Method and device for detecting weak and small ground target in complex scene

Abstract

The invention discloses a method and a device for detecting a weak and small ground target in a complex scene, and belongs to the technical field of computer vision. The method comprises the following steps: acquiring a satellite remote sensing image of a scene to be detected, and segmenting the satellite remote sensing image to obtain a plurality of small-size images; sequentially inputting all the small-size images into a pre-trained weak and small target detection model, and outputting to obtain a plurality of detection results; wherein the weak and small target detection model comprises a plurality of detection sub-models, the detection sub-models are obtained by performing specific training on an initial detection model by using an enhanced data set, and the enhanced data set is obtained by performing intelligent labeling and data enhancement on a cleaned multi-source scene data set; the initial detection model is obtained by performing architecture optimization on a target detection network; and filtering and fusing all the detection results to obtain a target detection result of the to-be-detected scene. According to the invention, the detection precision and speed of the weak and small ground target can be improved.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a method and apparatus for detecting small targets on the ground in complex scenes. Background Technology

[0002] In the current field of computer vision, the development of detection algorithms has formed two main branches: methods based on convolutional neural networks (CNNs) and methods based on Transformers. Within these two branches, CNN-based methods are constantly evolving and becoming more refined, primarily categorized into single-stage and two-stage detection methods. One representative of single-stage detection methods is the YOLO (You Only Look Once) series. The YOLO series algorithms have garnered significant attention for their excellent detection speed and high efficiency, capable of directly predicting the target's location and category in a single forward propagation, thus greatly improving the real-time performance of detection.

[0003] Among related technologies, the detection of small targets based on wide-swath remote sensing images is a highly challenging technical field. These targets, such as small vehicles or other small structures, occupy very few pixels due to their extremely small size relative to the vast observation area. At the same time, the complex and constantly changing background also greatly reduces the detection accuracy and processing speed of the detection algorithm.

[0004] Therefore, there is an urgent need for a method and device for detecting small ground targets in complex scenarios to solve the above-mentioned technical problems. Summary of the Invention

[0005] This invention provides a method and apparatus for detecting weak ground targets in complex scenes, which can improve the accuracy and speed of detecting weak ground targets. The technical solution is as follows: On the one hand, a method for detecting weak ground targets in complex scenes is provided, the method comprising: Acquire satellite remote sensing images of the scene to be detected, and segment the satellite remote sensing images to obtain multiple small-sized images; All the small images are sequentially input into the pre-trained weak target detection model, and multiple detection results are output. The weak target detection model includes multiple detection sub-models. The detection sub-models are obtained by specifically training the initial detection model using an augmented dataset. The augmented dataset is obtained by intelligently labeling and data augmenting a cleaned multi-source scene dataset. The initial detection model is obtained by optimizing the architecture of the target detection network. All detection results are filtered and fused to obtain the target detection results for the scene to be detected.

[0006] On the other hand, a detection device for weak ground targets in complex scenes is provided, the device comprising: The acquisition module is used to acquire satellite remote sensing images of the scene to be detected, and to segment the satellite remote sensing images into multiple small-sized images; The detection module is used to sequentially input all the small-sized images into a pre-trained weak target detection model and output multiple detection results. The weak target detection model includes multiple detection sub-models, which are obtained by specifically training an initial detection model using an augmented dataset. The augmented dataset is obtained by intelligently labeling and data augmenting a cleaned multi-source scene dataset. The initial detection model is obtained by optimizing the architecture of the target detection network. The fusion module is used to filter and fuse all detection results to obtain the target detection results for the scene to be detected.

[0007] On the other hand, a computer device is provided, the computer device including a memory and a processor, the memory for storing computer programs, and the processor for executing the computer programs stored in the memory to implement the steps of the method for detecting weak ground targets in complex scenarios described above.

[0008] On the other hand, a computer-readable storage medium is provided, wherein a computer program is stored therein, and when the computer program is executed by a processor, it implements the steps of the method for detecting weak ground targets in complex scenarios described above.

[0009] On the other hand, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the method for detecting weak ground targets in complex scenarios described above.

