A method and device for detecting an airborne target based on wavelet transform
Patent Information
- Application Number
- CN202510537264.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-27
- Publication Date
- 2026-10-27
AI Technical Summary
[0006]鉴于上述的分析，本发明实施例旨在提供一种基于小波变换的机载目标检测方法和装置，用以解决现有技术中无法对小目标进行高性能目标检测的问题
[0044]1、本发明基于无人机机载端目标检测任务，针对无人机视角下小目标占比大的问题，提出了通过主干网络的多个卷积块，生成多尺度特征图，并在部分层间引入小波变换模块(WTConv模块)，由此得到的高频流和低频流特征来强调物体边界和纹理，使目标检测模型具有良好的结合上下文信息的能力，提高对小目标的检测准确率。
Smart Images
Figure CN122990525A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image processing technology, and in particular to an airborne target detection method and apparatus based on wavelet transform. Background Technology
[0002] With the increasing prevalence of unmanned aerial vehicles (UAVs), camera-equipped drones have been rapidly deployed in a wide range of applications, including mountain rescue, border surveillance, sample collection, and agricultural operations. Therefore, automatically understanding and processing the visual data collected by these platforms through computer vision has become a crucial and indispensable task. Object detection, which aims to identify all objects of interest in an image and determine their category and location, is one of the core problems in computer vision. The core of object detection lies in solving four problems: classification, localization, size, and shape. Due to the different appearances, shapes, and poses of various objects, coupled with interference from factors such as lighting and occlusion during imaging, object detection has always been one of the most challenging problems in computer vision. It is the cornerstone of downstream applications (such as object tracking, action recognition, and other automated visual understanding tasks), and therefore has long been a focus of research in this field.
[0003] With advancements in deep learning, general object detection has achieved remarkable performance on public benchmark datasets such as MSCOCO, PASCAL, and ImageNet. However, applying object detection to images captured by drones presents unique challenges. Specifically, most targets captured from a drone's perspective are small objects. Small objects are often considered complex samples, and conventional object detection methods tend to perform poorly. This performance degradation is primarily attributed to two factors: the inability to capture fine-grained details of small objects and the underutilization of contextual information.
[0004] To address the problem of insufficient utilization of contextual information, researchers both domestically and internationally have conducted a series of studies. Various context-aware enhancement modules have been proposed to improve the model's utilization of contextual information, combining high-level and low-level contextual features or capturing global features. These methods enhance contextual relationships through various modules.
[0005] To address the lack of fine-grained information, an intuitive approach is to increase image resolution. However, cropping images to increase resolution leads to wasted computation on background regions. Taking the VisDrone dataset as an example, targets are sparsely distributed in drone-captured images, while small objects are typically concentrated in specific areas. Existing research focuses on extracting detailed features such as texture from images; however, it fails to address the problem of accurate detection and localization of small objects due to their weak feature responses. Summary of the Invention
[0006] In view of the above analysis, the present invention aims to provide an airborne target detection method and apparatus based on wavelet transform, so as to solve the problem that the prior art cannot perform high-performance target detection for small targets.
[0007] The objective of this invention is mainly achieved through the following technical solutions:
[0008] On one hand, embodiments of the present invention provide an airborne target detection method based on wavelet transform, comprising the following steps:
[0009] The image to be detected containing the target is input into the backbone network, and multi-scale feature extraction is performed using multiple convolutional blocks. The wavelet transform module is used to extract low-frequency flow features and high-frequency flow features of at least some scale feature maps respectively, so as to obtain low-frequency flow feature maps and high-frequency flow feature maps of the corresponding scales.
[0010] Multi-scale feature fusion is performed on the low-frequency flow feature map and the high-frequency flow feature map using the low-frequency branch and the high-frequency branch of the feature fusion network, respectively.
[0011] The detection head is used to predict the target by combining the multi-scale low-frequency flow fusion feature map and the high-frequency flow fusion feature image, and the predicted target type and bounding box location are obtained.
[0012] Furthermore, the wavelet transform module is used to extract the corresponding low-frequency flow features, including:
[0013] The feature map of the input wavelet transform module is filtered by low-frequency wavelet and high-frequency wavelet respectively to obtain the corresponding low-frequency components and high-frequency components.
[0014] Convolutional blocks are used to perform convolution operations and scale adjustments on the low-frequency and high-frequency components respectively to obtain the corresponding low-frequency enhancement features and high-frequency enhancement features.
