intuitively demonstrate the improvement in detection accuracy of this model, some representative scenes were selected for detection, such as... Figure 5 As shown in the diagram, a comparison of the visualization results clearly demonstrates that this invention performs more robustly in multi-target, crowded scenes, and low-contrast environments. Other methods generally suffer from missed detections (marked by yellow circles) and false detections (marked by red circles), while this invention can more accurately focus on the infrared target area, significantly reducing detection errors and ensuring the reliability of the detection results.

[0055] To more intuitively demonstrate the effectiveness of our improved method, each improvement was discussed in detail through ablation experiments. The results are shown in Table 2. The baseline model was YOLOv11n.

[0056] Table 2 Ablation experimental results on the HIT-UAV dataset The analysis is as follows: (1) Validity of the MDFB module After introducing MDFB, the validation set mAP@0.5 improved from 0.868 to 0.878, an increase of 1%; the test set mAP@0.5 improved from 0.846 to 0.875, an increase of 2.9%, which is the most significant improvement among single modules. At the same time, the recall rate improved significantly. This indicates that this module enhances feature discrimination ability while also taking into account detection coverage.

[0057] (2) Effectiveness of the MPA module After introducing the MPA module, the validation set mAP@0.5 increased from 0.868 to 0.873, an increase of 0.5%; the test set increased from 0.846 to 0.862, an increase of 1.6%.

[0058] (3) MFCM module effectiveness After introducing the MFCM module, the validation set mAP@0.5 improved to 0.871, a 0.3% improvement compared to the baseline; the test set mAP@0.5 was 0.862, an improvement of 1.6%. Although the gain at mAP@0.5 was not as significant as MDFB, the improvement in mAP@0.5:0.95 on the test set was more pronounced, increasing from 0.561 to 0.576 (+1.5%), indicating that the module enhanced the model's localization accuracy and robustness at high IoU thresholds through multi-frequency context modeling.

[0059] (4) Validity of ADRF-Det When all three modules were incorporated into the model simultaneously, optimal overall performance was achieved. On the validation set, mAP@0.5 reached 0.894, a 2.6% improvement over the baseline; mAP@0.5:0.95 improved to 0.602, an increase of 1.8%, and P-value improved to 0.91, a 2.9% improvement, both the highest values ​​among all combinations. On the test set, mAP@0.5 improved to 0.887 (+4.1%), slightly lower than the 0.890 of the MDFB and MFCM combination, but improved to 0.584 (+2.3%) at mAP@0.5:0.95, while the accuracy P-value reached 0.911 (+4.1%), both the highest among all combinations. This phenomenon indicates that the combination of the three modules can effectively reduce the false detection rate, thereby significantly improving accuracy; it also performs better at high IoU thresholds, verifying the model's enhancement in target localization accuracy and robustness. In contrast, the combination of MDFB and MFCM is more inclined to improve overall recall, while the three-module combination achieves a better trade-off between recall and precision.

[0060] The above ablation experiments verified the effectiveness and necessity of each module in the proposed ADRF-Det, providing an important basis for further optimization and improvement of the model.

[0061] To further verify the generalization ability of the proposed model, experiments were conducted on the M3FD and DroneVehicle datasets, and comparisons were made with several mainstream detection models. The experimental results are shown in Tables 3 and 4. Representative scenes from the M3FD and DroneVehicle datasets were also selected for visualization, as shown below. Figure 6 and Figure 7 As shown in the figure. Experimental results demonstrate that the proposed method exhibits good generalization ability on both the M3FD and DroneVehicle datasets. On the M3FD dataset, the method achieves the best detection performance in major categories such as Person, Car, and Truck, with an overall mAP of 0.712 (mAP 0.5), representing improvements of 1.4% and 1.8% compared to YOLOv11n and YOLOv12n, respectively, and a 2.1% improvement over RT-DETR-L, while requiring only 7.61M parameters, significantly lower than the RT-DETR series and other models. On the DroneVehicle dataset, the method maintains high accuracy in the Car, Truck, and Bus categories, with an overall mAP of 0.551 (mAP 0.5-0.95), representing a 1.6% improvement over YOLOv11n, and an mAP of 0.761 (mAP 0.5), representing a 1.3% improvement over YOLOv11n. These results indicate that the present invention can achieve superior performance to existing detection models on different scenarios and datasets, possessing strong generalization ability and application potential.

[0062] Table 3 Comparative experiments on the M3FD dataset Table 4 Comparative experiments on the DroneVehicle dataset Figure 8 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 8As shown, the electronic device may include: a processor 801, a communication interface 802, a memory 803, and a communication bus 804. The processor 801, communication interface 802, and memory 803 communicate with each other via the communication bus 804. The processor 801 can call logical instructions in the memory 803 to execute an adaptive dynamic receptive field infrared target detection method for multi-view scenarios of unmanned aerial vehicles (UAVs). This method includes: constructing a detection network, including: replacing the C3K2 module in YOLOv11 with a multi-dilation rate convolutional module (MFDB), adding a multi-scale pooling attention module (MPA) after the C2PSA module, and adding a multi-scale feature and channel modulation module (MFCM) at the end of the original neck network; inputting the infrared image to be detected into the detection network to obtain detection results, the detection results including the selected target, the target category, and the corresponding probability.

[0063] Furthermore, when the logical instructions in the aforementioned memory 803 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0064] This invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions, and when the program instructions are executed by a computer, the computer can execute an adaptive dynamic receptive field infrared target detection method for UAV multi-view scenarios provided in the above-described method embodiments.

