Unmanned aerial vehicle infrared small target detection method and device based on artificial intelligence technology

Abstract

The invention provides an unmanned aerial vehicle infrared small target detection method and device based on an artificial intelligence technology. The technical key points are as follows: obtaining target feature data in a target scene; performing feature processing from an initial convolutional layer, optimizing input features through channel adjustment and spatial dimensionality reduction, and then dividing a feature map into a plurality of independent regions; processing the plurality of independent regions through multiple stages, capturing cross-region dependence by using a multi-scale attention mechanism, and deepening feature expression through residual connection; and integrating the features output in multiple stages, performing channel fine tuning and output adaptation through a final convolutional layer, and obtaining processed target feature data, thereby improving the small target detection precision and optimizing the calculation efficiency.

Description

Technical Field

[0001] This application relates to the field of computer vision technology, and in particular to a method and apparatus for detecting small infrared targets on unmanned aerial vehicles based on artificial intelligence technology. Background Technology

[0002] The rapid development of unmanned aerial vehicle (UAV) technology has enabled it to demonstrate broad application prospects in key areas such as military reconnaissance, geographic information systems, environmental monitoring, disaster assessment, and urban planning. Infrared small target detection, as a key technology in UAV vision systems, supports functions such as target tracking, area monitoring, and emergency response by identifying targets with weak thermal signals.

[0003] However, this task faces multiple challenges: small target size, weak thermal radiation signal, complex background noise, significant scale variations, and stringent requirements for real-time performance and computational efficiency from the UAV platform. These factors result in insufficient performance of traditional target detection algorithms in infrared scenes, where target features are easily obscured by complex backgrounds.

[0004] In recent years, deep learning-based real-time object detection frameworks have offered new solutions to these challenges. RT-DETR (Real-Time Detection Transformer), with its end-to-end Transformer architecture and efficient decoding mechanism, has performed well in general object detection tasks. However, for the low signal-to-noise ratio and complex background characteristics of UAV infrared small target detection, the feature extraction, interaction, and fusion modules of RT-DETR urgently need optimization.

[0005] Research in the field of infrared small target detection has made some progress in recent years. Traditional methods, such as Scale Invariant Feature Transform (SIFT), Histogram of Oriented Gradients (HOG), and Deformable Part Model (DPM), rely on manual feature extraction, which is computationally complex and sensitive to scale changes, making it difficult to efficiently process infrared small targets. Moreover, existing research mostly focuses on the optimization of single scenes, making it difficult to comprehensively address the needs of multi-scale feature extraction, robustness to complex backgrounds, and efficient feature fusion in infrared small target detection.

[0006] Therefore, there is an urgent need for a solution to address the problems of insufficient feature representation, difficulty in adapting to scale changes, and low efficiency of multi-scale feature fusion in RT-DETR for small target detection in complex environments. Summary of the Invention

[0007] The embodiments of this application provide a method and apparatus for detecting small targets in infrared from a drone based on artificial intelligence technology, which at least have the technical effects of improving the accuracy of small target detection and optimizing computational efficiency.

[0008] To address the aforementioned technical problems, according to one aspect of this application, a method for detecting small infrared targets on a drone based on artificial intelligence technology is provided, comprising:

[0009] Acquire target feature data in the target scene;

[0010] The features are processed from the initial convolutional layer, combined with spatial compression to optimize the features and reduce redundancy, and then split into query, key and value components, and divided into regions corresponding to the target signal and background noise according to the symbol.

[0011] By processing the multiple independent regions in multiple stages, a multi-scale attention mechanism is used to capture cross-regional dependencies, and feature representation is deepened through residual connections.

[0012] The feature data of the multi-stage output is integrated to achieve scale alignment. Attention weights are generated by global average pooling and multilayer perceptron to suppress noise. Then, higher-order associations are strengthened by hypergraph calculation. Convolutional layers are used for channel fine-tuning and output adaptation to obtain the processed target feature data.

[0013] According to another aspect of this application, a drone infrared small target detection device based on artificial intelligence technology is also claimed, comprising:

[0014] The acquisition module is configured to acquire target feature data images in the target scene and perform preprocessing.

[0015] The polarization space-channel feature interaction module is configured to process features from the initial convolutional layer, combine spatial compression to optimize features and reduce redundancy, and split them into query, key and value components, and divide them into regions corresponding to the target signal and background noise according to the symbol.

[0016] The multi-scale attention feature enhancement module is configured to process the multiple independent regions in multiple stages, capture cross-regional dependencies using a multi-scale attention mechanism, and deepen feature representation through residual connections.

[0017] The multi-feature fusion module is configured to integrate features from multiple stages to achieve scale alignment. It uses global average pooling and multilayer perceptron to generate attention weights to suppress noise. Then, it uses hypergraph calculation to strengthen higher-order associations. The convolutional layer performs channel fine-tuning and output adaptation to obtain the processed target feature data.

[0018] The module is configured to use processed target feature data for target tracking, area monitoring, or emergency response.

