Infrared dim drone detection method and system based on fusion of geometric perception and denoising

Abstract

This invention belongs to the field of computer vision and infrared image processing, and relates to a method and system for detecting small infrared drones based on geometric perception and denoising fusion. The method includes acquiring an infrared drone image dataset and preprocessing it. The preprocessed dataset is then input into a preset geometric perception denoising detection model for detection. First, shallow features are extracted. These shallow features are then processed through a windmill-shaped convolution to obtain enhanced features. The enhanced features are then processed through feature fusion and attention focusing to obtain attention-focused features. The attention-focused features and enhanced features are selectively weighted and fused to obtain fused features. Based on the fused features, intermediate layer features and large target features are obtained. Detection is performed at different scales on the original fused features, intermediate layer features, and large target features to obtain detection results. This invention solves the problems of insufficient feature extraction, poor background noise suppression, and low positioning accuracy in existing infrared small target detection technologies.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and infrared image processing, specifically relating to an infrared weak drone detection method and system based on geometric perception and denoising fusion. Background Technology

[0002] With the booming development of the low-altitude economy, unmanned aerial vehicles (UAVs) are increasingly used in both civilian and military fields. However, unauthorized low-altitude, slow, and small UAVs pose a serious threat to airspace security and privacy, urgently requiring efficient monitoring and countermeasure systems. Compared to visible light sensors, infrared imaging technology relies on thermal radiation and has inherent advantages such as all-weather operation and resistance to light interference, making it the preferred method for UAV detection at night and in adverse weather conditions.

[0003] However, infrared detection of small drones faces significant technical challenges. First, infrared drone targets are typically extremely small, often occupying less than 32×32 or even 8×8 pixels, and lack texture and color information, appearing as faint, dot-like targets. Second, infrared images have complex backgrounds, containing a large amount of high-frequency noise (such as cloud edges, leaf textures, and ground heat sources), resulting in an extremely low signal-to-noise ratio and a high likelihood of false alarms.

[0004] Existing detection methods are mainly divided into two categories: model-driven and data-driven. Traditional model-driven methods (such as filtering and local contrast) rely on handcrafted features and have poor generalization ability. In deep learning methods, segmentation-based networks have high computational costs and are difficult to implement in real time; while general-purpose object detectors (such as the YOLO series) are fast, they suffer from a serious "scale mismatch" problem: depth downsampling (such as 32 times) causes the disappearance of features of small targets; simple feature fusion mechanisms introduce shallow noise into deep layers, leading to false alarms; and the IoU loss function is extremely sensitive to the bounding box deviation of small targets, resulting in unstable regression.

[0005] Therefore, there is an urgent need for an infrared detection method for small unmanned aerial vehicles that can simultaneously solve the problems of feature loss, background interference, and inaccurate positioning. Summary of the Invention

[0006] The purpose of this invention is to provide an infrared weak UAV detection method and system based on geometric perception and denoising fusion, so as to solve the problems of insufficient feature extraction, poor background noise suppression and low positioning accuracy in the existing technology of infrared weak target detection.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for detecting small infrared unmanned aerial vehicles (UAVs) by fusing geometric perception and noise reduction, characterized in that it includes: Acquire an infrared UAV image dataset and preprocess it to obtain a preprocessed dataset; The preprocessed dataset is input into a pre-defined geometry-aware denoising detection model for detection, and the detection results are obtained; specifically including: Shallow features are extracted from the preprocessed dataset. These shallow features are then processed by windmill-shaped convolution to obtain enhanced features. The enhanced features are then processed by feature fusion and attention focusing to obtain attention-focused features. After aligning the number of channels through convolution, the attention focus feature and the enhancement feature are summed. Global average pooling and max pooling are then applied to the summed feature, which is then concatenated to obtain the concatenated feature. Channel attention weights are generated using a multilayer perceptron. These channel attention weights are then used to selectively weight and fuse the attention focus feature and the enhancement feature to obtain the fused feature. The fused feature consists of an original fused feature and intermediate fused features. The intermediate fused feature is weighted and fused with the fused feature to obtain the intermediate layer feature. Finally, the intermediate layer feature is weighted and fused with the fused feature to obtain the large target feature. Detection results were obtained by performing detection at different scales on the original fusion features, intermediate layer features, and large target features.

[0008] Optionally, the preprocessing specifically involves processing the infrared UAV image dataset using the Copy-Paste data augmentation method to obtain a preprocessed dataset.

[0009] Optionally, the preset geometry-aware denoising detection model includes: A geometrically perceptive backbone network is used to extract shallow features based on a preprocessed dataset. The shallow features are then processed by windmill-shaped convolution to obtain enhanced features. The enhanced features are then processed by feature fusion and attention-focusing to obtain attention-focusing features. A denoising feature fusion neck network is used to align the attention focus features and enhancement features through convolution, calculate the sum of the attention focus features and enhancement features, perform global average pooling and max pooling on the sum of the sum features, and then concatenate them to obtain the concatenated features. The concatenated features are then used to generate channel attention weights through a multilayer perceptron. These channel attention weights are then used to selectively weight and fuse the attention focus features and enhancement features to obtain the fused features. The fused features are divided into original fused features and intermediate fused features. The intermediate fused features are weighted and fused with the fused features to obtain intermediate layer features. The intermediate layer features are weighted and fused with the fused features to obtain the large target features. A high-resolution detection head is used to detect original fused features, intermediate layer features, and large target features at different scales to obtain detection results; The geometric perception backbone network, the denoising feature fusion neck network, and the high-resolution detection head are integrated into a single-stage target detection basic model to form a geometric perception denoising detection model.

[0010] Optionally, the geometry-aware backbone network is constructed by introducing windmill-shaped convolutional modules in the shallow and middle layers of the backbone network. The windmill-shaped convolutional modules extract the geometric contour features of the target through parallel asymmetric filling and strip convolution in several directions, which is used to prevent the features of weak targets from being smoothly lost in standard convolution operations. The denoising feature fusion neck network is obtained by replacing the traditional channel splicing module with a modulation fusion module; The high-resolution detection head is obtained by removing the deep detection layer and adding a shallow detection layer.