[0010] The technical solution provided by this invention can bring at least the following beneficial effects: by annotating and augmenting multi-source scene data, an augmented dataset covering various application scenarios and needs is obtained; then, by optimizing the architecture of the target detection network and training the optimized detection network using the augmented dataset, a weak target detection model is obtained; finally, the satellite remote sensing image of the scene to be detected is segmented and input into the weak target detection model, and the output results are fused to obtain the detection results of weak targets in complex scenes. This method solves the problems of weak targets, low contrast, and background interference in satellite imagery by intelligently annotating images and by optimizing the target detection network and fusing multiple models, thereby improving the detection accuracy of weak targets in complex scenes. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present 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 the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart of a method for detecting weak ground targets in complex scenarios according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a data processing and annotation process provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the spdconv operator design provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of a high-resolution detection head design provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the reasoning result provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the model evaluation results provided in an embodiment of the present invention; Figure 7 This is a structural diagram of a detection device for weak ground targets in complex scenarios provided by an embodiment of the present invention; Figure 8 This is a hardware architecture diagram of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0014] As mentioned earlier, in the prior art, due to the small size of the target and the complex and varied background conditions, it is difficult to distinguish between the target and the background, resulting in low detection accuracy for small targets.

[0015] Based on this, the present invention obtains a multi-source scene dataset and annotates and enhances it to obtain an augmented dataset for training the detection model. At the same time, it performs specific optimizations on the object detection network to detect different types of scenes. Finally, by screening the training results, it obtains several single models with the highest detection accuracy to form a weak object detection model.

[0016] The following describes the specific implementation of the above concept.

[0017] Please refer to Figure 1 This invention provides a method for detecting weak ground targets in complex scenes, the method comprising: Step 100: Acquire satellite remote sensing images of the scene to be detected, and segment the satellite remote sensing images to obtain multiple small-sized images; Step 102: Input all the small-sized images into the pre-trained weak target detection model in sequence to output multiple detection results; wherein, the weak target detection model includes multiple detection sub-models, the detection sub-models are obtained by specifically training the initial detection model using an augmented dataset, the augmented dataset is obtained by intelligent annotation and data augmentation of a cleaned multi-source scene dataset, and the initial detection model is obtained by optimizing the architecture of the target detection network; Step 104: Filter and fuse all detection results to obtain the target detection results for the scene to be detected.

[0018] In this embodiment of the invention, by annotating and augmenting multi-source scene data, an augmented dataset covering various application scenarios and needs is obtained. Then, the target detection network architecture is optimized, and the optimized detection network is trained using the augmented dataset to obtain a weak target detection model. Finally, the satellite remote sensing image of the scene to be detected is segmented and input into the weak target detection model, and the output results are fused to obtain the detection results of weak targets in complex scenes. This method solves the problems of weak targets, low contrast, and background interference in satellite imagery by intelligently annotating images and optimizing the target detection network and fusing multiple models, thus improving the detection accuracy of weak targets in complex scenes.

[0019] The following description Figure 1 The execution method for each step is shown.

[0020] First, for step 100, a satellite remote sensing image of the scene to be detected is acquired, and the satellite remote sensing image is segmented into multiple small-sized images.

[0021] When processing satellite remote sensing image data, the resolution is typically high, and the relative size of objects within the image is small. Directly inputting the entire image into the network for object detection would result in enormous memory consumption, making this strategy impractical in real-world applications. Therefore, to effectively process these high-resolution images, a block-based inference approach is usually adopted. The image is divided into several small blocks, each of moderate size suitable for network processing. Then, object detection is performed on each small block separately. Finally, the detection results from each small block are fused to obtain the object detection result for the entire image.

[0022] Then, for step 102, all the small-sized images are sequentially input into the pre-trained weak target detection model, and multiple detection results are output.

[0023] In this embodiment of the invention, the weak target detection model includes multiple detection sub-models. The detection sub-models are obtained by specifically training an initial detection model using an augmented dataset. The augmented dataset is obtained by intelligently labeling and data augmenting a cleaned multi-source scene dataset. The initial detection model is obtained by optimizing the architecture of the target detection network.