[0065] This invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the adaptive dynamic receptive field infrared target detection method for UAV multi-view scenarios provided in the above-described method embodiments.

[0066] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0067] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An adaptive dynamic receptive field infrared target detection method for multi-view scenarios of unmanned aerial vehicles (UAVs), characterized in that, include: The detection network is constructed by replacing the C3K2 module in yolov11 with the multi-dilation rate convolution module MFDB, adding a multi-scale pooling attention module MPA after the C2PSA module, and adding a multi-scale feature and channel modulation module MFCM at the end of the original neck network. The infrared image to be detected is input into the detection network to obtain the detection result, which includes the selected target, the target category, and the corresponding probability.

2. The adaptive dynamic receptive field infrared target detection method for multi-view scenarios of unmanned aerial vehicles according to claim 1, characterized in that, The feature map processing procedure of the MDFB module includes: Input feature map First pass To expand the channel dimension, a split operation is then used to divide the expanded feature map into two feature maps along the channel dimension. and ; Feature map First pass To expand the channel dimension, a three-branch dilated convolutional feature extraction module (TDF) is used to achieve information mixing of the token path. Then, the SE module adaptively recalibrates the fused feature map in the channel dimension, and then... Projection back and feature map With the same channel dimensions, the feature map is finally connected via residual connections. Weighted fusion is performed to obtain feature maps ; Feature map The feature map is obtained by first passing it through layer normalization and bottleneck MLP layers. Then, through residual connections and feature maps... Weighted fusion is performed to obtain feature maps ; Feature map , and First, stitch the components along the channel dimension, then... Obtain the input feature map .

3. The adaptive dynamic receptive field infrared target detection method for multi-view scenarios of unmanned aerial vehicles (UAVs) according to claim 2, characterized in that, The TDF includes sequentially connected multi-scale dilated convolutional blocks, adaptive weight fusion layers, and... and The function layer; wherein the multi-scale dilated convolutional block is used to process the input feature map using three parallel, depthwise separable convolutions with different dilation rates; the adaptive weight fusion layer is used to weight and fuse the three feature maps output by the multi-scale dilated convolutional block according to learnable weights.

4. The adaptive dynamic receptive field infrared target detection method for multi-view scenarios of unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, The feature map processing procedure of the MPA module includes: Applying adaptive average pooling operations at multiple scales to the input feature map yields pooling results at multiple scales. The size is ;in, This indicates the batch size, and C represents the number of channels. Represents the spatial dimension of the feature map. This represents the pooling scale for the i-th adaptive average pooling operation; Pooling results at all scales are spatially compressed to... The channel compression vector is obtained. ; An adaptive weight fusion layer is used to weight and fuse all channel compressed vectors according to learnable weights to obtain the fused channel attention. ; The fused channel attention Channel attention weights are obtained through an attention generation network. ; Input feature map and channel attention weights Perform channel-by-channel multiplication to obtain the output feature map.

5. The adaptive dynamic receptive field infrared target detection method for multi-view scenarios of unmanned aerial vehicles according to claim 1, characterized in that, The feature map processing procedure of the MFCM module includes: The input feature map is first normalized by two-dimensional grouping, and then activated by Sigmoid to generate a learnable channel gated weight map. The learnable channel gated weight map is then multiplied element-wise with the input feature map to obtain the reweighted features. The reweighted features are processed by the Multi-Scale Spatial Awareness Module (MSAB) to enhance their ability to represent contextual information, resulting in enhanced features. ; Enhanced features The feature maps are divided into two sets along the channel dimension. One set of feature maps enters the semantic context enhancement branch (CEB) for full-channel feature fusion and enhancement, resulting in the final feature map. Another set of feature maps enters the detail supplementation branch (DRB) to model local details and low-response region features, resulting in feature maps. ; Enhanced features ,feature and The features are added together to obtain the output feature map.

6. The adaptive dynamic receptive field infrared target detection method for multi-view scenarios of unmanned aerial vehicles according to claim 5, characterized in that, The MSAB comprises five parallel branches; three branches employ 3×3 dilated convolutions with different dilation rates, one branch employs 3×3 convolution and AvgPool operation, and the last branch employs 3×3 convolution and MaxPool operation; correspondingly, the feature map processing procedure of the MSAB includes: processing the input feature map in parallel using the five branches, and adding the outputs of the five branches to obtain the output feature map.

7. The adaptive dynamic receptive field infrared target detection method for multi-view scenarios of unmanned aerial vehicles according to claim 5, characterized in that, The process of feature fusion and enhancement across the entire channel range in the CEB includes: The input feature map first passes through Compression involves performing pointwise convolution and segmentation operations on the compressed features. The compressed features are divided into two parts by a segmentation operation, and these two parts are then fed into independent... Parallel processing is performed, and the results of the two parallel processing paths are spliced ​​along the channel dimension. The compressed features are then added to the concatenated result after pointwise convolution to obtain the final features. ; feature After passing through the AvgPool operation and the Softmax function in sequence, the channel attention distribution is obtained. The input feature map is then weighted using this channel attention distribution to obtain the output feature map.

8. The adaptive dynamic receptive field infrared target detection method for multi-view scenarios of unmanned aerial vehicles according to claim 5, characterized in that, The process of modeling local details and low-response region features using DRB includes: The input feature map first passes through Compression is performed on the compressed features, followed by pointwise convolution and depthwise separable convolution. The features after both convolutions are then summed to obtain the final feature value. ; feature The concatenated features are then processed by the AvgPool operation and the Softmax function to obtain the channel attention distribution. This channel attention distribution is then used to weight the input feature map to obtain the output feature map.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 7.

Images