[0011] Optionally, the shallow features are processed by windmill-shaped convolution to obtain enhanced features, specifically including: The shallow features are divided into two parts: one part is the original features, and the other part is used for depth calculation to obtain deep nonlinear features. The deep spliced ​​features are obtained by concatenating the original features with the deep nonlinear features; The geometric contour features of different directions are extracted by parallel asymmetric filling and strip convolution in several directions for deep splicing features; the geometric contour features of different directions are spliced ​​and aggregated in the channel dimension to obtain enhanced features.

[0012] Optionally, after aligning the number of channels through convolution, the attention-focused features and enhancement features are summed, and the sum features are concatenated after global average pooling and max pooling respectively to obtain concatenated features. Channel attention weights are generated using a multilayer perceptron, and the attention-focused features and enhancement features are selectively weighted and fused using these channel attention weights to obtain fused features. The fused features are divided into original fused features and intermediate fused features. The intermediate fused features are weighted and fused with the fused features to obtain intermediate layer features. The intermediate layer features are weighted and fused with the fused features to obtain large target features. Specifically, this includes: After upsampling the attention-focused features, they are convolved with the enhanced features to align the number of channels, and then added element by element to obtain the sum features. After performing global average pooling and max pooling on the features respectively, we obtain two one-dimensional features. We then concatenate the two one-dimensional features to obtain the concatenated features. The splicing features are generated by a multilayer perceptron to generate channel attention weights. The channel attention weights are adaptively divided into two sub-weights. The two sub-weights are used to selectively weight and fuse the attention-focused features and the enhancement features to obtain the fused features. The fusion features are divided into original fusion features and intermediate fusion features. After downsampling, the intermediate fusion features are convolved with the fusion features to align the number of channels and then added element by element to obtain the first sum feature. After performing global average pooling and max pooling on the first feature and the feature respectively, two one-dimensional features are obtained. The two one-dimensional features are then concatenated to obtain the first concatenated feature. The first concatenation feature generates channel attention weights through a multilayer perceptron. The channel attention weights are adaptively divided into two sub-weights. The two sub-weights are used to selectively weight and fuse the intermediate fusion feature and the fusion feature to obtain the intermediate layer feature. After downsampling the intermediate layer features, they are aligned with the fused features through convolution and then added element by element to obtain the second sum feature; The second feature is obtained by performing global average pooling and max pooling on the first and second features respectively. The two one-dimensional features are then concatenated to obtain the second concatenated feature. The second concatenation feature generates channel attention weights through a multilayer perceptron. These channel attention weights are adaptively divided into two sub-weights. The two sub-weights are then used to selectively weight and fuse the intermediate layer features and the fused features to obtain the large target feature.

[0013] Optionally, the detection of the original fused features, intermediate layer features, and large target features at different scales to obtain detection results includes: For the original fusion features, intermediate layer features, and large target features, the P2, P3, and P4 detection heads were used for detection, and the detection results were obtained.

[0014] In a second aspect, the present invention provides an infrared weak unmanned aerial vehicle (UAV) detection system that integrates geometric perception and noise reduction, comprising: The data acquisition unit is used to acquire infrared UAV image datasets and perform preprocessing to obtain preprocessed datasets. A geometry-aware denoising detection model is used to input a preprocessed dataset for detection and obtain detection results; the geometry-aware denoising detection model specifically includes: A geometrically perceptive backbone network is used to extract shallow features based on a preprocessed dataset. The shallow features are then processed by windmill-shaped convolution to obtain enhanced features. The enhanced features are then processed by feature fusion and attention-focusing to obtain attention-focusing features. A denoising feature fusion neck network is used to align the attention focus features and enhancement features through convolution. The sum of the attention focus features and enhancement features is then calculated. Global average pooling and max pooling are applied to the sum of these features, and they are concatenated to obtain the concatenated features. Channel attention weights are generated using a multilayer perceptron. These channel attention weights are then used to selectively weight and fuse the attention focus features and enhancement features to obtain the fused features. The fused features consist of original fused features and intermediate fused features. The intermediate fused features are weighted and fused with the fused features to obtain intermediate layer features. Finally, the intermediate layer features are weighted and fused with the fused features to obtain the large target features. A high-resolution detection head is used to detect original fused features, intermediate layer features, and large target features at different scales to obtain detection results.

[0015] In a third aspect, the present invention provides an electronic device including a processor and a memory, the processor being configured to execute a computer program stored in the memory to implement the infrared weak UAV detection method based on geometric perception and denoising fusion as described in any one of the preceding claims.

[0016] In a fourth aspect, the present invention provides a computer-readable storage medium storing at least one instruction, which, when executed by a processor, implements an infrared weak unmanned aerial vehicle detection method based on geometric perception and denoising fusion as described in any of the preceding claims.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: This application discloses an infrared weak UAV detection method that integrates geometric perception and denoising. It processes shallow features through windmill-shaped convolution, preserving the geometric contours of weak infrared targets while avoiding the loss of subtle features caused by standard convolution, thus achieving feature preservation. By calculating the sum of attention-focused features and enhanced features, and selectively weighting and fusing the two features based on global average pooling, max pooling, and multi-layer perceptron channel attention weights, it actively suppresses shallow background clutter and false alarms at the feature level, achieving active denoising. Furthermore, by splitting the fused features into original fused features and intermediate fused features, and weighting and fusing the intermediate fused features with the fused features to obtain intermediate layer features, and then weighting and fusing the intermediate layer features with the fused features to obtain large target features, combined with multi-scale detection of the original fused features, intermediate layer features, and large target features, it can retain and transmit small target information from multiple scale levels, preventing them from disappearing in deep networks. Simultaneously, through multi-level feature weighted fusion and the collaboration of multi-scale detection heads, it achieves accurate target localization.

[0018] Feature preservation: By using the windmill-shaped receptive field of the windmill-shaped convolution, the geometric contours of infrared point targets are effectively extracted, solving the problem of standard convolution smoothing out weak features in shallow layers.

[0019] Physical anti-loss: By introducing a high-resolution detection head at the P2 layer and removing the redundant P5 layer, the feature loss of small targets in deep networks is prevented at the physical scale, which significantly improves the recall rate.

[0020] Active denoising: Through the channel attention mechanism of the MFM module, active denoising at the feature level is achieved, which effectively suppresses false alarms caused by shallow background clutter and improves the precision.

