Micro-target detection method based on multi-scale multi-receptive field geometric semantic synchronous enhancement

Abstract

This application discloses a method for small target detection based on multi-scale, multi-receptive-field geometric and semantic synchronous enhancement, belonging to the field of remote sensing image target recognition technology. The method includes: acquiring remote sensing images; scaling multi-layer backbone features, sequentially performing pooling, convolution, and activation, multiplying them, performing wavelet transform, truncating a portion of the channels, and finally sequentially performing depthwise separable convolution, batch normalization, and activation before concatenating to form a multi-scale, multi-receptive-field feature map; utilizing a neck network for feature extraction and fusion, nonlinear mapping, and channel compression, followed by sigmoid normalization and filtering, weighted fusion with the multi-scale, multi-receptive-field feature map, and finally residual combination to obtain an enhanced feature map; using a region candidate network to generate RoI features, calculating classification and regression weights to obtain classification and regression features. The method of this application effectively enhances the network's performance in remote sensing image classification and regression.

Description

Technical Field

[0001] This application relates to the field of remote sensing image target recognition technology, and in particular to a method for detecting small targets in remote sensing images using multi-scale, multi-receptive field and geometric semantic simultaneous enhancement techniques. Background Technology

[0002] In the field of remote sensing image analysis, small target detection has become a research hotspot in recent years. This task has significant practical implications and application value in numerous application scenarios such as autonomous driving, defect detection, and aerospace remote sensing analysis. However, compared with natural scene images, small target identification in remote sensing images is significantly more difficult because small targets typically refer to target instances occupying an area of ​​no more than 32×32 pixels in the image. This characteristic poses considerable challenges to detectors in feature extraction, localization, and classification. Nevertheless, given the urgent need for small target detection in practical applications, research in this direction not only has important engineering practicality but also contains rich theoretical research value. With the rapid development of deep learning technology, the performance of target detection in natural images has been significantly improved, but there is still considerable room for improvement in small target detection in remote sensing scenarios, requiring further exploration and optimization.

[0003] From a visual observation perspective, natural images typically feature large-scale targets, while remote sensing scenes often contain smaller targets and multiple foreground objects. For deep learning-based detectors, this small target scale presents numerous detection challenges. For example, during downsampling, details of the target object may be lost, making it difficult to distinguish between foreground and background, and potentially causing the detector to overlook targets that should be detected. Furthermore, small targets may gradually lose distinguishable features within complex networks, further increasing the difficulty of detection. To address these issues, researchers typically introduce constraints specifically designed for small target recognition into the detection network to improve detection accuracy.

[0004] To address the issue of information loss, some methods opt for higher resolution detection. While this larger-scale approach can mitigate the loss of information about small targets to some extent, simply upsampling the feature map without further processing, while reducing network information loss, still fails to effectively supplement more information. Meanwhile, other methods improve the performance of small target detection by optimizing the detector's head. However, these methods only optimize the head's extraction capabilities without tailoring optimization to the specific task, thus failing to maximize detection accuracy. Summary of the Invention

[0005] This application provides a method for detecting small targets based on multi-scale, multi-receptive-field geometric and semantic synchronization enhancement, in order to solve the problems in the prior art.

[0006] This application provides a method for detecting small targets based on multi-scale, multi-receptive-field geometric-semantic synchronization enhancement, including:

[0007] Acquire the remote sensing image to be detected;

[0008] The scale of the multi-layer backbone features of the remote sensing image extracted by the backbone network is adjusted to form a uniform scale feature map. The uniform scale feature map is then pooled, convolved and activated sequentially to obtain channel weights. The channel weights are multiplied with the uniform scale feature map to obtain channel feature maps. Wavelet transform is performed on the channel feature maps to obtain wavelet transform output. A portion of the channels is truncated from the wavelet transform output to form the final output. The final output is then subjected to depthwise separable convolution, batch normalization and activation sequentially and then concatenated with the final output to form a multi-scale multi-receptive field feature map.

