A remote sensing small target detection method driven by frequency domain decoupling using Gaussian blur

Abstract

This invention discloses a remote sensing small target detection method using Gaussian blur-driven frequency domain decoupling. This method can be adapted to embed into various classic target detection frameworks to effectively solve the problem of small target detection in remote sensing images. First, an adaptive frequency decoupling module is proposed to dynamically and adaptively separate high-frequency and low-frequency components in features, effectively decoupling the foreground and background. Furthermore, a simple and lightweight background smoothing reconstruction module is designed. By suppressing low-frequency background interference, it drives adaptive frequency domain segmentation, thereby preserving more discriminative target features in the high-frequency components. Finally, a novel multi-scale high-frequency enhancement module is constructed to refine the feature representation of remote sensing small targets within the model's neck region, thereby extracting more detailed information from the high-frequency features.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of remote sensing and computer vision, specifically relating to a remote sensing small target detection method that utilizes Gaussian blur to drive frequency domain decoupling. Background Technology

[0002] Remote sensing small target detection is an important branch of general target detection, focusing on identifying extremely small objects in images. This technology has key applications in urban planning, agricultural monitoring, and maritime rescue. According to industry standards, target size is divided into three levels: 2-8 pixels is "very small," 8-16 pixels is "small," and 16-32 pixels is "small." Due to the extremely low pixel count of objects, many high-performing general detection models struggle in such tasks. Therefore, the core challenge of current research lies in effectively narrowing the significant performance gap between small and regular targets to improve the accuracy of identifying extremely small targets.

[0003] Compared to general object detection tasks, remote sensing small object detection faces two core challenges. First, small objects in remote sensing images typically occupy only a few pixels (e.g., 2-32 pixels), their size is much smaller than that of conventional objects, and they are often located in complex and variable natural or urban environments. For example, in farmland monitoring, tiny spots of pests and diseases are easily obscured by vegetation texture; in maritime rescue, lifeboats may be highly blended into the sea background due to wave reflections or cloud cover. This combination of extremely small size and background interference makes it difficult for models to locate and separate the target region from the global image, let alone extract discriminative key features that can distinguish the target from the background. Second, in the feature extraction stage, mainstream deep learning models (such as convolutional neural networks) rely on downsampling operations to expand the receptive field and improve computational efficiency. However, for tiny targets with scarce pixels, repeated downsampling will progressively compress their limited pixel information, even causing the target to disappear completely in high-level feature maps. For example, after a 4x downsampling, a target that was originally 16 pixels becomes only 1×1 pixels, and crucial details such as shape and texture are almost completely lost. This prevents the model from learning sufficient feature representations, ultimately leading to serious false negatives or missed detections. These two challenges are intertwined, making remote sensing small target detection a critical technical bottleneck that urgently needs to be overcome in the field of computer vision. Summary of the Invention

[0004] In view of this, the purpose of this invention is to propose a remote sensing small target detection method that utilizes Gaussian blur to drive frequency domain decoupling, so as to effectively solve the problem of small target detection in remote sensing images.

[0005] The technical solution of this invention is as follows: A remote sensing small target detection method utilizing Gaussian blur-driven frequency domain decoupling includes the following steps: Step 1: Construct a novel adaptive frequency decoupling module. This module transforms the input features from the spatial domain to the frequency domain and adaptively separates high-frequency and low-frequency components using a dynamic learning mechanism without relying on a fixed threshold, thereby effectively decoupling the foreground small target region from the background interference in remote sensing images. Step 2: Construct a simple and lightweight background smoothing and reconstruction module to suppress low-frequency background interference. By smoothing and reconstructing low-frequency background information, this module drives an adaptive frequency decoupling module to dynamically adjust the frequency band decomposition threshold, thereby more accurately preserving the discriminative features related to small target regions within high-frequency components. Step 3: Construct a multi-scale high-frequency enhancement module to refine the feature representation of small targets within the neck of the model. This module is constructed by integrating multi-scale convolution and difference maps to further extract detailed information from high-frequency features, thereby enhancing the feature response of small objects.

[0006] Step 1, constructing the adaptive frequency decoupling module, is as follows: Step 1: Adaptive Frequency Decoupling Module. This module consists of two parallel paths: a spatial frequency path and a channel frequency path. For the spatial frequency path, a simple spatial attention mechanism is introduced to process input features in order to better preserve information about small objects. For the channel frequency path, spatial information is compressed by combining the channel attention mechanism with global average pooling (GAP) and global max pooling (GMP), thereby focusing on key channel features.