[0021] Precise positioning: By using the Focal-EIoU loss function, the aspect ratio regression is decoupled, eliminating the oscillation of small targets. Combined with an optimized label allocation strategy, high-precision positioning is achieved. Attached Figure Description

[0022] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention; Figure 2 This is a diagram of the geometry-aware denoising detection model architecture according to an embodiment of the present invention; Figure 3 This is a schematic diagram of windmill convolution according to an embodiment of the present invention; Figure 4 This is a flowchart of the MFM data processing according to an embodiment of the present invention; Figure 5 This is a detection effect diagram of an embodiment of the present invention; Figure 6 This is a data processing flowchart for C3K2_Pconv according to an embodiment of the present invention; Figure 7 This is a system structure block diagram according to an embodiment of the present invention; Figure 8 This is a structural block diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0023] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0024] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terminology used in this invention is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.

[0025] To address the shortcomings of existing infrared weak target detection technologies, such as insufficient feature extraction, poor background noise suppression, and low positioning accuracy, this invention provides an infrared weak UAV detection method that integrates geometric perception and denoising. (See [link to relevant documentation]). Figure 1 ,include: Acquire an infrared UAV image dataset and preprocess it to obtain a preprocessed dataset; The preprocessed dataset is input into a pre-defined geometry-aware denoising detection model for detection, and the detection results are obtained; specifically including: Shallow features are extracted from the preprocessed dataset. These shallow features are then processed by windmill-shaped convolution to obtain enhanced features. The enhanced features are then processed by feature fusion and attention focusing to obtain attention-focused features. After aligning the number of channels through convolution, the attention focus feature and the enhancement feature are summed. Global average pooling and max pooling are then applied to the summed feature, which is then concatenated to obtain the concatenated feature. Channel attention weights are generated using a multilayer perceptron. These channel attention weights are then used to selectively weight and fuse the attention focus feature and the enhancement feature to obtain the fused feature. The fused feature consists of an original fused feature and intermediate fused features. The intermediate fused feature is weighted and fused with the fused feature to obtain the intermediate layer feature. Finally, the intermediate layer feature is weighted and fused with the fused feature to obtain the large target feature. Detection results were obtained by performing detection at different scales on the original fusion features, intermediate layer features, and large target features.

[0026] In some embodiments, the preprocessing specifically involves processing the infrared UAV image dataset using the Copy-Paste data augmentation method to obtain a preprocessed dataset.

[0027] In some embodiments, the preset geometry-aware denoising detection model includes: A geometrically perceptive backbone network is used to extract shallow features based on a preprocessed dataset. The shallow features are then processed by windmill-shaped convolution to obtain enhanced features. The enhanced features are then processed by feature fusion and attention-focusing to obtain attention-focusing features. A denoising feature fusion neck network is used to align the attention focus features and enhancement features through convolution, calculate the sum of the attention focus features and enhancement features, perform global average pooling and max pooling on the sum of the sum features, and then concatenate them to obtain the concatenated features. The concatenated features are then used to generate channel attention weights through a multilayer perceptron. These channel attention weights are then used to selectively weight and fuse the attention focus features and enhancement features to obtain the fused features. The fused features are divided into original fused features and intermediate fused features. The intermediate fused features are weighted and fused with the fused features to obtain intermediate layer features. The intermediate layer features are weighted and fused with the fused features to obtain the large target features. A high-resolution detection head is used to detect original fused features, intermediate layer features, and large target features at different scales to obtain detection results; The geometric perception backbone network, the denoising feature fusion neck network, and the high-resolution detection head are integrated into a single-stage target detection basic model to form a geometric perception denoising detection model.

[0028] In some embodiments, the geometry-aware backbone network is constructed by introducing windmill-shaped convolutional modules in the shallow and middle layers of the backbone network. The windmill-shaped convolutional modules extract the geometric contour features of the target through parallel asymmetric filling and strip convolution in several directions, which is used to prevent the features of weak targets from being smoothly lost in standard convolution operations. The denoising feature fusion neck network is obtained by replacing the traditional channel splicing module with a modulation fusion module; The high-resolution detection head is obtained by removing the deep detection layer and adding a shallow detection layer.

[0029] In some embodiments, the shallow features are processed by windmill-shaped convolution to obtain enhanced features, specifically including: The shallow features are divided into two parts: one part is the original features, and the other part is used for depth calculation to obtain deep nonlinear features. The deep spliced ​​features are obtained by concatenating the original features with the deep nonlinear features; The geometric contour features of different directions are extracted by parallel asymmetric filling and strip convolution in several directions for deep splicing features; the geometric contour features of different directions are spliced ​​and aggregated in the channel dimension to obtain enhanced features.

[0030] In some embodiments, after aligning the number of channels through convolution, the attention focus feature and the enhancement feature are summed, and the summed feature is concatenated after performing global average pooling and max pooling on the summed feature to obtain the concatenated feature. The concatenated feature generates channel attention weights through a multilayer perceptron, and the attention focus feature and the enhancement feature are selectively weighted and fused using the channel attention weights to obtain the fused feature. The fused feature is divided into original fused feature and intermediate fused feature. The intermediate fused feature is weighted and fused with the fused feature to obtain intermediate layer feature. The intermediate layer feature is weighted and fused with the fused feature to obtain large target feature. Specifically, this includes: After upsampling the attention-focused features, they are convolved with the enhanced features to align the number of channels, and then added element by element to obtain the sum features. After performing global average pooling and max pooling on the features respectively, we obtain two one-dimensional features. We then concatenate the two one-dimensional features to obtain the concatenated features. The splicing features are generated by a multilayer perceptron to generate channel attention weights. The channel attention weights are adaptively divided into two sub-weights. The two sub-weights are used to selectively weight and fuse the attention-focused features and the enhancement features to obtain the fused features. The fusion features are divided into original fusion features and intermediate fusion features. After downsampling, the intermediate fusion features are convolved with the fusion features to align the number of channels and then added element by element to obtain the first sum feature. After performing global average pooling and max pooling on the first feature and the feature respectively, two one-dimensional features are obtained. The two one-dimensional features are then concatenated to obtain the first concatenated feature. The first concatenation feature generates channel attention weights through a multilayer perceptron. The channel attention weights are adaptively divided into two sub-weights. The two sub-weights are used to selectively weight and fuse the intermediate fusion feature and the fusion feature to obtain the intermediate layer feature. After downsampling the intermediate layer features, they are aligned with the fused features through convolution and then added element by element to obtain the second sum feature; The second feature is obtained by performing global average pooling and max pooling on the first and second features respectively. The two one-dimensional features are then concatenated to obtain the second concatenated feature. The second concatenation feature generates channel attention weights through a multilayer perceptron. These channel attention weights are adaptively divided into two sub-weights. The two sub-weights are then used to selectively weight and fuse the intermediate layer features and the fused features to obtain the large target feature.

