A remote sensing image target detection method based on a multi-scale cross-stage network model

Abstract

This invention relates to the field of remote sensing image processing technology, specifically disclosing a remote sensing image target detection method based on a multi-scale, cross-stage network model. The method includes the following steps: acquiring the remote sensing image to be tested; inputting the image into the backbone network after passing through an input layer, extracting multi-scale and multi-level semantic features step-by-step, and obtaining feature maps of different sizes through multi-scale downsampling; inputting the feature maps of different sizes into a feature integration network for cross-stage fusion and lightweight optimization to obtain multiple fused feature maps; and inputting the multiple fused feature maps into a prediction head for classification and detection to obtain detection results at multiple scales. This invention, by constructing a novel multi-scale, cross-stage network model, enhances the capabilities of multi-scale feature extraction, fusion, and target recognition, enabling accurate detection of multi-scale targets, especially small and complex-shaped targets, in remote sensing images. It balances detection accuracy and inference efficiency, making it suitable for remote sensing applications such as marine monitoring and urban planning.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, and in particular to a remote sensing image target detection method based on a multi-scale cross-stage network model. Background Technology

[0002] Target detection in remote sensing images is crucial for applications such as maritime surveillance, environmental monitoring, and urban planning, making performance improvement a key research focus. Traditional feature extraction methods typically rely on a series of fundamental geometric features, such as edges, textures, and colors, achieved through techniques like Harris corner detection, Histogram of Oriented Gradients (HOG), Haar features, and Scale Invariant Feature Transform (SIFT). For instance, Chinese invention patent application CN107992818A discloses a method for detecting ship targets on the sea surface in optical remote sensing images. By judging geometric features such as the target's aspect ratio and symmetry, it further excludes features like waves and clouds, and obtains the geometric information of the ship target, thereby detecting the ship. Subsequently, template matching or machine learning methods are typically employed to perform object detection tasks. For example, Chinese invention patent application CN114022812A discloses a DeepSort multi-target tracking method for floating objects on the water surface based on a lightweight SSD. This method uses a lightweight SSD detection algorithm to obtain the target detection boxes for floating objects on the water surface in the current frame, and performs state prediction based on the target detection boxes in the current video frame to obtain the target tracking boxes for floating objects on the water surface in the current video frame. However, remote sensing images typically have high resolution and large amounts of data, and traditional methods have certain limitations in efficiency, resulting in high time and labor costs. Therefore, improving the efficiency of target detection in remote sensing images and reducing time and labor costs is one of the urgent needs in this field.

[0003] Compared to traditional remote sensing image target detection methods, deep learning methods eliminate the need for tedious manual feature extraction and prior information setup. Instead, deep neural network models directly participate in the target detection task, simplifying the entire detection process. This new method not only significantly improves detection speed but also automatically learns more complex target features, resulting in a significant improvement in detection performance. Therefore, the application of deep learning technology in remote sensing image analysis is gradually increasing, bringing important progress to the field of remote sensing image target detection. However, due to the complexity of remote sensing images and some limitations of general-purpose deep neural networks, existing methods still have some problems, especially when dealing with remote sensing images of targets with large scale variations, small targets, and complex shapes; their detection performance and robustness remain unsatisfactory.

[0004] Existing remote sensing image target detection models suffer from problems such as insufficient diversity in feature extraction, inadequate feature interaction between channels, need to improve feature fusion efficiency, and limited accuracy in detecting small targets and complex-shaped targets. Summary of the Invention

[0005] This invention aims to address the shortcomings of existing remote sensing image target detection methods when handling scenes with densely packed small targets and complex backgrounds. These shortcomings include insufficient feature representation capabilities, poor multi-scale adaptability, high computational redundancy, and uneven attention paid to different dimensions of information by the detection head. To this end, this invention provides a remote sensing image target detection method based on a multi-scale cross-stage network model (MRMF-CFF). By constructing a novel MRMF-CFF model, it enhances multi-scale feature extraction, fusion, and target recognition capabilities, enabling accurate detection of multi-scale targets, especially small and complex-shaped targets, in remote sensing images. This improves the detection accuracy of small targets in complex backgrounds and the model inference efficiency, making it suitable for remote sensing applications such as marine monitoring and urban planning.

[0006] This invention provides a remote sensing image target detection method based on a multi-scale, cross-stage network model, and the technical solution adopted is as follows: including the following steps:

[0007] S1: Acquire the remote sensing image to be tested;

[0008] S2: The remote sensing image to be tested is input into the backbone network after passing through the input layer. The backbone network extracts multi-scale and multi-level semantic features from the input image step by step, and obtains feature maps of different sizes through multi-scale downsampling.

