Lightweight remote sensing target detection method based on feature focus diffusion

Abstract

The application discloses a lightweight remote sensing target detection method based on feature focusing diffusion, carries out data preprocessing on the obtained remote sensing image; a residual multi-scale convolution attention module (RMSCA) is used to construct a model backbone network, multi-scale features are extracted, background noise is suppressed, and the spatial positioning capability of the model for detecting targets is improved; a feature focusing module (FFM) and a feature focusing diffusion mechanism (FFDM) are used to construct a model neck (Neck) structure, different scale features are aggregated and separated, effective feature weights are enhanced, and feature maps rich in context information are diffused to each detection scale to obtain an MFFDNET detection model, namely a multi-scale feature focusing diffusion model; the model is trained based on a data set, parameters are optimized, and performance is evaluated. Compared with the prior art, the multi-scale features are aggregated, separated and diffused by the FFDM mechanism, the detection precision of multi-scale targets in the remote sensing image is effectively improved, and the method can be applied to remote sensing field target detection.

Description

Technical Field

[0001] This invention relates to the field of computer vision target detection, and in particular to a lightweight remote sensing target detection method based on feature focusing diffusion. Background Technology

[0002] With the continuous advancement of urbanization, urban transportation systems are becoming increasingly complex, and traffic flow on highways, urban expressways, and main roads continues to grow. To improve traffic efficiency and management, traffic management departments are gradually introducing intelligent analysis methods based on remote sensing images and aerial monitoring images to monitor and assess road operating conditions, vehicle distribution, and traffic congestion areas.

[0003] In practical applications, remote sensing images in traffic scenarios typically cover large traffic areas, characterized by wide coverage and significant differences in target scale. Traditional multi-scale fusion methods often employ simple feature overlay or weighting to address this significant scale difference, failing to adequately consider the correlation between features at different scales and easily introducing redundant feature information. When there are many vehicles in a traffic scene, their distribution is dense, or their scale is small, target features can easily be overwhelmed by complex background information, leading to missed detections, false detections, or inaccurate localization during vehicle detection, thus affecting the accuracy of traffic flow statistics and congestion area identification.

[0004] Furthermore, traffic monitoring systems typically require large-scale, long-term continuous operation under limited computing resources, which places stringent demands on the lightweight design and computational efficiency of target detection models. Simply increasing the number of network layers or the scale of model parameters to improve detection accuracy often leads to a significant increase in computational load, hindering the deployment and application of the model in practical traffic monitoring systems. Therefore, how to enhance the expressive power of multi-scale target features in remote sensing images and improve the accuracy of target detection from a remote sensing perspective, while ensuring model lightweight design, remains a pressing problem to be solved in current technologies. Summary of the Invention

[0005] Purpose of the Invention: To address the issues of low accuracy in multi-scale target detection from a remote sensing perspective and the difficulty in balancing lightweight design and detection performance in existing models, this invention proposes a lightweight remote sensing target detection method based on feature focusing diffusion. This method constructs the model backbone network by concatenating an RMSCA module with standard convolutions, concatenating an FFM module, and using the feature focusing diffusion mechanism FFDM to construct the model neck structure. This achieves a synergistic balance between low-parameter, lightweight design and high-precision multi-scale target detection in remote sensing images, effectively overcoming the shortcomings of existing technologies and meeting the application needs of practical remote sensing traffic monitoring scenarios.

[0006] Technical solution: This invention discloses a lightweight remote sensing target detection method based on feature focusing diffusion, comprising the following steps:

[0007] S1: Collect remote sensing image data and preprocess it to create image datasets and label datasets respectively, and integrate the image datasets and label datasets into the original dataset;

[0008] S2: Construct the MFFDNET detection model;

[0009] S2.1: Concatenate the RMSCA module with standard convolution to construct the backbone network, extract features from the preprocessed remote sensing image, and obtain feature maps at different scales;

[0010] S2.2: Two FFM modules are connected in series, and a model neck structure is constructed based on the feature focusing and diffusion mechanism FFDM. The neck structure receives the multi-scale feature map output by the backbone network, outputs the feature after processing by the first FFM module, focuses and diffuses it based on the feature focusing and diffusion mechanism FFDM, and then processes it through the second FFM module. The multi-scale features of the neck network are obtained through the mapping convolution of the corresponding scale.