[0031] In some embodiments, the detection of the original fused features, intermediate layer features, and large target features at different scales to obtain detection results includes: For the original fusion features, intermediate layer features, and large target features, the P2, P3, and P4 detection heads were used for detection, and the detection results were obtained.

[0032] Example 1 A method for detecting small, infrared drones by fusing geometric perception and denoising includes the following steps: Step 1: Acquire and preprocess the infrared drone image dataset. A data augmentation strategy is employed, randomly copying and pasting targets into other locations within the images to increase target density in individual images, forcing the model to learn target features rather than background texture. Step 2: Construct a geometry-aware denoising detection model, the design of which includes: geometry-aware backbone network module design, denoising feature fusion module design, high-resolution small target detection head design, and loss function and label assignment strategy design.

[0033] The geometry-aware backbone network module employs a windmill-shaped convolution (PConv). PConv utilizes asymmetric padding and multi-branch convolution operations to extract the geometric contour features of the target. Specifically, four parallel convolutional branches are defined, corresponding to the geometric perception in the upper left, upper right, lower left, and lower right directions, respectively. For the i-th branch, an asymmetric padding operation is first performed, followed by a convolution operation to obtain the branch features. The features from the four branches are concatenated along the channel dimension and aggregated through a fusion convolutional layer.

[0034] The denoising feature fusion module design employs a Modulation Fusion Mechanism (MFM) to replace the traditional channel concatenation operation. MFM first aligns the number of channels of the input features using 1x1 convolutions; then it calculates the sum of the input features; next, it performs global average pooling and max pooling on the sum, and generates channel attention weights using a multilayer perceptron (MLP); finally, it uses these weights to selectively weight and fuse the input features to suppress high-frequency background noise channels introduced by shallow layers.

[0035] The high-resolution micro-target detection head design reconstructs the feature pyramid architecture, removes the detection output of the P5 layer with its excessively large receptive field, and adds a high-resolution P2 detection layer (4x downsampling). A multi-scale detection structure containing three scales, P2, P3, and P4, is constructed, where the P2 layer retains a high-resolution feature of 160×160 (for 640 input), ensuring the visibility of micro-targets from a physical perspective.

[0036] The loss function design adopts Focal-EIoU as the bounding box regression loss function. This loss function includes an EIoU term and a Focal term: the EIoU term directly calculates the independent difference in width and height between the predicted box and the ground truth box, solving the gradient oscillation problem caused by the aspect ratio sensitivity of small infrared targets; the Focal term uses the IoU value as a weighting coefficient to reduce the weight of high-quality samples, so that the model focuses on samples that are difficult to regress.

[0037] The label allocation strategy optimizes and adjusts the hyperparameters of the task alignment allocator. The IoU weight coefficient (beta) is reduced to 4.0, and the classification score weight coefficient (alpha) is increased to 1.0, thereby relaxing the geometric matching threshold for small targets while maintaining classification confidence and improving the recall rate of small targets.

[0038] Step 3: Integrate the geometric-aware backbone network module, the denoising feature fusion module, and the high-resolution micro-target detection head into the general single-stage target detection basic model to construct the GD-MFM-YOLO (Geometric-aware and Denoising Network with Modulation Fusion Mechanism based YOLO) model.

[0039] Step 4: Train and validate the improved model, and evaluate the improved method.

[0040] Step 5: Input the infrared UAV image to be tested into the trained model to obtain the detection results.

[0041] Example 2 A method for detecting small, infrared drones by fusing geometric perception and denoising includes: Acquire an infrared UAV image dataset and preprocess it to obtain a preprocessed dataset; See Figure 2 The preprocessed dataset first enters the backbone network on the left for feature extraction. The data flow here is top-down: Shallow feature extraction: The image is first subjected to consecutive Conv (standard convolution) to extract basic edge and texture information from the infrared image.

[0042] Lightweight deep feature extraction (C3K2_Pconv): As the network deepens, the data enters the customized C3K2_Pconv module.

[0043] The C3K2 module is based on the CSP architecture, which stands for Cross-Stage Partial Network. Instead of sending all channel features for complex non-linear extraction, the network extracts only a local (half) of the channels for depth calculation. This local processing of feature channels ensures depth of feature extraction while significantly reducing computational cost. The Bottleneck or C3K module within C3K2 uses standard convolutional kernels (such as 3x3 convolutions). As the convolutional kernel slides across the image or feature map, it only sees a very limited area (i.e., the local pixel neighborhood) at a time. Its task is to focus on extracting texture, edges, corners, or minute gradient changes within this small local area.

[0044] See Figure 6 Within these modules, the data is first split into two paths (CSP structure idea). One path retains the original features, while the other path undergoes deep computation through multiple Bottleneck or C3K modules to obtain higher-order, more abstract deep nonlinear features. Finally, after concatenation with the original basic features, it is processed by a Pconv (windmill-shaped rotational convolution) to obtain enhanced features.

[0045] The windmill-shaped convolution is mainly deployed in the shallow and middle layers (P3, P4 stages) of the backbone network. To extract the geometric edges of targets as much as possible while the feature map is still relatively large and point-like targets haven't been downsampled, data X simultaneously enters four parallel branches. In each branch, the data first undergoes a special padding operation, Pad_i. These four branches perform asymmetric padding in the upper left, upper right, lower left, and lower right directions of the feature map, respectively. The padded data then passes through a strip convolution (e.g., a 1*k or k*1 kernel) in its respective branch. The outputs of these four branches are Y_1, Y_2, Y_3, and Y_4, respectively. The core function here is that the asymmetric processing in four directions, combined with the strip convolution, forms a "windmill"-shaped receptive field, which can very sensitively capture minute gradient changes in the up, down, left, and right directions of small infrared targets, without smoothing out isolated infrared bright spots like a standard 3*3 convolution. After the four branches are extracted, the four data sets Y_1 to Y_4 with different geometric features in different directions will be stacked together in the channel dimension (Concat operation). Finally, the stacked feature maps will be passed through a fusion convolutional layer Conv_{fuse} to deeply fuse the features in the four directions. After passing through a Pconv (windmill-shaped rotation convolution), the final enhanced features will be output.