[0015] Perform inverse wavelet transform on the low-frequency enhancement features and high-frequency enhancement features to obtain the inverse-transformed features;
[0016] The feature map of the input wavelet transform module is convolved, and the convolved feature map is concatenated with the inverse transform feature to obtain the low-frequency flow feature.
[0017] Furthermore, the low-frequency and high-frequency components are extracted in the following manner:
[0018] x LL ,x LH ,x HL ,x HH =WT(x,filters={ψ L ,ψ H}),
[0019] Where x is the feature map of the input wavelet transform module; ψ L and ψ H These are low-frequency and high-frequency wavelet filters, respectively; x LL Represents the low-frequency component of x; x LH x HL and x HH These are the high-frequency components corresponding to x in the horizontal, vertical, and diagonal directions, respectively.
[0020] Furthermore, the convolution operation is performed on the low-frequency components and high-frequency components based on the following formula:
[0021]
[0022] Where, x WT ∈R B×(C×4)×H×W It is x LL x LH x HL and x HH This is a combination of four-dimensional channel information; Conv2d is a depthwise separable convolution; γ is a learnable scale parameter; ⊙ represents element-wise product.
[0023] Furthermore, the high-frequency flow features are obtained by weighting and combining the high-frequency enhancement features based on the following formula:
[0024] x H =[x LH ,x HL ,x HH ]⊙ω T ,
[0025] Where, x H The high-frequency flow characteristics; x LH x HL x HH These represent the high-frequency enhancement features in the horizontal, vertical, and diagonal directions, respectively; ω T These are weighted parameters; ⊙ represents the Hadamard product.
[0026] Furthermore, the detection head is used to predict the target category and bounding box location; when predicting the target bounding box location, a combination of Focaler-IoU and CIoU is used to calculate the bounding box regression loss function based on the following formula:
[0027] L Box =L CIoU +IoU-IoU Focaler ,
[0028] Among them, L Box L is the bounding box regression loss function; CIoUCIoU is the loss function; IoU is the intersection-union ratio between the predicted bounding box and the ground truth bounding box; IoU Focaler This represents the crossover ratio loss after reconstruction.
[0029] Furthermore, the backbone network, feature fusion network, and detection head constitute the target detection model. The target detection model is trained using a joint loss function, the expression of which is:
[0030] L = L BCE +L Box +L DFL ,
[0031] Where L is the joint loss function; L BCE L is used to measure the classification loss for predicting the target category; DFL This is the distributed focus loss, used to optimize the regression of the bounding box.
[0032] Furthermore, the intersection-union ratio loss reconstructed through linear interval mapping is expressed as:
[0033]
[0034] Among them, IoU Focaler It is the reconstructed intersection-union ratio loss; d and u are the adjustment parameters, respectively.
[0035] Furthermore, the feature fusion network is obtained based on the two-stream feature pyramid network, including:
[0036] Using the upsampling layer of the centralized pyramid module of the low-frequency branch or high-frequency branch, the feature maps of the extracted low-frequency or high-frequency flow are sampled layer by layer upwards, and the output of the upsampling layer is fused with the feature maps of the corresponding low-frequency or high-frequency flow using the splicing layer respectively.
[0037] By utilizing the improved information fusion unit with low-frequency or high-frequency branches, features are passed up layer by layer from the bottom-level feature map, and fused with the low-frequency or high-frequency flow feature maps of the corresponding scale to obtain the low-frequency flow fusion feature map or the high-frequency flow fusion feature map.
[0038] On the other hand, embodiments of the present invention provide an airborne target detection device based on wavelet transform, comprising:
[0039] The image acquisition module is used to acquire the image to be detected;
[0040] The feature extraction module is used to take the image to be detected as input and perform multi-scale feature extraction using multiple convolutional blocks of the backbone network; the wavelet transform module is used to extract low-frequency flow features and high-frequency flow features of at least some scale feature maps respectively, to obtain low-frequency flow feature maps and high-frequency flow feature maps of the corresponding scales.
[0041] The feature fusion module is used to perform multi-scale feature fusion on the low-frequency flow feature map and the high-frequency flow feature map using low-frequency branch and high-frequency branch respectively;
[0042] The prediction module is used to predict targets from the fused multi-scale low-frequency flow fusion feature map and the high-frequency flow fusion feature map, and obtain the predicted target type and location results.