[0024] Specifically, the augmented dataset is obtained through the following process: the multi-source scene dataset is input into a pre-trained intelligent recognition model, which outputs key features of all weak targets in the image; weak targets in the multi-source scene dataset are labeled according to the key features, and the labeling results are reviewed and verified to obtain a target labeled dataset that meets the accuracy requirements; four images are randomly selected from the target labeled dataset and stitched together to obtain a composite image with complex background information; pixel filling is performed on composite images smaller than a preset fixed size to obtain an augmented dataset with the same image size.

[0025] For example, to ensure the accuracy of the detection model, it is necessary to select a high-quality training dataset for training.

[0026] For satellite remote sensing imagery data, especially for the detection of complex and small ground targets such as vehicles, preliminary data preparation work is first carried out by extensively collecting publicly available datasets. These datasets include, but are not limited to, various data sources such as DOTA2.0, DIOR, HRSC2016, SAMRS, and SODA-A, covering a wide range of application scenarios and needs. After collecting this data, meticulous cleaning and processing are performed to obtain multi-source scene datasets, ensuring data quality and usability.

[0027] Then, advanced VLM (Virtual Modeling) technology was used to train the basic model. The VLM model was mainly used for initial data labeling and cleaning. Through training the basic model, target objects could be effectively identified and labeled, resulting in a high-quality target labeling dataset. The dataset was carefully organized, ultimately yielding approximately 200,000 training samples. These samples covered various regions globally, including scene data from different time periods and under different weather conditions.

[0028] It's worth noting that a highly efficient AI-assisted processing workflow was implemented during the data annotation process. First, a basic model was trained using publicly available datasets. After thorough training, this basic model acquired preliminary recognition and analysis capabilities. Next, the trained model was applied to the processing of military target data. Through the model's inference capabilities, this data was automatically annotated, significantly reducing the workload of manual annotation. Specifically, the basic model could identify the key features of weak targets and label them in the dataset. Simultaneously, to ensure the accuracy and high quality of the annotations, a manual review and correction mechanism was introduced. Through meticulous manual review, errors or omissions that might occur during the model's inference process could be corrected, further improving the accuracy of the annotations.

[0029] By combining artificial intelligence with human review, the annotation workload was significantly reduced by over 60%. This means that more data can be processed in the same amount of time, or the same amount of data processing tasks can be completed in a shorter time. The data processing and annotation workflow is as follows: Figure 2 As shown.

[0030] Furthermore, in order to improve the quality of the dataset, data augmentation is required after obtaining the target labeled dataset.

[0031] This embodiment employs the Mosaic augmentation method from dynamic data augmentation techniques, a widely used strategy in image training. Specifically, this method is activated and applied during the first 90% of the training cycle. Its core idea is to randomly select four different images and cleverly stitch them together to form a new composite image. This stitching method not only enriches the background information of the image but also enhances the model's understanding of complex scenes, thereby improving the model's robustness and generalization performance in practical applications.

[0032] Furthermore, to adapt to input images of different sizes, an adaptive padding mechanism is introduced. When the input image is not square, for example, if the width and height are inconsistent, the system automatically pads the edges of the image with gray or black pixels to achieve a preset fixed size, such as 640×640 pixels. This padding operation ensures that the proportions of the original image are not changed, while allowing the image to adapt to the model's input requirements, thus ensuring the smooth progress of the training process.

[0033] For multi-scale enhancement, a simple yet effective method was employed. During training, the system randomly scales down and up the input image. This allows the model to learn to recognize targets at different scales, further improving its adaptability to changes in target size. This multi-scale training strategy helps the model maintain high detection accuracy when facing targets of various sizes in real-world applications.

[0034] In addition to the enhancement methods mentioned above, other common image enhancement techniques were employed to further improve the model's performance. For example, Flip enhancement increases the model's understanding of the left-right symmetry of the target by randomly flipping the image horizontally; noise enhancement improves the model's robustness to noise interference by introducing random noise into the image; blur enhancement trains the model to accurately detect targets even with poor image quality by blurring the image; and rotation enhancement enhances the model's understanding of the rotation invariance of the target by randomly rotating the image.

[0035] The datasets collected using the above methods are not only substantial in quantity but also rigorously screened and validated in quality, ensuring the accuracy and reliability of subsequent model training. These high-quality training samples can further improve the accuracy of object detection and recognition.

