Infrared anti-unmanned aerial vehicle detection method based on improved YOLO framework

Abstract

The invention discloses an infrared anti-unmanned aerial vehicle detection method based on an improved YOLO framework. The infrared anti-unmanned aerial vehicle detection method comprises the following steps: S1, preprocessing an infrared image and calculating energy distribution; s2, embedding the backbone network into an improved convolution module and adaptive wavelet basis convolution, and optimizing feature extraction in stages; s3, reinforcing local details and global contours of the feature maps by a multi-granularity pooling mechanism; s4, realizing cross-layer feature fusion through a feature aggregation propagation mechanism, and generating a comprehensive multi-scale feature map; s5, a lightweight detection channel is specially arranged to enhance the small target detection capability, and other standard detection heads are responsible for classified positioning of medium and large targets; and S6, outputting a final prediction result through a context sensing weighted frame fusion algorithm. Through adaptive wavelet feature enhancement, multi-granularity feature optimization, a lightweight detection head and an intelligent frame fusion technology, the problems of small target leak detection, background interference and multi-scale detection real-time performance in an infrared scene are solved, and the all-weather detection capability and the positioning precision are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of target detection and relates to an infrared anti-drone detection method based on an improved YOLO framework. Background Art

[0002] With the wide application of drones in the fields of military reconnaissance, commercial logistics, and people's livelihood services, security threats such as privacy infringement, airspace intrusion, and terrorist activities caused by them have become increasingly prominent. In this context, the core of anti-drone technology lies in building an all-weather and high-precision drone detection system. The detection technology based on infrared imaging realizes drone recognition by analyzing the thermal radiation characteristics of targets, has unique advantages such as penetrating smoke, haze, and anti-electromagnetic interference, and can effectively overcome the performance limitations of visible light sensors under low illumination and complex meteorological conditions. This technology has been deployed in scenarios such as key area air defense, border monitoring, airport clear sky protection, and critical infrastructure protection, providing reliable situation awareness and coordinate guidance for subsequent interception and disposal systems, and significantly enhancing the multi-dimensional space security protection ability.

[0003] Traditional infrared anti-drone detection technologies can be divided into single-frame and multi-frame methods: single-frame detection relies on threshold segmentation and morphological edge processing, is easily interfered by thermal noise and generates false edges; multi-frame methods are achieved through optical flow tracking and inter-frame difference, and have defects such as poor adaptability to motion mutations and high false alarm rates. Both types of methods are difficult to adapt to complex scenarios due to relying on artificial feature extraction. The current detection frameworks based on deep learning significantly enhance the detection robustness in complex environments through multi-scale feature fusion architectures. Their technical systems are mainly divided into two major branches: single-stage real-time methods (such as the YOLO series) and two-stage high-precision methods (such as Faster R-CNN). However, these methods still face the problem of mutual restriction between accuracy and speed in practical application scenarios with limited hardware. Especially under the weak texture characteristics and low signal-to-noise ratio conditions of infrared images, the target recall rate and positioning accuracy show a significant decline.

[0004] The present invention proposes an infrared anti-drone detection method (AW-YOLO) based on an improved YOLO framework. The core innovations include: proposing an adaptive wavelet basis selection algorithm based on relative energy distribution to achieve multi-scale feature perception and background interference decoupling through the optimal wavelet basis function; introducing a multi-granularity pooling mechanism to enhance the local detail and global contour representation capabilities of deep features; adopting a lightweight cross-layer detection head combined with a wavelet denoising module to reduce the missed detection rate of small targets while improving the inference efficiency; proposing a context-aware weighted box fusion algorithm to optimize the positioning accuracy by dynamically adjusting the boundary box weight parameters. This solution realizes the collaborative optimization of detection accuracy and real-time performance in infrared weak texture scenarios. Summary of the Invention

[0005] The present invention provides an infrared anti - UAV detection method based on an improved YOLO framework, aiming to solve the technical problems of high missed detection rate of small targets, large interference from complex backgrounds, and poor real - time performance of multi - scale target detection in the infrared scenario.

[0006] To achieve the above object, the following technical solutions are adopted:

[0007] The infrared anti - UAV detection method based on the improved YOLO framework includes the following steps:

[0008] Step 1: Perform standardized pre - processing on the input infrared image, calculate the relative energy distribution characteristics of each image block, and provide a basis for background noise suppression and target area enhancement for subsequent processing.

[0009] Step 2: Embed an improved convolutional module (StarReLU Convolution Module, SRConvModule) and an adaptive wavelet - based convolution (Adaptive Wavelet Base Convolution, AWBC) in the model backbone network for feature extraction and optimization at different stages. Among them, AWBC dynamically selects the optimal wavelet basis function based on the relative energy distribution, and implements multi - level wavelet decomposition to achieve high - frequency detail extraction and low - frequency noise suppression.

[0010] Step 3: Introduce a feature optimization mechanism based on multi - granularity pooling to hierarchically process the feature map. Synchronously capture local detail features and global contour information through pooling windows of different scales, strengthen the representation of the global contour and spatial position relationship while refining texture details, and construct a multi - level feature expression system.

[0011] Step 4: Achieve the fusion of multi - level features through a feature aggregation and propagation mechanism, and complete cross - level feature information interaction. After fusion, a feature map integrating information of different levels is obtained, providing complete multi - scale feature support for multiple detection heads.