[0007] Specifically, the following steps are included: Step 1.1: For the spatial frequency path, a simple spatial attention mechanism is introduced to process the input features, thereby better preserving the features of small targets.

[0008] Given an input feature tensor Where C represents the number of channels, and H and W represent the height and width of the input feature, respectively. The specific formula is as follows:

[0009] in Indicates the activation function Sigmoid. This represents a two-dimensional convolution operation.

[0010] Step 1.2: For the channel frequency path, a lightweight channel attention mechanism is introduced, which combines global average pooling (GAP) and global max pooling (GMP) to compress spatial information and thus focus on key channel features.

[0011] Given an input feature tensor Where C represents the number of channels, and H and W represent the height and width of the input feature, respectively. The specific formula is as follows:

[0012] Step 1.3: Utilize the output features obtained in Steps 1.1 and 1.2 and Then, convert it to the frequency domain for processing.

[0013] Given two feature tensors and The frequency domain is transformed using the two-dimensional discrete cosine transform (DCT). The specific formula is as follows:

[0014] Considering that the low-frequency and high-frequency components in the DCT spectrum are concentrated in the upper left and lower right corners respectively, a hyperparameter r The effective bandwidth of the filter is introduced to dynamically control it. Specifically, it is determined based on a specified ratio parameter. r Calculate the frequency-weighted mask The specific formula is as follows:

[0015] If and only if the hyperparameter r When the value is between 0 and 1, the frequency-weighted mask M Only then will it be activated, thereby separating and extracting the high-frequency and low-frequency components. Next, the frequency domain features will be analyzed separately. and Multiplying the components by their respective frequency-weighted masks, and then transforming them back to the spatial domain using the inverse discrete cosine transform (IDCT), the high-frequency and low-frequency components on the spatial and channel paths are finally obtained. The specific formulas are as follows:

[0016]

[0017] Step 1.4: Further aggregate the output features obtained in Step 1.3 to produce the output.

[0018] Given two sets of feature tensors and This is processed through a shared refinement module. The specific formula is as follows:

[0019]

[0020] Step 2: Construct a background smoothing reconstruction module: This module suppresses low-frequency background interference and drives adaptive adjustment of the frequency domain decoupling threshold, thereby enabling more key information of the micro-target region to be retained in the high-frequency components.

[0021] Specifically, the following steps are included: Step 2.1: Since low-frequency components correspond to the overall contour and smooth areas of the image, while the details and edge information of small targets are concentrated in high-frequency components, suppressing low-frequency components can effectively highlight small targets, thereby improving their detection performance. Therefore, a lightweight and simple upsampling module is adopted, consisting of only a small number of convolutional layers and transposed convolutional layers.

[0022] Given a low-frequency spatial feature tensor Where C represents the number of channels, and H and W represent the height and width of the input feature, respectively. The specific formula is as follows:

[0023]

[0024] in, T represents the activation function ReLU, and T Conv (・) represents the transposed convolutional layer. Step 2.2: Reconstruct the image obtained in Step 2.1 ,use Loss function for reconstructed image Reconstruction constraints are applied to reconstruct a degraded image with Gaussian background blur.

[0025] Given a reconstructed image Where C represents the number of channels, and H and W represent the height and width of the input feature, respectively. Using... Loss function for reconstructed image Apply reconstruction constraints. The specific formula is as follows:

[0026] in, This refers to a degraded image generated by applying Gaussian blur to the original image.

[0027] Step 3: Multi-scale high-frequency enhancement module: Based on the high-frequency components obtained by frequency domain decoupling of the neck features of the model in Step 1. A multi-scale high-frequency enhancement module was constructed. By integrating multi-scale convolution and difference map construction mechanisms, this module can make fuller use of high-frequency details to enhance the feature representation of small target regions, thereby optimizing the representation of small targets in the neck features of the original model.

[0028] Specifically, the following steps are included: Step 3.1: In order to comprehensively capture high-frequency patterns at different granularities, the multi-scale high-frequency enhancement module adopts a parallel multi-branch convolutional structure to capture features at multiple scales.