[0009] Feature extraction and fusion of multi-layer backbone features are performed using a neck network to obtain a fused feature map. The fused feature map is then subjected to nonlinear mapping and channel compression to obtain a single-channel soft response map. This single-channel soft response map is then normalized and filtered using Sigmoid to obtain a spatial mask. This spatial mask is then weighted and fused with multi-scale, multi-receptive-field feature maps. The result of this weighted fusion is then compared with the fused feature map... By combining feature residuals, an enhanced feature map is obtained;

[0010] The enhanced feature map is used to generate RoI (Region of Interest) features using a region candidate network. Classification weights and regression weights are calculated in the classification branch and regression branch respectively. The final classification features and regression features are obtained by weighting the residuals.

[0011] The small target detection method based on multi-scale, multi-receptive-field geometric-semantic synchronization enhancement in this application has the following advantages:

[0012] By supplementing the network with missing information through a multi-scale, multi-receptive-field feature extraction (MMF) module, enriching the geometric and semantic information of the neck network through a geometric-semantic synchronization enhancement (GSSE) module, and finally using a detection head with a subtask channel separation (SCSH) module, the network's performance in remote sensing image classification and regression is effectively enhanced. The designed micro-target detection framework effectively improves the network's detection accuracy for micro-target foregrounds in remote sensing images. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is an architecture diagram of a method for detecting small targets based on multi-scale, multi-receptive-field geometric and semantic synchronization enhancement provided in an embodiment of this application.

[0015] Figure 2 This is an architecture diagram of the multi-scale multi-receptive field feature extraction module (MMF) provided in the embodiments of this application.

[0016] Figure 3 This is an architecture diagram of the Geometric Semantic Synchronization Enhancement Module (GSSE) provided in an embodiment of this application.

[0017] Figure 4 This is an architecture diagram of the Subtask Channel Separation Module (SCSH) provided in an embodiment of this application. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] Figure 1 This is an architecture diagram of a small target detection method based on multi-scale, multi-receptive-field geometric-semantic synchronization enhancement provided in an embodiment of this application. The embodiment of this application provides a small target detection method based on multi-scale, multi-receptive-field geometric-semantic synchronization enhancement, including:

[0020] Step 1: Acquire the remote sensing image to be detected.

[0021] Step 2: Scale the multi-layer backbone features of the remote sensing image extracted by the backbone network to form a uniform scale feature map. Pooling, convolution, and activation are performed sequentially on the uniform scale feature map to obtain channel weights. The channel weights are multiplied by the uniform scale feature map to obtain channel feature maps. Wavelet transform is performed on the channel feature maps to obtain wavelet transform output. A portion of the channels is truncated from the wavelet transform output to form the final output. The final output is then sequentially subjected to depthwise separable convolution, batch normalization, and activation, and then concatenated with the final output to form a multi-scale, multi-receptive-field feature map.

[0022] For example, this step includes the following sub-steps:

[0023] Step 2.1: Using the multi-layer backbone features extracted from the backbone network as input, the Multi-Scale Multi-Receptive Field Feature Extraction (MMF) module adjusts the number of channels in these multi-layer backbone features using group convolutions, compressing the information to the same dimension. Then, linear interpolation of different orders is applied to these feature maps at different scales to adjust them to match the dimensions in the neck network. The feature maps are of the same size, and then they are merged into a feature map of a uniform scale. In the embodiments of this application, the backbone network has 4 layers. Each of the 4 layers uses group convolution to adjust the corresponding backbone features to the same dimension. Then, these feature maps of different scales but the same dimension are adjusted using linear interpolation of the appropriate number of times to match the feature maps in the neck network. Feature maps of the same size are ultimately merged into a unified-scale feature map after scaling.