[0012] Step 5: Combine with an optimized multi - detection - head design to perform classification and bounding - box regression prediction on feature maps of different scales. Among them, a lightweight detection head H2 is designed for small - target detection tasks, and the detection ability for small targets is optimized to further improve real - time performance; standard detection heads H3, H4, and H5 respectively process medium - and large - target detection tasks, and comprehensively predict the category, position, and confidence of the target. Design a heterogeneous multi - detection - head architecture, adopt feature - map enhancement processing and parameter - optimization techniques for small - target detection to improve detection sensitivity, and the standard detection heads achieve accurate positioning of medium - and large - targets.

[0013] Step 6: Implement Context-Aware Weighted Box Fusion (CAWBF) to screen and fuse the output results of multiple detection heads. Optimize the bounding box positioning by combining the confidence, geometric information, and background influence factor to generate the final object detection result.

[0014] Among them, the specific operations of preprocessing and calculating the relative energy distribution in Step 1 are as follows:

[0015] Step 1.1: Perform equal ratio scaling according to the original aspect ratio of the input infrared image and adjust it to the preset input size. When the size of the scaled image is smaller than the target size, symmetric filling is performed using the gray value 128 to maintain the geometric features without deformation.

[0016] Step 1.2: Perform normalization processing on the scaled image to linearly map the pixel values to the interval [0, 1].

[0017] Step 1.3: Construct a feature optimization matrix through multi-level energy analysis:

[0018] Step 1.3.1: Calculate the global energy reference value: Among them, E global is the global energy reference value, and W and H are the width and height of the image respectively.

[0019] Step 1.3.2: Divide the image into an m×n uniform grid (m represents the number of sub-regions in the vertical direction of the grid, and n represents the number of sub-regions in the horizontal direction of the grid). The size of each sub-region R(i, j) is adaptively set according to the imaging characteristics of typical UAV targets. R(i, j) represents the sub-region in the i-th row and j-th column of the divided grid.

[0020] Step 1.3.3: Use the integral image acceleration technique to calculate the local energy mean and construct the integral image: Among them, I int (x, y) is the value of the integral image at the coordinate (x, y), and (x′, y′) is the coordinate variable traversed during the summation process. The value range of x′ is from 0 to x. For the sub-region R(i, j) with boundary coordinates from (x1, y1) to (x2, y2), the local energy mean is: Among them, E local (i, j) is the local energy mean of the sub-region R(i, j), (x1, y1) is the coordinate of the upper left corner of the sub-region, and (x2, y2) is the coordinate of the lower right corner of the sub-region.

[0021] Step 1.3.4: Calculate the relative energy distribution matrix: E rel (i,j) = E local (i,j) - E global where E rel (i,j) is the relative energy value of sub-region R(i,j), and E local (i,j) is as explained in Step 1.3.3, and E global is as explained in Step 1.3.1

[0022] Step 1.4 (optional): During the model training phase, data augmentation strategies can be selectively implemented, including: thermal noise simulation, contrast adjustment and brightness perturbation, translation and rotation operations, small object scaling, and random occlusion.

[0023] Among them, the specific operations for feature extraction and optimization in the backbone part of the model in Step 2 are as follows:

[0024] Step 2.1: The backbone network stem layer (Stem Layer) receives the preprocessed normalized image and performs preliminary feature extraction through the SRConvModule, which includes:

[0025] Step 2.1.1: The two-dimensional convolution (Conv2d) uses a 3×3 depthwise separable convolutional layer. Assuming the input image dimension is H×W×C and the convolutional kernel size is k, the calculation follows the following formula: where is the feature map of the input image, is the output feature map of the convolution operation, is the weight of the convolutional kernel, is the bias term, and c m is the number of input channels.

[0026] Step 2.1.2: The normalization layer (BatchNorm2d) normalizes the features of each channel. Calculate the mean and variance of all elements within the channel, and then normalize each feature value according to the mean and variance: where is as explained in Step 2.1.1, where γ and β are learnable parameters used to adjust the normalized feature distribution.

[0027] Step 2.1.3: The StarReLU activation function, and its non-linear transformation formula is: StarReLU(x) = s·(ReLU(x)) 2 +b where s ∈ R and b ∈ R are set as globally shared learnable parameters with initial values set respectively, and ReLU(x) = max(0, x).

[0028] Step 2.2: Stage Layer 1 generates the feature map P2 by combining SRConvModule and Cross Stage Partial Network (C2f). The feature map P2 is directly fed into the small object detection head H2 or passed to the neck of the model.

[0029] Step 2.3: Stage Layer 2 receives the output feature map of Stage Layer 1, reuses the SRConvModule + C2f architecture of Stage Layer 1 to extract the feature P3, and feeds the extraction result into the next layer.

[0030] Step 2.4: Stage Layer 3 receives the output feature map of Stage Layer 2. First, it performs a channel downsampling operation through the SCDown module to reduce the spatial resolution of the feature map. Subsequently, it uses C2f to extract the feature P4 and passes the result to the next layer.

[0031] Step 2.5: Stage Layer 4 receives the output feature map of Stage Layer 3. First, it performs a channel downsampling operation through the SCDown module to reduce the spatial resolution of the feature map. Subsequently, it uses Adaptive Wavelet Base Convolution (AWBC) to extract stage features.