[0046] Global Features and Attention: Data flows to the bottom of the backbone, where it undergoes SPPF (Spatial Pyramid Pooling) to fuse features from different receptive fields, and finally enters C2PSA. C2PSA is a module with an attention mechanism that forces the network to focus its attention on heat sources suspected to be drones in a noisy infrared background.

[0047] Feature fusion is achieved using a geometry-aware backbone network, which incorporates bidirectional flows from bottom to top and from top to bottom. FPN path: Data flows out from the lowest level C2PSA and is then upsampled (upsampled / enlarged in size).

[0048] The magnified feature map meets and merges with the deep side branches from the backbone network in the MFM (Multi-Feature Fusion Module).

[0049] The merged data is processed by C3K2, upsampled again, and then merged with the middle-layer side branches of the trunk in the next MFM.

[0050] This process continues upwards until it merges with the shallow lateral branches of the main trunk.

[0051] MFM (Multi-Feature Fusion Module) replaces the traditional direct concatenation. Traditional concatenation simply stitches together features from deep and shallow layers, which is fatal in infrared images because shallow features contain a large amount of high-frequency background noise. MFM introduces an attention mechanism for intelligent filtering. When data flows into MFM, there are two different data sources: The first is the upsampled features F_high (which understands semantics and knows roughly what the target is) from deep layers, which flow upwards along the FPN path. It initially flows out from the C2PSA module at the bottom of the backbone network. To fuse with the shallower features with higher resolution, it must first undergo an upsample operation to enlarge its size before entering MFM. The second is the features F_low (which has detail but also contains background noise) from shallow layers, which come directly from the lateral branches of the backbone network. As F_high propagates upwards layer by layer, it encounters data "transmitted laterally" from the corresponding levels of the backbone network at different scale nodes (deep, mid, and shallow). This characteristic, directly transmitted from the left-side backbone network, is F_low.

[0052] See Figure 4 The two data streams are first adjusted to have the same number of channels through 1*1 convolutions, and then directly added element-wise to obtain the mixed feature F_sum. To determine which channels are important, the data F_sum is processed by Global Average Pooling (GAP) and Global Max Pooling (GMP). This step compresses the massive spatial feature map into two one-dimensional vectors, extracting channel-level statistical features. The vectors extracted by GAP and GMP are concatenated and fed into a Multilayer Perceptron (MLP) network, and finally normalized using the Softmax function. This calculates a channel attention weight vector W. The calculated referee weight W is split into two halves, called W_high and W_low. Finally, the original input data F_high and F_low are multiplied element-wise with their respective weights (i.e., weighted modulation), and then summed to output the final fused feature F_fused.

[0053] PANet Path: After the data reaches the top of FPN, one path goes directly to the prediction head, while the other path starts to go downwards.

[0054] The data is downsampled (reduced in size) by Conv, then fused again in the intermediate layer features corresponding to the MFM and FPN stages, and finally passed through C3K2.

[0055] The Conv downsampling continues downwards, fusing with the features from the lower layers. This step passes the precise location and boundary information from the top layer to the deep feature map, enhancing the network's ability to locate the target.

[0056] The fully fused feature streams are then fed into three detection heads of different scales for prediction: P2 Detect (160*160): Killer for tiny targets Data source: Received feature maps from the top layer of Neck, with minimal downsampling and the highest resolution (160x160).

[0057] Function: This is a "super-large feature map detection head" that is not typically found in standard YOLO models. In infrared images, distant drones may only occupy a few pixels. The P2 head preserves extremely rich spatial geometric details and is specifically designed to capture these faint drone hotspots.

[0058] P3 Detect (80*80): Medium target detection Data source: Feature maps received from the Neck intermediate layer.

[0059] Function: Responsible for detecting UAV targets at a moderate distance with certain outline features.

[0060] P4 Detect (40*40): Large target detection Data source: Received from the lowest layer of the Neck, which has the largest receptive field.

[0061] Function: Responsible for detecting drone targets that are close to the camera and occupy a large proportion of the frame.

[0062] Infrared video stream - backbone network extracts features - Neck network mixes deep and shallow features through MFM - data is split to three detection heads P2 / P3 / P4 - outputs the bounding box and confidence score of the UAV.

[0063] Example 3 A method for detecting weak infrared drones by fusing geometric perception and noise reduction, the specific implementation steps of which are as follows: Step 1: Data Preparation and Augmentation. Image data from the ANTI-UAV410 infrared UAV was collected. The background environment is highly diverse, encompassing complex elements such as buildings, cities, mountains, clouds, and water surfaces under different conditions (autumn and winter, day and night). The dataset was divided into 200 training sequences, 90 validation sequences, and 120 test sequences. We used the first 290 sequences for training and validation, and retained 120 completely independent sequences for testing. The image resolution is 640×512 pixels, and the proportion of small-sized targets is significantly increased compared to similar datasets, making it extremely challenging. To address the sparse infrared small target samples, a copy-paste data augmentation strategy was adopted. Specifically, image patches containing targets were randomly cropped from the training set and randomly pasted into arbitrary non-target regions of other training images to generate new training samples. This operation significantly increased the positive sample density in a single image, forcing the model to focus on the features of the target itself rather than relying on background context for inference, thus alleviating the overfitting problem. Furthermore, Mosaic data augmentation was employed to simulate long-distance and small target environments.

[0064] Step 2: Construction of Geometric Perception Denoising Detection Model See Figure 2 This invention improves upon the YOLOv11 architecture, constructing the GD-MFM-YOLO model. This model mainly consists of three parts: 2.1 Geometric Perception Backbone Network In the shallow and middle layers of the backbone network (specifically P3 and P4 stages, i.e., 8x and 16x downsampling layers), the C3k2_PConv module replaces the standard C3k2 module. The core of the C3k2_PConv module is the integration of pinwheel convolution (PConv). Figure 4 As shown, PConv contains four parallel branches, each used to extract geometric features in different directions. Let the input feature map be... Define asymmetric fill operations in four directions. (Corresponding to top left, top right, bottom left, and bottom right respectively). For the i-th branch, its feature extraction process can be represented as follows:

[0065] Here, Conv stands for strip convolution. Subsequently, the features from the four branches are concatenated along the channel dimension (Concat), and then aggregated through a fusion convolutional layer (Convfuse) to obtain the geometrically enhanced output feature Flow, i.e., the enhanced feature:

[0066] This design enables the network to accurately capture geometric gradient changes of infrared point targets even in shallow layers, preventing feature loss during standard 3x3 convolutions. For deeper layers (P5, 32x downsampling), only semantic extraction capabilities are retained.