[0019] According to another aspect of this application, a computer-readable storage medium is also claimed, the computer-readable storage medium including a stored program wherein, when the program is executed, it controls the device on which the computer-readable storage medium is located to perform any of the methods described above.

[0020] According to another aspect of this application, an electronic device is also claimed, comprising: one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs including methods for performing any of the methods described above.

[0021] This application has the following beneficial effects:

[0022] This application demonstrates stronger generalization ability when processing small targets in UAV infrared images. It maintains high detection accuracy while achieving lightweight deployment, making it more competitive in practical applications and providing new insights for research and practice in the field of UAV infrared image target detection. The improved model provides more accurate target region localization, a more concentrated and stronger thermal response, and significantly reduces attention to irrelevant background areas; while the baseline model's attention to target areas is relatively scattered, with a more pronounced background noise response, resulting in a higher risk of missed or false detections. The above heatmap analysis results clearly show that the improved method proposed in this invention can effectively enhance the model's feature capture ability for small target regions, and the visualization verifies the significant advantages of this method in improving detection accuracy and suppressing background interference. The model proposed in this invention can achieve accurate, stable, and fast small target detection under conditions such as complex backgrounds, low thermal contrast, and multi-target interference. The visualization results verify the robustness and effectiveness of the method in actual infrared image scenarios. Attached Figure Description

[0023] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0024] Figure 1 A flowchart illustrating a method for detecting small infrared targets from a drone based on artificial intelligence technology, according to an embodiment of this application, is shown.

[0025] Figure 2 A structural block diagram of an infrared small target detection method device for unmanned aerial vehicles based on artificial intelligence technology is shown according to an embodiment of this application.

[0026] Figure 3 This is a schematic diagram of the multi-scale attention feature enhancement module proposed in this invention;

[0027] Figure 4 This is a schematic diagram of the polarization space-channel feature interaction module proposed in this invention;

[0028] Figure 5This is a schematic diagram of the hypergraph-driven multi-feature fusion module proposed in this invention;

[0029] Figure 6 This is a schematic diagram comparing thermograms from an experiment of the present invention;

[0030] Figure 7 This is a schematic diagram showing the comparison of test results in an experiment of the present invention. Detailed Implementation

[0031] As introduced in the background section, the introduction of deep learning technology has significantly improved detection performance. Algorithms based on convolutional neural networks, such as Faster R-CNN, RetinaNet, and the YOLO series, have achieved effective recognition of targets at different scales by automatically learning multi-level features. The introduction of the Transformer model has further promoted the development of object detection. DETR simplifies the detection process by removing anchors and non-maximum suppression (NMS) through ensemble prediction. However, the high computational complexity and long training time of DETR limit its real-time application. RT-DETR, by optimizing the Transformer structure, balances accuracy and speed, becoming an ideal choice for real-time detection.

[0032] To address the challenges of feature extraction and insufficient detection accuracy in small target detection in UAV infrared images due to the small target size, low signal-to-noise ratio, and interference from complex backgrounds such as clouds and thermal noise, this application proposes an improved method. Embodiments of this application provide a UAV infrared small target detection method, apparatus, computer, readable storage medium, and electronic device based on artificial intelligence technology.

[0033] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0034] Figure 1 This is a flowchart illustrating a method for detecting small infrared targets from a drone based on artificial intelligence technology, according to an embodiment of this application. Figure 1 As shown, the method includes the following steps:

[0035] Step S101: Obtain target feature data in the target scene;

[0036] Step S102: Process features from the initial convolutional layer, combine spatial compression to optimize features and reduce redundancy, split into query, key and value components, and divide them into regions corresponding to the target signal and background noise according to the symbol.

[0037] Step S103: The multiple independent regions are processed in multiple stages, a multi-scale attention mechanism is used to capture cross-regional dependencies, and feature representation is deepened through residual connections.

[0038] Step S104: Integrate the features of multi-stage output to achieve scale alignment, use global average pooling and multilayer perceptron to generate attention weights to suppress noise, then use hypergraph calculation to strengthen higher-order associations, and use convolutional layers to fine-tune channels and adapt outputs to obtain the processed target feature data.

[0039] Furthermore, this application also provides an apparatus for detecting small infrared targets in unmanned aerial vehicles (UAVs) based on artificial intelligence (AI) technology. It should be noted that this apparatus can be used to execute the UAV infrared small target detection method based on AI technology provided in this application. This apparatus is used to implement the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the apparatus described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0040] The following describes the method and apparatus for detecting small infrared targets of unmanned aerial vehicles based on artificial intelligence technology provided in the embodiments of this application.

[0041] Figure 2 This is a structural block diagram of an infrared small target detection device for unmanned aerial vehicles based on artificial intelligence technology, according to an embodiment of this application.

[0042] like Figure 2 As shown, the device includes:

[0043] The acquisition module is configured to acquire target feature data images in the target scene and perform preprocessing.