[0032] Step 2.5.1: According to the relative energy distribution matrix E rel (i, j) calculated in Step 1.3, set the dynamic thresholds: T h = μ rel + σ rel , T l = μ rel - σ rel where μ rel , σ rel are the mean and standard deviation of the relative energy of all sub-regions respectively, T h is the high dynamic threshold, and T l is the low dynamic threshold. Region classification rule: where, E rel(i,j) is the same as the explanation in Step 1.3.4 Wavelet basis dynamic allocation strategy: Among them, the wavelet basis selection for each area is adjusted according to the actual hardware level, and the principle of selecting high-frequency wavelet basis in the high-frequency area, medium-frequency or low-frequency wavelet basis in the medium-frequency area, and skipping the subsequent wavelet transform in the low-frequency area can be satisfied.

[0033] Step 2.5.2: Adaptive wavelet decomposition, perform two-dimensional discrete wavelet transform of the selected wavelet basis on each sub-region R(i,j): Among them, DWT ψ(i,j) represents the discrete wavelet transform operator, are the four components obtained after the discrete wavelet transform, namely the low-frequency component, the horizontal high-frequency component, the vertical high-frequency component, and the diagonal high-frequency component.

[0034] Step 2.5.3: Multi-frequency feature processing, perform lightweight convolution on each decomposed component: Among them, are the four components obtained by wavelet decomposition in Step 2.5.2, are the feature maps obtained after lightweight convolution processing, corresponding to different wavelet decomposition components.

[0035] Step 2.5.4: Complexity-driven depth control, calculate the complexity index C of the current feature map rel : Among them, E rel is the element in the relative energy distribution matrix calculated in Step 1.3, is the average value of all elements in the relative energy distribution matrix E rel (i,j), and N is the total number of sub-regions. Dynamically adjust the wavelet decomposition depth, set the maximum decomposition depth D according to the scene requirements max , and adjust the decomposition depth D with the complexity index C rel : Among them, γ is the depth adjustment coefficient.

[0036] Step 2.5.5: Cascade decomposition, recursively perform wavelet decomposition: Among them, are the four components in the k-th layer wavelet decomposition, The low-frequency component obtained from the (k-1)-th layer wavelet decomposition serves as the input for the k-th layer wavelet decomposition.

[0037] Step 2.5.6 (optional): Wavelet component weight optimization. According to the current decomposition layer k and complexity index C rel , dynamically calculate the fusion weights of each frequency component: where is the fusion weight corresponding to each frequency component in the k-th layer wavelet decomposition. The optimized component expression is: where is the optimized component in the k-th layer, is the feature map obtained after performing lightweight convolution (Step 2.5.3) on each frequency component in the k-th layer wavelet decomposition.

[0038] Step 2.5.7: Depthwise convolution and inverse transform. Perform depthwise separable convolution on the optimized component and perform inverse discrete wavelet transform (IDWT): where is the feature map obtained after depthwise separable convolution and inverse discrete wavelet transform in the k-th layer.

[0039] Step 2.5.8: Cascade feature fusion. Fuse the decomposition features of each level obtained recursively to construct the final feature map Z (i) : Recursive termination condition: When i = D, where Z (i) is the feature map obtained after cascade feature fusion at the i-th level, is the low-frequency feature map obtained after depthwise convolution and inverse transform (Step 2.5.7) in the i-th layer, is the low-frequency part in the feature map obtained after cascade feature fusion at the (i + 1)-th level, is the high-frequency feature map obtained after depthwise convolution and inverse transform in the i-th layer.

[0040] where the specific operation of optimizing features by the multi-granularity pooling mechanism in Step 3 is:

[0041] Step 3.1: Input the stage feature P5 into the Multi-Granularity Pooling Optimization Module (MGPO) and perform Global Aggregation Pooling (GAP). The calculation formula is as follows: where is the input feature map, is the global feature vector, H and W are the height and width of the input feature map F respectively, c is the channel index of the feature map F and the global feature vector g, and the same definitions apply to the subsequent steps.

[0042] Step 3.2: Broadcast the global feature vector to the original spatial dimension to generate a global feature map Preferably, the cross-attention mechanism can be used instead of the broadcast method. The cross-attention obtains the enhanced global feature map G' by calculating the interaction between the global feature g c and the input feature map F: G' = Attention(g c , F) where Attention is the cross-attention mechanism.

[0043] Step 3.3: Perform Segment Max-Pooling (SMP): Divide the feature map into a K×K grid, and the size of each sub-region is: where H s , W s are the height and width of each sub-region after dividing the feature map into a K×K grid. Extract the sub-region features: where s k,c is the maximum eigenvalue of the c-th channel in the k-th sub-region, and segment k is the k-th sub-region. Restore the spatial dimension by nearest neighbor interpolation upsampling: where is the segment-level feature after interpolation, and Upsample nearest is the nearest neighbor interpolation upsampling operation.

[0044] Step 3.4: Local Context Pooling (LCP) uses sliding window max pooling: Among them, L[c,i,j] is the element in the c-th channel, the i-th row, and the j-th column of the local context feature map L, and F[c,u,v] is the element in the c-th channel, the u-th row, and the v-th column of the input feature map F.

[0045] Step 3.5: Multi-granularity feature fusion: Among them, P is the feature map after multi-granularity fusion, is the channel dimension concatenation operation. If cross-attention is adopted, it is: Among them represents the channel dimension concatenation operation.

