YOLOv10-based optical remote sensing image vehicle target detection method
By improving the backbone network, neck, and head structure of the YOLOv10 model, the problem of insufficient vehicle target detection accuracy in remote sensing images was solved, achieving higher detection accuracy and robustness, making it suitable for traffic monitoring and urban planning.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHANGCHUN UNIV OF SCI & TECH
- Filing Date
- 2026-04-03
- Publication Date
- 2026-06-09
AI Technical Summary
Existing remote sensing image vehicle target detection methods struggle to effectively capture the multi-scale features of small vehicle targets in complex backgrounds, and target occlusion is frequent in dense scenes, resulting in insufficient detection accuracy.
The backbone network, neck, and head structures of the YOLOv10 model are improved by introducing the PSA-INEMA module, DIDA module, and GMI-IoU loss function to optimize feature extraction and loss function, thereby enhancing the model's adaptability and robustness in complex remote sensing scenarios.
It significantly improves the detection accuracy and robustness of vehicle targets in remote sensing images, reduces the false negative rate and the false positive rate, maintains real-time detection capability, and is suitable for scenarios such as traffic monitoring and urban planning.
Smart Images
Figure CN122176652A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for detecting vehicle targets in optical remote sensing images based on an improved version of YOLOv10, belonging to the field of target detection in optical remote sensing images. Background Technology
[0002] With rapid modernization, automobiles have become commonplace in ordinary households, an indispensable consumer good. However, the development of transportation infrastructure cannot meet the ever-increasing traffic pressure, leading to increasingly prominent traffic congestion problems. With the continuous maturation of remote sensing technology, it has become possible to obtain road surface information using remote sensing images.
[0003] Because remote sensing images are unaffected by other ground-based factors and can capture large areas in a single image, they can quickly capture vast amounts of urban traffic information. Furthermore, due to the unique imaging angle, remote sensing images clearly reveal the relative positions of objects, such as the spatial arrangement of vehicles, roads, and buildings. Based on vehicle detection within remote sensing images, low-cost traffic congestion detection can be achieved, providing timely feedback to vehicle owners and traffic control departments. This also helps in identifying vehicle distribution, pinpointing traffic congestion locations, and locating accident sites, demonstrating significant practical application value.
[0004] Remote sensing images are characterized by large format, high resolution, and high contrast. Among them, optical remote sensing imaging has the most diverse application scenarios and target categories, making target detection in remote sensing images inherently challenging. Compared to other remote sensing targets, such as aircraft and ships, cars are smaller in size, numerous and densely distributed, have limited texture features, complex and varied distribution scenes, and significant background interference, making car target detection in remote sensing images particularly difficult.
[0005] Existing target detection techniques for detecting vehicles in remote sensing images are not entirely satisfactory. These methods typically rely on single-scale feature representations, making it difficult to effectively capture the multi-scale features and complex details of small vehicles, especially in complex backgrounds. In dense scenes, target occlusion is frequent, and traditional loss functions struggle to stably handle the resulting harmful gradients from low-quality samples, thus limiting the model's localization accuracy. Therefore, in-depth research on vehicle target detection in optical remote sensing images has significant academic value and practical application implications.
[0006] This invention, taking into account the characteristics of small vehicle targets in remote sensing images, proposes an improved detection algorithm based on YOLOv10, aiming to enhance detection accuracy and robustness, and achieve more precise vehicle target identification and localization. This method not only optimizes the model's perception ability for small vehicle targets but also enhances its adaptability in complex remote sensing scenarios, which is of significant value in promoting the development of vehicle target detection technology in remote sensing images. Summary of the Invention
[0007] This invention provides a vehicle target detection method based on improved YOLOv10 for optical remote sensing images, aiming to address the insufficient detection accuracy caused by complex backgrounds, small target sizes, dense distribution, and frequent target occlusion in dense scenes. Existing single-stage detectors face three core challenges in remote sensing scenarios: First, the normalization strategy and global attention mechanism of the backbone network are difficult to adapt to the fine-grained feature extraction requirements of small vehicle targets in remote sensing images; second, feature aliasing occurs in the multi-scale feature fusion stage, and background noise is insufficient to suppress target features; third, traditional IoU series loss functions tend to generate a large number of harmful gradients when dealing with densely arranged and mutually occluded scenes of small targets, which restricts the model's convergence quality. This invention effectively addresses the above challenges by synergistically improving the three stages of YOLOv10—the backbone, neck, and head—achieving a dual improvement in accuracy and real-time performance on the DOTAv2.0 benchmark dataset.
[0008] To address the shortcomings of existing technologies, this application proposes an improved method for vehicle target detection in optical remote sensing images, optimized based on YOLOv10. The overall process includes three stages: dataset construction, model building and training, and model testing. In the dataset construction stage, based on the DOTAv2.0 large-scale remote sensing dataset, preprocessing and enhancement operations are sequentially performed on the original high-resolution remote sensing images: preprocessing includes removing overexposed, blurred, and duplicate images, and cropping large-format images into 640×640 standard sub-images using a sliding window strategy to ensure consistent input resolution; enhancement operations include Gaussian noise addition, color dithering, geometric rotation, and random scaling, expanding sample diversity from multiple dimensions and significantly improving the model's adaptability to complex lighting and shooting angle changes. Subsequently, the LabelImg tool is used to accurately annotate vehicle targets with rectangular boxes, recording target category and location information. Finally, the dataset is divided into training, validation, and test sets in a 7:2:1 ratio to ensure the scientific rigor of model training, tuning, and evaluation.