[0044] The polarization space-channel feature interaction module is configured to process features from the initial convolutional layer, combine spatial compression to optimize features and reduce redundancy, and split them into query, key and value components, and divide them into regions corresponding to the target signal and background noise according to the symbol.

[0045] The multi-scale attention feature enhancement module is configured to process the multiple independent regions in multiple stages, capture cross-regional dependencies using a multi-scale attention mechanism, and deepen feature representation through residual connections.

[0046] The multi-feature fusion module is configured to integrate features from multiple stages to achieve scale alignment. It uses global average pooling and multilayer perceptron to generate attention weights to suppress noise. Then, it uses hypergraph calculation to strengthen higher-order associations. The convolutional layer performs channel fine-tuning and output adaptation to obtain the processed target feature data.

[0047] The module is configured to use processed target feature data for target tracking, area monitoring, or emergency response.

[0048] The aforementioned UAV infrared small target detection method and apparatus based on artificial intelligence technology includes a processor and a memory. The acquisition module, polarization space-channel feature interaction module, multi-feature fusion module, multi-scale attention feature enhancement module, and determination module are all stored as program units in the memory. The processor executes these program units stored in the memory to achieve the corresponding functions. All of the above modules are located in the same processor; alternatively, the modules may be located in different processors in any combination.

[0049] In one embodiment, reference Figures 1-3 In the RT-DETR-R18 feature extraction framework, the performance of small target detection is limited by the volatile effect of weak features in complex background noise. This invention proposes a multi-scale attention feature enhancement module (MAFE). This module accurately enhances weak features, highlights the thermal radiation signal of small targets, and suppresses background interference through a dynamic multi-scale attention mechanism, thus solving the problem of weak feature overwhelming at the feature mining level.

[0050] MAFE structure as follows Figure 3 As shown in (a), feature processing begins with the initial convolutional layer. Input features are optimized through channel adjustment and spatial dimensionality reduction. The feature map is then divided into multiple independent regions, providing a foundation for subsequent enhancement stages. The FAT-Block unit processes these regions in multiple stages, employing a multi-scale attention mechanism to capture cross-region dependencies and deepening feature representation through residual connections. Finally, the features from the multi-stage outputs are integrated and fine-tuned through the final convolutional layer to ensure compatibility with subsequent tasks. Figure 3 This enhanced process is clearly demonstrated, highlighting the synergistic effect of region processing and feature fusion.

[0051] The FAT-Block module structure is as follows: Figure 3As shown in (b), this module integrates three key mechanisms: Frequency Decomposition Attention (FDWA), Frequency Modulation Feedforward Network (FMFFN), and Residual Fusion, forming a complete link for preprocessing, enhancement, and output. FDWA borrows from multi-window attention (LL-WA, HH-WA, etc.) designs to adapt to the multi-scale characteristics of small targets and accurately capture long-distance dependencies. FMFFN modulates the decomposed frequency components, effectively suppressing background redundancy. Residual Fusion preserves the original feature information, optimizes gradient propagation, and strengthens the correlation of small target features, thereby ensuring information integrity. These three mechanisms work together to construct a hierarchical collaborative enhancement system, significantly improving detection performance.

[0052] As the core unit of MAFE, FAT-Block first takes input features X∈R as input. C×H×W (Where H×W is the spatial resolution and C is the number of channels), first, the redundant channels are compressed using Conv1×1, nonlinear activation is used to enhance the nonlinear expression of features, and the thermal radiation profile of small targets is focused by region division. Then, the data is converted to the frequency domain by fast Fourier transform, and a learnable filter is introduced to initially suppress background noise.

[0053]

[0054] In the formula, GELU(·): applies the GELU activation function to the convolution result, BlockPartitioning(·): divides the activated result into blocks, FFT(·): performs a Fast Fourier Transform on the block-partitioned result, W⊙·: multiplies the FFT result element-wise with the weight matrix W, X freq This represents the frequency domain enhancement features obtained after a series of processing steps.

[0055] Secondly, the FAT-Block module relies on Frequency Decomposition Attention (FDWA) plus Frequency Modulation Feedforward Network (FMFFN) to enhance features in stages. FDWA adopts a multi-window attention strategy, dividing the features into four multi-scale windows: LL-WA, HH-WA, HL-WA, and LH-WA. The sizes (height × width) of these four windows are 4s × 4s, s × s, s × 4s, and 4s × s, respectively, where s is the basic window size. Long-distance dependencies are accurately captured through a process of Head Split to multi-head attention and then Concat. Specifically, after the features are linearly projected to K heads, they are evenly divided into 4 groups according to the number of heads K, that is, each group has K / 4 heads, corresponding to the four types of window attention. The processing flow is as shown in formula (2):

[0056]

[0057] Without loss of generality, taking LH-WA as an example, for instance: LH-WA. X is uniformly divided into non-overlapping windows [X1, ..., X...].M , where each window contains 4s × s tokens. Assume that the projected queries, keys, and values of the K-th head (3 / 4K < k ≤ K) are dk-dimensional tensors, then the LH-WA output of the K-th head is as follows:

[0058]

[0059] where, X i ∈R C×4s×s , M = H × W / 4s × s, i = 1, ···, M, W k Q 、W k K 、W k V ∈R C×dk are the projection matrices of the query, key, and value of the k-th head respectively, d k = C / k.