[0046] Among them, the specific operation of feature aggregation and propagation in Step 4 is:

[0047] Step 4.1: Bottom-up feature aggregation. Starting from the deepest feature map of the backbone part, upsampling is used to increase the resolution, and feature maps of different levels are fused through channel concatenation (Concat). After the fused feature map is processed by the C2f module, feature P3 is generated and input into the corresponding detection head. Preferably, to meet the small target detection requirement, the stage feature P3 is upsampled again and concatenated with the feature of stage layer 2 to generate feature P2, and the fused feature is input into the small target detection head H2 to enhance the multi-scale information support.

[0048] Step 4.2: Top-down feature propagation. Starting from the shallow feature map, downsampling (SCDown) is used to reduce the resolution, and it is fused with the next-layer feature map layer by layer through channel concatenation (Concat). During the propagation process, the C2f and AWBC modules are used to extract features from stage layer 4 and stage layer 5 respectively to generate features P4 and P5, which are input into the corresponding detection heads respectively.

[0049] Among them, the specific operation of performing classification and bounding box regression prediction on multiple detection heads in Step 5 is:

[0050] Step 5.1: Perform two-dimensional discrete wavelet transform (DWT) on the feature map P2 to decompose the feature F of each channel c as: [LL (c) ,LH (c) ,HL (c) ,HH (c) = DWT(Fc ) Among them, the low-frequency component LL (c) remains unchanged, and only the high-frequency components LH (c) , HL (c) , HH (c) are processed.

[0051] Step 5.2: Dynamically divide the noise area, and calculate the local noise variance σ 2 (x, y) of the high-frequency component using an n×n sliding window as follows: where μ window is the average value of the pixels within the window, and the boundary area is processed using mirror filling; n is the size of the sliding window, x and y are the coordinates of the pixels in the high-frequency component, i and j are the coordinates for traversing within the n×n sliding window, and X(i, j) is the pixel value of the high-frequency component at the coordinate (i, j). Generate a dynamic threshold based on the characteristics of the noise variance distribution: T h = Φ h (σ 2 ), T l = Φ l (σ 2 ) where Φ h , Φ l is a threshold generation function constructed according to the noise statistical characteristics, T h is the dynamic high threshold, and T l is the dynamic low threshold.

[0052] Step 5.3: Direction-aware threshold processing, perform differential denoising on each high-frequency component. For the horizontal component LH: The vertical component HL and the diagonal component HH adopt the same strategy, and the threshold is scaled according to the component characteristics; After the horizontal component LH undergoes direction-aware threshold processing, LH(x, y) is the original pixel value at the coordinate (x, y) in the horizontal component LH, and sign(LH(x, y)) is the sign function of LH(x, y), which is used to retain the sign of the pixel value and maintain its positive or negative nature when the pixel value is shrunk.

[0053] Step 5.4: Feature map reconstruction, perform inverse discrete wavelet transform (IDWT) on the denoised high-frequency components: where, The feature after denoising of the c-th channel is reconstructed to obtain the feature. The feature map is obtained after reconstruction They are the results after the horizontal, vertical, and diagonal high-frequency components of the c-th channel are processed (denoised) by the direction perception threshold

[0054] Step 5.5: Parallel prediction by multiple detection heads. Input the feature map into the detection head network in the following way: Detection head H2: → (class probability, bounding box, confidence) Detection head H3: P3 → (class probability, bounding box, confidence) Detection head H4: P4 → (class probability, bounding box, confidence) Detection head H5: P5 → (class probability, bounding box, confidence)

[0055] Among them, the specific operation of context-aware weighted box fusion in step 6 is as follows:

[0056] Step 6.1: Initial screening of bounding boxes. Set a confidence threshold τ to screen the confidence C of each bounding box i : B′ = {B i |C i ≥ τ} Among them, B' is the set of bounding boxes obtained after screening, and B i represents the i-th bounding box. Select the bounding box with the highest confidence as the reference box B ref : Among them, is to find the bounding box B that obtains the maximum value i .

[0057] Step 6.2: Geometric consistency evaluation. Calculate the CIOU between the candidate box and the reference box: Among them, ρ is the Euclidean distance between the center points of the bounding boxes, c is the length of the diagonal of the minimum enclosing box, v is the difference measurement factor of the aspect ratio of the bounding boxes, and α is a hyperparameter; IoU is the ratio of the intersection area to the union area of the candidate box and the reference intersection, and CIoU(B i , B ref ) represents the Complete Intersection over Union (CIoU) between the candidate box and the reference box. It comprehensively considers the overlapping area, center point distance, and aspect ratio of the bounding boxes.

[0058] Step 6.3: Background influence quantification. Based on the background feature map F extracted by the backbone network bg , calculate the background influence factor BG i : Among them, F bg (h, w) represents the feature response intensity at position (h, w), is an indicator function indicating whether the pixel (h, w) falls within the bounding box B i inside, H and W are the height and width of the background feature map F bg .

[0059] Step 6.4: Adaptive weight calculation, the fusion weight formula: w i = (C i ·β)·exp(-(α·CIoU(B i , B ref ) + γ·BG i )) Among them, w i is the fusion weight of the i-th bounding box, β and γ are hyperparameters, exp is the exponential function, and BG i is the background influence factor.