[0067] 2.2 Denoising Feature Fusion Neck Network The PANet (Path Aggregation Network) architecture is adopted, but its fusion mechanism is improved. The traditional Concat (channel splicing) operation is completely replaced by a modulation fusion module (MFM).

[0068] like Figure 4 As shown, the specific calculation process of MFM is as follows: Let the attention-focused features from the deep layer be represented as Fhigh, and the enhancement features from the shallow layer be represented as Flow.

[0069] First, align the two channels to the same number using a 1×1 convolution; then calculate the sum of the fused features, Fsum = Fhigh + Flow. Next, perform global average pooling (GAP) and global max pooling (GMP) on Fsum to extract channel statistics, and generate channel attention weight vectors W using a multilayer perceptron (MLP).

[0070]

[0071] Where [;] denotes concatenation, the weight vector W is divided into two sub-weights Whigh and Wlow, which are then used to weight and modulate the attention-focused feature and the enhancement feature respectively, resulting in the final fused feature Ffused.

[0072] in This indicates element-wise multiplication. This mechanism uses deep semantic information to guide the selection of shallow features, automatically suppressing high-frequency background noise channels introduced by shallow layers and refining features.

[0073] 2.3 High-resolution micro-target detection head For tiny targets, this invention reconstructs the detection head architecture.

[0074] First, the P5 detection layer (Stride32) with an excessively large receptive field was removed because small targets were lost in that layer.

[0075] Secondly, a top-down path leading to the P2 layer (Stride4) and a corresponding bottom-up path were added to the neck network, and a high-resolution P2 detection layer was added.

[0076] The final detection scales are P2, P3, and P4. The P2 layer retains a high-resolution feature of 160×160 (for 640 inputs), which physically ensures the visibility of small targets (less than 8 pixels) and significantly improves the recall rate.

[0077] Step 3: Loss Function Optimization To further improve positioning accuracy and recall, this invention optimizes the training process.

[0078] 3.1 Regression loss: Focal-EIoU Loss was used.

[0079] The aspect ratio penalty term in traditional CIoU loss is extremely sensitive to pixel deviations of small targets, leading to gradient oscillations. This invention uses EIoU to directly regress the width and height difference between the predicted and ground truth bounding boxes:

[0080] Based on this, a Focal focusing coefficient γ=0.5 is introduced to construct the final loss, which is used to mine difficult regression samples.

[0081]

[0082] Step 4: Train and validate the improved model, and evaluate the improved method. The improved model is systematically trained and validated, and comparative evaluation experiments are conducted. Specifically, to verify the advancement of the method of this invention, the basic network model YOLOv11, as well as mainstream classic and advanced object detection algorithms in the industry, are selected as comparison benchmarks, including YOLOv8, YOLOv9, YOLOv10, YOLOv12, Gold-YOLO, and RT-DETR-ResNet50. All the above models are trained on the Anti-UAV410 infrared UAV dataset using a unified experimental configuration.

[0083] During the model evaluation phase, a multi-dimensional evaluation index system was constructed: Precision, Recall, and Mean Precision (mAP@50 and mAP@50-95) were used as evaluation indicators for detection accuracy; Floating-point operations (GFLOPs) and frame rate (FPS) were used as evaluation indicators for computational efficiency.

[0084] The quantitative evaluation metrics for each model are shown in Table 1. Experimental results show that, compared with general object detection models, the GD-MFM-YOLO model based on geometric perception and denoising fusion proposed in this invention achieves significant performance improvements in key metrics such as Recall and mAP@50-95. Specific analysis is as follows: The significant improvement in accuracy fully demonstrates the effectiveness of the P2 high-resolution detection head and windmill-shaped convolution (PConv) introduced in this invention in preserving small target features at the physical scale and preventing missed detections; the improvement in mAP@50-95 strongly verifies the advantages of the Focal-EIoU loss function in optimizing bounding box regression quality and improving localization accuracy.

[0085] In summary, the experimental data fully verify the superiority of the method of the present invention in infrared weak UAV target detection tasks, especially in the precise acquisition of small targets, suppression of complex background noise and high-precision positioning, which have significant technical advantages.

[0086] Table 1. Quantitative experimental results of the basic model and the model improved based on the method of this invention on the ANTI-UAV410 dataset.

[0087]

[0088] Step 5: Input the infrared UAV imagery to be tested into the trained model to obtain the detection results. Specifically, the trained GD-MFM-YOLO model is used to perform inference on the real-time acquired infrared video stream or image. Its detection effect is as follows: Figure 5 As shown, analysis of detection results under various complex backgrounds, including sky, mountains, and urban buildings, demonstrates that the model applying the method of this invention can accurately detect tiny UAV targets in infrared images. The model effectively captures point targets occupying only a few pixels using the high-resolution features of the P2 layer, while dynamically suppressing false alarm interference from cloud edges and ground heat sources using the MFM module. The detection results fully demonstrate the advantages of this invention in improving the recall rate and localization accuracy of tiny targets. Through a geometric perception and denoising fusion strategy, this invention focuses solely on detecting infrared UAV target areas while eliminating interference from complex backgrounds, significantly improving detection robustness while ensuring real-time detection performance.

[0089] In summary, this invention relates to a method for detecting infrared-based weak UAV targets based on geometric perception and denoising fusion, mainly comprising the following steps: First, acquiring an infrared UAV target detection dataset and preprocessing it using data augmentation strategies. Next, constructing a geometric perception-based denoising detection model, the design of which includes: integrating a windmill-shaped convolution (PConv) module in the shallow layers of the backbone network to extract geometric contours; employing a modulation fusion module (MFM) in the neck network for denoising feature fusion; constructing a high-resolution multi-scale detection head containing a P2 layer to prevent feature vanishing; and designing a Focal-EIoU loss function. Then, porting the above improved modules to the general single-stage target detection model YOLOv11. Next, training and validating the improved model, and evaluating the improved method. Finally, inputting infrared images into the trained model to obtain detection results. This invention solves the problems of feature vanishing during deep network downsampling of infrared-based weak UAVs, and the high false alarm rate and low positioning accuracy caused by complex thermal background noise interference. This invention designs a detection architecture specifically for the characteristics of infrared micro-targets. By physically preserving high-resolution features through the P2 layer and combining the geometric perception capabilities of PConv with the feature denoising capabilities of MFM, accurate capture of faint targets is achieved. While ensuring detection accuracy, the regression gradient oscillation problem is solved by optimizing the loss function, thereby improving detection efficiency and positioning accuracy.