[0011] S2.3: Input the multi-scale features of the neck network after feature focusing and diffusion processing into the detection head network to perform category prediction and location regression of targets in remote sensing images, and output the target detection results;

[0012] S3: Train the MFFDNET detection model using the training dataset, select the optimal model weights, obtain the final MFFDNET detection model, and use it for remote sensing target detection.

[0013] Furthermore, the specific steps of step S1 are as follows:

[0014] S1.1: The acquired image data is preprocessed using translation scaling, image rotation, and brightness enhancement methods;

[0015] S1.2: Use the data annotation tool Labelimg to annotate different target types, integrate the image dataset and the label dataset to create the original dataset; and divide it into training set, validation set and test set in a ratio of 8:1:1.

[0016] Furthermore, the backbone network is configured with two RMSCA modules and two layers of standard convolutions. The input image data is first processed through two layers of 3×3 standard convolutions to extract features, and then two layers of RMSCA modules are connected in series to construct the backbone network. The backbone network outputs three scale features, namely the features initially extracted by the second layer of 3×3 standard convolutions and the features output by the two RMSCA modules.

[0017] Furthermore, the RMSCA comprises three stages: multi-scale weight generation, cross-feature fusion, and channel recalibration. The specific methods for each stage are as follows:

[0018] Multi-scale weight generation: for input feature maps First, feature space compression is performed using traditional convolution. Then, an MSCA module is introduced. The MSCA module divides the feature map compressed by traditional convolution into two groups along the channel dimension. , Spatial attention weights are generated by using 3×3 and 5×5 depthwise separable convolutions, respectively, followed by a sigmoid function. , ;

[0019] Cross-feature fusion: combining the first set of feature maps Spatial attention weights generated by themselves respectively The second set of generated spatial attention weights Perform element-wise multiplication to obtain the weighted features of this group. , Add the two together to obtain the first set of feature outputs. ; the second set of feature maps Spatial attention weights generated by themselves respectively The first set of generated spatial attention weights Perform element-wise multiplication to obtain the weighted features of this group. , Add the two together to obtain the second set of feature outputs. The two sets of features after cross-fusion , The concatenation is then added to the original input features, which are weighted by global basic weights generated through a 1×1 convolution, to obtain... ;

[0020] Channel recalibration: For the original input Average pooling is performed, and the channel dependencies are learned through two fully connected layers. Channel weights are then generated using a sigmoid function. ,Will Expand to the same size as the input features, using channel weights. For the original input respectively and fusion features After weighting, the two are concatenated and then subjected to a 1×1 convolution to obtain the final output. .

[0021] Furthermore, the FFM module comprises three stages: multi-scale feature extraction, adaptive path selection, and multi-dimensional feature fusion. The specific methods for each stage are as follows:

[0022] Multi-scale feature extraction: The FFM module takes feature maps from three different channels as input. , , These correspond to the three different scales of feature maps in the backbone network, from shallowest to deepest; for the shallowest feature map... Downsampling is achieved using depthwise separable convolution to obtain feature maps. ; Middle layer feature map and After element-wise addition, feature maps are obtained by depthwise separable convolution downsampling. Then, the spatial size is restored through upsampling to obtain the feature map. Depth feature map and After element-wise addition, the result is processed in parallel with 1×1 convolution and upsampling operations to obtain the feature map. ;

[0023] Adaptive path selection: for intermediate layer feature maps Information weights are generated using average pooling and the sigmoid function, serving as a gating mechanism for global feature quality path selection. ;right and After normalization, the sigmoid function is used for activation to generate the corresponding weight information. , Based on the weight information generated above, the feature quality is graded: path selection is gated. respectively with weight information , The comparison generates four sets of Boolean feature masks, which are two sets of high-threshold masks. , With two sets of low threshold masks , The mask shape is consistent with the input features; the generated high-threshold mask is multiplied with the intermediate layer features respectively, and then the two multiplication results are added element by element to obtain the high-weight features. The generated low-threshold mask is multiplied by the intermediate layer features, and then the two multiplication results are added element-wise to obtain the low-weight features. ;

[0024] Multidimensional feature fusion: for high-weight features Use 1×1 convolution to fuse semantic information; for low-weight features The process employs two parallel parts. The first part involves 3×3 depthwise separable convolution, while the second part sequentially performs global pooling, ReLU activation, and Softmax normalization. The results from both parts are then multiplied element-wise to obtain the processed low-weight features. Finally, the processed high-weight features are added to the low-weight features, and a 1×1 convolution is applied to obtain the final output features. .