[0029] Given a high-frequency input feature Where C represents the number of channels, and H and W represent the height and width of the input features, respectively. This module consists of four different convolutional layers processed in parallel: a 1 x 1 convolution for pointwise feature transformation, standard 3 x 3 and 5 x 5 convolutions for local context aggregation, and a 3 x 3 dilated convolution with a dilation rate of 2, which captures a wider range of contextual information without increasing the number of parameters. The specific formula is as follows:

[0030]

[0031] Here, i represents the kernel size. For example, i=1 corresponds to a 1 x 1 kernel size, and so on. Also, This represents a dilated convolution with a kernel size of 3 x 3.

[0032] Step 3.2: In order to make fuller use of the detailed information in the high-frequency components, a difference map activation mechanism was developed. This mechanism uses the Sobel operator to calculate the gradient magnitude map, which has the dual advantages of differentiation and smoothing. It is not overly affected by noise and is not prone to edge blurring.

[0033] Given a high-frequency input feature Where C represents the number of channels, and H and W represent the height and width of the input features, respectively. The Sobel operator is applied in the horizontal and vertical directions. After obtaining the gradient response, features are fused through 1x1 convolution to finally generate a difference feature map. The specific formula is as follows:

[0034] in, and These represent the horizontal and vertical convolution kernels of the Sobel operator, respectively. These are small positive numbers used to ensure numerical stability.

[0035] Step 3.3: Based on the results of steps 3.1 and 3.2, adaptively recalibrate the importance of different spatial locations through a convergent attention mechanism.

[0036] First, the channel dimension statistics of each feature branch are aggregated using global average pooling. These pooled vectors are then concatenated and processed by a multilayer perceptron containing a hidden layer and employing the sigmoid activation function to generate a spatial attention map. M The specific formula is as follows:

[0037]

[0038] Subsequently, a scalar parameter is introduced. Dynamically learned modulated aggregated features F' Neck features compared to the original model N The optimal fusion weights between them. Final output. The calculation method is as follows:

[0039] . Attached Figure Description

[0040] Figure 1 The overall flowchart of this method; Figure 2 A remote sensing method for small target detection using Gaussian blur-driven frequency domain decoupling; Figure 3 Adaptive frequency decoupling module; Figure 4 Background smoothing reconstruction module; Figure 5 Multi-scale high-frequency enhancement module; Figure 6 Small target detection results image.

[0041] Beneficial effects 1. Dynamic adaptive frequency domain decoupling of features is achieved, significantly improving the separation between foreground and background: This invention abandons the traditional strategy of relying on fixed thresholds for high- and low-frequency separation and innovatively constructs an adaptive frequency decoupling module. This module combines spatial frequency paths and channel frequency paths, utilizing a dynamic learning mechanism to adaptively separate high-frequency (details / edges) and low-frequency (background / contour) components in the frequency domain. This mechanism can dynamically adjust the frequency band decomposition threshold based on input features, thereby more accurately decoupling small foreground target regions from complex background interference in remote sensing images, solving the problem that traditional fixed thresholds are difficult to adapt to varied remote sensing scenarios. 2. By using Gaussian blur to drive background suppression, the key features of small targets are effectively highlighted: To address the problem of small targets being easily obscured by complex backgrounds in remote sensing images, this invention designs a background smoothing reconstruction module. This module actively suppresses low-frequency background interference by constraining the reconstructed image to approximate the degraded image after Gaussian blurring. This "blur-driven sharpness" strategy forces the network to retain more discriminative target features in high-frequency components, thereby significantly enhancing the feature response of small objects in complex backgrounds and improving detection robustness. 3. Enhanced the ability to extract multi-scale high-frequency features, solving the problem of information loss caused by downsampling: This invention constructs a multi-scale high-frequency enhancement module specifically designed to refine features at the neck of the model. This module utilizes parallel multi-branch convolutions (covering 1x1, 3x3, 5x5, and dilated convolutions) to fuse high-frequency patterns of different granularities, and combines this with a difference map activation mechanism (Sobel operator) to further extract edge details. This not only compensates for the compression loss of small target pixel information during downsampling in deep neural networks, but also adaptively calibrates spatial location weights through a convergent attention mechanism, thereby greatly enriching the feature representation of small remote sensing targets and effectively reducing the false negative rate. Detailed Implementation

[0042] The method of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0043] Example To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment uses Cascade R-CNN (ResNet-50 + FPN backbone network) as the baseline object detection framework, taking the AI-TOD dataset for remote sensing images used for detecting small objects as an example, and setting the input image size to the default AI-TOD dataset size—800×800×3. The implementation process of the proposed method is described in detail. All experiments were performed on an NVIDIA L40S (48GB) GPU, using PyTorch 2.0 + MMDetectionv3.0 as the software environment.

