Remote sensing target recognition method, terminal equipment and storage medium
By improving the feature extraction and fusion parts of the YOLOv8 network, the problems of complex background, multi-scale targets and target occlusion in remote sensing images are solved, and the detection accuracy and robustness of small targets are improved, making it suitable for environmental monitoring and disaster assessment.
Patent Information
- Application Number
- CN202510953592.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-10
- Publication Date
- 2025-09-23
AI Technical Summary
The existing YOLO algorithm has problems in remote sensing images such as complex background interference, multi-scale changes of targets, dense distribution and occlusion of targets, and imbalance between model lightweight and detection efficiency, resulting in insufficient detection accuracy and robustness, especially poor performance in extreme environments.
An improved YOLOv8 network is adopted. By adding depthwise separable convolution and residual connection mechanism to the Backbone part, the FPN network of the Neck part is replaced with the BiFPN network. The P2 layer feature map is introduced into BiFPN, the convolution layer weights are dynamically adjusted, and an adaptive weighting mechanism is used for feature fusion to improve the model's detection ability in complex backgrounds and multi-scale targets.
It significantly improves the detection accuracy and robustness of small targets in remote sensing images, enhances adaptability, and is suitable for fields such as environmental monitoring and disaster assessment.
Smart Images
Figure CN120689725A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of image processing, and in particular to a remote sensing target recognition method, terminal equipment and storage medium. Background Art
[0002] In recent years, with the rapid development of deep learning and computer vision technologies, small object detection in remote sensing images has gradually become a research hotspot. Small objects in remote sensing images often present significant challenges to traditional object detection algorithms due to their low resolution, complex backgrounds, and small size. To address these challenges, deep learning-based object detection methods, particularly the YOLO family of algorithms, have been widely used for object detection in remote sensing images. Among these methods, YOLO achieves significant performance improvements by transforming the object detection problem into a regression problem, enabling real-time detection while maintaining high accuracy.
[0003] With the continuous iteration of YOLO, algorithms such as YOLOv2 and YOLOv3 have been released, further improving detection accuracy and computational speed. In particular, YOLOv4 and YOLOv5, compared to their predecessors, place greater emphasis on small object detection, incorporating techniques such as multi-scale feature fusion and data augmentation, significantly improving small object detection performance. In recent years, improved versions of YOLO, such as YOLOv8 and YOLOv10, have also been introduced, adapting to the stringent computational requirements of small object detection in remote sensing imagery.
[0004] In summary, the application of the YOLO family of algorithms in small target detection in remote sensing imagery continues to make progress. Researchers have continuously improved detection accuracy and efficiency through various optimization strategies, such as multi-scale feature fusion, attention mechanisms, and generative adversarial networks. However, faced with challenges such as complex backgrounds, low contrast, and extremely small targets, the existing YOLO algorithm still has certain limitations. Therefore, how to further improve the robustness of the algorithm and enhance its adaptability and detection accuracy in complex environments, especially its application in extreme environments, remains a current research hotspot.
[0005] In summary, the existing technology has the following problems:
[0006] 1. Complex background interference. Remote sensing images often contain rich ground information, but they also contain a large amount of redundant background information. This background interference can significantly affect the extraction of target features, increasing the difficulty of target detection. Especially in complex terrain and environments, background information may be confused with the target, resulting in reduced detection performance. In addition, changes in lighting, weather, and other factors in remote sensing images can also blur the target's salient features, further increasing the detection difficulty.
[0007] 2. Multi-scale variations in targets. Objects in remote sensing images vary greatly in size, potentially ranging from a few meters to hundreds of meters. This scale variation requires detection algorithms to accurately identify targets at multiple scales. Furthermore, since remote sensing images are typically captured from overhead satellites or drones, adjustments to the shooting angle and focal length can cause significant variations in the pixel size of the same target within the image. Therefore, algorithm design must fully consider the detection requirements for both small and large targets, and address the problem of feature loss for small targets due to downsampling.