[0025] Furthermore, the specific method of the Feature Focused Diffusion Mechanism (FFDM) is as follows:

[0026] After multiple feature maps of different scales are input into the FFM module, a single-scale feature map is output. This single-scale feature map is then used as a mesoscale feature map. Bilinear interpolation and convolution operations are used to diffuse the single-scale feature map into large-scale and small-scale feature maps, respectively. Together with the original mesoscale feature map, these three scale feature maps are formed. These three scale feature maps are then input into the FFM module again to perform feature focusing processing. The processed output feature maps are then diffused into the detection heads of each detection scale through corresponding scale mapping convolutions.

[0027] Furthermore, the training parameters in S3 are as follows: the input image size is 640; the initial learning rate is set to 0.01; the learning rate momentum is set to 0.937; the weight decay coefficient is set to 0.0005; the number of training iterations is set to 200; the batch size is set to 64; and the training optimizer is SGD.

[0028] Beneficial effects:

[0029] 1. This invention uses a series of RMSCA modules and standard convolutions to construct the backbone network of the model. The RMSCA module adopts a multi-branch structure and captures local details of different receptive fields through two different sizes of depth separable convolutions, 3×3 and 5×5. It enhances information interaction through feature cross-fusion between different groups, so that the output feature map has cross-scale semantic information. On this basis, global weights of the original input are introduced to realize the fusion of shallow and deep semantic information, avoid information loss during feature interaction, and alleviate the gradient vanishing problem in deep network training, which is beneficial to the development of downstream tasks.

[0030] 2. The FFM module designed in the neck region of this invention can input feature maps of three different scales. It solves the problem of incomplete expression of single-path features through multi-dimensional feature processing. The module adds shallow and mid-level features element by element to form global basic features. By performing average pooling and sigmoid activation on the global basic features, it achieves channel-dimensional information compression, generates path selection gating, fully explores the local details and global correlation information of features, and filters out high-weight features and low-weight features for differentiated processing to enhance key information and suppress redundant noise. The core operation of the module is mainly 1×1 convolution and depthwise separable convolution. Compared with the traditional 3×3 convolution-dominated enhancement module, the computational cost is significantly reduced. The computational cost of the FFM module is only 65% ​​of that of the ResNet bottleneck module with the same number of channels.

[0031] 3. The main process of the Feature Focusing Diffusion (FFDM) mechanism designed in the neck region of this invention is as follows: three feature maps of different scales are input into the backbone network to the FFM module, outputting a single-scale feature map (medium scale), which is then diffused into large-scale and small-scale feature maps through bilinear interpolation, forming three scale feature maps together with the original medium-scale feature map; these three scale feature maps are then input into the FFM module again for feature focusing processing, and the processed feature maps are transmitted to the detection heads of each detection scale after corresponding scale mapping convolution; the first focusing diffusion integrates shallow and deep features to enhance semantic information; the second focusing diffusion further refines the feature expression, so that each scale contains richer contextual information, and the enhanced features are redistributed to each scale to improve the expressive ability of small targets on large-scale feature maps. Attached Figure Description

[0032] Figure 1 This is a flowchart of the method of the present invention;

[0033] Figure 2 This is a structural diagram of the MFFDNET model of the present invention;

[0034] Figure 3 This is a structural diagram of the RMSCA of the present invention;

[0035] Figure 4 This is a structural diagram of the FFM of the present invention;

[0036] Figure 5 This is a visual comparison of the features of the neck FFDM mechanism and FPN structure of the present invention;

[0037] Figure 6 This is a comparison chart of the performance parameters of this invention with other models;

[0038] Figure 7 This is a comparison chart of the actual test output results. Detailed Implementation

[0039] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0040] This embodiment provides a lightweight remote sensing target detection method based on feature focusing diffusion. The method includes acquiring the image to be identified, processing the image, and constructing an original dataset. A backbone network is constructed using RMSCA concatenated with the original convolutional layer, and a network neck is constructed using an FFM module and the feature focusing diffusion mechanism FFDM. The output features are then diffused to detection heads of different scales through bilinear interpolation and convolution operations. The processed image dataset is input into the MFFDNET network model for training, and the model performance is evaluated. Figure 1 As shown.