[0024] This process can be expressed by the following formula:

[0025]

[0026] in, X Features of Bacbone GC Representative group convolution, Up For upsampling, Conv It is a convolution. Sum For addition, F s This is to create a unified scale feature map. The scale of the final unified scale feature map is similar to that in the neck network. The feature maps are of uniform size. Group convolution reduces the number of connections between the input feature maps and the convolution kernel by splitting them into groups, thereby significantly reducing the number of multiplication and accumulation operations and thus reducing the computational cost of multichannel convolution.

[0027] Furthermore, the four layers in the backbone network are layers B1, B2, B3, and B4. Layers B1, B2, B3, and B4 undergo dimension adjustment through group convolutions with 64, 32, 16, and 8 groupings, respectively. After dimension adjustment, layers B2, B3, and B4 undergo scaling through 1st, 2nd, and 3rd linear interpolation, respectively. Figure 2 As shown.

[0028] Step 2.2 uses a channel attention mechanism to allow the network to automatically focus on important channels and suppress useless channels. Pooling, convolution, and activation functions are used, and finally, the channel weights are broadcast to each weight using the sigmoid function. This process can be expressed by the following formula:

[0029]

[0030] in, GAP This is average pooling (AvgPooling). Conv 16 This means the number of channels will be compressed to 16. Conv 256 This represents expanding the number of channels to 256. ReLU It is a linear rectified function. Sigmoid The Sigmoid activation function is used, and the result of Sigmoid activation is the channel weight. For multiplication, This is a channel feature map.

[0031] Step 2.3, perform wavelet transform on the channel feature map, including the following steps:

[0032] Construct the kernel matrix; when constructing the kernel matrix, use the normalized matrix to initialize the wavelet basis, retain the low-frequency components obtained by performing basic wavelet decomposition using the wavelet basis, discard the high-frequency components, and obtain the Haar matrix. Repeatedly splice the Haar matrix to form the kernel matrix.

[0033] The channel feature map and kernel matrix are respectively decomposed into C low Each independent group;

[0034] For each feature map group, a kernel group is used to perform a 1×1 convolution operation;

[0035] The results of all grouped convolution operations are concatenated according to channels to obtain the wavelet transform output.

[0036] Specifically, wavelet transform is performed along the channel direction to separate the signal and noise. The low-frequency components retain the overall structure of the signal, while removing noise or irrelevant details represented by the high-frequency components. In this step, the wavelet basis is first initialized using a normalized matrix. h :

[0037]

[0038] Using wavelet bases h After performing a basic wavelet decomposition, four components are generated. In this embodiment, only the low-frequency components are retained, while the horizontal high-frequency, vertical high-frequency, and diagonal high-frequency components are discarded, resulting in a Haar matrix. The Haar matrix is ​​then repeated to obtain the complete kernel matrix. K :

[0039]

[0040] in, The identity matrix is ​​used to complete the spatial dimensions. Finally, wavelet transform is implemented through grouped convolution: the input feature map is split into... There are 4 independent groups, each with 4 channels, corresponding to the input group. The kernel matrix is ​​then... K Also split into There are three groups, each kernel group is responsible for processing one feature map group, and each feature map group interacts with the kernel group. Convolution operation, by concatenating all grouped outputs according to channels, yields the complete wavelet transform output. Y :

[0041]

[0042] in, Indicates the number of groups.

[0043] Finally in Y Extract the first part of each group One channel is needed to obtain the final output. Z :

[0044]

[0045] in, Indicates splicing.

[0046] Step 2.4: Multiple depthwise separable convolutions with different kernel sizes are used in parallel to process the final output. Then, the feature maps processed by multiple receptive fields are concatenated to obtain multi-scale, multi-receptive-field feature maps.

[0047]

[0048] in, DW This represents depthwise separable convolution. BN For batch normalization, Con This represents the stitching process, resulting in a multi-scale, multi-receptive-field feature map. F x This will serve as input to the geometric semantic synchronization enhancement mechanism to strengthen the detector's feature representation of the foreground target.