[0060] The processes of the other three types of window attention are also derived as in formula (3), and their outputs for the k-th head are denoted as LL-WAk(X), HH-WAk(X), HL-WAk(X) respectively. Finally, formula (4) is used to concatenate the outputs of these four parallel groups to form the overall output.

[0061] FDWA(X) = Concat[head1, ···, head K W O (4) After the FDWA output stabilizes the feature distribution through LayerNorm, formula (5) is used to further strengthen the feature correlation and suppress background redundancy, and then the original features are fused through residual connection to retain the original information and optimize the gradient transmission. On this basis, according to formula (6), the enhanced frequency-domain features are converted back to the spatial domain through inverse fast Fourier transform, and after integrating multi-region information, the channel dimension is fine-tuned through Conv1×1 to adapt to the input of the subsequent detection head. This operation not only retains the sharpness and clarity of the edge features, but also enhances the model's perception ability of the overall shape and contour of the target object, effectively improving the model's ability to capture the edge detail information of the target object in the UAV infrared image.

[0062]

[0063] FMFFN(·) represents applying inverse fast Fourier transform to the frequency-domain features, FDWA(·) represents the frequency modulation feed-forward network, BlockMerging(·) represents integrating multi-region spatial-domain features, Conv1×1 represents adjusting the channel dimension through 1×1 convolution, and X out represents the fine-tuned features, adapting to the input of the detection head.

[0064] In the neck module of RT-DETR, the accuracy of feature interaction directly affects the quality of transmitting weak features of small targets. While the original AIFI module performs feature processing based on the Transformer encoder layer, its ability to distinguish between positive and negative sign dependencies and multi-scale spatial-channel associations is limited, leading to blurred coupling of thermal radiation signals and noise in complex backgrounds, thus affecting subsequent detection. To address this issue, this invention proposes a Polarization Spatial-Channel Feature Interaction Module (Pola-SEFN) to replace AIFI, employing a dual-path design of "symbol-aware attention + gated multi-scale enhancement" to optimize the transmission of small target features in the neck area. Its structure is as follows: Figure 4 As shown.

[0065] Pola-SEFN replaces traditional multi-head self-attention with polarized linear attention (PolaAttention) to capture feature symbolic dependencies; and optimizes the feedforward network with deep feature enhancement network (SEFN) to strengthen multi-scale spatial channel fusion.

[0066] like Figure 4 As shown, after initial convolutional adjustment of the input features, the PolaAttention branch decomposes positive and negative dependencies, the SEFN branch enhances local details, and finally, residual connections and layer normalization output optimize features to suit neck requirements. PolaAttention is the core component in the Pola_SEFN module responsible for fine-grained interactions, specifically designed to solve the problem that weak features of small targets are easily drowned out by background noise. Its workflow is as follows: After the input features are split into query Q, key K, and value V, K and Q are first symbolically split, and K is divided into positive components K0 and K1. + The corresponding positively correlated signals and negative component K, such as thermal radiation from small targets. - For random negatively correlated signals such as background noise, Q-synchronous decomposition is performed on Q-mode. + Q - V is then assigned to the branch with the same sign. s and the opposite branch V o Then proceed to the two-branch attention calculation: the same-signed branch [K] + ,K + ] and [Q + Q + ]、[K - ,K - ] and [Q - Q - ] and opposite sign branches [K + ,K - ] and [Q + Q -All features first undergo a (d:.) transformation, i.e., dimension adaptation and feature fine-tuning, to ensure that different features can interact within a unified dimension. Then, a linear attention Matmul operation is used to strengthen the self-association of small targets by branches with the same sign, and to distinguish small targets from background noise by branches with opposite signs. At the same time, linear attention reduces the complexity from O(N) to O(N). 2 The complexity is reduced to O(N) (N = H × W) to meet real-time detection requirements. Finally, the dual-branch output is gated by G. s G o The contribution is dynamically adjusted, and after concatenation and reshaping to restore the spatial dimensions, the output features containing symbolic interaction information are obtained, providing more accurate small target feature support for subsequent detection. This symbolic dependency interaction process can be formally represented as follows:

[0067] A=s+Dropout(PolaAttn(s)) (7)

[0068] In the formula, s represents the input features, PolaAttn(s) performs symbolic dependency decomposition and linear attention calculation, Dropout() is used for regularization, and residual connections s+· ensure that the original feature information is not lost. After this operation, the features are stably distributed through layer normalization, as shown in formula (8):

[0069] N = LayerNorm(A) (8)

[0070] In the formula, A is the output of formula (7) (features enhanced by PolaAttention and residual connection); LayerNorm(·): normalizes the channel dimension of A so that the mean of each channel approaches 0 and the variance approaches 1.