[0036] In this embodiment of the invention, the architecture optimization of the target detection network includes image processing optimization, convolution calculation optimization, and detection resolution optimization. Specifically: image processing optimization of the target detection network is performed based on the rotating target box technique to reduce orientation detection errors; convolution calculation optimization of the network backbone of the target detection network is performed based on the preset SPDCONV operator to preserve fine-grained features of weak targets when downsampling the image to be identified; and detection resolution optimization of the target detection network is performed based on the preset high-resolution detection module to obtain detailed information of weak targets.

[0037] Specifically, for satellite data processing, a series of improvements were made based on the YOLOv11 model, including rotating the target bounding box, using the spdconv operator, and employing a high-resolution detection head, to improve the accuracy and efficiency of detection.

[0038] First, when processing satellite imagery, Rotated Object Bounding (OBB) technology is employed. This technology better fits the actual shape of objects. For example, in satellite remote sensing images, when detecting long, narrow objects such as ships, OBB can accurately reflect the orientation of these objects, such as the direction of the bow. This significantly reduces detection errors caused by orientation deviations, thereby improving detection accuracy.

[0039] Secondly, in the backbone of the model, the spdconv operator is used to replace the traditional pooling layer. Traditional pooling layers often lose some fine-grained information when using convolutional strides, resulting in inefficient feature representation. The spdconv operator effectively solves this problem, improving feature extraction efficiency while preserving feature information. The optimization includes: splitting the original feature map input to the network backbone into multiple feature map sub-blocks containing local information of the original spatial dimensions at a fixed ratio; concatenating these feature map sub-blocks along the channel dimension to obtain a concatenated feature map; and performing feature fusion on the concatenated feature map to obtain a downsampled high-resolution feature map.

[0040] Specifically, the design of the spdconv operator is as follows: Figure 3 As shown.

[0041] Traditional convolutional neural network (CNN) pooling layers are prone to losing fine-grained features (especially key textures of small objects) during downsampling. However, the spdconv operator improves computational efficiency while preserving feature information by combining space-to-depth transformation with depthwise convolution.

[0042] Its working process consists of three stages. The first stage is spatial decomposition. SPDCONV first decomposes the input feature map (e.g., of size H×W×C) into multiple sub-blocks. Unlike the direct compression of traditional pooling, it adopts a space-to-depth transformation: recombining information from the spatial dimension (H×W) into the channel dimension (C). For example, a 4×4 feature map can be decomposed into four 2×2 sub-blocks, each containing the original local information, and then concatenated along the channel dimension to form a new concatenated feature map (e.g., the size becomes 2×2×4C). This avoids information loss and preserves pixel-level details of weak targets.

[0043] Next comes the feature transformation stage, where the concatenated feature maps are fused using lightweight convolutional operations (such as depthwise separable convolution). This step focuses on extracting cross-channel correlations without reducing the spatial size to ultimately output a high-resolution feature map, but achieves "implicit downsampling" by expanding the channel dimension. For example, a 640×640 input feature map processed by SPDCONV may output a 320×320 feature map, but the number of channels is doubled (e.g., C becomes 2C), ensuring that the information density remains unchanged.

[0044] Finally, in the head part of the model, only three detection heads are typically used. To further improve the detection capability of small targets, a high-resolution detection head is added to the existing structure. This additional detection head can better capture the detailed information of small targets, thereby improving the detection accuracy. The optimization includes: directly obtaining the high-resolution feature map from the shallow network of the network backbone; fusing deep semantic features with the high-resolution feature map based on adaptive upsampling to enhance the semantic information of the feature map; and using the high-resolution detection module to detect the fused result and identify weak small targets that meet the preset requirements.

[0045] Specifically, the design of a high-resolution detection head is as follows: Figure 4 As shown.

[0046] This embodiment adds a high-resolution branch, directly extracting high-resolution feature maps (such as the original) from the shallow layers of the backbone (e.g., stage 2 or 3). Figure 1 The new branch uses a 1 / 4 scale instead of the original 1 / 32. For example, if the input image is 640×640, the minimum feature map of a traditional detection head is 20×20, while the new branch can retain a resolution of 160×160. Then, adaptive upsampling is used to fuse deep semantic features (e.g., 1 / 16 scale) with the new high-resolution feature map to enhance semantic information. Simultaneously, lightweight convolutional layers (e.g., 1×1 Conv) are used to adjust the number of channels, avoiding computational redundancy. Finally, an independent detection layer (classification + regression) is deployed at the end of the new branch to focus on recognizing small-scale targets.