[0060] Step 6.5: Bounding box fusion, the final fused bounding box B fused and the corresponding confidence C fused are obtained from the weighted fusion formula: where N is the number of bounding boxes participating in the weighted fusion. Description of the Drawings

[0061] Figure 1 The overall flowchart of the present invention is shown as follows.

[0062] Figure 2 The AW-YOLO network structure diagram of the present invention is shown as follows.

[0063] Figure 3 The AWBC structure diagram (adaptive wavelet basis convolution structure diagram) in the AW-YOLO network of the present invention is shown as follows.

[0064] Figure 4 The MGPO Module structure diagram (multi-granularity pooling optimizer structure diagram) in the AW-YOLO network of the present invention is shown as follows. Detailed Embodiments

[0065] The technical solutions of the present invention will be further described in detail below with reference to the drawings and embodiments, but the present invention is not limited to the following embodiments.

[0066] As Figures 1 to 4 shown:

[0067] Embodiment: The present invention provides an infrared anti - UAV detection method (AW - YOLO) based on an improved YOLO framework, and its specific implementation process is as follows:

[0068] Step 1: Infrared image pre - processing and feature optimization

[0069] Step 1.1: Image standardization processing:

[0070] Step 1.1.1: Input source adaptation, receive the real - time video stream of the infrared camera or the labeled infrared data set, and uniformly convert it into a tensor format.

[0071] Step 1.1.2: Isometric scaling, scale the image while maintaining the original aspect ratio, and set the target size to S w ×S h .

[0072] Step 1.1.3: Symmetric padding, when the size of the scaled image is smaller than the target size, use neutral gray values for edge padding, satisfying:

[0073] Step 1.2: Pixel value normalization, perform linear normalization processing: where k is the bit - depth of the infrared sensor (usually k = 8, I raw ∈[0,255]).

[0074] Step 1.3: Multi - level energy analysis

[0075] Step 1.3.1: Global energy benchmark calculation:

[0076] Step 1.3.2: Dynamic grid division, divide the image into m×n blocks according to the imaging size of typical targets:

[0077] Step 1.3.3: Integral image acceleration calculation, construct the integral image I int and calculate the local energy:

[0078] Step 1.3.4: Relative energy matrix generation: E rel (i,j)=E local (i,j)-E global

[0079] Step 1.4 (Optional): Robustness enhancement, implemented during the training phase: Noise injection, adding random noise that conforms to the physical characteristics of infrared; photometric distortion, randomly adjusting contrast and brightness, geometric transformation: implementing a combined transformation of translation, rotation, and scaling; target occlusion, simulating local occlusion to enhance the robustness of small target detection.

[0080] Step 2: Multi-scale feature extraction and optimization of the backbone network

[0081] Step 2.1: Stem layer feature extraction, performed using the improved SRConvModule:

[0082] Step 2.1.1: Input tensor Convolution kernel Output: Y = DepthwiseConv2D(X; W) + b

[0083] Step 2.1.2: Batch normalization processing, performing standard batch normalization on the output of depthwise separable convolution:

[0084] Step 2.1.3: Parametric activation, introducing learnable parameters Perform: StarReLU(Z) = s·(ReLU(Z)) 2 +b

[0085] Step 2.2: Stage layer 1 feature extraction: Receive the output feature F of the stem layer stem , generate the feature map P2 through the SRConvModule and the C2f module, and perform path decision-making: Parameter definition: T process is the maximum allowable processing delay for a single frame, τ realtime Real-time threshold (default value: 30ms).

[0086] Step 2.3: Stage layer 2 feature extraction, reuse the output feature map P3 of the stage layer 1 architecture, and reduce the size at the same time.

[0087] Step 2.4: Stage layer 3 feature extraction, compress the feature map through the spatial-channel downsampling module (SCDown): Then generate the feature map P4 through the C2f module

[0088] Step 2.5: Adaptive wavelet processing for stage layer 4:

[0089] Step 2.5.1: Frequency domain region division Dynamic threshold calculation: T h = μ rel + σ rel , T l = μ rel - σ rel Region classification and basis function mapping: The default high-frequency wavelet basis is Haar, and the default medium-low frequency wavelet basis is Daubechies3, which can be modified according to actual needs.

[0090] Step 2.5.2: Adaptive wavelet decomposition, perform two-dimensional discrete wavelet transform: [X LL , X LH , X HL , X HH = DWT ψ (R(i,j))

[0091] Step 2.5.3: Multi-frequency feature processing, component convolution processing: Y ** = DepthwiseConv2D(X ** ), **∈{LL,LH,HL,HL}

[0092] Step 2.5.4: Decomposition depth control Calculation complexity index: Dynamic depth adjustment:

[0093] Step 2.5.5: Recursive feature processing Recursive decomposition: Convolution enhancement:

[0094] Step 2.5.6: Frequency domain weight optimization Dynamic weight allocation: w = Softmax([1 - C rel , C rel , C rel , C rel ) Component fusion:

[0095] Step 2.5.7: Inverse wavelet transform:

[0096] Step 2.5.8: Cross-cascade feature fusion: Output the stage features. The finally generated feature map P5 is used as the output of stage layer 4 and fed into the subsequent feature fusion module.

[0097] Step 3: Multi-granularity feature enhancement.