[0071] Layer normalization targets the feature channel dimension, standardizing the mean and variance of each channel to a reasonable range, eliminating feature distribution shifts caused by symbolic dependency decomposition and attention calculation, creating a stable input environment for subsequent multi-scale enhancement of the SEFN branch, and enabling SEFN to focus more on the semantic features of small targets rather than being disturbed by numerical fluctuations.

[0072] SEFN, as a spatial channel enhancement unit, employs conventional Conv3×3 convolution and depthwise separable convolution DWConv3×3 to extract multi-scale features. Upsampling aligns the scales, and gated activation (GELU+gate) dynamically adjusts the fusion weights. GELU, with its smooth nonlinear characteristics, injects nonlinear transformation into feature fusion, enhancing the model's expressive power. The gate mechanism intelligently allocates feature weights at different scales based on feature importance, effectively suppressing background noise and highlighting small target features. Finally, a residual fusion strategy fuses the enhanced features with the original input features, preserving the basic information of the original features while overlaying the multi-scale enhancement results, thus strengthening feature discriminative power.

[0073] E = SEFN(N,A) (9)

[0074] In the formula, SEFN(·) integrates multi-scale convolution and gating mechanisms to enhance features, while residual connections ensure the retention of basic features. Finally, the output features are optimized through a second residual connection and layer normalization:

[0075] O=LayerNorm(S+Dropout(E)) (10)

[0076] Formula (10) again uses residual connection to preserve the original feature information in depth. Combined with layer normalization to stabilize the feature distribution, the output features contain key information of small targets captured by symbolic dependence and enhanced by multiple scales, and have a stable and regular distribution shape, providing high-quality input for subsequent neck detection processes.

[0077] In the neck module of RT-DETR, the quality of feature fusion directly determines the accuracy of small target detection. Traditional methods rely on simple feature concatenation, which suffers from core defects such as cross-scale semantic fragmentation, missing higher-order dependencies, and fixed fusion weights. This leads to small target features being easily submerged by noise in complex backgrounds, and difficulty in capturing multi-body associations (such as dense small target occlusion). To overcome these limitations, this invention proposes a Hyper-MFM hypergraph-driven multi-feature fusion module. Through a three-stage process of "multi-scale semantic collection, dynamic weight fusion, and higher-order association enhancement," it reconstructs the neck feature fusion logic, providing the detection head with more discriminative multi-scale input. Its structure is as follows: Figure 5 As shown.

[0078] The Hyper-MFM module reconstructs the feature fusion process through a multi-stage, sequential design. For example... Figure 5 As shown in (a), the module's functional mechanism consists of three steps. First, for the multi-scale features P2 / P3 / P4 / P5 output by the backbone network, differential pooling (P2 is averaged by 8×8, P3 by 4×4, and P4 by 2×2) and upsampling (P5 is upsampled after adjusting channels via 1×1 convolution) are used to achieve scale alignment between shallow high-resolution details and deep strong semantic information. The above process can be formally expressed as:

[0079]

[0080] Avgpool(·) represents pooling the features, P align These are the features after scale alignment.

[0081] This operation aligns the small target detail features (P2 / P3) with the global semantic features (P4 / P5) in terms of spatial resolution and channel dimension, solving the problem of small target feature overwhelmed by cross-scale semantic fragmentation in traditional methods, and providing multi-scale feature input for subsequent hypergraph convolution.

[0082] Secondly, to adapt to the dynamic changes in background interference in UAV infrared scenarios, the module introduces a modulation fusion (MFM) module to achieve dynamic weighted fusion. Aligned multi-scale features (such as P2 / P3 pooling results and P5 upsampling results) are first unified by a 1×1 convolution to unify the channel dimensions; subsequently, global average pooling (GAP) is performed on the preprocessed features to compress spatial information, and attention weights A are generated through multilayer perceptron (MLP) mapping and Softmax normalization. r,c This weight reflects the importance of features in the small target detection task; finally, through element-wise multiplication operation ☉, high weights are assigned to small target features to suppress background noise (such as cloud interference). The dynamic weighted fusion process can be formally expressed as:

[0083] F fusion =A r,c ΘF shallow +A r,c ΘF deep (12)

[0084] In the formula, F shallow For shallow detail features, F deep This is for deep semantic features. This mechanism enables the module to adaptively adapt to complex background interference and enhance the feature representation of small targets.

[0085] Furthermore, higher-order association enhancement is achieved by constructing a dynamic association graph between features based on feature similarity thresholds using the HyperComputeModule, capturing the higher-order dependencies of "small target-background-semantics". This process can be formally expressed as:

[0086]

[0087] In the formula, Nv(e) represents the set of vertices contained in hyperedge e, and |Nv(e)| is the number of vertices contained in hyperedge e, realizing the aggregation of "multi-vertex features - hyperedge features"; Ne(v) represents the set of hyperedges containing vertex v, and |Ne(v)| is the number of hyperedges associated with vertex v, realizing the update of "hyperedge features - vertex features" and strengthening the association and discrimination between small targets and background.