[0009] The backbone network optimizes feature representation capabilities and enhances feature interaction between channels by integrating the MRMF module and the CFFtc-CR module. The MRMF module has a multi-branch structure, including max pooling branches and multiple convolutional branches with different kernel sizes. Multi-dimensional features are obtained by concatenating and fusing features from multiple branches to enhance the diversity of feature extraction. The CFFtc-CR module achieves parallel extraction of local details and contextual information through parallel multi-scale convolution, and refines local features to global features through global average pooling and fully connected layers.

[0010] S3: Input feature maps of different sizes into the feature integration network, perform cross-stage fusion and lightweight optimization, and obtain multiple fused feature maps;

[0011] S4: Input multiple fused feature maps into the prediction head for classification and detection to obtain detection results at multiple scales.

[0012] Furthermore, the specific processing flow of the backbone network is as follows: the input image passes through 2 MRMF modules, 3 CFFtc modules, 1 MRMF module, and 6 CFFtc modules in sequence to obtain the first feature map; the first feature map passes through the MRMF module and the CFFtc-CR module to obtain the second feature map; the second feature map passes through the MRMF module, the CFFtc-CR module, and the PMPM module to obtain the third feature map;

[0013] The backbone network outputs a first feature map, a second feature map, and a third feature map.

[0014] Furthermore, the specific structure of the MRMF module is as follows: the first branch uses a 5×5 depthwise separable convolution with a stride of 2 followed by a 1×1 convolution; the second branch uses a 7×7 depthwise separable convolution with a stride of 2 followed by a 1×1 convolution; and the third branch uses max pooling followed by a 1×1 convolution. The features output by the three branches are concatenated through channels to obtain multi-dimensional features.

[0015] All three branches use the XCON adaptive activation function.

[0016] Furthermore, the processing flow of the CFFtc-CR module is as follows: perform a convolution operation on the input feature map, split the convolutional feature map into two parallel branches, one branch passes through the CR layer and is concatenated with the other branch, and then perform a convolution operation on the concatenated feature map.

[0017] The processing flow of the CR layer is as follows: the input feature map is processed by 1×1 convolution, and then input by 3×3 convolution and 3×3 dilated convolution respectively. The output feature maps are concatenated and then processed by batch normalization and PReLU activation function to obtain intermediate feature maps. After the intermediate feature maps are processed by global average pooling and two fully connected layers, they are multiplied element-wise with the intermediate feature maps.

[0018] Furthermore, the feature integration network constructs a bidirectional feature pyramid structure through the CFF module, upsampling layer, and concatenation layer. The CFF module optimizes feature transfer efficiency by using a cross-stage partial connection strategy. The upsampling and concatenation operations upsample the low-level high-resolution feature map through bilinear interpolation and concatenate it with the high-level semantic feature map element by element, thereby realizing a feature pyramid structure that combines top-down and bottom-up approaches.

[0019] The specific process of the feature integration network is as follows: The third feature map is upsampled and then concatenated with the second feature map, and processed by the CFFtc-CR module to obtain the fourth feature map; the fourth feature map is upsampled and then concatenated with the first feature map, and processed by the CFFtc-CR module to obtain the first fused feature map; the first fused feature map is processed by the MRMF module, then concatenated with the fourth feature map, and processed by the CFF module to obtain the second fused feature map; the second fused feature map is processed by the MRMF module, then concatenated with the third feature map, and processed by the CFF module to obtain the third fused feature map.

[0020] The feature integration network outputs a first fused feature map, a second fused feature map, and a third fused feature map.

[0021] Furthermore, the processing flow of the CFF module is as follows: the input feature map is divided into two channel dimension branches along the channel dimension: the first channel dimension branch is compressed by 1×1 convolution; the second channel dimension branch is enhanced and transformed by 1×1 convolution, reparameterized convolution and 3×3 convolution in sequence; the output feature maps of the two channel dimension branches are fused by channel splicing.

[0022] Furthermore, the prediction head includes multiple MD modules, which are used to classify and detect the fused feature map; the MD modules apply attention modulation to the fused feature map in three dimensions: hierarchical, spatial and channel.

[0023] Furthermore, the processing flow of the MD module is as follows:

[0024] The fused feature map is subjected to global average pooling, followed by 1×1 convolution, ReLU activation and Hard Sigmoid function to generate a weight map; the weight map and the fused feature map are multiplied element-wise to obtain the scale feature map.