[0043] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:
[0044] 1. This invention is based on the target detection task on the UAV airborne end. In view of the large proportion of small targets in the UAV perspective, it proposes to generate multi-scale feature maps by using multiple convolutional blocks of the backbone network, and introduce wavelet transform module (WTConv module) between some layers. The high-frequency and low-frequency features obtained by this invention emphasize the object boundary and texture, so that the target detection model has a good ability to combine contextual information and improve the detection accuracy of small targets.
[0045] 2. The WTConv module significantly expands the receptive field of the model by combining wavelet transform with convolutional blocks, while preserving high-frequency information in the wavelet transform process and extracting weighted high-frequency flow feature maps to preserve detailed information in the image and improve the visibility of small targets.
[0046] 3. For low-frequency and high-frequency feature maps, a dual-stream feature pyramid network is designed. Each branch stream includes feature fusion of top-down and bottom-up paths. More upsampling layers are introduced for feature fusion. At the same time, the same convergence method as low-frequency features is used to supplement fine-grained details for high-frequency features to provide higher resolution feature maps.
[0047] 4. In the loss function module, Focaler-IoU is introduced. By combining CIoU and Focaler-IoU, the existing detection head is improved to enhance the ability to extract difficult samples, thereby increasing the network's attention to challenging targets.
[0048] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description
[0049] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.
[0050] Figure 1 This is a flowchart of an airborne target detection method based on wavelet transform, according to an embodiment of the present invention.
[0051] Figure 2 This is a schematic diagram of the wavelet transform target detection model according to an embodiment of the present invention;
[0052] Figure 3 This is a schematic diagram of the workflow of the WTConv module in an embodiment of the present invention;
[0053] Figure 4a This is a schematic diagram of adding WTConv pre-receptive field transform to the backbone network of the wavelet transform target detection model in an embodiment of the present invention;
[0054] Figure 4b This is a schematic diagram of the receptive field transformation after adding WTConv to the backbone network of the wavelet transform target detection model in an embodiment of the present invention.
[0055] Figure 5 This is a schematic diagram of the structure of a dual-stream FPN according to an embodiment of the present invention;
[0056] Figure 6 This is a comparative schematic diagram of the detection results of the airborne target detection method based on wavelet transform according to an embodiment of the present invention. Detailed Implementation
[0057] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.
[0058] Example 1
[0059] A specific embodiment of the present invention discloses an airborne target detection method based on wavelet transform, such as... Figure 1 As shown, it includes the following steps:
[0060] Step S1: Input the image to be detected containing the target into the backbone network and perform multi-scale feature extraction using multiple convolutional blocks; use the wavelet transform module to extract low-frequency flow features and high-frequency flow features of at least some scale feature maps respectively, to obtain low-frequency flow feature maps and high-frequency flow feature maps of the corresponding scales.
[0061] Step S2: Perform multi-scale feature fusion on the low-frequency flow feature map and the high-frequency flow feature map using the low-frequency branch and the high-frequency branch of the feature fusion network, respectively;
[0062] Step S3: Use the detection head to predict the target by combining the multi-scale low-frequency flow fusion feature map and the high-frequency flow fusion feature image, and obtain the predicted target type and bounding box position results.
[0063] The above method processes the input detection image through convolutional blocks, expands the receptive field through a wavelet transform module, and extracts low-frequency and high-frequency flow features. For the two types of features, the parallel low-frequency and high-frequency branches of the feature fusion network are used to fuse the features respectively. The detection head performs target detection based on the corresponding low-frequency and high-frequency flow fusion features, enabling the target detection model to have a good ability to combine contextual information and achieve high-performance detection of small targets.
[0064] It should be noted that the wavelet transform target detection model uses the YOLOV8 model as the basic model, which includes a backbone network, a feature fusion network, and a head network (detection head). Among them, the WTConv module is introduced on the basis of the backbone network of the basic model to enhance the detection effect of small targets through signal processing methods.
[0065] Specifically, in step S1, the backbone network includes multiple convolutional blocks (Conv), a wavelet transform module (WTConv module), a fast implementation of CSP Bottleneck with 2convolutions (C2f module), and a fast spatial pyramid pooling module (SPPF module).