[0090] Example 4 like Figure 7 As shown, based on the same inventive concept as the above embodiments, the present invention also provides an infrared weak unmanned aerial vehicle (UAV) detection system that fuses geometric perception and noise reduction, comprising: The data acquisition unit is used to acquire infrared UAV image datasets and perform preprocessing to obtain preprocessed datasets. A geometry-aware denoising detection model is used to input a preprocessed dataset for detection and obtain detection results; the geometry-aware denoising detection model specifically includes: A geometrically perceptive backbone network is used to extract shallow features based on a preprocessed dataset. The shallow features are then processed by windmill-shaped convolution to obtain enhanced features. The enhanced features are then processed by feature fusion and attention-focusing to obtain attention-focusing features. A denoising feature fusion neck network is used to align the attention focus features and enhancement features through convolution. The sum of the attention focus features and enhancement features is then calculated. Global average pooling and max pooling are applied to the sum of these features, and they are concatenated to obtain the concatenated features. Channel attention weights are generated using a multilayer perceptron. These channel attention weights are then used to selectively weight and fuse the attention focus features and enhancement features to obtain the fused features. The fused features consist of original fused features and intermediate fused features. The intermediate fused features are weighted and fused with the fused features to obtain intermediate layer features. Finally, the intermediate layer features are weighted and fused with the fused features to obtain the large target features. A high-resolution detection head is used to detect original fused features, intermediate layer features, and large target features at different scales to obtain detection results.

[0091] Example 5 like Figure 8 As shown, the present invention also provides an electronic device 100 for realizing an infrared weak UAV detection method that integrates geometric perception and noise reduction; The electronic device 100 includes a memory 101, at least one processor 102, a computer program 103 stored in the memory 101 and executable on at least one processor 102, and at least one communication bus 104.

[0092] The memory 101 can be used to store the computer program 103. The processor 102 implements the steps of the infrared weak UAV detection method by running or executing the computer program stored in the memory 101 and calling the data stored in the memory 101.

[0093] The memory 101 may primarily include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created based on the use of the electronic device 100 (such as audio data), etc. In addition, the memory 101 may include non-volatile memory, such as hard disk, RAM, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other non-volatile solid-state storage device.

[0094] At least one processor 102 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Processor 102 may be a microprocessor or any conventional processor. Processor 102 is the control center of electronic device 100, connecting various parts of electronic device 100 via various interfaces and lines.

[0095] The memory 101 in the electronic device 100 stores multiple instructions to implement a geometric perception and noise reduction fusion infrared weak drone detection method, and the processor 102 can execute multiple instructions to achieve the following: Acquire an infrared UAV image dataset and preprocess it to obtain a preprocessed dataset; The preprocessed dataset is input into a pre-defined geometry-aware denoising detection model for detection, and the detection results are obtained; specifically including: Shallow features are extracted from the preprocessed dataset. These shallow features are then processed by windmill-shaped convolution to obtain enhanced features. The enhanced features are then processed by feature fusion and attention focusing to obtain attention-focused features. After aligning the number of channels through convolution, the attention focus feature and the enhancement feature are summed. Global average pooling and max pooling are then applied to the summed feature, which is then concatenated to obtain the concatenated feature. Channel attention weights are generated using a multilayer perceptron. These channel attention weights are then used to selectively weight and fuse the attention focus feature and the enhancement feature to obtain the fused feature. The fused feature consists of an original fused feature and intermediate fused features. The intermediate fused feature is weighted and fused with the fused feature to obtain the intermediate layer feature. Finally, the intermediate layer feature is weighted and fused with the fused feature to obtain the large target feature. Detection results were obtained by performing detection at different scales on the original fusion features, intermediate layer features, and large target features.

[0096] Example 6 If the modules / units integrated in the electronic device 100 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, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, and read-only memory (ROM).

[0097] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0098] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0099] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0100] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0101] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0102] 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 it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for detecting small infrared unmanned aerial vehicles (UAVs) by fusing geometric perception and denoising, characterized in that, include: Acquire an infrared UAV image dataset and preprocess it to obtain a preprocessed dataset; The preprocessed dataset is input into a preset geometric perception denoising detection model for detection, and the detection results are obtained. Specifically, it includes: Shallow features are extracted from the preprocessed dataset. These shallow features are then processed by windmill-shaped convolution to obtain enhanced features. The enhanced features are then processed by feature fusion and attention focusing to obtain attention-focused features. After aligning the number of channels through convolution, the attention focus feature and the enhancement feature are summed. Global average pooling and max pooling are then applied to the summed feature, which is then concatenated to obtain the concatenated feature. Channel attention weights are generated using a multilayer perceptron. These channel attention weights are then used to selectively weight and fuse the attention focus feature and the enhancement feature to obtain the fused feature. The fused feature consists of an original fused feature and intermediate fused features. The intermediate fused feature is weighted and fused with the fused feature to obtain the intermediate layer feature. Finally, the intermediate layer feature is weighted and fused with the fused feature to obtain the large target feature. Detection results were obtained by performing detection at different scales on the original fusion features, intermediate layer features, and large target features.

2. The infrared weak UAV detection method based on geometric perception and denoising fusion according to claim 1, characterized in that, The preprocessing specifically involves using the Copy-Paste data augmentation method to process the infrared UAV image dataset, resulting in a preprocessed dataset.