[0025] For the scale feature map, key spatial locations are extracted through indexing operations, and a spatial attention map is generated by 3×3 convolution and the Sigmoid function. The spatial attention map and the offset mechanism are then applied to the scale feature map to obtain the spatial feature map.

[0026] For the spatial feature map, after global average pooling, the first fully connected layer, the ReLU activation function, the second fully connected layer, and normalization, channel-level weights are generated. The channel-level weights are added element-wise to a fixed vector and then multiplied element-wise with the spatial feature map to obtain the detection result.

[0027] Furthermore, S5: Select the target to be optimized from the detection results, optimize it according to the probability of its candidate box category, and obtain the optimized result map.

[0028] Furthermore, the multi-scale cross-stage network model includes an input layer, a backbone network, a feature integration network, and a prediction head. The multi-scale cross-stage network model is a pre-trained model, and its loss function is a weighted sum of IoU loss, distance loss, and aspect ratio loss.

[0029] The above-described one or more technical solutions in the embodiments of the present invention have at least one of the following technical effects:

[0030] (1) This invention significantly enhances the ability to extract multi-scale features of small targets in remote sensing images by using parallel depth convolution and pooling paths of different sizes, effectively alleviating the problem of missed detection caused by drastic changes in target scale.

[0031] (2) The present invention proposes an MRMF module, which combines parallel 5×5 and 7×7 depth separable convolution and max pooling paths with feature extraction of different receptive fields and XCON adaptive activation function to achieve effective fusion of multi-scale features, significantly improve the detection capability of multi-scale small targets in remote sensing images, and reduce the false negative rate caused by scale changes.

[0032] (3) The present invention designs the CFFtc-CR module, which integrates local details and global context information, and combines channel weighting mechanism to significantly improve the robustness of the network to complex backgrounds while maintaining lightweight design.

[0033] (4) The present invention designs a CFF module, which uses feature segmentation and multi-branch aggregation strategy to reduce feature redundancy and computational burden, and achieve synergistic optimization of accuracy and efficiency.

[0034] (5) The present invention designs an MD module that performs fine feature modulation in the hierarchical, spatial and channel dimensions, and adaptively responds to different task requirements such as multi-scale, spatial location and classification and localization, significantly improving the target recognition accuracy in complex scenarios.

[0035] (6) This invention proposes the PMPM module, which achieves efficient global context modeling by performing single-branch convolution preprocessing followed by parallel multi-scale max pooling operations, significantly enhancing the network's ability to perceive a wide range of contextual information and effectively improving the accuracy of target localization and recognition in complex scenarios.

[0036] In summary, this invention is particularly suitable for remote sensing image analysis tasks involving high resolution, dense small targets, and complex backgrounds, and has broad application prospects in fields such as UAV aerial photography, satellite remote sensing monitoring, and urban planning.

[0037] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in this invention 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 some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0039] Figure 1 This is a flowchart of the method provided by the present invention.

[0040] Figure 2 This is a structural diagram of the multi-scale, cross-stage network model provided by the present invention.

[0041] Figure 3 This is a structural diagram of the MRMF module provided by the present invention.

[0042] Figure 4 This is a structural diagram of the CFFtc-CR module provided by the present invention.

[0043] Figure 5 This is a structural diagram of the PMPM module provided by the present invention.

[0044] Figure 6 This is a structural diagram of the CFF module provided by the present invention.

[0045] Figure 7 This is a structural diagram of the MD module provided by the present invention. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention. The following embodiments are used to illustrate this invention but should not be used to limit the scope of this invention.

[0047] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0048] The following is combined with Figures 1 to 7 The present invention will be further described in detail below, which describes a remote sensing image target detection method based on a multi-scale cross-stage network model:

[0049] In this embodiment, as Figure 1 As shown, a remote sensing image target detection method based on a multi-scale, cross-stage network model is provided, including the following steps:

[0050] S1: Acquire the remote sensing image to be tested.

[0051] S2: The remote sensing image to be tested is input into the backbone network after passing through the input layer. The backbone network extracts multi-scale and multi-level semantic features from the input image step by step, and obtains feature maps of different sizes through multi-scale downsampling. The backbone network optimizes the feature representation capability by integrating the MRMF module and the CFFtc-CR module, and enhances the feature interaction between channels.

[0052] like Figure 2 As shown, the specific processing flow of the backbone network is as follows: the input image passes through 2 MRMF modules, 3 CFFtc modules, 1 MRMF module, and 6 CFFtc modules in sequence to obtain the first feature map; the first feature map passes through the MRMF module and the CFFtc-CR module to obtain the second feature map; the second feature map passes through the MRMF module, the CFFtc-CR module, and the PMPM module to obtain the third feature map.