[0066] For example, such as Figure 2 As shown, based on convolutional blocks, the backbone network is divided into 5 parts, namely:
[0067] The first part consists of one convolutional block and a C2f module. This convolutional block has 64 output channels, a kernel size of 3, and a stride of 2. The input image size is 640×640×3, and the output feature map size after this layer is 320×320×64, where 320×320 represents the height and width of the feature map, and 64 represents the number of channels, denoted as resolution L1.
[0068] The second part consists of one convolutional block, a WTConv module, and a C2f module. The convolutional block in this part has 128 output channels, a kernel size of 3, and a stride of 2. The input feature map size is 320×320×64, and the output feature map size is 160×160×128, denoted as resolution L2.
[0069] The third part consists of one convolutional block, a WTConv module, and a C2f module. The convolutional block in this part has 256 output channels, a kernel size of 3, and a stride of 2. The input feature map size is 160×160×128, and the L3 size of the output feature map is 80×80×256.
[0070] Part Four: This section includes one convolutional block, the WTConv module, and the C2f module. The convolutional block in this part has 512 output channels, a kernel size of 3, and a stride of 2. The input feature map size is 80×80×256, and the L4 size of the output feature map is 40×40×512.
[0071] Part 5 includes one convolutional block, a WTConv module, a C2f module, and an SPPF module. The convolutional block in this part has 1024 output channels, a kernel size of 3, and a stride of 2. The input feature map size is 40×40×512, and the output feature map size L5 is 20×20×1024.
[0072] First, each part uses convolutional blocks to adjust the resolution and number of channels of the input image. Each convolutional block consists of a convolutional layer, batch normalization, and activation function. A sliding window is used to perform convolution operations on the input image to extract local spatial features. Weight sharing reduces the number of parameters and improves computational efficiency. The extracted features are normalized using activation functions (such as ReLU) and batch normalization to enhance the model's expressive power.
[0073] Secondly, the WTConv module is used to further analyze the feature maps output by the second to fifth convolutional blocks to better expand the model's receptive field and focus on local features of a specific region. Each WTConv module includes wavelet transform, convolutional blocks, inverse wavelet transform, and a concatenation layer. By combining wavelet transform with convolutional blocks, low-frequency and high-frequency flow features corresponding to at least a portion of the scale feature maps are extracted. The workflow is as follows: Figure 3 As shown, Conv is the convolutional block, WT represents wavelet transform, IWT represents inverse wavelet transform, Output High represents the output result after high-frequency information merging, and Output represents the low-frequency stream output result (i.e., low-frequency stream features), specifically including:
[0074] S11. Perform wavelet transform on the feature map X of the input WTConv module based on the following formula to obtain the low-frequency component and the high-frequency component:
[0075] x LL ,x LH ,x HL ,x HH =WT(x,filters={ψ L ,ψ H}),
[0076] Where, ψ L and ψ H These are low-frequency and high-frequency wavelet filters, x LL This represents the low-frequency component of x, while x LH x HL and x HH These correspond to the high-frequency components in the horizontal, vertical, and diagonal directions, respectively.
[0077] S12. Using convolutional blocks, perform convolution operations and scale adjustments on these four components respectively, rescaling and supplementing fine-grained details to obtain the corresponding processed features. The convolution operation is represented as follows:
[0078]
[0079] Where, x WT ∈R B×(C×4)×H×W It is made by x LL x LH x HL and x HH This is composed of four-dimensional channel information, where C represents the feature dimension; B, H, and W represent the Batch Size, the height of the feature map, and the width of the feature map, respectively. The result after convolution (including low-frequency enhanced features x) LL 'Horizontal high-frequency incremental characteristics x' LH Vertical high-frequency incremental features x HL 'and diagonal high-frequency incremental feature x HH Conv2d is a depthwise separable convolution that processes all channels. γ is a learnable scaling parameter, ⊙ represents element-wise multiplication (or Hadamard product), which rescales the features of each channel to enhance representational power.
[0080] S13. Obtain the low-frequency feature x through inverse wavelet transform. IWT , represented as:
[0081]
[0082] S14. Perform convolution processing on the feature map of the input wavelet transform module, and use a concatenation layer to fuse the convolutionally processed feature map and the inverse wavelet transform feature map to obtain the low-frequency flow feature, represented as:
[0083] x out =x IWT +Conv(x),
[0084] Where, x outThe low-frequency stream information, i.e., low-frequency features, represents the output of the WTConv module; Conv(x) represents the input image after convolution processing.
