Real-time small target detection method with direction sensitivity
By employing orthogonal direction enhancement and large convolutional kernel gating in remote sensing images, the features of small targets are enhanced and background interference is suppressed, thus solving the problem of inaccurate detection of small targets in remote sensing images and achieving accurate detection in complex backgrounds.
Patent Information
- Application Number
- CN202511413893.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-29
- Publication Date
- 2026-02-17
Smart Images
Figure CN121544862A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a real-time small target detection method with direction sensitivity, belonging to the field of remote sensing image processing. Background Technology
[0002] Remote sensing target detection aims to locate and classify targets of interest in remote sensing images. It has important applications in many fields, such as environmental monitoring, disaster control, traffic management, and urban planning.
[0003] Currently, deep learning technology has become the mainstream method in the field of remote sensing target detection due to its powerful feature extraction capabilities. Based on whether candidate regions are generated, deep learning-based target detection can be divided into two-stage methods and one-stage methods. Compared to two-stage methods, the good balance between speed and detection accuracy of one-stage methods makes them more suitable for real-time detection. Among one-stage methods, YOLOv10 has become a leading example in recent years due to its excellent performance.
[0004] Small object detection is a significant challenge in object detection tasks. Small objects in remote sensing images refer to objects with a pixel area smaller than 32×32. These small objects often exhibit a large aspect ratio, leading to significant variability in their features across different directions. However, current state-of-the-art methods employ a structure composed of square convolutional kernels, which treat feature extraction indiscriminately across all directions. This results in excessive background interference when the network attempts to capture small objects with highly directional feature variations, ultimately preventing the network from accurately locating and classifying these small objects.
[0005] To address the issue of inaccurate detection caused by redundant backgrounds introduced by square convolutional networks, resulting in significant feature variability of small targets with large aspect ratios in different directions, this patent proposes a direction-sensitive real-time small target detection method based on YOLOv10. This method enhances small target features from multiple directions and suppresses background interference, thereby accurately perceiving small targets with significant feature variability in different directions even in complex backgrounds. The model comprises two key parts: an orthogonal direction enhancement method and a background suppression method based on large convolutional kernel gating. The orthogonal direction enhancement method designs multiple parallel branches containing multi-directional strip convolutions to extract the variability features of small targets in different directions. Then, features along the main direction are integrated through strip convolutions orthogonal to the branch to enhance feature robustness. Furthermore, dilated convolutions are introduced at the end of each branch to establish a correlation between the small target and the background. Finally, the features of all branches are fused to achieve synergistic enhancement of the varied features in different directions. The background suppression method based on large convolutional kernel gating includes two sub-methods: context-guided background suppression and target density-gated feature preservation. Both methods use the contextual semantic information extracted by the large convolutional kernel as weights to guide the channel attention mechanism and self-attention mechanism respectively to suppress the background while preserving the features of small targets during downsampling. Summary of the Invention
[0006] The technical problem solved by this invention is to overcome the shortcomings of existing technologies, namely, that current target detection algorithms based on square convolution introduce redundant background interference when detecting small targets, resulting in an inability to accurately perceive the features of small targets. This invention proposes a direction-sensitive real-time small target detection method (such as...). Figure 1 (As shown). In this method, an orthogonal direction enhancement method is designed to perceive and synergistically enhance the variability features of small targets in different directions. A background suppression method based on large convolutional kernel gating is designed to suppress background interference while preserving small target features during downsampling. The technical solution of this invention is: A direction-sensitive real-time small target detection method includes: The first step involves using orthogonal direction enhancement methods to improve the network's accurate perception of features in different directions for small targets. First, bar convolutions in different directions are designed in parallel branches to perceive the variability of small targets in different directions. Then, bar convolutions in directions orthogonal to these directions are cascaded to integrate the features in that direction, making the features of small targets more robust (e.g., ...). Figure 2 (As shown). Assuming the network input is feature F, and the module has 5 parallel branches, each with a different bar convolution direction, the process can be represented as: ; ; ; ; ; in, - The horizontal bar convolution represents the features obtained after each branch. The vertical bar convolution is represented as The strip convolution along the main diagonal (from top left to bottom right) is represented as: The strip convolution along the secondary diagonal direction (from the upper right to the lower left) is represented as: It is worth noting that each direction of the strip convolution is obtained by rotating and transforming a 3×1 convolution. This represents a 3×3 convolution. After the above steps, each branch is designed with cascaded orthogonal convolutions. Most of these convolutions are strip-shaped, which can fit the shape of small targets with large aspect ratios, so that the network will not mix in redundant background interference when extracting features of small targets, unlike square convolutional networks. Furthermore, these directional convolutions can accurately perceive the differences in features of small targets in different directions. Finally, after further integrating these features with these square orthogonal convolutions, the small target features extracted by the network are more robust.