[0041] The overall architecture of this embodiment is divided into four parts: input layer, backbone, neck, and output layer. To address the issue of insufficient representational power of lightweight models for multi-scale target features in remote sensing images, the MFFDNET network model was constructed, as follows: Figure 2 As shown. The specific steps of this embodiment are as follows:

[0042] Step 1: Collect remote sensing image data and preprocess it to create image datasets and label datasets respectively. Then, integrate the image datasets and label datasets into the original dataset.

[0043] Step 1.1: Image Data Acquisition: Use a drone to acquire remote sensing data; collect remote sensing images from the internet, including independent remote sensing images and publicly available datasets. Scale the data to 640×640 pixels, perform data augmentation operations such as translation, rotation, and brightness enhancement, and create an image dataset.

[0044] Step 1.2: Use the LabelImg software to annotate the image dataset. Annotate the location and category of the targets to create a labeled dataset. Integrate the image dataset and the labeled dataset to create the original dataset, which is then divided into training, validation, and test sets in an 8:1:1 ratio.

[0045] Step 2: Input the dataset into the MFFDNET network model. The backbone network, composed of RMSCA and the original convolutional layers, will sequentially extract semantic features at different scales. The input image data first undergoes two layers of 3×3 standard convolutions for initial feature extraction, and then two layers of RMSCA modules are concatenated to construct the backbone network. RMSCA consists of three stages: multi-scale weight generation, cross-feature fusion, and channel recalibration. The specific structure diagram of RMSCA is shown below. Figure 3 As shown.

[0046] Step 2.1: In the multi-scale weight generation stage, in order to effectively integrate the spatial features of different receptive fields and enhance the inter-channel dependencies, the input feature map is... First, feature space compression is performed using traditional convolution. Then, the MSCA module is introduced. This module adopts a multi-branch structure, dividing the feature map compressed by traditional convolution into two groups along the channel dimension. , Spatial attention weights are generated by using 3×3 and 5×5 depthwise separable convolutions, respectively, followed by a sigmoid function. , The formula is expressed as follows:

[0047]

[0048] in: For the sigmoid function, This indicates that an n×n depthwise separable convolution process is performed.

[0049] Step 2.2: In the cross-feature fusion stage, cross-weight the different groups of features obtained in Step 2.1: Combine the first group of feature maps... Spatial attention weights generated by themselves respectively The second set of generated spatial attention weights Perform element-wise multiplication to obtain the weighted features of this group. , Add the two together to obtain the first set of feature outputs. ; the second set of feature maps Spatial attention weights generated by themselves respectively The first set of generated spatial attention weights Perform element-wise multiplication to obtain the weighted features of this group. , Add the two together to obtain the second set of feature outputs. The two sets of features after cross-fusion , The concatenation is then added to the original input features, which are weighted by global basic weights generated through a 1×1 convolution, to obtain the feature map. The formula is expressed as follows:

[0050]

[0051] in: This is the feature splicing function.

[0052] Step 2.3: In the channel recalibration stage, different channels of the input features correspond to different semantic information, and some channels may contain redundancy or noise. To enhance the response of key channels, a channel calibration branch is introduced to recalibrate the original input. Average pooling is performed, and the channel dependencies are learned through two fully connected layers. Channel weights are then generated using a sigmoid function. ,Will Expand to the same size as the input features, using channel weights. For the original input respectively and fusion features After weighting, the two are concatenated and then subjected to a 1×1 convolution to obtain the final output. , With initial feature input Element-wise addition forms residual connections, which fuses shallow and deep semantics, avoiding the gradient vanishing problem caused by complex interactions in network layers. The formula is expressed as follows:

[0053]

[0054]

[0055] in: Indicates average pooling. This indicates a fully connected layer.