[0085] To refine fine-grained details, the three high-frequency components, namely x, are combined after convolution. LH '、x HL 'and x HH 'Retain high-frequency information for further processing in the next stage.'
[0086] x H =[x LH ',x HL ',x HH ']⊙ω T ,
[0087] Where, ω T These are weighted parameters, such as ω T Defined as
[0088] Then, the C2f module is used to transform the features after processing the convolutional blocks or the WTConv module, which reduces the amount of computation and enhances the feature extraction capability.
[0089] Finally, the SPPF module is used to perform pooling operations on feature maps of different scales to generate multi-scale feature representations, thereby enhancing the semantic information of the features. The output multi-scale feature maps include: feature maps at resolution L1, low-frequency flow feature maps at resolutions L2-L5, and high-frequency flow feature maps at resolutions L2-L5.
[0090] like Figure 4a and 4b As shown, by introducing the WTConv module into the backbone network, high-frequency features are preserved during wavelet transform, the receptive field is expanded, more detailed information in the image is retained, and the visibility of small targets is improved, laying the foundation for improving prediction accuracy.
[0091] Specifically, in step S2, a dual-stream feature pyramid network is used for feature fusion, including parallel low-frequency and high-frequency branches, which process low-frequency and high-frequency stream feature maps of different scales respectively. The low-frequency stream branch is used to fuse low-frequency feature images of different scales, and the high-frequency branch is used to fuse high-frequency stream feature images of different scales to obtain their respective fused feature representations.
[0092] The feature pyramid network constructed within each flow includes an Improved Information Fusion Unit (IIFU) and a Cross Feature Pyramid (CFP). The IIFU starts from deep-scale features and upsamples the extracted low-frequency or high-frequency flow features layer by layer, converting the extracted feature map into a high-resolution feature map to achieve effective fusion of features at different levels. Then, the CFP starts from the bottom-scale features and upsamples the features layer by layer to recover the high-density information lost during downsampling.
[0093] For example, IIFU for low-frequency streams includes an upsampling layer, a concat layer, and a C2f module to complete feature fusion along a top-down path. The upsampling layer doubles the width and height of the extracted feature map, and the concatenation layer merges the output of the upsampling layer with the corresponding lower-level feature map to supplement spatial information and enhance semantic information. Specifically, this includes:
[0094] The first layer is an upsampling layer, which doubles the length and width of the low-frequency flow feature map with resolution L5 extracted by the backbone network, while keeping the number of channels unchanged.
[0095] The second layer is the Concat layer, which concatenates and fuses the upsampled feature map with the low-frequency stream feature map of resolution L4 output by the backbone network.
[0096] The third layer C2f module further processes the spliced and fused feature map, with an output channel count of 512.
[0097] The fourth layer is an upsampling layer, which again doubles the length and width of the low-frequency flow feature map.
[0098] The fifth layer is the Concat layer, which concatenates and fuses the upsampled feature map with the low-frequency stream feature map of resolution L3 output by the backbone network.
[0099] The 6th layer C2f module further processes the spliced and fused feature map, with 256 channels.
[0100] The 7th layer is an upsampling layer, which again doubles the length and width of the low-frequency flow feature map.
[0101] The 8th layer is the Concat layer, which concatenates and fuses the upsampled feature map with the low-frequency flow feature map of resolution L2 output by the backbone network.
[0102] The 9th layer, C2f module, further processes the stitched and fused feature map, with 128 channels. The low-frequency stream fused feature map with resolution L2 output from this layer is denoted as F2 and used as one of the inputs to the detection head. Figure 2As shown, layer C2f is labeled as Detected, indicating that the low-frequency stream fusion feature map F2 output by this layer is used for target detection;
[0103] The 10th layer is an upsampling layer, which again doubles the length and width of the low-frequency flow feature map.
[0104] The 11th layer is the Concat layer, which concatenates and fuses the upsampled feature map with the low-frequency stream feature map of resolution L1 output by the backbone network.
[0105] In addition, CFP is used to complete feature fusion along the bottom-up path, further enriching multi-scale features, including convolutional blocks and C2f modules, which are as follows:
[0106] The 12th layer C2f module further processes the feature map after the 11th layer is stitched and fused, with 64 channels; and the low-frequency stream fused feature map with resolution L1 output by this layer is denoted as F1, which is used as one of the inputs of the detection head. This layer C2f is labeled as Detected for target detection.
[0107] The 13th layer is a convolutional block that performs a convolution operation on the output of the 12th layer, with 64 output channels.