[0049] Specifically, in step 2.4, the final output is divided into three branches, which are then subjected to depthwise separable convolution, batch normalization, and activation sequentially. In the first branch, a 3×3 kernel depthwise separable convolution, batch normalization, and ReLU activation are performed sequentially. In the second branch, a 7×7 kernel depthwise separable convolution, batch normalization, and ReLU activation are performed sequentially. In the third branch, a 13×13 kernel depthwise separable convolution, batch normalization, and ReLU activation are performed sequentially. The outputs of these three branches are concatenated and then concatenated with the final output to form a multi-scale, multi-receptive-field feature map. F x .

[0050] In the embodiments of this application, the overall network is built based on the MMDetection framework, and the backbone network adopts the residual network ResNet50 or ResNet101. When using ResNet50, the specific network parameters are as follows: the network depth is 50 layers, including 1 convolutional preprocessing layer and 4 sets of residual modules. Each set of residual modules contains 3, 4, 6 and 3 residual units respectively. The convolutional kernel size is 3×3, the initial number of channels is 64, and the number of channels increases to 2048 as the network depth increases. The activation function is ReLU, and the pooling method is max pooling. The gradient vanishing problem in the training process of deep networks is solved by the residual structure. This network is used to extract multi-scale deep semantic features of images.

[0051] The neck network employs a Feature Pyramid Network (FPN). Specific network parameters are as follows: the input consists of four sets of feature maps at different scales (scales of 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the input image, respectively) output by the backbone network. Channel dimensionality reduction is achieved through 1×1 convolutions. A top-down upsampling path and lateral connections are used to fuse shallow texture features and deep semantic features, outputting five sets of multi-scale fused feature maps with a uniform number of channels (256 dimensions). The fusion method employs element-wise addition. FPN effectively fuses multi-scale features, improving the robustness of detecting both small and large targets.

[0052] The detection head employs a two-stage detection algorithm: the first stage is a Region Proposal Network (RPN), which generates foreground / background classification scores and target candidate box coordinates based on the multi-scale fusion feature map output by the neck network, filtering out high-precision target candidate regions; the second stage is a Region of Interest Pooling / Region of Interest Alignment (ROI Pooling) module, which maps the target candidate regions onto the feature map to extract fixed-dimensional features, and finally achieves target category classification and accurate regression of candidate boxes through a fully connected layer. This two-stage detection algorithm improves the localization accuracy and classification accuracy of target detection by first filtering candidate regions and then performing precise detection.

[0053] Step 3: The neck network is used to extract and fuse features from the multi-layer backbone, resulting in a fused feature map. This fused feature map undergoes non-linear mapping and channel compression to obtain a single-channel soft response map. The single-channel soft response map is then normalized and filtered using Sigmoid to obtain a spatial mask. This spatial mask is then weighted and fused with the multi-scale, multi-receptive-field feature map. The result of this weighted fusion is then compared with the fused feature map... The feature residuals are combined to obtain an enhanced feature map.

[0054] For example, such as Figure 3 As shown, this step includes the following sub-steps:

[0055] Step 3.1: For the features output by the FPN, i.e., the multi-layer backbone features, the Geometric Semantic Synchronous Enhancement (GSSE) module first uses a lightweight sub-network consisting of two layers of convolution and ReLU activation to perform nonlinear mapping and channel compression on the input features, reducing the 256-dimensional features to a single-channel response map to extract salient regions that may correspond to small targets. Subsequently, a spatial smoothing operation is performed on this response map using a 3×3 average convolution kernel with fixed parameters.

[0056]

[0057] in, For FPN output Feature map and They are 3×3 convolution and 1×1 convolution, respectively. For smoothing operations using a 3×3 average convolution kernel, M This is a single-channel soft response plot.