[0007] The second step is to establish a relationship between the small target and the background and to fuse and enhance the extracted features of the small target from all directions.
[0008] This process can be represented as: ; Here, concat indicates a feature concatenation and fusion operation. This represents dilated convolution. When using dilated convolution, the receptive field of the network expands accordingly. The network is no longer limited to grasping target features, but combines target features with their surrounding environment to construct a relationship between the target and the background. After concatenating and fusing features from all directions, the final output is the enhanced small target feature obtained from the fusion of features from all directions. At this point, the features of each direction of the small target no longer exist in isolation but complement and synergistically enhance features from other directions. This is beneficial for the network to accurately locate and classify small targets.
[0009] The third step is to extract the contextual information of the target over a large range using large convolutional kernels, thereby establishing the relationship between the target and its surrounding environment.
[0010] In small object detection tasks, establishing a correlation between the target and the background is crucial because small objects occupy a very small pixel area, leading to blurred features and hindering accurate classification. However, once a correlation is established between the target and its surrounding background, contextual information can be used to assist the network's classification. For example, when the network needs to distinguish between a car and a boat with very similar shapes, the small object's features alone are insufficient. However, when the contextual information of the small object is water, the network can classify it as a boat instead of a car. Therefore, this design utilizes large convolutional kernels to extract contextual information from small objects, establishing a correlation between the small object and the background over a larger scope. This information is then used as network weights to prepare for subsequent steps. It is worth noting that directly using N×N large convolutional kernels would significantly increase the number of network parameters and computational cost. To ensure real-time performance, this design employs an N×1 convolutional layer cascaded with 1×N convolutions, effectively creating an N×N convolution.
[0011] The fourth step involves fusing the contextual semantic information extracted by the large convolutional kernel as guiding information into two sub-methods: context-guided background suppression and target density gating feature preservation. This achieves suppression of background interference while preserving the small target features mentioned earlier during downsampling.
[0012] Specifically, the two sub-methods function as parallel branches. In the context-guided background suppression branch, contextual semantic information is used as weights for the channel attention mechanism. This guides the channel attention to focus on channels that encode significant background interference and assigns low weights to these channels during subsequent channel selection, thereby suppressing background interference (e.g., ...). Figure 3 (As shown). In this process, contextual information extracted using two large convolutional kernels is utilized. Between these two extractions, to ensure the network's real-time performance, this design employs the classic YOLOv8 design, Bottleneck C2f, which maintains the model's representational power while reducing computational cost and parameter count. Assume the model's input features at this point are... This process can then be represented as: ; in, This represents the features obtained after context-guided background suppression branch processing. GMP represents global max pooling, and GMP represents global average pooling. For the feature preservation branch of target density gating, contextual semantic information is used as the weight of V in the self-attention mechanism. This guides the self-attention to focus on spatial features that encode a large number of small targets and assigns high weights to these spatial features in subsequent processes, thereby achieving the goal of preserving small target features during downsampling (e.g., ...). Figure 4 (As shown). This process can be represented as: ; in, This represents the features obtained after processing the feature branches of target density gating. Finally, fusing the features obtained from the two branches achieves the goal of preserving small target features while suppressing the background.
[0013] Beneficial effects 1. This invention addresses the real-time small target detection task in remote sensing by proposing a direction-sensitive real-time small target detection method. This method can enhance small target features from multiple directions and suppress background interference, achieving foreground perception in complex backgrounds. It improves upon existing advanced small target detection methods that suffer from missed detections and classification errors in complex backgrounds, representing a significant step forward in realizing a real-time small target detection model for remote sensing and providing an effective technical means for intelligent interpretation of remote sensing images.