[0098] Step 3.1: Global aggregation pooling (GAP), input the stage features To the multi-granularity pooling optimizer module (MGPO), perform global average pooling: Generate the global feature vector where F is the input feature map, and H, W, and C are the height, width, and number of channels respectively.

[0099] Step 3.2: Spatial attention enhancement (optional embodiment), interact the global feature vector g with the input feature map through the cross-attention mechanism.

[0100] Step 3.2.1: Generate the query vector: is a learnable weight matrix

[0101] Step 3.2.2: Extract key-value pairs:

[0102] Step 3.2.3: Calculate the attention weights:

[0103] Step 3.3: Segment maximum pooling (SMP)

[0104] Step 3.3.1: Spatial partitioning, evenly divide the feature map into K×K rectangular grids, and the sub-region size is calculated as:

[0105] Step 3.3.2: Segment pooling, perform maximum pooling on each sub-region: Generate the regional feature map

[0106] Step 3.3.3: Use bilinear interpolation to upsample S to the original size: S′ = Upsample(S; scale factor= K, mode = 'bilinear') Output the reconstructed feature map

[0107] Step 3.4: Local context pooling (LMP), using sliding window max pooling to capture fine-grained features. Define the window radius p (default p = 1), and perform the following operations for each spatial position (i, j): For boundary handling, use the mirror padding strategy to output the local feature map

[0108] Step 3.5: Multi-granularity feature fusion, concatenate the features of each granularity in the channel dimension: P = Concat dim＝1 (F, G (′) , S′, L) where G (′) represents either the original global feature G or the attention-enhanced feature (from Step 3.2).

[0109] Step 4: Feature aggregation and propagation

[0110] Step 4.1: Bottom-up feature aggregation. Start from the feature map of the deepest layer of the backbone network (denoted as F5), perform a 2x upsampling operation; concatenate the upsampling result with the adjacent shallow feature map (F4) in the channel dimension; process the concatenated features through the standard C2f module to output the intermediate fusion feature; repeat the (upsampling - concatenation - C2f) processing flow, and finally generate the feature P3 and input it into the detection head H3; (optionally) perform a secondary upsampling on P3 and concatenate it with the shallower feature F2 to generate the small object detection feature P2.

[0111] Step 4.2: Top-down feature propagation. Start from the effective feature map (F2) of the shallowest layer of the backbone network, perform downsampling using strided convolution; concatenate the downsampled feature with the deep feature in the channel dimension; process it sequentially through the C2f module and the standard AWBC module to generate the features P4 and P5 respectively; the output features of each stage are input into the corresponding detection heads H4 and H5.

[0112] Step 5: Multi-detection head object detection

[0113] Step 5.1: Wavelet domain feature decomposition, perform two-dimensional discrete wavelet transform on the input feature map channel by channel: The low-frequency component LL (c) keeps the original value unchanged.

[0114] Step 5.2: Adaptive noise modeling

[0115] Step 5.2.1: Calculate the spatial domain noise variance of the high-frequency component using a 5×5 sliding window (n = 5):

[0116] Step 5.2.2: Dynamic threshold generation High noise threshold: T h = Q 90 (σ 2 ) + 1.5(Q 90 - Q 10 ) Low noise threshold: T l = Q 50 (σ 2 ) where is the global noise variance matrix.

[0117] Step 5.3: Direction-aware denoising

[0118] Step 5.3.1: Horizontal component processing (LH):

[0119] Step 5.3.2: Vertical component (HL) and diagonal component (HH) processing Threshold scaling factor: α HL = 1.2, α HH = 0.8 Adjust the threshold:

[0120] Step 5.4: Feature map reconstruction

[0121] Step 5.4.1: Perform the inverse discrete wavelet transform:

[0122] Step 5.4.2: Post-processing, add a 1×1 convolutional layer for channel calibration:

[0123] Step 5.5: Parallel prediction of the detection head Process the small object detection head H2, input the denoised feature map into the small object detection head H2, and generate the class probability, bounding box coordinates, and confidence of the small object through a lightweight convolutional network. Process the large and medium target detection heads H3, H4, and H5. Input the feature maps P3, P4, and P5 into the corresponding detection heads H3, H4, and H5 respectively. Each detection head independently processes feature maps of different scales and synchronously outputs class probabilities, bounding box parameters, and confidence levels through convolutional layers.

[0124] Step 6: Context-aware weighted box fusion implementation and result output

[0125] Step 6.1: Confidence threshold screening and reference box selection

[0126] Step 6.1.1: Confidence screening. Sort all bounding boxes by confidence and filter out bounding boxes with a confidence lower than the threshold τ to obtain a candidate set: B′ = {B i |C i ≥ τ} Reference box selection. Among the filtered bounding boxes, select the bounding box B ref with the highest confidence as the reference box, and the remaining bounding boxes form the set of bounding boxes to be weighted and fused {B1, B2,..., B N}, where.

[0127] Step 6.1.2: Reference box determination. Select the bounding box with the highest confidence in the candidate set as the reference box:

[0128] Step 6.2: Geometric consistency evaluation. For each candidate box B i ∈ B′, calculate its CIOU value with the reference box:

[0129] Step 6.3: Background influence factor calculation

[0130] Step 6.3.1: Background feature extraction. Calculate the global threshold T = median(F) of the feature map to generate the background feature map:

[0131] Step 6.3.2: Region response calculation. For each bounding box B i , extract the background feature region F bg (B i ) it covers and calculate the background influence factor:

[0132] Step 6.4: Adaptive weight calculation. Integrate confidence, geometric consistency, and background influence to calculate the weight of each box:: w i = (C i·β)·exp(-(α·CIoU(B i ,B ref )+γ·BG i )) where α, β, and γ are adjustable hyperparameters.