[0008] 3. Densely distributed targets and occlusion in complex backgrounds. In remote sensing images, targets may be densely distributed and partially occluded. This is especially true in urban areas or complex terrain. Overlap and occlusion between targets can affect detection. Furthermore, changes in target orientation and scale can complicate detection. Therefore, the YOLO algorithm needs to be optimized to address issues such as densely distributed targets, complex backgrounds, and occlusion, in order to improve the accuracy and robustness of target detection.
[0009] Solving the above technical problems has the following technical difficulties:
[0010] 1. Feature enhancement and noise suppression under complex background interference
[0011] Feature confusion problem: The complex background and target in remote sensing images are highly similar in low-level features such as color and texture. Traditional convolution operations cannot effectively distinguish the target from background noise.
[0012] Robustness to dynamic interference: Lighting changes and weather conditions cause the apparent characteristics of the target to be unstable, and an adaptive feature enhancement mechanism needs to be designed to cope with dynamic interference.
[0013] 2. Information Preservation and Feature Fusion in Multi-Scale Object Detection
[0014] Small object feature loss: Traditional downsampling operations result in the loss of detailed information about small objects, resulting in insufficient resolution of shallow feature maps, making it difficult to support accurate positioning. Cross-scale feature conflict: When multi-scale features are fused, there is information conflict between deep semantic features and shallow detail features. Direct concatenation or addition can easily lead to feature confusion.
[0015] 3. Improved detection accuracy in dense target and occlusion scenarios
[0016] Overlapping objects: Densely distributed objects lead to high bounding box overlap, making traditional non-maximum suppression algorithms prone to mistakenly deleting real objects. Occlusion and feature incompleteness: Partially occluded objects have incomplete appearance features, making it difficult for the model to infer the overall object category and location based on local information.
[0017] 4. Balance between model lightweighting and detection efficiency
[0018] Accuracy-speed trade-off: Lightweight design may lead to a decrease in the ability to express small object features, especially in complex scenarios, with a significant reduction in detection accuracy. Edge device adaptability: Existing models rely on GPU computing power, making it difficult to achieve real-time inference on edge computing platforms such as satellites and drones. Summary of the Invention
[0019] In order to solve the above problems, the present invention proposes a remote sensing target recognition method, a terminal device and a storage medium.
[0020] The specific plan is as follows:
[0021] A remote sensing target recognition method comprises: forming a training set based on remote sensing images annotated with target recognition results; constructing a remote sensing target recognition model and training the model using the training set; adopting YOLOv8 as the network structure of the model, and modifying the third layer C2f of the Backbone part of YOLOv8: adding a depthwise separable convolution before each Bottleneck unit contained in C2f, and introducing a residual connection mechanism to add input features and output features of the Bottleneck unit; and using the trained model for target recognition in remote sensing images.
[0022] Furthermore, the model also includes replacing the FPN network used in the Neck part of YOLOv8 with the BiFPN network, and additionally introducing the P2 layer feature map output by the Backbone part into BiFPN as the bottom-level node in BiFPN for fusion, and then adding a detection head corresponding to the fusion result of the bottom-level node in the Head part.
[0023] Furthermore, it also includes: dynamically adjusting the weight initialization parameters and regularization strategy of the convolutional layer in the Bottleneck unit according to the spatial distribution density of the target in the remote sensing image.
[0024] Furthermore, the method further includes: in the BiFPN network, adding a corresponding weight parameter as a hyperparameter for each fusion node.
[0025] A remote sensing target recognition terminal device includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the method described above are implemented.
[0026] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the method described above in an embodiment of the present invention.
[0027] The present invention adopts the above technical solution to significantly improve the detection accuracy and robustness of small targets in complex backgrounds. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] Figure 1 Shown is a flow chart of a method according to a first embodiment of the present invention.
[0029] Figure 2 Shown is a schematic diagram of the network structure of the model in this embodiment.
[0030] Figure 3 The figure shows a schematic diagram comparing the structures of the C2f module and the CNF module in this embodiment.
[0031] Figure 4 FIG. 4 is a schematic diagram of the improvement process of the BiFPN network in this embodiment. DETAILED DESCRIPTION
[0032] To further illustrate various embodiments, the present invention provides accompanying drawings. These drawings form part of the present disclosure and are primarily used to illustrate the embodiments and, in conjunction with the relevant description in the specification, to explain the operating principles of the embodiments. By referring to these drawings, those skilled in the art will be able to understand other possible implementations and the advantages of the present invention.