[0058] Step 3.2, generate the single-channel soft response map. M Upsampled to the same resolution as the original image, and the pixel-level reconstruction error between the upsampled image and the original image is calculated. This loss term, after being scaled by weight coefficients, is optimized together with other detection losses. In the embodiments of this application, backpropagation is performed using a stochastic gradient descent optimizer based on the total loss. The specific parameter adjustment process is as follows: 1. Forward propagation: The original image is input into the detection network, passing sequentially through the backbone network ResNet, the neck network FPN, and the detection head, outputting the detection result and a single-channel soft response map M. 2. Loss calculation: The detection core loss and reconstruction loss are calculated in parallel and fused to obtain the total loss. 3. Backpropagation: The gradient value is calculated based on the total loss, and the gradient is propagated backward layer by layer in the order of "detection head → neck network FPN → backbone network ResNet → upsampling module". 4. Parameter update: The parameter is updated according to the optimizer update rules. ,in For learning rate, The gradient of the total loss with respect to the parameters. For the total loss, For parameters before the update, For the updated parameters, adjust the parameters for each module. θ 5. Iterative Convergence: Repeat the above forward propagation, loss calculation, backpropagation, and parameter update process until the total loss converges, completing the parameter optimization of all network modules. This guides the entire network to generate a more accurate spatial distribution and a more continuous structure during training.

[0059]

[0060]

[0061] in, This is the single-channel soft response plot after upsampling. The binary foreground mask (GTMask) is the one that is actually labeled. For single-channel soft response plots in ( i , j The pixel value at () For activation function, coordinates ( i , j The truth label at ) H and W These are the height and width of the input image, respectively. N The number of pixels in the input image. The binary cross-entropy loss is applied to the single-channel soft response map and the ground truth labeled binary foreground mask. Meanwhile, for the single-channel soft response map... M Sigmoid normalization is performed, and a learnable threshold is set using nn.Parameter for filtering, thus obtaining a spatial mask for the foreground region of small targets. This mask is then used in conjunction with multi-scale, multi-receptive-field feature maps. Add them together, and then combine the fused result with the FPN. By combining the feature residuals, an enhanced feature map can be obtained. This operation, while preserving the high-level semantic discriminative ability, explicitly compensates for the geometric structural information of small targets, achieving synergistic enhancement of semantic and geometric features, thereby effectively improving the localization accuracy and boundary awareness capability of the network model in small target detection.

[0062] In this embodiment, the network model optimizes its parameters during training by using the pixel-level reconstruction error between the upsampled single-channel soft response map and the original image as the loss term. The training set used during training includes sample images and annotation files, which are prepared as follows:

[0063] Collect sample images and labeled files containing tiny targets;

[0064] All sample images and labeled files are proportionally divided into training and test sets;

[0065] The image is converted to an OpenCV BGR format array with dimensions H×W×C, where H, W, and C are the height, width, and number of channels, respectively. The image is then scaled to 800×800 pixels, and the image and bounding boxes are randomly flipped horizontally and vertically. After these processing steps, the training set is loaded into the initial network model for training.

[0066] Step 4: Generate RoI features from the enhanced feature map using the Region Candidate Network (RPN). Calculate the classification weights and regression weights in the classification and regression branches respectively, and obtain the final classification and regression features using residual weighting.

[0067] For example, such as Figure 4 As shown, this step includes the following sub-steps:

[0068] Step 4.1: In the classification branch and the regression branch, after applying Dropout layer, fully connected layer and Softmax activation in sequence, the weights are scaled using classification scaling factor and regression scaling factor respectively to obtain classification weights and regression weights.

[0069] Specifically, for RoI features, they are first compressed into two-dimensional RoI features. To avoid over-reliance on redundant dimensions in the features, regularization (Dropout) is first applied to prevent model overfitting. Then, two fully connected layers are passed, followed by softmax activation and weight scaling, resulting in RoI features with scales of [B,D,2] representing single-channel RoI maps for classification and regression, respectively, used for the final classification and regression.