[0047] Through the aforementioned improvements, the designed algorithm model can more accurately detect various targets when processing satellite remote sensing image data, achieving better detection results for both elongated objects and small targets. These improvements not only enhance the model's performance but also make it more robust and reliable in practical applications.

[0048] After processing the dataset and the initial detection model, the initial detection model can be trained using the dataset. The training process includes: optimizing the target detection network based on the characteristics of weak targets and scenes to obtain multiple initial detection sub-models for detecting weak targets under different conditions; dividing all weak target images in the augmented dataset into blocks, inputting the processed images into the corresponding initial detection models, and fusing the output results to obtain the initial detection result for each image; filtering the initial detection results according to preset prior conditions, and determining the initial detection sub-models corresponding to the initial detection results that meet the prior conditions as the detection sub-models; and establishing a model set based on the detection sub-models to obtain a weak target detection model for complex scenes.

[0049] Specifically, in complex scenarios, the detection accuracy of a single model is often affected by background interference, leading to less than ideal detection results. Furthermore, there are certain correlations between different models and data source scenarios. Therefore, to further improve the accuracy of object detection, a multi-model fusion strategy was adopted. This involves training different models separately for different object types, optimizing each model for specific targets and scenarios, resulting in multiple detection sub-models. By comprehensively considering the advantages of each model, a more accurate and comprehensive detection result is ultimately obtained.

[0050] Next, the images in the dataset are processed using the same block-based method as in step 100. This method can not only effectively reduce memory consumption, but also improve inference speed through parallel inference scheme, thereby improving overall processing efficiency while ensuring detection accuracy.

[0051] Furthermore, in the target detection process, a large number of detection results are usually obtained, but not all of them are valid. To improve the accuracy and efficiency of detection, prior knowledge can be used to filter the detection results. This method can not only improve the efficiency of detection but also effectively reduce false positives and false negatives, thereby further improving the overall detection performance.

[0052] This embodiment combines the resolution of satellite remote sensing image data and sets some prior conditions, such as the maximum and minimum values ​​of target size, aspect ratio, length, width, and area. Then, using these prior conditions, target detection results that clearly do not conform to reality are filtered out, thereby reducing the burden of subsequent processing and improving the reliability of the detection results. A model set is established by constructing detection sub-models that meet the detection prior conditions, resulting in a weak target detection model for complex scenes.

[0053] Finally, these test results that meet the prior conditions are merged to obtain more accurate and comprehensive test results.

[0054] In this embodiment, after the trained model is obtained, a test dataset can be used to evaluate its performance.

[0055] By plotting accuracy and recall curves, you can visually observe the model's performance at different thresholds. Examples of inference results are shown below. Figure 5 As shown, an example of the model evaluation results is as follows. Figure 6 As shown.

[0056] As shown in the figure, the model's accuracy and recall both exceeded 80%, indicating that the model performed well in detecting positive samples and making correct predictions. Specifically, accuracy reflects the proportion of correctly predicted samples out of the total predicted samples, while recall measures the proportion of correctly detected positive samples out of the total number of actual positive samples. Both exceeding 80% demonstrate that the model achieved a good balance between accuracy and coverage.

[0057] Step 104 is used to filter and fuse all detection results to obtain the target detection results for the scene to be detected.

[0058] Similar to the training process, the outputs of the weak target detection model are fused to obtain more accurate and comprehensive detection results.

[0059] Please refer to Figure 7 This invention provides a device for detecting weak ground targets in complex scenarios. The device includes: The acquisition module 700 is used to acquire satellite remote sensing images of the scene to be detected, and to segment the satellite remote sensing images into multiple small-sized images; The detection module 702 is used to sequentially input all the small-sized images into a pre-trained weak target detection model and output multiple detection results; wherein, the weak target detection model includes multiple detection sub-models, the detection sub-models are obtained by specifically training an initial detection model using an augmented dataset, the augmented dataset is obtained by intelligent annotation and data augmentation of a cleaned multi-source scene dataset, and the initial detection model is obtained by optimizing the architecture of the target detection network; The fusion module 704 is used to filter and fuse all detection results to obtain the target detection results of the scene to be detected.