[0053] The backbone network outputs a first feature map, a second feature map, and a third feature map.

[0054] The MRMF module has a multi-branch structure, including a max pooling branch and multiple convolutional branches with different kernel sizes. It outputs multi-dimensional features by splicing and fusing features from multiple branches to enhance the diversity of feature extraction.

[0055] like Figure 3As shown, in this embodiment, the MRMF module includes three parallel branches, specifically structured as follows: the first branch uses a 5×5 depthwise separable convolution with a stride of 2 followed by a 1×1 ordinary convolution; the second branch uses a 7×7 depthwise separable convolution with a stride of 2 followed by a 1×1 ordinary convolution; and the third branch uses max pooling followed by a 1×1 ordinary convolution. Each branch employs the XCON adaptive activation function. The features output from the three branches are concatenated through channels to obtain multi-dimensional features. The MRMF module achieves multi-receptive field feature fusion.

[0056] The XCON activation function uses learnable parameters to control the intensity of nonlinearity. Its calculation method is as follows: the input feature is multiplied by a first learnable scalar parameter, and then multiplied by a result obtained by multiplying the input feature by a second learnable scalar parameter and processing it with a sigmoid function. The two scalar parameters are automatically optimized during training using a backpropagation algorithm, enabling the activation function to adaptively adjust the balance between linear transformation and nonlinear mapping, thereby enhancing the model's expressive power and convergence stability during training.

[0057] The CFFtc-CR module achieves parallel extraction of local details and contextual information through parallel multi-scale convolutions, while refining local features to global features through global average pooling and fully connected layers. By combining parallel extraction of local details and contextual information with local-to-global feature refinement, the CFFtc-CR module achieves an efficient multi-scale feature fusion and adaptive feature enhancement mechanism.

[0058] like Figure 4 As shown, in this embodiment, the CFFtc-CR module includes stacked convolutional layers, segmentation layers, CR layers, concatenation layers, and convolutional layers. The processing flow is as follows: The input feature map is convolved to extract basic features; the convolved feature map is split into two parallel branches through the segmentation layer; one branch is passed through the CR layer and then concatenated with the other branch to fuse the features of the two branches; the concatenated feature map is convolved again to integrate the features; finally, the integrated feature map is output.

[0059] The CR layer contains two parallel paths: one path extracts local detail features through 1×1 convolutions and 3×3 standard convolutions, while the other path expands the receptive field to capture contextual information through 3×3 dilated convolutions. The two feature paths are concatenated and then processed by batch normalization and PReLU activation. Global average pooling, fully connected layers, and a weighting mechanism are further introduced to complete global context aggregation and feature refinement. The processing flow is as follows: the input feature map is processed by 1×1 convolutions to adjust the number of channels for dimensionality adaptation. Then, 3×3 standard convolutions and 3×3 dilated convolutions are input separately. The output feature maps of the 3×3 standard convolutions and 3×3 dilated convolutions are concatenated and then processed by batch normalization and PReLU activation to obtain an intermediate feature map. The intermediate feature map is then processed by global average pooling and two fully connected layers, and finally multiplied element-wise with the intermediate feature map to achieve adaptive weighting of the features.

[0060] In this embodiment, the CFFtc-CR module replaces the Bottleneck module of the CFFtc module with a CR layer. Compared to the CFFtc module, the CFFtc-CR module, which uses a CR layer, enhances the network's ability to process information at various scales, achieving a balance between local details and global context. Due to the use of channel convolution, the computational cost across channels is reduced, resulting in a certain degree of lightweighting.

[0061] The PMPM module employs single-branch convolution preprocessing followed by parallel multi-scale max pooling operations to achieve efficient global context modeling. For example... Figure 5 As shown, in this embodiment, the processing flow of the PMPM module is as follows: the input feature map first passes through the CBS module for feature extraction and nonlinear transformation, the output of the CBS module passes through three max pooling layers in sequence, and the outputs of the CBS module and the max pooling layers are concatenated and then input into another CBS module.

[0062] S3: Input feature maps of different sizes into the feature integration network, perform cross-stage fusion and lightweight optimization, and obtain multiple fused feature maps rich in contextual information.

[0063] The feature fusion network constructs a bidirectional feature pyramid structure through a CFF module, an upsampling layer, and a concatenation layer to achieve effective fusion of multi-level features. Specifically, the CFF module optimizes feature transfer efficiency using a cross-stage partial connectivity strategy. The upsampling and concatenation operations upsample the low-level high-resolution feature maps using bilinear interpolation and then concatenate them element-wise with the high-level semantic feature maps, achieving a feature pyramid structure that combines top-down and bottom-up approaches, thus enhancing the collaborative expression of fine-grained details and high-level semantics.