[0088] To reduce computational complexity, the matrix form of hypergraph convolution is defined as:

[0089] HyperConv(X,H)=X+D v -1 HD e -1 H T XΘ (14)

[0090] Where X is the vertex feature matrix, i.e., the features after multi-scale fusion, H is the hypergraph correlation matrix, and D...v and D e are the degree matrices of the vertices and hyperedges, respectively, and are learnable parameters.

[0091] Finally, the multi-branch enhancement network MANet is used to achieve direct transformation of features and local detail enhancement through RepC3 module and 3×3 convolution. After upsampling and secondary MFM fusion, multi-path feature iterative optimization is completed, which significantly improves the discriminativeness of small target features.

[0092] F output =MANet(Hyper(X,H))+F fusion (15)

[0093] To comprehensively evaluate the effectiveness of the model in infrared small target detection for UAVs, five metrics were selected to assess the model performance: precision, recall, average precision (mAP), number of parameters, and frames per second (FPS).

[0094] The formulas for calculating Precision, Recall, and mAP are as follows:

[0095]

[0096] Where TP represents the number of positive samples that are correctly detected, i.e., true positives; FP represents the number of false samples that are correctly detected; FN represents the number of true samples that are falsely detected; and N represents the number of categories.

[0097] The above experiments verified the beneficial effects of the present invention, and the experimental results are as follows:

[0098] Table 1 Ablation Experiment

[0099]

[0100] Table 2 Comparison of Mainstream Models

[0101]

[0102] Table 3 shows the experimental results for the M3FD dataset.

[0103]

[0104] This embodiment uses the HIT-UAV dataset for detailed ablation experiments. Table 1 shows the ablation experiment results using the RT-DETR-r18 model as the baseline model. Experiment A shows the experimental results of RT-DETR, and Experiment B shows the baseline experimental results of RT-DETR-r18. Experiments B and C show that adding the MAFE and Pola-SEFN modules to the backbone network enhances the model's feature extraction and symbolic dependency capture capabilities. Experiment B shows that MAFE significantly improves P and significantly reduces the number of parameters. Experiment C shows that Pola-SEFN further improves P, with a slight increase in the number of parameters. Experiment D shows that the improvement to the Hyper-MFM module effectively enhances the multi-feature fusion capability and improves detection accuracy. Experiment D shows that Hyper-MFM improves P, with a slight increase in the number of parameters. Experiment E combines MAFE and Pola-SEFN, improving P and reducing the number of parameters. Experiment F shows that improving the neck network by adding Pola-SEFN and Hyper-MFM modules enhances the model's ability to capture small targets, with a slight increase in the number of parameters. Experiments G and H further validated the effectiveness of the module combination. Model H, integrating all improvements, achieved a maximum P-value of 92.4% and a significant improvement in mAP0.5 to 83.7%. Compared to the RT-DETR baseline, the improved model had an average parameter count reduction of approximately 6.5% in model H, enhancing deployability. Ablation experiments demonstrated that the improved method enhanced the model's learning ability while maintaining strong practicality. When using an RTX4090 for FPS testing, experimental results showed a slight decrease in detection speed for the improved model.

[0105] In summary, the ablation experiments demonstrate that while the improved MPH-DETR model experiences a decrease in detection speed, it delivers a significant improvement in accuracy (e.g., P increases to 92.4%) and optimization of parameters. These improvements significantly enhance the performance of UAV image target detection tasks, making it more suitable for complex real-world application scenarios.

[0106] To verify the detection performance of the improved RT-DETR algorithm model compared to other mainstream algorithms, parameter count, precision, recall, and mean precision of various metrics were used as evaluation indicators. Comparative experiments were conducted on the HIT-UAV dataset with mainstream algorithms YOLOv5s, YOLOv5l, YOLOv8s, YOLOv8l, YOLOv10s, YOLOv10l, YOLOv11s, YOLOv11l, RT-DETR-R34, RT-DETR-R50, and RT-DETR-R18. The performance comparison results of different network models are shown in Table 2.

[0107] The experimental results are shown in Table 2. Under the condition that the number of model parameters and computational complexity are basically the same, the MPH-DETR algorithm performs well in key performance indicators such as accuracy, recall, and mAP@0.5, with particularly outstanding overall performance. Specifically, in terms of accuracy, the MPH-DETR algorithm achieves an accuracy of 92.4%, which is significantly higher than RT-DETR-R34 (88.6%) and RT-DETR-R50 (85.1%), which have larger parameter numbers and higher computational complexity. This fully demonstrates that it can still achieve high-precision detection results with smaller model size and lower computational complexity. In terms of recall, MPH-DETR reaches 78.8%, which is better than YOLOv8m (75.1%) and YOLOv5m (75.2%), indicating that it can effectively reduce false negatives in detection tasks. In terms of mean precision (mAP@0.5), MPH-DETR achieves 83.7%, surpassing mainstream detection algorithms such as YOLOv8m (82.8%) and YOLOv5l (82.2%), demonstrating excellent overall performance. Specifically, MPH-DETR outperforms YOLOv8m (87.5% precision, 75.1% recall) in both detection precision (92.4%) and recall (78.8%), and its parameter count is only 18.8MB, which is not only significantly lower than large models such as YOLOv5l (53.1MB) and YOLOv8l (43.7MB), but also about 27.4% lower than YOLOv8m (25.9MB). This synergistic advantage enables MPH-DETR to maintain strong generalization ability for complex scenes when processing small targets in UAV infrared images of varying difficulty levels (such as smoke interference and scale-changing scenes), while also adapting to the computing power constraints of UAV platforms, thus demonstrating comprehensive performance that is more in line with practical application needs.