[0014] 2. This invention proposes an orthogonal direction enhancement method to perceive and collaboratively enhance the variability of small targets in different directions from multiple directions. This avoids the problem of current advanced methods using square convolution kernels that treat features in all directions indiscriminately when extracting target features, thus enhancing the network's ability to collaboratively perceive the multi-directional features of small targets.
[0015] 3. This invention designs a background suppression method based on large convolutional kernel gating, which suppresses background interference while preserving small target features during downsampling. This improves upon the problem of current state-of-the-art methods introducing significant background interference when extracting small target features, achieving the goal of preserving target features while suppressing background interference. Attached Figure Description
[0016] Figure 1 , one A diagram illustrating the overall framework of a direction-sensitive real-time small target detection method; Figure 2 Schematic diagram of orthogonal direction enhancement method; Figure 3 A schematic diagram of a context-guided background suppression method; Figure 4 A schematic diagram of the feature preservation method for target density gating; Figure 5 Compared with other advanced methods, this invention can more accurately perceive small targets in different directions; Figure 6 Compared with other advanced methods, this invention can perceive the foreground more accurately. Detailed Implementation
[0017] The present invention will be further described below with reference to the embodiments.
[0018] Example A direction-sensitive real-time small target detection method, taking real-time remote sensing small target detection as an example, is illustrated in the following diagram: Figure 1 As shown, the steps of this method include: The first step involves using orthogonal direction enhancement methods to improve the network's accurate perception of the directional features of small targets. For example... Figure 2 The diagram first designs strip convolutions in different directions within parallel branches to perceive the variability of small targets in different directions. Then, it cascades strip convolutions in directions orthogonal to these directions to integrate the features in that direction, making the features of small targets more robust. Assuming the network input is feature F, and the module has 5 parallel branches, each with a different strip convolution direction, this process can be represented as: ; ; ; ; ; in, - The horizontal bar convolution represents the features obtained after each branch. The vertical bar convolution is represented as The strip convolution along the main diagonal (from top left to bottom right) is represented as: The strip convolution along the secondary diagonal direction (from the upper right to the lower left) is represented as: It is worth noting that each direction of the strip convolution is obtained by rotating and transforming a 3×1 convolution. This represents a 3×3 convolution. After the above steps, each branch is designed with cascaded orthogonal convolutions. Most of these convolutions are strip-shaped, which fits the large aspect ratio of small targets, preventing redundant background interference from being mixed in when extracting small target features, unlike square convolutional networks. Furthermore, these directional convolutions can accurately perceive the differences in features exhibited by small targets in different directions. Finally, by further integrating these features with these square orthogonal convolutions, the small target features extracted by the network are more robust, such as... Figure 5 As shown, this step enables the network to accurately perceive small targets in different directions.
[0019] The second step is to establish a relationship between the small target and the background and to fuse and enhance the extracted features of the small target from all directions.
[0020] This process can be represented as: ; Here, concat indicates a feature concatenation and fusion operation. This represents dilated convolution. When using dilated convolution, the receptive field of the network expands accordingly. The network is no longer limited to grasping target features, but combines target features with their surrounding environment to construct a relationship between the target and the background. After concatenating and fusing features from all directions, the final output is the enhanced small target feature obtained from the fusion of features from all directions. At this point, the features of each direction of the small target no longer exist in isolation but complement and synergistically enhance features from other directions. This is beneficial for the network to accurately locate and classify small targets.
[0021] The third step is to extract the contextual information of the target over a large range using large convolutional kernels, thereby establishing the relationship between the target and its surrounding environment.
[0022] In small object detection tasks, establishing a correlation between the target and the background is crucial because small objects occupy a very small pixel area, leading to blurred features and hindering accurate classification. However, once a correlation is established between the target and its surrounding background, contextual information can be used to assist the network's classification. For example, when the network needs to distinguish between a car and a boat with very similar shapes, the small object's features alone are insufficient. However, when the contextual information of the small object is water, the network can classify it as a boat instead of a car. Therefore, this design utilizes large convolutional kernels to extract contextual information from small objects, establishing a correlation between the small object and the background over a larger scope. This information is then used as network weights to prepare for subsequent steps. It is worth noting that directly using a 7×7 large convolutional kernel would significantly increase the number of network parameters and computational cost. To ensure real-time performance, this design uses a 7×1 convolution cascaded with a 1×7 convolution, effectively creating a 7×7 convolution.