[0064] The specific process of the feature integration network is as follows: After the first, second, and third feature maps are input into the feature integration network, the network upsamples the third feature map and then concatenates it with the second feature map. This concatenation is then processed by the CFFtc-CR module to obtain the fourth feature map. The fourth feature map is then upsampled and concatenated with the first feature map, and processed by the CFFtc-CR module to obtain the first fused feature map. The first fused feature map is processed by the MRMF module, then concatenated with the fourth feature map, and processed by the CFF module to obtain the second fused feature map. The second fused feature map is processed by the MRMF module, then concatenated with the third feature map, and processed by the CFF module to obtain the third fused feature map.

[0065] The feature integration network outputs a first fused feature map, a second fused feature map, and a third fused feature map.

[0066] The CFF module is based on a cross-stage partial connection strategy, such as... Figure 6 As shown, the processing flow is as follows: the input feature map is divided into two channel-dimensional branches along the channel dimension; the first channel-dimensional branch undergoes channel compression through a 1×1 convolution; the second channel-dimensional branch sequentially performs feature enhancement and transformation through a 1×1 convolution, a 3×3 reparameterized convolution, and a 3×3 standard convolution; the output feature maps of the two channel-dimensional branches are then fused by channel concatenation. The CFF module effectively alleviates the gradient redundancy problem, enhancing feature diversity while reducing computational overhead.

[0067] S4: Input multiple fused feature maps into the prediction head for classification and detection, obtaining detection results at multiple scales. The detection results include the location of the target in the image to be classified and the category corresponding to each target.

[0068] The prediction head is used to perform classification and regression tasks on the fused feature map, outputting the final detection results. The prediction head includes multiple MD modules, which are used to perform classification and detection on the fused feature map. The MD modules are designed based on deformable convolution mechanisms, performing fine-tuning on the multi-scale fused feature map output by the feature integration network. The MD modules employ a hierarchical, task-driven, multi-dimensional separable attention mechanism. Through cascaded scale, spatial, and channel-aware modules, attention modulation is applied to the fused feature map sequentially in the hierarchical, spatial, and channel dimensions, achieving comprehensive and fine-tuning from macro-contextuality to micro-discriminability, and finally to task-specific requirements.

[0069] like Figure 7 As shown, the specific processing flow of the MD module is as follows:

[0070] The first step, scale-sensing attention: global average pooling is performed on the fused feature map, followed by 1×1 convolution and Hard Sigmoid function to generate a weight map, realizing adaptive weighted fusion of cross-level features and enhancing the model's ability to perceive multi-scale targets; the weight map and the fused feature map are multiplied element-wise to obtain the scale feature map.

[0071] The second step is spatial sensing attention: For the scale feature map, key spatial locations are extracted through indexing operations, and a spatial attention map is generated by 3×3 convolution and the Sigmoid function. The offset mechanism is then used to focus on the target discriminative region and suppress interference from complex backgrounds. The spatial attention map and the offset are then applied to the scale feature map to obtain the spatial feature map.

[0072] The third step is task-sensing attention: For the spatial feature map, after global average pooling, the first fully connected layer, the ReLU activation function, the second fully connected layer, and normalization, channel-level weights are generated. Channels are activated for the classification and localization sub-tasks respectively to achieve task-oriented feature selection and optimization. The channel-level weights are added element-wise to the fixed vector [1,0,0,0] and then multiplied element-wise with the spatial feature map to obtain the detection result.

[0073] In this embodiment, the MD module is used to classify and locate targets in the first fused feature map, the second fused feature map, and the third fused feature map in the prediction head; finally, the bounding box coordinates, category confidence, and precise location information of each type of target are output.

[0074] In this embodiment, in step S1, a 3-channel color image (RGB) with a size of 640×640 is input, and the prediction head outputs detection results at three scales: 80×80, 40×40, and 20×20.

[0075] The multi-scale, multi-stage network model enhances multi-scale feature extraction capabilities by integrating the MRMF module, optimizes feature representation with the CFFtc-CR module, and improves feature fusion efficiency through the CFF module. In addition, the MD module significantly improves detection accuracy, effectively enhancing the detection performance of multi-scale targets in remote sensing images, especially small targets and objects with complex shapes.

[0076] S5: Select the target to be optimized from the detection results, optimize it according to the probability of its candidate box category, and obtain the optimized result map.