[0044] The three modules of the invention (adaptive frequency decoupling module, background smoothing reconstruction module, and multi-scale high-frequency enhancement module) are sequentially embedded into the latter part of the FPN neck network of Cascade R-CNN. Specifically: Step 1 Adaptive Frequency Decoupling Module (AFD): (e.g.) Figure 3 As shown, this module consists of two parallel paths: a spatial frequency path and a channel frequency path. For the spatial frequency path, a simple spatial attention mechanism is introduced to process input features in order to better preserve information about small objects. For the channel frequency path, spatial information is compressed by combining the channel attention mechanism with global average pooling (GAP) and global max pooling (GMP), thereby focusing on key channel features.

[0045] Step 1.1: For the spatial frequency path, the input is the feature map X from the P2 layer of the FPN, with dimensions B × 256 × 200 × 200, where B is the batch size, set to 2 here. Then, a 3 × 3 convolutional kernel is applied to generate a spatial attention map, which is then multiplied by the original features after sigmoid activation. The specific representation is as follows:

[0046] Step 1.2: For the channel frequency path, perform global average pooling and global max pooling on the input feature map X of layer P2, respectively, to obtain two vectors of size B × 256 × 1 × 1. These vectors are then added together and passed through a fully connected layer (with the number of output channels remaining unchanged) to generate channel attention weights, which are then multiplied by X. The specific representation is as follows:

[0047] Step 1.3: Combine the output features obtained in Step 1.1 and Step 1.2 and Frequency domain features were obtained by transforming the data to the frequency domain using two-dimensional discrete cosine transform (DCT). Then, the hyperparameter r = 0.25 was determined through grid search. This mask marked the upper left corner (0.25*H, 0.25*W) region of the DCT spectrum as low frequency, and the remaining region as high frequency. The high-frequency and low-frequency components were then extracted and restored to the spatial domain using IDCT, as follows:

[0048]

[0049]

[0050]

[0051] Step 1.4: Pass the output features obtained in Step 1.3 through a shared 3×3 convolutional layer (256 output channels) to obtain the final high-frequency output. Similarly, after adding the low-frequency components, the same convolution is performed to obtain... Specifically, it is expressed as follows:

[0052]

[0053] Step 2: Construct the Background Smoothing Reconstruction (BSR) module: (e.g.) Figure 4 As shown, this module suppresses low-frequency background interference and drives adaptive adjustment of the frequency domain decoupling threshold, thereby enabling more key information of the micro-target region to be retained in the high-frequency components.

[0054] Step 2.1: The low-frequency component obtained through the adaptive frequency decoupling module (AFD) in Step 1 above. Its dimensions are B × 256 × 200 × 200, where B is the batch size, set to 2 here. Subsequently, a lightweight upsampler is constructed, specifically including: Conv2d(256, 128, kernel=3, stride=1, padding=1) + ReLU ConvTranspose2d(128, 64, kernel=4, stride=2, padding=1) ConvTranspose2d(64, 3, kernel=4, stride=2, padding=1) The final reconstructed image is the same size as the input image. Its dimensions are 800 × 800 × 3.

[0055] Step 2.2: Reconstruct the image obtained in Step 2.1 The original input image was Gaussian blurred (kernel size 21×21, standard deviation σ=1.5) to obtain a degraded image of the same size. Then, [the following steps were performed]. Loss function for reconstructed image Reconstruction constraints are applied to ensure that the reconstructed image approximates the blurred image. This is specifically expressed as follows: .

[0056] Step 3: Multi-scale High-Frequency Enhancement Module (MHFE): Based on the high-frequency components obtained by frequency domain decoupling of the neck features of the model in Step 1. A multi-scale high-frequency enhancement module was constructed, such as Figure 5 As shown, this module, by integrating multi-scale convolution and difference map construction mechanisms, can make fuller use of high-frequency details to enhance the feature representation of small target regions, thereby optimizing the representation of small targets in the neck features of the original model.