[0108] MPH-DETR demonstrates stronger generalization ability when processing small targets in UAV infrared images. Overall, the MPH-DETR model maintains high detection accuracy while achieving lightweight deployment, making it more competitive in practical applications and providing new insights for research and practice in the field of UAV infrared image target detection.

[0109] To evaluate the detection performance and generalization ability of the improved algorithm proposed in this invention on different small target datasets, the generalization experiment results of this invention on the M3FD thermal imaging dataset are presented. As shown in Table 3, the experiments demonstrate that the improved algorithm is also effective on other datasets. It can be clearly observed that the accuracy P of the improved MPH-DETR model is improved by 0.5%, P is improved by 1.2%, the mAP@0.5 index is improved by 1.9%, and the mAP@0.5-0.95 index is improved by 0.9%, proving the effectiveness of the improved method.

[0110] To more intuitively demonstrate the feature focusing capability of the method of this invention for key targets in UAV infrared small target detection tasks, we conducted a visual comparative analysis of the baseline model and the improved model using heatmaps. The results are as follows: Figure 6 As shown in the figure, the improved model is more accurate in locating the target area, with a more concentrated and stronger thermal response, significantly reducing its focus on irrelevant background areas. In contrast, the baseline model's focus on the target area is relatively scattered, with a more pronounced background noise response, resulting in a higher risk of missed or false detections. The above heatmap analysis results clearly demonstrate that the improved method proposed in this invention can effectively enhance the model's feature capture ability for small target areas, and the visualization verifies the significant advantages of this method in improving detection accuracy and suppressing background interference.

[0111] To further verify the model performance, this invention selected infrared images from multiple scenes in the dataset for testing and compared the detection effects of the baseline model and the method of this invention. Some visualization examples of the detection results before and after the improvement are provided, as shown in the following figures. Figure 7 As shown.

[0112] from Figure 7 As can be seen, the proposed model can accurately identify small infrared targets in various scenarios, with clear target localization boundaries and concentrated response areas, demonstrating good robustness and generalization ability. In contrast, the baseline model is prone to missed detections, false detections, and low confidence levels under complex backgrounds and low contrast conditions, as indicated by the red box in the figure. The results show that the model proposed in this invention can achieve accurate, stable, and fast small target detection under conditions such as complex backgrounds, low thermal contrast, and multi-target interference. The visualization results verify the robustness and effectiveness of the proposed method in real-world infrared image scenarios.

[0113] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application 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. Such 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 this application.