[0077] The specific process is as follows: S51: Select any target from the detection results as the target to be optimized, and obtain the probability of each candidate box category;

[0078] S52: Calculate the detection score of each candidate box by category probability using nonmaximum suppression;

[0079] S53: Delete all candidate boxes except the one corresponding to the highest detection score to obtain the optimized result image.

[0080] The MRMF-CFF used in this embodiment is a trained model, and its training process is as follows:

[0081] Step 1: Obtain the remote sensing image dataset and preprocess the remote sensing image dataset.

[0082] Step 11: This embodiment uses a publicly available remote sensing image dataset. Specifically, this embodiment uses the remote sensing image dataset DOTA and selects 2500 original remote sensing images from it.

[0083] Step 12: Preprocess the original remote sensing images, specifically: resample all images (2500 images) to a given resolution, enhance image contrast using histogram equalization, crop high-resolution images using a sliding window strategy, and generate label files containing target category labels and precise bounding box coordinates through manual annotation. Specifically:

[0084] Step 121: Sample the original remote sensing images to the same resolution, and perform histogram equalization on the remote sensing images after unification to enhance image contrast and highlight target details and edges;

[0085] Step 122: Use a sliding window strategy to crop and equalize the image, segmenting the large-size remote sensing image into smaller images that meet the needs of subsequent processing, while preserving local ground feature details;

[0086] Step 123: Generate label information containing target category and bounding box through manual annotation.

[0087] Step 13: Divide the labeled data into training, validation, and test sets in an 8:1:1 ratio. That is, select 2000 images as the training set, 250 images as the validation set, and 250 images as the test set.

[0088] Step 2: Construct a deep learning model, the architecture of which is as follows: Figure 2 As shown. Deep learning models include an input layer ( Figure 2 (Not shown in the diagram), backbone network, feature integration network, and prediction head.

[0089] Step 3: Train the deep learning model using the preprocessed remote sensing image dataset to obtain MRMF-CFF. Specifically:

[0090] Using 2000 training samples as input to the deep learning model, the training batch size was set to 64, and the maximum number of iterations was set to 300. The input was passed through the backbone network, the feature integration network, and the prediction head in sequence. The gradient was obtained through label matching and loss calculation, and the model parameters were updated through backpropagation.

[0091] The loss function for object detection consists of IoU loss, distance loss, and aspect ratio loss. IoU loss measures the overlap between the predicted and ground truth bounding boxes; distance loss calculates the normalized Euclidean distance between the center points of the predicted and ground truth boxes; and aspect ratio loss penalizes inconsistencies in the aspect ratio between the predicted and ground truth boxes. These three losses work synergistically to accelerate model convergence and significantly improve object localization accuracy.

[0092] loss function Loss of IoU Distance loss Aspect Ratio Loss The weighted sum is calculated using the following formula:

[0093]

[0094] in, Weights are lost for IoU. For distance loss weights, The aspect ratio loss weight is used.

[0095] The IoU loss is calculated based on the Intersection over Union (IoU). The IoU loss is 0 when the predicted box completely overlaps with the ground truth box, and 1 when there is no overlap. This part directly reflects the accuracy of the box's position. The distance loss reduces scale sensitivity through normalization (e.g., using the diagonal length of the smallest bounding rectangle), ensuring fast convergence of the center point. The aspect ratio loss measures the difference in width and height between the predicted and ground truth boxes, i.e., shape similarity. This part penalizes shape mismatch by comparing the aspect ratios of the predicted and ground truth boxes, helping the model generate bounding boxes that better match the real-world shape. The overall loss function is a weighted sum of these three parts, balancing overlap, distance, and shape factors to allow the model to more stably optimize bounding box parameters during training.

[0096] After the above training is completed, the final weights of the deep convolutional network model for target detection in remote sensing images are obtained. The weight file is then loaded into the deep learning model to obtain a multi-scale, multi-stage network model (target detection model).

[0097] The testing steps for the object detection model are as follows:

[0098] (1) Obtain any remote sensing image from the remote sensing dataset. The range and resolution of the test sample images are not required, nor are their size and shape restricted. No additional preprocessing is required for any test sample images.

[0099] (2) After all the steps of training the detection model are completed, load the optimal weights, adjust the input image to 640×640, extract the threshold of 0.45, and retain the bounding boxes with an intersection-union ratio greater than 0.45.

[0100] (3) The detection box fusion algorithm is used to process all the detection boxes in the test sample to obtain the final remote sensing image target detection result.