[0060] In this embodiment of the invention, an enhanced dataset is obtained by intelligently labeling and data augmenting a cleaned multi-source scene dataset. This includes: inputting the multi-source scene dataset into a pre-trained intelligent recognition model to output key features of all weak targets in the image; labeling the weak targets in the multi-source scene dataset based on the key features, and reviewing the labeling results to obtain a target-labeled dataset that meets accuracy requirements; randomly selecting four images from the target-labeled dataset and stitching them together to obtain a composite image with complex background information; and filling pixels in all composite images smaller than a preset fixed size to obtain an enhanced dataset with the same image size.

[0061] In this embodiment of the invention, the architecture optimization of the target detection network includes image processing optimization, convolution calculation optimization, and detection resolution optimization. Specifically: image processing optimization of the target detection network is performed based on the rotating target box technique to reduce orientation detection errors; convolution calculation optimization of the network backbone of the target detection network is performed based on the preset SPDCONV operator to preserve fine-grained features of weak targets when downsampling the image to be identified; and detection resolution optimization of the target detection network is performed based on the preset high-resolution detection module to obtain detailed information of weak targets.

[0062] In this embodiment of the invention, the optimization of the network backbone of the target detection network based on the preset SPDCONV operator includes: splitting the original feature map of the input network backbone into multiple feature map sub-blocks containing local information of the original spatial dimension by a fixed ratio; concatenating the feature map sub-blocks along the channel dimension to obtain a concatenated feature map; and performing feature fusion on the concatenated feature map to obtain a downsampled high-resolution feature map.

[0063] In this embodiment of the invention, the target detection network is optimized based on a preset high-resolution detection module, including: directly obtaining the high-resolution feature map from the shallow network of the network backbone; fusing deep semantic features with the high-resolution feature map based on adaptive upsampling to enhance the semantic information of the feature map; and using the high-resolution detection module to detect the fusion result and identify weak targets that meet the preset requirements.

[0064] In this embodiment of the invention, the weak target detection model is trained as follows: the target detection network is specifically optimized according to the characteristics of the weak target and the scene to obtain multiple initial detection sub-models for detecting weak targets under different conditions; all weak target images in the augmented dataset are processed into blocks, the processed images are input into the initial detection models corresponding to the conditions, and the output results are fused to obtain the initial detection result corresponding to each image; the initial detection results are filtered according to preset prior conditions, and the initial detection sub-models corresponding to the initial detection results that meet the prior conditions are determined as the detection sub-models; a model set is established based on the detection sub-models to obtain a weak target detection model for complex scenes.

[0065] It should be noted that the detection device for weak ground targets in complex scenes provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the detection device for weak ground targets in complex scenes and the method embodiment for detecting weak ground targets in complex scenes provided in the above embodiments belong to the same concept, and the specific implementation process is detailed in the method embodiment, which will not be repeated here.

[0066] Embodiments of this application also provide a computer device, please refer to... Figure 8 The computer device includes a processor and a memory, the memory storing at least one instruction, at least one program, code set or instruction set, the at least one instruction, at least one program, code set or instruction set being loaded and executed by the processor to implement the method for detecting weak ground targets in complex scenarios provided in the above-described method embodiments.

[0067] The embodiments of this application also provide a computer-readable storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, at least one program, code set, or instruction set is loaded and executed by a processor to implement the method for detecting weak ground targets in complex scenarios provided in the above-described method embodiments.

[0068] Embodiments of this application also provide a computer program product, which includes a computer program. A processor of a computer device reads the computer program from a computer-readable storage medium and executes the computer program, causing the computer device to perform the detection method for weak ground targets in complex scenarios as described in any of the above embodiments.

[0069] For ease of description, the above systems or devices are described separately as various modules or units based on their functions. Of course, in implementing this application, the functions of each unit can be implemented in one or more software and / or hardware components.

[0070] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of this application.