Claims

1. A method for detecting an infrared small target of a UAV based on artificial intelligence technology, characterized in that, Comprise: Acquiring target feature data in a target scene; Processing features from an initial convolutional layer, combining spatial compression to optimize features and reduce redundancy, splitting into query, key, and value components, and dividing into regions corresponding to target signals and background noise according to symbols; Processing the multiple independent regions through multiple stages, capturing cross-region dependencies using a multi-scale attention mechanism, and deepening feature expression through residual connections; Integrating the features output by the multiple stages to achieve scale alignment, generating attention weights using global average pooling and a multi-layer perceptron to suppress noise, and then strengthening high-order correlations through hypergraph calculation, adjusting channels using a convolutional layer, and outputting to obtain processed target feature data. 2.The UAV infrared small target detection method based on artificial intelligence technology according to claim 1, characterized in that, The step of processing the multiple independent regions through multiple stages, capturing cross-region dependencies using a multi-scale attention mechanism, and deepening feature expression through residual connections, specifically includes: Using a multi-window attention design to adapt to the multi-scale characteristics of small targets and capture long-range dependencies; Modulating the decomposed frequency components to suppress background redundant information; Residual fusion retains original feature information, optimizes gradient transmission, and strengthens the relevance of small target features, including: Input feature X ∈ R C×H×W (Where H × W is spatial resolution, C is the number of channels), first compressed redundant channels by Conv1 × 1, enhanced feature nonlinear expression by nonlinear activation, focused on small target thermal radiation profile by region division, and then converted to frequency domain by fast Fourier transform, and introduced a learnable filter to preliminarily suppress background noise. In the formula, GELU(·) represents applying a GELU activation function to the convolution result, BlockPartitioning(·) represents block processing of the activated result, FFT(·) represents fast Fourier transform on the block-processed result; W represents element-wise multiplication of the FFT result and the weight matrix W, X freq represents the frequency domain enhanced feature obtained after a series of processing. 3.The UAV infrared small target detection method based on artificial intelligence technology according to claim 2, characterized in that, Further comprising: After the features are projected linearly to K heads, they are evenly divided into 4 groups according to the number K, i.e. K / 4 heads per group, which correspond to four types of window attention respectively, including: Without loss of generality, take LH-WA as an example, such as: LH-WA; X is uniformly divided into non-overlapping windows [X1, ···, X M ], where each window contains 4s×s tokens; assuming that the projection query, key and value of the Kth head (3 / 4K < k ≤ K) are dk-dimensional tensors, the LH-WA output of the Kth head is as follows: where X i ∈ R C×4s×s , M = H x W / 4s x s, i = 1, ···, M, W k V ∈ R C×dk are the projection matrices of the query, key and value of the kth head, respectively, d k = C / k; The processes of the other three types of window attention are also derived from formula (3), and their outputs for the kth head are denoted as LL-WAk(X), HH-WAk(X), and HL-WAk(X) respectively. The outputs of the four parallel groups are concatenated using formula (4) to form the overall output; FDWA(X) = Concat [head1, ···, head K ]W O (4). 4.The method of claim 3, wherein, After stabilizing the feature distribution, the step of constructing a neighborhood-based joint tomographic imaging model includes: Further strengthening feature relevance and suppressing background redundancy using formula (5), and fusing original features through residual connection to retain original information and optimize gradient transmission; According to formula (6), the strengthened frequency domain features are converted back to the spatial domain through inverse fast Fourier transform, and after integrating multi-region information, the channel dimension is fine-tuned through Conv1x1 to adapt to the input of the subsequent detection head where FMFFN(·) denotes applying inverse fast Fourier transform to the frequency domain feature, FDWA(·) denotes a frequency modulation feedforward network, BlockMerging(·) denotes integrating multi-region spatial domain features, Conv1x1 denotes adjusting channel dimension by 1x1 convolution, X out denotes the fine-tuned feature, which is adapted to the detector input. 5.The UAV infrared small target detection method based on artificial intelligence technology according to claim 1, characterized in that, Further comprising a step to prevent small target weak features from being overwhelmed by background noise, specifically including: A = s + Dropout(PolaAttn(s)) (7) In the formula, s is the input feature, PolaAttn(s) realizes symbol-dependent decomposition and linear attention calculation, Dropout() is used for regularization, and the residual connection s+ ensures that the original feature information is not lost; After this operation, the features are stably distributed through layer normalization, as shown in formula (8): N = LayerNorm(A) (8) In the formula, A is the output of formula (7) (enhanced features after PolaAttention and residual connection); LayerNorm(·): normalizes the channel dimension of A, making the mean of each channel approach 0 and the variance approach 1; Through residual fusion strategy, the enhanced features and original input features are fused to retain the original feature basis information, superimpose multi-scale enhancement results, and strengthen feature discriminability; E = SEFN(N, A) (9) In the formula, SEFN(·) integrates multi-scale convolution, gate mechanism to complete feature enhancement, and residual connection to ensure basic feature retention; finally, through the second residual connection and layer normalization output optimization feature: O=LayerNorm(S+Dropout(E)) (10) Formula (10) applies residual connection again, preserves original feature information, combines layer normalization to stabilize feature distribution, and outputs features containing key information captured by symbol-dependent small targets and multi-scale enhancement; For the output multi-scale features, differential pooling and upsampling operations are used to obtain scale alignment of shallow high-resolution details and deep strong semantic information, and the above process can be formalized as: where Avgpool(·) denotes pooling the features, P align is the scale-aligned feature.

6. An unmanned aerial vehicle infrared small target detection device based on artificial intelligence technology, characterized in that, Including: The acquisition module is configured to acquire target feature data images in the target scene and perform preprocessing; The polarization space-channel feature interaction module is configured to process features from the initial convolutional layer, optimize features with spatial compression and reduce redundancy, split into query, key, and value components, and divide into regions corresponding to target signals and background noise according to symbols; The multi-scale attention feature enhancement module is configured to process the multiple independent regions through multiple stages, capture cross-region dependencies using multi-scale attention mechanisms, and deepen feature expression through residual connections; The multi-feature fusion module is configured to integrate features output by multiple stages to achieve scale alignment, generate attention weights using global average pooling and multi-layer perceptron to suppress noise, and then strengthen high-order correlations through hypergraph calculation, perform channel fine-tuning and output adaptation through convolutional layers, and obtain processed target feature data. The determination module is configured to use the processed target feature data for target tracking, region monitoring, or emergency response.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a program stored thereon, which, when executed, controls the device where the computer-readable storage medium is located to perform the unmanned aerial vehicle infrared small target detection method based on artificial intelligence technology according to any one of claims 1-5.

8. An electronic device, comprising: Including: one or more processors, memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs including a program for executing the unmanned aerial vehicle infrared small target detection method based on artificial intelligence technology according to any one of claims 1-5.

Images