[0033] The present invention will now be further described with reference to the accompanying drawings and specific embodiments.
[0034] Example 1:
[0035] The embodiment of the present invention provides a remote sensing target recognition method, such as Figure 1 As shown, the method includes the following steps:
[0036] S1: Build a training set based on remote sensing images annotated with target recognition results.
[0037] The remote sensing images used to construct the training set can adopt existing public datasets, such as VisDrone-2019, DOTA dataset, etc.
[0038] S2: Build a remote sensing target recognition model and train the model using the training set.
[0039] like Figure 2 As shown in FIG, the main network used by the model in this embodiment is the YOLOv8 network. In order to enhance the model's adaptability to the category diversity and spatial distribution complexity in remote sensing images, this embodiment improves the feature extraction part (i.e., the Backbone part) in the YOLOv8 network, that is, replaces its third-layer C2f module with a CNF (ConvNextBlock) module.
[0040] like Figure 3The figure shows a structural comparison between the C2f module and the CNF module. The CNF module adds a depthwise separable convolution before each Bottleneck unit in the original C2f module and introduces a residual connection mechanism after the depthwise separable convolution to add the input features of the Bottleneck unit to the output features. Therefore, the CNF module contains multiple feature extraction units, each of which consists of a depthwise separable convolution, a Bottleneck unit, and a residual connection structure. The input of the CNF module is the input of the depthwise separable convolution, and the output of the depthwise separable convolution serves as the input of the Bottleneck unit. The residual connection structure reintroduces the output of the Bottleneck unit into the input of the depthwise separable convolution, and the output of the Bottleneck unit serves as the output of the CNF module.
[0041] By adding a depthwise separable convolutional structure, spatial and channel-domain features of the image can be extracted, reducing parameter redundancy and resource consumption during the convolution process. Residual connections preserve shallow edge features, thereby extracting multi-level feature information, including shallow details and deep semantics, enhancing feature expression capabilities and improving the real-time performance of the model in actual deployment. Therefore, through the aforementioned CNF module, edge texture information and semantic abstraction information in the image can be extracted, and feature maps of different resolutions and semantic levels can be generated at different hierarchical output nodes. This forms a combination of shallow detail feature maps and deep semantic feature maps, constructing a multi-scale feature map group for subsequent fusion processing.
[0042] Furthermore, this embodiment also includes dynamically adjusting the weight initialization parameters and regularization strategy of the convolutional layer (Conv 1×1) in the Bottleneck unit according to the spatial distribution density of the target in the remote sensing image, thereby improving the adaptability and generalization ability of the model in complex scenarios such as uneven distribution of multiple types of targets and large size differences.
[0043] Furthermore, in order to achieve fine positioning of small targets in the model and robust representation of complex backgrounds, and to enhance the comprehensive perception capability of the detection network for multi-scale targets, this embodiment also includes replacing the original FPN network used in the Neck part of YOLOv8 with a BiFPN network to introduce a bidirectional fusion path (including a top-down path and a bottom-up path), and realize efficient transmission and complementation of features through layer-by-layer connection. At the same time, the P2 layer feature map (i.e., shallow feature map) output by the Backbone part is additionally introduced into BiFPN as the lowest node in BiFPN for fusion, and a new detection head (EODL) corresponding to the fusion result of the lowest node is added to the Head part. The detection head has the same structure as the original detection head (Detect) contained in the Head part, such as Figure 4This embodiment replaces the FPN network with a BiFPN network to enhance the information interaction capability between multi-scale features. By introducing shallow feature maps as high-resolution input and adding a detection head, the resolution and key information extraction capabilities of the feature maps can be improved, the hierarchical expression structure of the feature maps can be optimized, and the model's ability to capture target features in complex scenes can be enhanced, thereby significantly improving overall detection accuracy and small target recognition.