[0070]

[0071] in, RoI features D For Dropout operations, F c It is a fully connected layer. This is a temperature coefficient used to enhance the sharpness of adjusting attention weights. For the Softmax function, This represents the RoI features processed by Dropout and fully connected layers, categorized by temperature coefficient. τ After scaling, a Softmax operation is performed, and the feature of the 0th channel of the output tensor is taken to obtain the attention weight distribution for the classification task. This is used to highlight the feature dimensions that are more critical to the classification task. This means performing the same operation on the same set of features, taking the first channel feature of the output tensor, and obtaining the attention weight distribution for the regression task. This is used to highlight the feature dimensions that are more critical to the bounding box regression task. and These are the classification scaling factor and the regression scaling factor, respectively. and These are the classification weights and regression weights, respectively.

[0072] Step 4.2: Obtain the final classification and regression features through residual weighting.

[0073]

[0074] in, and These are classification features and regression features, where classification features correspond to the semantic information of the category of the micro-target, and regression features correspond to the spatial location and scale information of the micro-target.

[0075] To verify the effectiveness of the proposed method, training and testing were performed on three widely used benchmark datasets. The results are shown in Tables 1-3, with the last row representing the evaluation metrics of the proposed method. Experimental results show that the proposed method outperforms existing small target detection methods in all evaluation metrics, exhibiting high detection performance across various datasets. Among the evaluation metrics, mAP and AP are... 50 AP 75 AP vt AP t and AP s These are the average accuracy, average accuracy with an IoU threshold of 0.5, average accuracy with an IoU threshold of 0.75, average accuracy of medium target, average accuracy of large target, and average accuracy of small target, respectively. Faster R-CNN is a faster region convolutional neural network; Query Det is a query detector; Cascade R-CNN is a cascaded region convolutional neural network; CEASC is a global context-enhanced adaptive sparse convolutional network; Detecto RS is a detector fusion network; ORFENet is an open region feature enhancement network; NWD is normalized Gaussian Walsh-Tenby distance; HANet is a heatmap-assisted network / hierarchical attention network; RFLA w / Detecto RS is a Gaussian receptive field-based label assignment combined with detector fusion network; SR-TOD w / Detecto RS is a self-reconstructed small target detection combined with detector fusion network; HS-FPN is a hierarchical scale feature pyramid network; Salience DETR is a saliency Transformer detector; FCOS is a fully convolutional single-stage target detector; AI-TOD is an aerial image small target detection dataset; Visdrone dataset is a Tianjin University UAV dataset; and Dior dataset is an optical recognition UAV dataset.

[0076] Table 1 Evaluation metrics of different methods on the AI-TOD dataset

[0077]

[0078] Table 2 Evaluation metrics of different methods on the Visdrone dataset

[0079]

[0080] Table 3 Evaluation metrics of different methods on the Dior dataset

[0081]

[0082] As can be seen from the above results, unlike existing methods for reducing information loss in networks, the method in this application improves the detection accuracy of small targets by supplementing the network with new information, and at the same time, it adopts a detection head feature screening mechanism to further enhance the network's ability to capture small target foregrounds.

[0083] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0084] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for detecting micro-objects based on multi-scale multi-receptive field geometric semantic synchronous enhancement, characterized in that, include: Acquire the remote sensing image to be detected; The scale of the multi-layer backbone features of the remote sensing image extracted by the backbone network is adjusted to form a uniform scale feature map. The uniform scale feature map is then subjected to pooling, convolution, and activation sequentially to obtain channel weights. The channel weights are multiplied by the uniform scale feature map to obtain channel feature maps. Wavelet transform is performed on the channel feature maps to obtain wavelet transform output. A portion of the channels is truncated from the wavelet transform output to form the final output. The final output is then subjected to depthwise separable convolution, batch normalization, and activation sequentially and then concatenated with the final output to form a multi-scale, multi-receptive-field feature map. The neck network is used to extract and fuse features from the multi-layer backbone features to obtain a fused feature map. The fused feature map is then subjected to nonlinear mapping and channel compression to obtain a single-channel soft response map. This single-channel soft response map is then normalized and filtered using Sigmoid to obtain a spatial mask. This spatial mask is then weighted and fused with the multi-scale, multi-receptive-field feature map. The result of this weighted fusion is then compared with the fused feature map. Feature residuals are combined to obtain enhanced feature maps; the backbone network has four layers, and each of the four layers uses group convolutions to adjust the corresponding backbone features to the same dimension. Then, these feature maps of the same dimension but different scales are adjusted using linear interpolation of the appropriate number of iterations to match the features in the neck network. The feature maps are of the same size, and the scaled feature maps are finally merged into the unified scale feature map. The enhanced feature map is used to generate RoI features using a region candidate network. Classification weights and regression weights are calculated in the classification branch and regression branch, respectively. The final classification features and regression features are obtained by weighting the residuals.