[0056] Step 3: Input the feature maps of different scales output from the second layer of original convolution and the two layers of RMSCA in Step 2 into FFM for multi-scale feature focusing. The FFM module includes three stages: multi-scale feature extraction, adaptive path selection, and multi-dimensional feature fusion. The FFM structure diagram is shown below. Figure 4 As shown.

[0057] Step 3.1: In the multi-scale feature extraction stage, FFM receives feature maps from different scales of the backbone network. , , , where feature map Feature maps derived from shallower layers of the network have high spatial resolution but less semantic information, making them suitable for detecting small objects in remote sensing images. Feature maps derived from deeper layers of the network have lower spatial resolution but richer semantic information, making them suitable for detecting large objects in remote sensing images. For the shallowest feature maps... Downsampling is achieved using depthwise separable convolution to obtain feature maps. ; Middle layer feature map and After element-wise addition, feature maps are obtained by depthwise separable convolution downsampling. Then, the spatial size is restored through upsampling to obtain the feature map. Depth feature map and After element-wise addition, the result is processed in parallel with 1×1 convolution and upsampling operations to obtain the feature map. Intermediate layer feature map The formula is expressed as follows:

[0058]

[0059] in: This indicates an upsampling operation.

[0060] Step 3.2: In the adaptive path selection stage, the intermediate layer feature map obtained in step 3.1 is processed. Information weights are generated using average pooling and the sigmoid function, serving as a gating mechanism for global feature quality path selection. .right and After normalization, the sigmoid function is used for activation to generate the corresponding weight information. , Based on the weight information generated above, the feature quality is graded: path selection is gated. respectively with weight information , The comparison generates four sets of Boolean feature masks, which are two sets of high-threshold masks. , With two sets of low threshold masks , The mask shape is consistent with the input features; the generated high-threshold mask is multiplied with the intermediate layer features respectively, and then the two multiplication results are added element by element to obtain the high-weight features. The generated low-threshold mask is multiplied by the intermediate layer features, and then the two multiplication results are added element-wise to obtain the low-weight features. The relevant steps and formulas are expressed as follows:

[0061]

[0062]

[0063]

[0064]

[0065]

[0066]

[0067] in: This indicates normalization processing. For model learning parameters, Let these represent the mean and variance of the feature, respectively. This indicates a weight comparison operation.

[0068] Step 3.3: In the multi-dimensional feature fusion stage, the high-weight features obtained in step 3.2 are processed... With low-weight features Differentiated enhancement: High-weight features It incorporates more effective semantic information, using 1×1 convolution to fuse channel semantics, avoiding information loss due to overprocessing while controlling computational cost; it also addresses low-weight features. The process employs a two-part parallel processing. The first part performs a 3×3 depthwise separable convolution to extract local spatial detail features from each channel without increasing the number of channels, compensating for the lack of detail in low-quality features. Based on this, a 1×1 convolution is used to fuse channel information, associating the local detail features of a single channel with all channels, thus improving the channel correlation of low-weight features. The second part sequentially performs global pooling, ReLU activation, and Softmax normalization to generate channel-level weights, filtering effective detail features after convolution and suppressing redundant noise. The results from the first two parts are then multiplied element-wise, using gated weights as a mask to weight low-weight enhanced features, preserving key details. Finally, the processed high-weight output and low-weight output are added element-wise, and a 1×1 convolution is used for feature fusion, achieving complementarity between features of different quality, and outputting the final feature map. The formula is expressed as follows:

[0069]

[0070] in: This indicates Softmax normalization.

[0071] Step 4: Extract the final output feature map from Step 3. As mesoscale feature maps, bilinear interpolation and convolution operations were used respectively to... The feature maps are diffused into large-scale and small-scale feature maps, which, together with the original medium-scale feature map, constitute three-scale feature maps. These three-scale feature maps are then input into the FFM module for feature focusing processing. The processed output feature maps are then diffused into the detection heads at each detection scale through corresponding scale mapping convolutions. The above "multi-scale input focusing into a single scale, and single-scale diffusion into multiple scales" is the core idea of ​​this invention's FFDM. Unlike the traditional Feature Pyramid (FPN) construction, the FFDM in this paper has a better feature focusing effect on remote sensing targets. A feature visualization comparison between the neck FFDM mechanism and the FPN structure is shown in the figure below. Figure 5 As shown.