[0101] To verify the performance advantage of this method in specific complex scenarios, two representative challenging subsets were selected from the test dataset for comparative experimental analysis. These two subsets correspond to "high-density small target scenarios" and "complex background interference scenarios," respectively, and are used to evaluate the robustness and detection accuracy of this method under conditions of densely packed small targets and strong background interference.

[0102] Experiment 1: Performance Evaluation in High-Density Small Target Scenarios

[0103] This experiment selected typical high-density small target scenarios, such as parking lots and ports, as the test subset to compare the present invention with existing mainstream target detection models. Evaluation metrics included mAP@0.5, false negative rate, and false positive rate. Experimental results are shown in Table 1.

[0104] Table 1. Performance Comparison of Various Models in High-Density Small Target Scenarios

[0105]

[0106] As shown in Table 1, in high-density small target scenarios, our method achieves significant improvements over the comparison model in all metrics, especially in reducing the false negative rate. These results demonstrate that the MRMF and CFF modules used in our method effectively enhance the feature representation and discrimination capabilities for densely arranged and interconnected small targets, thereby significantly reducing false negatives.

[0107] Experiment 2: Performance Evaluation under Complex Background Interference Scenarios

[0108] This experiment selected scenes with significant background interference as the test subset, such as vehicles in the shadow areas of urban buildings and targets under vegetation cover, to focus on evaluating the model's performance in suppressing background false detections. Evaluation metrics included mAP@0.5 and the average number of background false detections per image. Experimental results are shown in Table 2.

[0109] Table 2 Performance Comparison of Various Models in Complex Background Interference Scenarios

[0110]

[0111] As shown in Table 2, our proposed method maintains optimal detection accuracy even under complex background interference conditions, and the number of false detections due to background interference is significantly lower than that of the comparative model. This result verifies that the CFFtc-CR and MD modules introduced in our method can effectively integrate multi-scale contextual information, enhance the ability to distinguish between targets and backgrounds, and thus effectively suppress false detections caused by complex backgrounds.

[0112] This embodiment uses MRMF-CFF to detect vehicles in remote sensing images of complex urban road scenes. This method can accurately locate and identify high-density, multi-scale vehicle targets. It can effectively detect vehicles regardless of whether they are occluded, in shadow, or under different lighting conditions. The detection bounding boxes are compact and accurate, maintaining a high recall rate even for small vehicles in densely trafficked areas, effectively suppressing false positives and false negatives.

[0113] This embodiment uses MRMF-CFF to perform port detection tasks, detecting multiple types of targets in remote sensing images of large port scenes. This method successfully identifies multiple targets with vastly different shapes and sizes, including large ships, small tugboats, container trucks, port buildings, and densely stacked containers. For closely packed ships and containers facing different directions, this method can clearly distinguish each individual object and accurately label its location and category, demonstrating excellent ability to resolve complex scenes and adapt to multi-scale targets.

[0114] In summary, the present invention demonstrates significantly superior detection performance compared to existing technologies in both high-density small target scenarios and complex background interference scenarios, fully reflecting its applicability and technological advancement in real-world complex environments.