[0071] Finally, it should be noted that in this document, relational terms such as first, second, third, and fourth are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0072] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for detecting weak ground targets in complex scenes, characterized in that, The method includes: Acquire satellite remote sensing images of the scene to be detected, and segment the satellite remote sensing images to obtain multiple small-sized images; All the small images are sequentially input into the pre-trained weak target detection model, and multiple detection results are output. The weak target detection model includes multiple detection sub-models. The detection sub-models are obtained by specifically training the initial detection model using an augmented dataset. The augmented dataset is obtained by intelligently labeling and data augmenting a cleaned multi-source scene dataset. The initial detection model is obtained by optimizing the architecture of the target detection network. All detection results are filtered and fused to obtain the target detection results for the scene to be detected.

2. The method as described in claim 1, characterized in that, The cleaned multi-source scene dataset is intelligently labeled and augmented to obtain an augmented dataset, including: The multi-source scene dataset is input into the pre-trained intelligent recognition model, and the key features of all weak targets in the image are output. Based on the key features, weak targets in the multi-source scene dataset are labeled, and the labeling results are reviewed and verified to obtain a target labeled dataset that meets the accuracy requirements. Four images are randomly selected from the target annotation dataset and stitched together to obtain a composite image with complex background information; Pixel padding is performed on all composite images smaller than a preset fixed size to obtain an augmented dataset with the same image size.

3. The method as described in claim 1, characterized in that, Optimizing the architecture of object detection networks includes image processing optimization, convolution computation optimization, and detection resolution optimization, among which: Image processing optimization of the target detection network is performed based on the rotating target box technique to reduce orientation detection error; The target detection network backbone is optimized by performing convolution calculation based on the preset SPDCONV operator in order to preserve the fine-grained features of weak targets when downsampling the image to be identified. The target detection network is optimized based on a pre-set high-resolution detection module to obtain detailed information about small targets.

4. The method as described in claim 3, characterized in that, The optimization of the target detection network backbone by performing convolution calculations based on the preset SPDCONV operator includes: The original feature map input to the network backbone is split into multiple feature map sub-blocks containing local information of the original spatial dimensions by a fixed ratio. The feature map sub-blocks are stitched together along the channel dimension to obtain a stitched feature map; Feature fusion is performed on the stitched feature maps to obtain a downsampled high-resolution feature map.

5. The method as described in claim 4, characterized in that, The target detection network is optimized based on a preset high-resolution detection module, including: The high-resolution feature map is obtained directly from the shallow network of the network backbone; Based on adaptive upsampling, deep semantic features are fused with the high-resolution feature map to enhance the semantic information of the feature map; The high-resolution detection module is used to detect the fusion results and identify weak targets that meet the preset requirements.

6. The method as described in claim 1, characterized in that, The weak target detection model was trained in the following way: The target detection network is specifically optimized based on the characteristics of weak targets and scenes to obtain multiple initial detection sub-models for detecting weak targets under different conditions; All weak target images in the augmentation dataset are divided into blocks. The processed images are input into the initial detection model with corresponding conditions, and the output results are fused to obtain the initial detection result for each image. The initial detection results are filtered according to preset prior conditions, and the initial detection sub-model corresponding to the initial detection results that meet the prior conditions is determined as the detection sub-model. A model set is established based on the detection sub-model to obtain a weak target detection model for complex scenarios.

7. A detection device for weak ground targets in complex scenes, characterized in that, The device includes: The acquisition module is used to acquire satellite remote sensing images of the scene to be detected, and to segment the satellite remote sensing images into multiple small-sized images; The detection module is used to sequentially input all the small-sized images into a pre-trained weak target detection model and output multiple detection results. The weak target detection model includes multiple detection sub-models, which are obtained by specifically training an initial detection model using an augmented dataset. The augmented dataset is obtained by intelligently labeling and data augmenting a cleaned multi-source scene dataset. The initial detection model is obtained by optimizing the architecture of the target detection network. The fusion module is used to filter and fuse all detection results to obtain the target detection results for the scene to be detected.

8. A computer device, characterized in that, The computer device includes a memory and a processor. The memory is used to store computer programs, and the processor is used to execute the computer programs stored in the memory to implement the steps of the method according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the method described in any one of claims 1-6.

10. A computer program product, characterized in that, Includes a computer program, which, when executed by a processor, implements the steps of the method according to any one of claims 1-6.

Images