[0057] Step 3.1: High-frequency components obtained through the adaptive frequency decoupling module (AFD) in Step 1 above. Its dimensions are B × 256 × 200 × 200, where B is the batch size, set to 2 here. Subsequently, four convolutional layers are used in parallel, each with 64 output channels, specifically including: Conv2d(256, 64, kernel=1) Conv2d(256, 64, kernel=3, padding=1) Conv2d(256, 64, kernel=5, padding=2) Conv2d(256, 64, kernel=3, padding=2, dilation=2) Step 3.2: Use the Sobel operators (horizontal kernel Gx and vertical kernel Gy) to process the high-frequency components respectively. Convolution is performed to generate difference feature maps. The specific process is as follows:

[0058] Step 3.3: Subsequently, I concatenated the above features along the channel dimension using the Concat function to obtain an aggregated feature map with a size of 2 × 320 × 200 × 200. Then, I used an average pooling layer to reduce its dimension to 2 × 320 × 1 × 1, and passed it through a multilayer perceptron and a sigmoid function to obtain a spatial attention map. M Subsequently, a scalar parameter is introduced. (Ultimately converges to approximately 0.6–0.8), dynamically learning the modulated aggregated features. F' The optimal fusion weights between the input features of layer P2 and the input features are used to obtain the final output.

[0059] To effectively demonstrate the effectiveness of this invention, the following comparative experiments were conducted to verify and analyze it. As shown in Table 1, several advanced related methods in the field were selected for comparison, with Cascade R-CNN used as the baseline model for all. It can be observed that the method presented in this paper achieves the best accuracy across all metrics to date.

[0060] Table 1: Performance Comparison with Mainstream Methods (AI-TOD Test Set)

[0061] To further demonstrate the effectiveness of the modules proposed in this invention, the following ablation experiments were conducted to verify and analyze the invention. In the ablation experiments, as shown in Table 2, Cascade R-CNN was still used as the baseline. It can be observed that with the addition of each module, the indicators gradually improved.

[0062] Table 2: Ablation Experiments – Validating the Effectiveness of Each Module

[0063] This invention may have many other embodiments, and the above are merely preferred embodiments of the invention and are not intended to limit the scope of protection of the invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A remote sensing small target detection method driven by frequency domain decoupling with Gaussian blur, characterized in that include: Step 1: Construct a novel adaptive frequency decoupling module; This module effectively decouples the foreground small target region from the background interference part in the remote sensing image by transforming the input features from the spatial domain to the frequency domain and adaptively separating the high frequency and low frequency components using a dynamic learning mechanism without relying on a fixed threshold. Step 2: Construct a simple and lightweight background smoothing and reconstruction module to suppress low-frequency background interference; by smoothing and reconstructing low-frequency background information, this module drives an adaptive frequency decoupling module to dynamically adjust the frequency band decomposition threshold, thereby more accurately preserving the discriminative features related to small target regions within high-frequency components. Step 3: Construct a multi-scale high-frequency enhancement module to refine the feature representation of small targets within the neck of the model. This module is constructed by integrating multi-scale convolution and difference maps to further extract detailed information from high-frequency features, thereby enhancing the feature response of small objects.

2. The remote sensing small target detection method driven by Gaussian blur for frequency domain decoupling according to claim 1, wherein, Step 1, constructing the adaptive frequency decoupling module, includes: Step 1.1, for spatial frequency path, a simple spatial attention mechanism is introduced to process the input features, so as to better retain small target features; given an input feature tensor where C represents the number of channels, H and W represent the height and width of the input feature respectively; the formula is: ； wherein denotes the activation function Sigmoid, denotes a two-dimensional convolution operation; Step 1.2: For the channel frequency path, a lightweight channel attention mechanism is introduced, which combines global average pooling (GAP) and global max pooling (GMP) to compress spatial information, thereby focusing on key channel features; given an input feature tensor Where C represents the number of channels, and H and W represent the height and width of the input feature, respectively; the formula is: ； Step 1.3: Utilize the output features obtained in Steps 1.1 and 1.