[0072] Step 5: Train the MFFDNET network model using the dataset processed in Step 1. Set the training parameters as follows: input image size is 640; initial learning rate is 0.01; learning rate momentum is 0.937; weight decay coefficient is 0.0005; number of training iterations is 200; batch size is 64; and SGD is selected as the training optimizer.

[0073] Step 6: Evaluate the training results. The specific evaluation metrics are: Precision (P), Recall (R), and Mean Accuracy (mAP). The specific formulas are as follows:

[0074]

[0075]

[0076]

[0077] Where: TP represents the number of samples that are positive and predicted as positive, FP represents the number of samples that are negative and predicted as positive, FN represents the number of samples that are positive and predicted as negative, and n represents the number of target classes detected. This represents the average precision of the i-th target class.

[0078] The experimental environment parameters in this embodiment are shown in Table 1 below.

[0079] Table 1

[0080] Environmental parameters operating system Windows 11 CPU AMD Ryzen 9 7945HX GPU NVIDIA GeForce RTX 4060 GPU RAM 16G ROM 1T Python 3.9 Torch 2.2.1

[0081] This embodiment includes a comparative experiment to verify the effectiveness of the proposed method. The MFFDNET detection model is compared with Yolov8n and Yolov10s on the same dataset. The experimental results are as follows: Figure 6 As shown in the figure, the recognition accuracy (mAP) is significantly improved compared to other models. The specific remote sensing target detection output results are shown in the figure below. Figure 7As shown in the figure. The experimental results above demonstrate that the present invention achieves a balance between lightweight design and high precision through the collaborative design of the RMSCA module, FFM module, and FFDM mechanism, effectively solving the existing problems of multi-scale target detection in remote sensing images and meeting the needs of practical remote sensing traffic monitoring scenarios.

[0082] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent transformations or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A lightweight remote sensing target detection method based on feature-focused diffusion, characterized in that, Includes the following steps: S1: Collect remote sensing image data and preprocess it to create image datasets and label datasets respectively, and integrate the image datasets and label datasets into the original dataset; S2: Construct the MFFDNET detection model; S2.1: Concatenate the RMSCA module with standard convolution to construct the backbone network, extract features from the preprocessed remote sensing image, and obtain feature maps at different scales; S2.2: Two FFM modules are connected in series, and a model neck structure is constructed based on the feature focusing and diffusion mechanism FFDM. The neck structure receives the multi-scale feature map output by the backbone network, outputs the feature after processing by the first FFM module, focuses and diffuses it based on the feature focusing and diffusion mechanism FFDM, and then processes it through the second FFM module. The multi-scale features of the neck network are obtained through the mapping convolution of the corresponding scale. S2.3: Input the multi-scale features of the neck network after feature focusing and diffusion processing into the detection head network to perform category prediction and location regression of targets in remote sensing images, and output the target detection results; S3: Train the MFFDNET detection model using the training dataset, select the optimal model weights, obtain the final MFFDNET detection model, and use it for remote sensing target detection.

2. The lightweight remote sensing target detection method based on feature focusing diffusion according to claim 1, characterized in that, The specific steps of step S1 are as follows: S1.1: The acquired image data is preprocessed using translation scaling, image rotation, and brightness enhancement methods; S1.2: Use the data annotation tool Labelimg to annotate different target types, integrate the image dataset and the label dataset to create the original dataset; and divide it into training set, validation set and test set in a ratio of 8:1:

1.

3. The lightweight remote sensing target detection method based on feature focusing diffusion according to claim 1, characterized in that, The backbone network consists of two RMSCA modules and two layers of standard convolutions. The input image data is first processed through two layers of 3×3 standard convolutions to extract features, and then two layers of RMSCA modules are connected in series to construct the backbone network. The backbone network outputs three scale features, namely the features extracted by the second layer of 3×3 standard convolutions and the features output by the two RMSCA modules.