2. The multi-scale multi-receptive field geometric semantic synchronous enhancement micro target detection method based on claim 1, characterized in that, During training, the pixel-level reconstruction error between the upsampled single-channel soft response map and the original image is used as the loss term to optimize the model parameters.

3. The multi-scale multi-receptive field geometric semantic synchronous enhancement micro-object detection method according to claim 1, characterized in that, The backbone network consists of four layers: B1, B2, B3, and B4. Layers B1, B2, B3, and B4 are subjected to dimensionality adjustment through group convolution with group numbers of 64, 32, 16, and 8, respectively. After dimensionality adjustment, layers B2, B3, and B4 are subjected to scale adjustment through 1, 2, and 3 linear interpolations, respectively.

4. The multi-scale multi-receptive field geometric semantic synchronous enhancement micro-object detection method according to claim 1, characterized in that, The method for performing wavelet transform on the channel feature map is as follows: Construct the kernel matrix; The channel feature map and the kernel matrix are respectively split into C low Each independent group; For each feature map group, a kernel group is used to perform a 1×1 convolution operation; The results of all grouped convolution operations are concatenated according to channels to obtain the wavelet transform output.

5. The method for detecting small targets based on multi-scale, multi-receptive-field geometric-semantic synchronization enhancement according to claim 4, characterized in that, When constructing the kernel matrix, the wavelet basis is initialized using the normalized matrix. The low-frequency components obtained by performing basic wavelet decomposition using the wavelet basis are retained, while the high-frequency components are discarded to obtain the Haar matrix. The Haar matrix is ​​then repeatedly spliced ​​together to form the kernel matrix.

6. The multi-scale multi-receptive field geometric semantic synchronous enhancement micro-object detection method according to claim 4, characterized in that, Extracting the first C from the wavelet transform output low Each channel forms the final output.

7. The method for detecting small targets based on multi-scale, multi-receptive-field geometric-semantic synchronization enhancement according to claim 1, characterized in that, The final output is divided into three branches, which are then subjected to depthwise separable convolution, batch normalization, and activation in sequence. In the first branch, a 3×3 kernel depthwise separable convolution, batch normalization, and ReLU activation are performed in sequence. In the second branch, a 7×7 kernel depthwise separable convolution, batch normalization, and ReLU activation are performed in sequence. In the third branch, a 13×13 kernel depthwise separable convolution, batch normalization, and ReLU activation are performed in sequence. The outputs of the three branches are concatenated and then concatenated with the final output to form the multi-scale multi-receptive field feature map.

8. The multi-scale multi-receptive field geometric semantic synchronous enhancement micro-object detection method according to claim 1, characterized in that, In the classification branch and the regression branch, after sequentially using the Dropout layer, the fully connected layer and the Softmax activation, the weights are scaled using the classification scaling factor and the regression scaling factor, respectively, to obtain the classification weights and the regression weights.

9. The method for detecting small targets based on multi-scale, multi-receptive-field geometric-semantic synchronization enhancement according to claim 1, characterized in that, The backbone network uses ResNet, and the neck network uses FPN.

Images