[0023] The fourth step involves fusing the contextual semantic information extracted by the large convolutional kernel as guiding information into two sub-methods: context-guided background suppression and target density gating feature preservation. This achieves suppression of background interference while preserving the small target features mentioned earlier during downsampling.
[0024] Specifically, the two sub-methods are treated as two parallel branches. In the context-guided background suppression branch, such as... Figure 3As shown, contextual semantic information is used as weights in the channel attention mechanism. This guides the channel attention to focus on channels that encode a large amount of background interference and assigns these channels low weights in the subsequent channel selection process, thereby suppressing background interference. In this process, contextual information extracted by two large convolutional kernels is used. Between the two information extractions, to ensure the real-time performance of the network, this design adopts the classic Bottleneck C2f design from the YOLOv8 network, which maintains the model's representational power while reducing computation and parameter count. Assume the model's input features are... This process can then be represented as: ; in, This represents the features obtained after context-guided background suppression branch processing. GMP represents global max pooling, and GMP represents global average pooling. For the feature preservation branch of target density gating, such as... Figure 4 As shown, contextual semantic information is used as the weight of V in the self-attention mechanism. This guides the self-attention to focus on spatial features that encode a large number of small targets and assigns high weights to these spatial features in subsequent processes, thereby preserving the features of small targets during downsampling. As illustrated, this step enables the network to accurately perceive the background and target while reducing the network's allocation of weights to the background and retaining high weights for the target, thus suppressing the perception of the foreground from the background. This process can be represented as: ; in, This represents the features obtained after processing the feature branches of target density gating. Finally, fusing the features obtained from the two branches achieves the goal of preserving small target features while suppressing the background.
[0025] Taking real-time remote sensing small target detection as an example, a comparison of various existing methods on multiple public remote sensing datasets, as shown in Tables 1 and 2, reveals that the method proposed in this embodiment has better accuracy.
[0026] Table 1. Comparison of accuracy of various existing object detection methods on the SODA-A dataset.
[0027] Table 2. Comparison of accuracy of various existing target detection methods on the SAR-ship dataset.
[0028] In summary, the above are preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A real-time small target detection method with direction sensitivity, characterized in that, Comprise: The first step, the orthogonal direction enhancement method enhances the network's accurate perception of small target's different direction features; The second step, the relationship between small target and background is established and all the extracted direction features of small target are fused and enhanced; The third step, the context information of small target in a larger range is extracted by large convolution kernel to establish the correlation between small target and surrounding environment; The fourth step, the context semantic information extracted by large convolution kernel is taken as guide information to be respectively fused into the context-guided background suppression and target density gating feature reservation sub-methods to realize the suppression of background interference and the reservation of small target features in the process of down-sampling.
2. The method of claim 1, wherein, In the first step, the orthogonal direction enhancement method comprises: The different direction features of small target are perceived and integrated by orthogonal direction bar convolution, and the process can be expressed as: ； ； ； ； ； wherein, - represents the feature obtained after each branch, the horizontal direction bar convolution is represented as , the vertical direction bar convolution is represented as , the main diagonal direction (from the upper left to the lower right) bar convolution is represented as , the sub-diagonal direction (from the upper right to the lower left) bar convolution is represented as , it is worth noting that the bar convolution of each direction is obtained by rotating and transforming the 3x1 convolution, represents the 3x3 convolution.
3. The method of claim 1, wherein, In the second step, the method of establishing target background correlation and fusing different direction features comprises: ； wherein concat represents a concatenation operation on the features, denotes a dilated convolution.
4. The method of claim 1, wherein, In the third step, the method of extracting context semantic information of small target in a larger range by large convolution kernel comprises: The N×1 convolution cascade 1×N convolution is designed to equivalent N×N convolution to equivalent N×N large convolution kernel while avoiding increasing the parameter quantity and calculation quantity of network to ensure the real-time performance of network, and the context information extracted by large convolution kernel is used to establish the correlation between small target and background in a larger range to prepare for the subsequent method.
5. The method of claim 1, wherein, In the fourth step, the background suppression and feature reservation method based on large convolution kernel comprises: ； wherein, GMP represents a global maximum pooling, and GMP represents a global average pooling. ； wherein, represents the features obtained after processing the feature branch of the target density gating.