4. A lightweight remote sensing target detection method based on feature-focused diffusion according to claim 1 or 3, characterized in that, The RMSCA consists of three stages: multi-scale weight generation, cross-feature fusion, and channel recalibration. The specific methods for each stage are as follows: Multi-scale weight generation: for input feature maps First, feature space compression is performed using traditional convolution. Then, an MSCA module is introduced. The MSCA module divides the feature map compressed by traditional convolution into two groups along the channel dimension. , Spatial attention weights are generated by using 3×3 and 5×5 depthwise separable convolutions, respectively, followed by a sigmoid function. , ; Cross-feature fusion: combining the first set of feature maps Spatial attention weights generated by themselves respectively The second set of generated spatial attention weights Perform element-wise multiplication to obtain the weighted features of this group. , Add the two together to obtain the first set of feature outputs. ; the second set of feature maps Spatial attention weights generated by themselves respectively The first set of generated spatial attention weights Perform element-wise multiplication to obtain the weighted features of this group. , Add the two together to obtain the second set of feature outputs. The two sets of features after cross-fusion , The concatenation is then added to the original input features, which are weighted by global basic weights generated through a 1×1 convolution, to obtain... ; Channel recalibration: For the original input Average pooling is performed, and the channel dependencies are learned through two fully connected layers. Channel weights are then generated using a sigmoid function. ,Will Expand to the same size as the input features, using channel weights. For the original input respectively and fusion features After weighting, the two are concatenated and then subjected to a 1×1 convolution to obtain the final output. .

5. A lightweight remote sensing target detection method based on feature focusing diffusion according to claim 1, characterized in that, The FFM module comprises three stages: multi-scale feature extraction, adaptive path selection, and multi-dimensional feature fusion. The specific methods for each stage are as follows: Multi-scale feature extraction: The FFM module takes feature maps from three different channels as input. , , These correspond to the three different scales of feature maps in the backbone network, from shallowest to deepest; for the shallowest feature map... Downsampling is achieved using depthwise separable convolution to obtain feature maps. ; Middle layer feature map and After element-wise addition, feature maps are obtained by depthwise separable convolution downsampling. Then, the spatial size is restored through upsampling to obtain the feature map. Depth feature map and After element-wise addition, the result is processed in parallel with 1×1 convolution and upsampling operations to obtain the feature map. ; Adaptive path selection: for intermediate layer feature maps Information weights are generated using average pooling and the sigmoid function, serving as a gating mechanism for global feature quality path selection. ;right and After normalization, the sigmoid function is used for activation to generate the corresponding weight information. , Based on the weight information generated above, the feature quality is graded: path selection is gated. respectively with weight information , The comparison generates four sets of Boolean feature masks, which are two sets of high-threshold masks. , With two sets of low threshold masks , The mask shape is consistent with the input features; the generated high-threshold mask is multiplied with the intermediate layer features respectively, and then the two multiplication results are added element by element to obtain the high-weight features. The generated low-threshold mask is multiplied by the intermediate layer features, and then the two multiplication results are added element-wise to obtain the low-weight features. ; Multidimensional feature fusion: for high-weight features Use 1×1 convolution to fuse semantic information; for low-weight features The process employs two parallel parts. The first part involves 3×3 depthwise separable convolution, while the second part sequentially performs global pooling, ReLU activation, and Softmax normalization. The results from both parts are then multiplied element-wise to obtain the processed low-weight features. Finally, the processed high-weight features are added to the low-weight features, and a 1×1 convolution is applied to obtain the final output features. .

6. A lightweight remote sensing target detection method based on feature-focused diffusion according to claim 1 or 5, characterized in that, The Feature Focused Diffusion Mechanism (FFDM) is described in the following specific method: After multiple feature maps of different scales are input into the FFM module, a single-scale feature map is output. This single-scale feature map is then used as a mesoscale feature map. Bilinear interpolation and convolution operations are used to diffuse the single-scale feature map into large-scale and small-scale feature maps, respectively. Together with the original mesoscale feature map, these three scale feature maps are formed. These three scale feature maps are then input into the FFM module again to perform feature focusing processing. The processed output feature maps are then diffused into the detection heads of each detection scale through corresponding scale mapping convolutions.

7. A lightweight remote sensing target detection method based on feature focusing diffusion according to claim 1, characterized in that, The specific training parameters in S3 are as follows: the input image size is 640; the initial learning rate is set to 0.01; and the learning rate momentum is set to 0.

937. The weight decay factor is set to 0.0005; the number of training iterations is set to 200; and the batch size is set to 64. The training optimizer is SGD.

Images