3. The infrared weak UAV detection method based on geometric perception and denoising fusion according to claim 1, characterized in that, The preset geometric perception denoising detection model includes: A geometrically perceptive backbone network is used to extract shallow features based on a preprocessed dataset. The shallow features are then processed by windmill-shaped convolution to obtain enhanced features. The enhanced features are then processed by feature fusion and attention-focusing to obtain attention-focusing features. A denoising feature fusion neck network is used to align the attention focus features and enhancement features through convolution, calculate the sum of the attention focus features and enhancement features, perform global average pooling and max pooling on the sum of the sum features, and then concatenate them to obtain the concatenated features. The concatenated features are then used to generate channel attention weights through a multilayer perceptron. These channel attention weights are then used to selectively weight and fuse the attention focus features and enhancement features to obtain the fused features. The fused features are divided into original fused features and intermediate fused features. The intermediate fused features are weighted and fused with the fused features to obtain intermediate layer features. The intermediate layer features are weighted and fused with the fused features to obtain the large target features. A high-resolution detection head is used to detect original fused features, intermediate layer features, and large target features at different scales to obtain detection results; The geometric perception backbone network, the denoising feature fusion neck network, and the high-resolution detection head are integrated into a single-stage target detection basic model to form a geometric perception denoising detection model.

4. The infrared weak UAV detection method based on geometric perception and denoising fusion according to claim 3, characterized in that, The geometric perception backbone network is constructed by introducing windmill-shaped convolutional modules in the shallow and middle layers of the backbone network. The windmill-shaped convolutional modules extract the geometric contour features of the target through parallel asymmetric filling and strip convolution in several directions, which is used to prevent the features of weak targets from being smoothly lost in standard convolution operations. The denoising feature fusion neck network is obtained by replacing the traditional channel splicing module with a modulation fusion module; The high-resolution detection head is obtained by removing the deep detection layer and adding a shallow detection layer.

5. The infrared weak UAV detection method based on geometric perception and denoising fusion according to claim 1, characterized in that, The shallow features are processed through a windmill-shaped convolution to obtain enhanced features, specifically including: The shallow features are divided into two parts: one part is the original features, and the other part is used for depth calculation to obtain deep nonlinear features. The deep spliced ​​features are obtained by concatenating the original features with the deep nonlinear features; The geometric contour features of different directions are extracted by parallel asymmetric filling and strip convolution in several directions for deep splicing features; the geometric contour features of different directions are spliced ​​and aggregated in the channel dimension to obtain enhanced features.

6. The infrared weak UAV detection method based on geometric perception and denoising fusion according to claim 1, characterized in that, The attention focus feature and the enhancement feature are aligned with the number of channels through convolution. The sum of the attention focus feature and the enhancement feature is calculated. The sum feature is then concatenated after global average pooling and max pooling, respectively, to obtain the concatenated feature. The concatenated feature is used to generate channel attention weights through a multilayer perceptron. The attention focus feature and the enhancement feature are then selectively weighted and fused using the channel attention weights to obtain the fused feature. The fusion features are divided into original fusion features and intermediate fusion features. The intermediate fusion features are weighted and fused with the fusion features to obtain intermediate layer features. The large target feature is obtained by weighted fusion of intermediate layer features and fusion features; specifically including: After upsampling the attention-focused features, they are convolved with the enhanced features to align the number of channels, and then added element by element to obtain the sum features. After performing global average pooling and max pooling on the features respectively, we obtain two one-dimensional features. We then concatenate the two one-dimensional features to obtain the concatenated features. The splicing features are generated by a multilayer perceptron to generate channel attention weights. The channel attention weights are adaptively divided into two sub-weights. The two sub-weights are used to selectively weight and fuse the attention-focused features and the enhancement features to obtain the fused features. The fusion features are divided into original fusion features and intermediate fusion features. After downsampling, the intermediate fusion features are convolved with the fusion features to align the number of channels and then added element by element to obtain the first sum feature. After performing global average pooling and max pooling on the first feature and the feature respectively, two one-dimensional features are obtained. The two one-dimensional features are then concatenated to obtain the first concatenated feature. The first concatenation feature generates channel attention weights through a multilayer perceptron. The channel attention weights are adaptively divided into two sub-weights. The two sub-weights are used to selectively weight and fuse the intermediate fusion feature and the fusion feature to obtain the intermediate layer feature. After downsampling the intermediate layer features, they are aligned with the fused features through convolution and then added element by element to obtain the second sum feature; The second feature is obtained by performing global average pooling and max pooling on the first and second features respectively. The two one-dimensional features are then concatenated to obtain the second concatenated feature. The second concatenation feature generates channel attention weights through a multilayer perceptron. These channel attention weights are adaptively divided into two sub-weights. The two sub-weights are then used to selectively weight and fuse the intermediate layer features and the fused features to obtain the large target feature.

7. The infrared weak UAV detection method based on geometric perception and denoising fusion according to claim 1, characterized in that, The detection is performed at different scales on the original fused features, intermediate layer features, and large target features to obtain detection results, including: For the original fusion features, intermediate layer features, and large target features, the P2, P3, and P4 detection heads were used for detection, and the detection results were obtained.

8. A geometric perception and noise reduction fusion infrared weak UAV detection system, characterized in that, include: The data acquisition unit is used to acquire infrared UAV image datasets and perform preprocessing to obtain preprocessed datasets. A geometry-aware denoising detection model is used to input a preprocessed dataset for detection and obtain detection results. The geometric perception denoising detection model specifically includes: A geometrically perceptive backbone network is used to extract shallow features based on a preprocessed dataset. The shallow features are then processed by windmill-shaped convolution to obtain enhanced features. The enhanced features are then processed by feature fusion and attention-focusing to obtain attention-focusing features. A denoising feature fusion neck network is used to align the attention focus features and enhancement features through convolution. The sum of the attention focus features and enhancement features is then calculated. Global average pooling and max pooling are applied to the sum of these features, and they are concatenated to obtain the concatenated features. Channel attention weights are generated using a multilayer perceptron. These channel attention weights are then used to selectively weight and fuse the attention focus features and enhancement features to obtain the fused features. The fused features consist of original fused features and intermediate fused features. The intermediate fused features are weighted and fused with the fused features to obtain intermediate layer features. Finally, the intermediate layer features are weighted and fused with the fused features to obtain the large target features. A high-resolution detection head is used to detect original fused features, intermediate layer features, and large target features at different scales to obtain detection results.

9. An electronic device, characterized in that, It includes a processor and a memory, the processor being used to execute a computer program stored in the memory to implement an infrared weak UAV detection method based on geometric perception and denoising fusion as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one instruction, which, when executed by a processor, implements the infrared weak unmanned aerial vehicle detection method based on geometric perception and denoising fusion as described in any one of claims 1 to 7.

Images