[0108] The 14th layer is the Concat layer, which concatenates and fuses the convolutional feature map with the output of the 9th layer.
[0109] The 15th layer C2f module further processes the spliced and fused feature map, with an output channel count of 128.
[0110] The 16th layer is a convolutional block that performs a convolution operation on the output of the 15th layer, with 128 output channels.
[0111] The 17th layer is the Concat layer, which concatenates and fuses the convolutional feature map with the output of the 6th layer.
[0112] The 18th layer is the C2f module, which further processes the spliced and fused feature map. It has 256 channels, and the low-frequency stream fused feature map with resolution L3 output by this layer is denoted as F3. It is used as one of the inputs of the detection head. This layer C2f is labeled as Detected and used for target detection.
[0113] The 19th layer is a convolutional block that performs a convolution operation on the output of the 18th layer, with 256 output channels.
[0114] The 20th layer is the Concat layer, which concatenates and fuses the feature map after convolution with the output of the 3rd layer. The low-frequency stream fused feature map with resolution L4 output of this layer is denoted as F4 and used as one of the inputs of the detection head. This layer C2f is labeled as Detected and used for target detection.
[0115] For the high-frequency stream IIFU, the same architecture as the low-frequency stream is adopted (fusion method of upsampling layer, splicing layer connected to C2f module). Based on the high-frequency stream feature map with L5 resolution extracted by the backbone network, it is passed through 3 upsampling layers in sequence, and spliced with the high-frequency stream feature maps with resolutions of L4, L3 and L2 output by the backbone network respectively. After multiple convolutions by CFP, the feature maps with resolutions of L3, L4 and L5 are spliced and fused in sequence to obtain high-frequency fused feature maps with resolutions of L2-L5, which are denoted as H1-H4 respectively (the splicing channel is different from the low-frequency stream branch, so different resolution fused feature maps are obtained). This is used as another part of the input of the detection head for target detection.
[0116] Through the above methods, such as Figure 5 As shown, F1-F4 and H1-H5 are used as inputs to the detection head for target detection. Compared with the single-channel feature fusion network, the dual-stream feature pyramid network uses the same feature convergence method to supplement detailed information for both high-frequency and low-frequency branches. It performs multi-channel feature fusion of different sizes from the top-down path and the bottom-up path, respectively, so that the feature map contains richer contextual and semantic information, which enhances the model's ability to detect targets at different scales.
[0117] Specifically, in step S3, the fused feature map output by the dual-stream feature fusion network is input into the detection head for target detection to obtain the final detection result, including predicted information such as target category and bounding box location. The specific process includes:
[0118] The bounding box prediction branch is responsible for predicting the location and size of the target.
[0119] The category prediction branch is responsible for predicting the category of the target and outputs the category probability for each bounding box.
[0120] Each branch contains two convolutional blocks and a prediction layer. Each convolutional block consists of a convolutional layer, a batch normalization layer, and an activation function. Each convolutional block is used to further process the low-frequency flow fusion feature map and the high-frequency flow fusion feature map. The prediction layer is a deep convolutional layer. For the bounding box prediction branch, the prediction layer predicts the bounding box parameters of the target, including the center point coordinates, width, and height. For the class prediction branch, the prediction layer predicts the probability that the target belongs to each class.
[0121] It should be noted that the target category to be detected depends on the categories labeled in the training dataset, such as Figure 6 As shown, when the labeled categories include people, ships, and vehicles, the object detection model can only detect objects of these three categories. The object detection model is trained based on a joint loss function, which includes classification loss, bounding box regression loss, and dynamic feature loss, expressed as:
[0122] L = LBCE +L Box +L DFL ,
[0123] Where L is the total loss function of the model; L BCE L represents the classification loss, which uses the binary cross-entropy loss to calculate the difference between the predicted class probability and the true label, and is used to measure the accuracy of the model's target classification prediction; Box L is the bounding box loss function, used to measure the difference between the predicted bounding box and the ground truth bounding box; DFL This represents the distribution focus loss (domain feature loss), used to optimize bounding box regression. When dealing with class imbalance problems, it reduces excessive attention to common classes and improves the detection ability of rare classes. By adjusting the weights of the loss function, it reduces the loss contribution to easily predictable samples and increases the loss contribution to difficult-to-predict samples.