[0115] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A remote sensing image target detection method based on a multi-scale cross-stage network model, characterized in that, The method comprises the following steps: S1: obtaining a remote sensing image to be tested; S2: the remote sensing image to be tested is input into a backbone network after passing through an input layer, the backbone network extracts multi-scale and multi-level semantic features from the input image step by step, and different size feature maps are obtained through multi-scale down sampling; The backbone network optimizes the feature representation capability by integrating an MRMF module and a CFFtc-CR module, and enhances the feature interaction between channels; the MRMF module is a multi-branch structure, comprising a maximum pooling branch and a plurality of convolution branches with different convolution kernel sizes, multi-dimensional features are obtained by fusing and splicing the features of the multi-branch to enhance the diversity of feature extraction; the CFFtc-CR module realizes parallel extraction of local details and context information through parallel multi-scale convolution, and realizes feature refinement from local to global through global average pooling and a fully connected layer; The specific processing procedure of the backbone network is that the input image sequentially passes through two MRMF modules, three CFFtc modules, one MRMF module and six CFFtc modules to obtain a first feature map; the first feature map passes through an MRMF module and a CFFtc-CR module to obtain a second feature map; the second feature map passes through an MRMF module, a CFFtc-CR module and a PMPM module to obtain a third feature map; The backbone network outputs the first feature map, the second feature map and the third feature map; S3: inputting the feature maps of different sizes into a feature integration network to perform cross-stage fusion and lightweight optimization, and obtaining a plurality of fused feature maps; S4: inputting the plurality of fused feature maps into a prediction head to perform classification detection, and obtaining detection results of multiple scales. 2.The remote sensing image target detection method based on the multi-scale cross-stage network model of claim 1, wherein, The specific structure of the MRMF module is that the first branch adopts a 5*5 depth separable convolution followed by a 1*1 convolution, the second branch adopts a 7*7 depth separable convolution followed by a 1*1 convolution, and the third branch adopts a maximum pooling operation followed by a 1*1 convolution; the features output by the three branches are spliced in the channel to obtain multi-dimensional features; The three branches all adopt an XCON adaptive activation function. 3.The remote sensing image target detection method based on the multi-scale cross-stage network model of claim 1, wherein, The processing procedure of the CFFtc-CR module is that a convolution operation is performed on the input feature map, the feature map after the convolution is split into two parallel branches, one branch is processed through a CR layer and then spliced with the other branch, and a convolution operation is performed on the spliced feature map; The processing procedure of the CR layer is that the input feature map is processed through a 1*1 convolution, then input into a 3*3 convolution and a 3*3 atrous convolution respectively, the output feature maps are spliced, then processed through batch normalization and a PReLU activation function to obtain an intermediate feature map; the intermediate feature map is processed through global average pooling, two fully connected layers and then multiplied with the intermediate feature map element by element. 4.The remote sensing image target detection method based on the multi-scale cross-stage network model of claim 1, wherein, The feature integration network constructs a bidirectional feature pyramid structure through a CFF module, an upsampling layer and a splicing layer, wherein the CFF module optimizes the feature transmission efficiency through a cross-stage partial connection strategy, and the upsampling and splicing operations perform upsampling on the low-layer high-resolution feature map through bilinear interpolation and element-by-element channel splicing with the high-layer semantic feature map to realize a feature pyramid structure combining top-down and bottom-up. The specific process of the feature integration network is as follows: the third feature map is up-sampled, then spliced with the second feature map, and processed by the CFFtc-CR module to obtain a fourth feature map; the fourth feature map is up-sampled, then spliced with the first feature map, and processed by the CFFtc-CR module to obtain a first fusion feature map; the first fusion feature map is processed by the MRMF module, then spliced with the fourth feature map, and processed by the CFF module to obtain a second fusion feature map; the second fusion feature map is processed by the MRMF module, then spliced with the third feature map, and processed by the CFF module to obtain a third fusion feature map; The feature integration network outputs the first fusion feature map, the second fusion feature map and the third fusion feature map.

5. The remote sensing image target detection method based on the multi-scale cross-stage network model according to claim 4, wherein, The processing process of the CFF module is as follows: the input feature map is divided into two channel dimension branches along the channel dimension: the first channel dimension branch is compressed in channel by 1x1 convolution; the second channel dimension branch is sequentially subjected to 1x1 convolution, reparameterization convolution and 3x3 convolution to complete feature enhancement and transformation; the output feature maps of the two channel dimension branches are fused by channel splicing. 6.The remote sensing image target detection method based on the multi-scale cross-stage network model of claim 1, wherein, The prediction head includes a plurality of MD modules, which are used for classification detection of the fusion feature map; the MD modules sequentially apply attention modulation to the fusion feature map in the three dimensions of level, space and channel.

7. The remote sensing image target detection method based on the multi-scale cross-stage network model according to claim 6, wherein, The processing process of the MD module is as follows: The fusion feature map is globally averaged pooled, then subjected to 1x1 convolution, ReLU activation and Hard Sigmoid function to generate a weight map; the weight map and the fusion feature map are multiplied element by element to obtain a scale feature map; For the scale feature map, the key spatial positions are extracted by index operation, and a spatial attention map is generated by 3x3 convolution and Sigmoid function, and the spatial attention map and the offset are applied to the scale feature map to obtain a spatial feature map; For the spatial feature map, a channel-level weight is generated by global average pooling, a first fully connected layer, a ReLU activation function, a second fully connected layer and normalization; the channel-level weight and a fixed vector are added element by element, and then multiplied with the spatial feature map to obtain a detection result. 8.The remote sensing image target detection method based on the multi-scale cross-stage network model of claim 1, wherein, S5: Selecting an optimization target from the detection result, optimizing according to the class probability of the candidate box to obtain an optimized result image. 9.The remote sensing image target detection method based on the multi-scale cross-stage network model of claim 1, wherein, The multi-scale cross-stage network model includes an input layer, a backbone network, a feature integration network and a prediction head, and the multi-scale cross-stage network model is a pre-trained model, and the loss function thereof is the weighted sum of the IoU loss, the distance loss and the aspect ratio loss.

Images