2. and Transform it to the frequency domain for processing; given two feature tensors and The frequency domain is transformed using the two-dimensional discrete cosine transform (DCT); the formula is: ； Considering that the low-frequency and high-frequency components in the DCT spectrum are concentrated in the upper left and lower right corners respectively, a hyperparameter r The effective bandwidth of the filter is introduced to dynamically control the filter; specifically, it is determined according to a specified ratio parameter. r Calculate the frequency-weighted mask The formula is: ； If and only if the hyperparameter r When the value is between 0 and 1, the frequency-weighted mask M Only then will it be activated, thereby separating and extracting the high-frequency and low-frequency components; next, the frequency domain features will be... and Multiply by their respective frequency-weighted masks, and then transform them back to the spatial domain using the inverse discrete cosine transform (IDCT) to finally obtain the high-frequency and low-frequency components on the spatial path and the channel path; the formula is: ； ； Step 1.4: Further aggregate and output the output features obtained in Step 1.3; given two sets of feature tensors. and This is processed through a shared refinement module; the formula is: ； 。 3. The remote sensing small target detection method using Gaussian blur-driven frequency domain decoupling as described in claim 1, characterized in that, Step 2, constructing the background smoothing reconstruction module, includes: Step 2.1: Since low-frequency components correspond to the overall contour and smooth areas of the image, while the details and edge information of small targets are concentrated in high-frequency components, suppressing low-frequency components can effectively highlight small targets, thereby improving their detection performance. Therefore, a lightweight and simple upsampling module is adopted, consisting of only a few convolutional layers and transposed convolutional layers. Given a low-frequency spatial feature tensor... Where C represents the number of channels, and H and W represent the height and width of the input feature, respectively; the formula is: ； ； in, T represents the activation function ReLU, and T Conv (・) represents the transposed convolutional layer; Step 2.2: Based on the reconstructed image obtained in Step 2.1 ,use Loss function for reconstructed image Apply reconstruction constraints to reconstruct a degraded image with Gaussian background blur; given a reconstructed image Where C represents the number of channels, and H and W represent the height and width of the input feature, respectively; using Loss function for reconstructed image Apply reconstruction constraints; the formula is: ； in, This refers to a degraded image generated by applying Gaussian blur to the original image.

4. The remote sensing small target detection method using Gaussian blur-driven frequency domain decoupling as described in claim 1, characterized in that, Step 3, constructing the multi-scale high-frequency enhancement module, includes: Step 3.1: To comprehensively capture high-frequency patterns at different granularities, the multi-scale high-frequency enhancement module employs a parallel multi-branch convolutional structure to capture features at multiple scales; given a high-frequency input feature... Where C represents the number of channels, and H and W represent the height and width of the input features, respectively; this module is processed in parallel by four different convolutional layers: a 1 x 1 convolution for pointwise feature transformation, standard 3 x 3 and 5 x 5 convolutions for local context aggregation, and a 3 x 3 dilated convolution with a dilation rate of 2, which can capture a wider range of contextual information without increasing the number of parameters; the formula is: ； ； Here, i represents the kernel size; for example, i=1 corresponds to a 1 x 1 kernel size, and so on; at the same time, This represents a dilated convolution with a kernel size of 3 x 3; Step 3.2: To further and more fully utilize the detailed information in high-frequency components, a difference map activation mechanism was developed. This mechanism uses the Sobel operator to calculate the gradient magnitude map, combining the advantages of differentiation and smoothing. It is neither overly affected by noise nor prone to edge blurring. Given a high-frequency input feature... Where C represents the number of channels, and H and W represent the height and width of the input feature, respectively; the Sobel operator is applied in the horizontal and vertical directions, respectively; after obtaining the gradient response, feature fusion is performed through 1 x 1 convolution to finally generate the difference feature map; the specific formula is as follows: ； in, and These represent the horizontal and vertical convolution kernels of the Sobel operator, respectively. These are smaller positive constants used to ensure numerical stability; Step 3.3: Based on the results of steps 3.1 and 3.2, an aggregated attention mechanism is used to adaptively recalibrate the importance of different spatial locations. First, the channel dimension statistics of each feature branch are aggregated using global average pooling. These pooled vectors are then concatenated and processed by a multilayer perceptron containing a hidden layer and employing the sigmoid activation function to generate a spatial attention map. M The formula is: ； ； Subsequently, a scalar parameter is introduced. Dynamically learned modulated aggregated features F' Neck features compared to the original model N The optimal fusion weights between them; final output The calculation method is as follows: ； 。

Images