[0124] Typically, the bounding box (Bbox) loss is calculated based on CloU (Complete Intersection over Union), and the expression is as follows:
[0125]
[0126] Where IoU refers to the crossover-union ratio, b and b gt ρ represents the center of the two bounding boxes, v represents the Euclidean distance between them, v measures the consistency of their relative proportions, and α is a weighting factor.
[0127] This application addresses the hard sample problem by introducing Focaler-IoU into the bounding box loss function, thereby improving the accuracy of bounding box regression. Focaler-IoU does not explicitly define hard samples; instead, it utilizes changes in IoU to quantify the difficulty of objects using a linear method. Specifically, a linear interval mapping method is used to optimize the IoU loss, thereby enhancing edge regression, as shown in the following formula:
[0128]
[0129] Among them, IoU Focaler This is the reconstructed IoU, and [d,u]∈[0,1]. By adjusting the values of d and u, the IoU can be optimized. Focaler Focusing on different regression samples, the Focaler-IoU loss function is defined as follows:
[0130] L Focaler-IoU =1-IoU focaler ,
[0131] As can be seen, Focaler-IoU enhances the gradient contribution of complex samples by adjusting the loss weights, enabling the model to learn the features of these challenging samples more effectively during training.
[0132] Based on the above characteristics, combining CIoU and Focaler-IoU into the Box loss function slightly improves the model's performance, especially in small object detection. The bounding box loss function is obtained based on the following formula:
[0133]
[0134] Compared with existing technologies, this embodiment provides an airborne target detection method based on wavelet transform, which significantly improves the detection accuracy of small targets from the perspective of a UAV without excessively increasing the model size. On one hand, the method captures contextual information and extracts detailed information by combining low-level and high-level features; on the other hand, Focaler-IOU loss is introduced during model training to emphasize the role of complex samples in the detection process.
[0135] Example 2
[0136] A specific embodiment of the present invention discloses an airborne target detection device based on wavelet transform, comprising:
[0137] The image acquisition module is used to acquire the image to be detected;
[0138] The feature extraction module is used to take the image to be detected as input and perform multi-scale feature extraction using multiple convolutional blocks of the backbone network; the wavelet transform module is used to extract low-frequency flow features and high-frequency flow features of at least some scale feature maps respectively, to obtain low-frequency flow feature maps and high-frequency flow feature maps of the corresponding scales.
[0139] The feature fusion module is used to perform multi-scale feature fusion on the low-frequency flow feature map and the high-frequency flow feature map using low-frequency branch and high-frequency branch respectively;
[0140] The prediction module is used to predict targets from the fused multi-scale low-frequency flow fusion feature map and the high-frequency flow fusion feature map, and obtain the predicted target type and location results.
[0141] The device can perform target detection according to any of the wavelet transform-based target detection methods described in Embodiment 1. Related aspects can be referenced from each other, but are not repeated in this embodiment.
[0142] Compared with existing technologies, this embodiment provides an airborne target detection device based on wavelet transform. Based on an image acquisition module, a feature extraction module, a feature fusion module, and a prediction module, it proposes a wavelet transform-based convolutional approach to achieve high-performance target detection for small targets.
[0143] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.
[0144] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.
Claims
1. An airborne target detection method based on wavelet transform, characterized in that, Includes the following steps: The image to be detected containing the target is input into the backbone network, and multi-scale feature extraction is performed using multiple convolutional blocks. The wavelet transform module is used to extract low-frequency flow features and high-frequency flow features of at least some scale feature maps respectively, so as to obtain low-frequency flow feature maps and high-frequency flow feature maps of the corresponding scales. Multi-scale feature fusion is performed on the low-frequency flow feature map and the high-frequency flow feature map using the low-frequency branch and the high-frequency branch of the feature fusion network, respectively. The detection head is used to predict the target by combining the multi-scale low-frequency flow fusion feature map and the high-frequency flow fusion feature image, and the predicted target type and bounding box location are obtained.
2. The airborne target detection method based on wavelet transform according to claim 1, characterized in that, The wavelet transform module is used to extract the corresponding low-frequency flow features, including: The feature map of the input wavelet transform module is filtered by low-frequency wavelet and high-frequency wavelet respectively to obtain the corresponding low-frequency components and high-frequency components. Convolutional blocks are used to perform convolution operations and scale adjustments on the low-frequency and high-frequency components respectively to obtain the corresponding low-frequency enhancement features and high-frequency enhancement features. Perform inverse wavelet transform on the low-frequency enhancement features and high-frequency enhancement features to obtain the inverse-transformed features; The feature map of the input wavelet transform module is convolved, and the convolved feature map is concatenated with the inverse transform feature to obtain the low-frequency flow feature.