[0133] Step 6.5: Bounding box fusion and result output Location fusion: Confidence fusion: Generate the final detection result, including the target bounding box position (x, y, w, h), class label, and fused confidence C fused , for downstream tasks such as UAV monitoring and tracking.

[0134] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and without departing from the spirit or basic characteristics of the present invention, the present invention can be implemented in other specific forms. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be encompassed within the present invention. Any reference signs in the claims should not be construed as limiting the claims involved.

[0135] In addition, it should be understood that although this specification is described according to embodiments, not every embodiment only contains an independent technical solution. This narrative manner of the specification is only for clarity. Those skilled in the art should regard the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. An infrared anti-UAV detection method based on an improved YOLO framework, characterized in that, It includes the following steps: Step 1: Preprocess the input infrared image to generate standardized data that meets the model input requirements, and calculate the relative energy distribution of the image to provide a basis for subsequent feature extraction and optimization; Step 2: Embed an improved convolutional module (StarReLU Convolution Module, SRConvModule) and an Adaptive Wavelet Base Convolution (AWBC) including adaptive wavelet basis selection, complexity calculation, cascaded wavelet decomposition, and feature fusion in the model backbone network, which are respectively used for feature extraction and optimization in different stages; Step 3: Introduce a feature optimization mechanism based on multi-granularity pooling to capture local details and global contour information on the feature map through pooling windows of different scales, refine the detail features, and strengthen the expression of global contour and spatial position information; Step 4: Achieve feature fusion through the feature aggregation and propagation mechanism, complete the interaction of features at different levels, and provide complete multi-scale feature support for multiple detection heads; Step 5: Combine the optimized multi-detection head design, and through the lightweight detection branch designed for small targets, perform classification and bounding box regression prediction on feature maps of different scales, improving the detection accuracy of small targets while avoiding interference with other detection heads; Step 6: After the bounding box decoding operation, apply the Context-Aware Weighted Box Fusion (CAWBF) algorithm to screen and fuse the outputs of the detection heads through weighted calculations of confidence, geometric information, and background influence factors, optimize the bounding box positioning, and remove redundancy.

2. The infrared anti-UAV detection method according to claim 1, wherein Step 1 includes the following steps: Step 1.1: Scale the input infrared image proportionally to the preset input size. When the size of the scaled image is smaller than the target size, symmetric padding is performed using the gray value 128; Step 1.2: Normalize the scaled image to map the pixel values to the interval [0, 1]; Step 1.3: Calculate the global energy mean E of the normalized image global , divide the image into an m×n grid, and calculate the local energy mean E local (i, j) of each sub-region R(i, j), and finally calculate the relative energy distribution: E re l (i, j) = E local (i, j) - E global 3. The infrared anti-UAV detection method according to claim 1, wherein Step 2 includes the following steps: Step 2.1: In the Stem Layer of the backbone network, use the improved convolutional module (SRConvModule) for preliminary feature extraction. The SRConvModule is composed of a 3×3 depthwise separable convolutional layer, a batch normalization layer, and a StarReLU activation function connected in sequence; Step 2.2: In Stage Layer 1, extract the first-level feature map P2 by combining the SRConvModule and the C2f module, and send the feature map P2 to the neck network or the small target detection head H2; Step 2.3: In Stage Layer 2, reuse the combination of the SRConvModule and the C2f module in Stage Layer 1 to generate the second-level feature map P3; Step 2.4: In Stage Layer 3, generate a feature map with a halved spatial resolution through channel downsampling, and output the third-level feature map P4 after being processed by the C2f module; Step 2.5: In the Stage Layer 4, an Adaptive Wavelet Base Convolution (AWBC) operation is adopted, and after processing, the third-level feature map P5 is output.

4. The infrared anti-UAV detection method according to claim 1, wherein , The wavelet basis selection in the AWBC described in Step 2 includes: Dynamically select the optimal wavelet basis for each sub-region R(i, j) of the feature map according to the relative energy distribution and energy threshold allocation: (1) Select high-frequency wavelet bases in the high-frequency region to extract detailed information; (2) Select medium-frequency or low-frequency wavelet bases in the medium-frequency region to maintain edge and structural features; (3) Skip the subsequent wavelet transform operation in the low-frequency region; After the allocation is completed, each sub-region uses the corresponding wavelet basis for subsequent wavelet convolution operations.