[0044] Furthermore, the improved BiFPN network includes multiple bidirectional information flow paths and cross-level jump connection structures. To enable the network to autonomously select response channels for key target areas, this embodiment also includes an adaptive weighting mechanism during the feature fusion process, which dynamically assigns fusion weights to features of different scales based on factors such as the semantic strength and spatial resolution of the feature map. The adaptive weighting mechanism adds corresponding weight parameters as hyperparameters to each fusion node and learns these weight parameters during the training process.
[0045] This embodiment of the present invention improves the feature extraction and feature fusion components of the original YOLOv8, effectively enhancing the ability to detect small objects in complex backgrounds. Experiments demonstrate that this method achieves higher accuracy, robustness, and adaptability in small object detection in remote sensing images, and is widely applicable in fields such as environmental monitoring and disaster assessment.
[0046] Example 2:
[0047] The present invention also provides a remote sensing target identification terminal device, including a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the computer program, the steps in the above-mentioned method embodiment of embodiment 1 of the present invention are implemented.
[0048] Furthermore, as an executable solution, the remote sensing target recognition terminal device can be a computing device such as a desktop computer, laptop, PDA, or cloud server. The remote sensing target recognition terminal device may include, but is not limited to, a processor and memory. Those skilled in the art will appreciate that the above-described component structure of the remote sensing target recognition terminal device is merely an example of a remote sensing target recognition terminal device and does not constitute a limitation on the remote sensing target recognition terminal device. The remote sensing target recognition terminal device may include more or fewer components than those described above, or a combination of certain components, or different components. For example, the remote sensing target recognition terminal device may also include input and output devices, network access devices, buses, etc., but this is not limited in the present embodiment.
[0049] Furthermore, as an executable solution, the processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices. The general-purpose processor may be a microprocessor or any conventional processor. The processor is the control center of the remote sensing target recognition terminal device, and connects various parts of the remote sensing target recognition terminal device using various interfaces and lines.
[0050] The memory can be used to store the computer programs and / or modules, and the processor realizes the various functions of the remote sensing target identification terminal device by running or executing the computer programs and / or modules stored in the memory, and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system and at least one application required for a function; the data storage area can store data created according to the use of the mobile phone, etc. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage device.
[0051] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above method in the embodiment of the present invention are implemented.
[0052] If the module / unit integrated in the remote sensing target recognition terminal device is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention implements all or part of the process in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, it can implement the steps of the above-mentioned various method embodiments. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium can include: any entity or device that can carry the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM) and software distribution medium, etc.
[0053] Although the present invention has been particularly shown and described in conjunction with preferred embodiments, it will be understood by those skilled in the art that various changes in form and details may be made to the present invention without departing from the spirit and scope of the invention as defined in the appended claims, and all such changes are within the scope of protection of the present invention.
Claims
1. A remote sensing target recognition method, characterized in that: include: A training set is formed based on remote sensing images annotated with target recognition results; Build a remote sensing target recognition model and train the model using a training set; The model's network structure adopts YOLOv8, and the third layer C2f of the Backbone part of YOLOv8 is modified: for each Bottleneck unit contained in C2f, a depth-wise separable convolution is added before it, and a residual connection mechanism is introduced to add the input features and output features of the Bottleneck unit; The trained model is used for target recognition in remote sensing images.
2. The remote sensing target recognition method according to claim 1, wherein: The model also includes replacing the FPN network used in the Neck part of YOLOv8 with the BiFPN network, and additionally introducing the P2 layer feature map output by the Backbone part into BiFPN as the bottom-level node in BiFPN for fusion, and then adding a detection head corresponding to the fusion result of the bottom-level node in the Head part.
3. The remote sensing target recognition method according to claim 1, characterized in that: Also includes: According to the spatial distribution density of the target in the remote sensing image, the weight initialization parameters and regularization strategy of the convolutional layer in the Bottleneck unit are dynamically adjusted.
4. The remote sensing target recognition method according to claim 1, characterized in that: Also includes: In the BiFPN network, corresponding weight parameters are added to each fusion node as hyperparameters.
5. A remote sensing target recognition terminal device, characterized in that: The method comprises a processor, a memory, and a computer program stored in the memory and running on the processor, wherein the processor implements the steps of the method according to any one of claims 1 to 4 when executing the computer program.
6. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 4 are implemented.