3. The airborne target detection method based on wavelet transform according to claim 2, characterized in that, The low-frequency and high-frequency components are extracted using the following method: x LL ,x LH ,x HL ,x HH =WT(x,filters={ψ L ,ψ H }), Where x is the feature map of the input wavelet transform module; ψ L and ψ H These are low-frequency and high-frequency wavelet filters, respectively; x LL Represents the low-frequency component of x; x LH x HL and x HH These are the high-frequency components corresponding to x in the horizontal, vertical, and diagonal directions, respectively.
4. The airborne target detection method based on wavelet transform according to claim 3, characterized in that, The convolution operation is performed on the low-frequency and high-frequency components based on the following formula: Where, x WT ∈R B×(C×4)×H×W It is x LL x LH x HL and x HH This is a combination of four-dimensional channel information; Conv2d is a depthwise separable convolution; γ is a learnable scale parameter; ⊙ represents element-wise product.
5. The airborne target detection method based on wavelet transform according to claim 4, characterized in that, The high-frequency flow features are obtained by weighting and combining the high-frequency enhancement features according to the following formula: x H =[x LH ',x HL ',x HH ']⊙ω T , Where, x H The high-frequency flow characteristics; x LH '、x HL '、x HH 'represents the high-frequency enhancement features in the horizontal, vertical, and diagonal directions, respectively; ω T These are weighted parameters.
6. The airborne target detection method based on wavelet transform according to claim 1, characterized in that, The detection head is used to predict the target category and bounding box location; when predicting the target bounding box location, a combination of Focaler-IoU and CIoU is used to calculate the bounding box regression loss function based on the following formula: L Box =L CIoU +IoU-IoU Focaler , Among them, L Box L is the bounding box regression loss function; CIoU CIoU is the loss function; IoU is the intersection-union ratio between the predicted bounding box and the ground truth bounding box; IoU Focaler This represents the crossover ratio loss after reconstruction.
7. The airborne target detection method based on wavelet transform according to claim 6, characterized in that, The backbone network, feature fusion network, and detection head constitute the target detection model. The target detection model is trained using a joint loss function, the expression of which is: L=L BCE +L Box +L DFL , Where L is the joint loss function; L BCE L is used to measure the classification loss for predicting the target category; DFL This is the distributed focus loss, used to optimize the regression of the bounding box.
8. The airborne target detection method based on wavelet transform according to claim 6, characterized in that, The intersection-union ratio loss reconstructed by linear interval mapping is expressed as: Among them, IoU Focaler is the reconstructed intersection-union ratio loss; d and u are the adjustment parameters, respectively.
9. A wavelet transform-based airborne target detection method according to any one of claims 1-8, characterized in that, The feature fusion network is obtained based on a two-stream feature pyramid network, including: Using the upsampling layer of the centralized pyramid module of the low-frequency branch or high-frequency branch, the feature maps of the extracted low-frequency or high-frequency flow are sampled layer by layer upwards, and the output of the upsampling layer is fused with the feature maps of the corresponding low-frequency or high-frequency flow using the splicing layer. By utilizing the improved information fusion unit with low-frequency or high-frequency branches, features are passed up layer by layer from the bottom-level feature map, and fused with the low-frequency or high-frequency flow feature maps of the corresponding scale to obtain the low-frequency flow fusion feature map or the high-frequency flow fusion feature map.
10. An airborne target detection device based on wavelet transform, characterized in that, include: The image acquisition module is used to acquire the image to be detected; The feature extraction module is used to take the image to be detected as input and perform multi-scale feature extraction using multiple convolutional blocks of the backbone network; the wavelet transform module is used to extract low-frequency flow features and high-frequency flow features of at least some scale feature maps respectively, to obtain low-frequency flow feature maps and high-frequency flow feature maps of the corresponding scales. The feature fusion module is used to perform multi-scale feature fusion on the low-frequency flow feature map and the high-frequency flow feature map using low-frequency branch and high-frequency branch respectively; The prediction module is used to predict targets from the fused multi-scale low-frequency flow fusion feature map and the high-frequency flow fusion feature map, and obtain the predicted target type and location results.