5. The infrared anti-UAV detection method according to claim 1, characterized in that , The wavelet decomposition link in the Adaptive Wavelet Base Convolution (AWBC) adopted in Step 2 includes: (1) Apply the corresponding wavelet basis to each sub-region R(i, j) of the input feature map to perform wavelet transform, and decompose it into four frequency components: [X LL ,X LH ,X HL ,X HH =WT(X) where it is X LL low-frequency component, X LH , X HL , X HH are the high-frequency components in the horizontal, vertical, and diagonal directions respectively; (2) Perform lightweight convolution operations on the decomposed frequency components to extract multi-frequency features: [Y LL ,Y LH ,Y HL ,Y HH = Conv(W, [X LL ,X LH ,X HL ,X HH ); (3) Fuse the processed low-frequency and high-frequency components to generate an enhanced feature map: Z = Concat([Y LL , Y LH , Y HL , Y HH ) 6. The infrared anti-UAV detection method according to claim 1, wherein , The complexity calculation link in the AWBC described in Step 2 includes: Based on the divided sub-region R(i,j), through the relative energy distribution E rel (i,j), calculate the complexity index C rel : Among them The average relative energy value of all sub-regions, where N is the total number of sub-regions.

7. The infrared anti-UAV detection method according to claim 1, characterized in that , The cascaded wavelet decomposition link in the AWBC described in Step 2 includes: According to the scene, preset the maximum decomposition depth D max , and at the same time introduce a dynamic selection mechanism, using the complexity index C rel Dynamically adjust the depth D of the cascaded wavelet decomposition. The depth adjustment formula is: where γ is the depth adjustment coefficient, which is used to control the sensitivity of complexity to depth; Recursively perform cascaded wavelet decomposition, using the low-frequency component of the previous level as the input, and decompose it into four frequency components of the current layer: where WT(X) is the input multi-frequency component, and the decomposition layer number is controlled by the depth D.

8. The infrared anti-UAV detection method according to claim 1, wherein Step 3 includes the following steps: Step 3.1: Input the P5 feature map into the Multi-Granularity Pooling Optimization Module (MGPO), perform Global Aggregation Pooling (GAP), and perform global average pooling operation on the input feature map F. Its calculation formula is: where g c is the global feature vector of channel c, and H and W are the height and width of the feature map, respectively; Step 3.2: Broadcast the global feature, broadcasting the global feature vector g c to the same spatial dimension as the input feature map to generate a global feature map G, and its calculation formula is: Step 3.3: Segment Max-Pooling (SMP) and upsampling. Divide the input feature map F into K sub-regions, and the size of each sub-region is Perform a max-pooling operation within each segment to extract the segment feature S, and its calculation formula is: s k,c = max{F[c, i, j] | (i, j) ∈ segment k} Upsample the segment feature S using nearest neighbor interpolation to obtain a segment feature map S' that matches the scale of the input feature map, so as to extract and restore local features; Step 3.4: Local Context Pooling (LCP), perform sliding window max pooling on the input feature map F with a window radius of p to generate a local feature map The calculation formula is as follows: L[c,i,j] = max{F[c,u,v]|u∈[i-p,i+p],v∈[j-p,j+p]} where zeros are automatically filled when the sliding window boundary exceeds the feature map range; Step 3.5: Fuse the global feature map G, the interpolated segment-level feature map S', the local feature map L and the original input feature map F pixel by pixel to obtain the multi-granularity pooled feature map P.

9. The infrared anti-UAV detection method according to claim 1, wherein Step 5 includes the following steps: Step 5.1: Perform a two-dimensional discrete wavelet transform on the feature map P2, and decompose the feature F of each channel c into low-frequency components LL and high-frequency components LH, HL, HH. Only the high-frequency components will be processed subsequently, and the original low-frequency components will remain unchanged; Step 5.2: Calculate the local noise variance of the high-frequency components using the sliding window method, and define the noise variance matrix σ 2 (x, y), set a dynamic threshold based on the percentile of the noise variance distribution, and divide the high-frequency components into high-noise regions and low-noise regions; Step 5.3: Direction-aware thresholding, and perform the following differential denoising strategies on the high-frequency components LH, HL, HH respectively: where the thresholds of the vertical component HL and the diagonal component HH are scaled and optimized according to their direction characteristics. By dynamically judging the relative relationship between the local variance and the threshold, adaptive discrimination processing of high / low noise regions is realized. Step 5.4: For the denoised high-frequency components Perform an inverse discrete wavelet transform with the original low-frequency component LL to reconstruct the feature map of the high-frequency components and generate a denoised feature map Step 5.5: Input the feature maps P3, P4, and P5 into the detection head network, and respectively output class probabilities, bounding box information, and confidence scores through parallel branches.

10. The infrared anti-UAV detection method according to claim 1, characterized in that Step 6 includes the following steps: Step 6.1: Perform confidence threshold screening and reference box selection to remove bounding boxes with confidence lower than the set threshold. Select the bounding box with the highest confidence as the reference box B ref , and form a bounding box set with the remaining bounding boxes to participate in weighted fusion; Step 6.2: Calculate the CIOU value of each candidate bounding box B i and the reference bounding box B ref ; Step 6.3: Calculate the background influence factor, and calculate the background influence factor BG of each bounding box i , and the formula is: Among them, F bg (h, w) represents the feature response intensity at the position (h, w), is an indicator function, indicating whether the pixel (h, w) falls within the bounding box B i inside; Step 6.4: Calculate the weight w of each bounding box based on the CAWBF method i , and the formula is: w i = (C i ·β)·exp(-(α·CIoU(B i , B ref ) + γ·BG i )) where C i is the confidence of the bounding box, and α, β, γ are adjustment parameters; Step 6.5: According to the weight w i Calculate the fused bounding box B fused and the corresponding confidence as the output result. The formula is as follows: where N is the number of bounding boxes participating in weighted fusion.

Images