[0009] After preparing the dataset, an improved detection model was built and trained. The overall network structure follows the YOLOv10 Backbone-Neck-Head three-stage framework, with targeted improvements introduced in each of the three stages. In the Backbone part, an improved INEMA (Improved Efficient Multi-Scale Attention) unit was embedded into the PSA (Partial Self-Attention) module of the YOLOv10 backbone network to form a PSA-INEMA module, replacing the original Multi-Head Self-Attention (MHSA) module. The INEMA unit replaces Group Normalization (GN) with Layer Normalization (LN) to eliminate the problem of GN artificially amplifying statistical differences between groups in complex remote sensing backgrounds. At the same time, 1×1 convolutional residual connections are added to provide a fast propagation path for gradients, enhancing the ability to preserve small target features and integrate cross-channel information, effectively suppressing the interference of background noise on the expression of key features. In the Neck section, a DIDA (Dynamic Interactive Dual Attention) module is introduced at the key feature fusion node. Through bidirectional cross-modulation and dynamic weighted fusion between spatial and channel attention, the saliency of vehicle target features is enhanced, improving target localization accuracy while effectively reducing interference from complex backgrounds in remote sensing images. In the Head section, a GMIn-IoU hybrid loss function replaces the original CIoU loss function. Using Inner-IoU as the core framework and incorporating a dynamic non-monotonic focusing mechanism, the gradient allocation strategy is optimized to improve the performance and localization accuracy of small target detection.
[0010] The core improvement of the PSA-INEMA module lies in the design of the INEMA unit. The original EMA (Efficient Multi-Scale Attention) module, through feature grouping and the design of three parallel sub-networks, achieves the effect of capturing multi-scale spatial and channel attention information without dimensionality reduction. However, the group normalization (GN) used in the EMA module has obvious limitations: GN divides the feature channels into several groups of fixed size and normalizes the mean and variance of each group independently. This design is effective in tasks such as image classification, but in optical remote sensing images with complex backgrounds and extremely uneven target distribution, it artificially amplifies the differences in statistical characteristics between groups, leading to a significant decrease in the ability to capture cross-channel associations and weakening the model's ability to perceive sparsely distributed small targets. To address this, INEMA replaces GN in the three parallel branches of the EMA module with layer normalization (LN): LN jointly calculates the mean and variance of all channels of a single sample, avoiding the inter-group fragmentation problem of GN and more stably maintaining the channel consistency of key features of small targets. Meanwhile, INEMA adds a parallel 1×1 convolutional residual branch next to the three branches, providing a more direct backpropagation path for the gradient, effectively suppressing the gradient fluctuation and decay risk caused by LN global normalization of deep features, and improving the model training stability and sensitivity to small targets.
[0011] The DIDA dynamic interactive dual attention mechanism module is an interactive improvement on the DA-Net (Dual Attention Network). In the original DA-Net, the spatial attention branch and the channel attention branch are independent, and the outputs of the two branches are simply added and fused, lacking information interaction between the branches, thus limiting its ability to suppress complex backgrounds in remote sensing scenes. The core innovation of DIDA lies in introducing a bidirectional cross-modulation mechanism: the feature map output from the spatial attention module's adder is introduced into the multiplication operation position of the channel attention module, enabling the channel attention to simultaneously perceive spatial location information when aggregating the global responses of each channel; simultaneously, the output feature map of the channel attention module is introduced into the corresponding multiplication position of the spatial attention module, allowing spatial attention to benefit from the prior guidance of channel weights when modeling long-range dependencies between pixels. This bidirectional cross-modulation mechanism achieves dynamic synergy between spatial attention and channel attention, guiding the network to focus on weak vehicle target features while effectively suppressing redundant background information. In the final fusion stage, the channel branch output... and spatial branch output The corresponding dynamic weight maps are generated through lightweight 1×1 convolutional layers. and After Softmax normalization along the branch dimension, the weighted average is adjusted using an adaptive weighting formula. Fusion is achieved to replace the equal-weight summation of the original DA-Net, significantly enhancing the detailed response of the road area and suppressing the noise interference in the empty area. The improved DIDA module is embedded into the key feature fusion nodes of the upsampling and downsampling stages at the Neck end, effectively solving the problem of multi-scale feature aliasing, weakening the feature weights of complex background information in remote sensing images, and making the model more focused on the automotive target area.
[0012] The GMIn-IoU hybrid loss function is composed of three components working together and is designed layer by layer according to the special requirements of small automotive target detection in remote sensing images. The first component is the Inner-IoU core framework: The traditional IoU loss directly optimizes the regression with the intersection over union of the GT box and the predicted box. For extremely small automotive targets, the small area of the GT box makes the IoU value highly sensitive to the offset height of the predicted box, prone to training instability problems. Inner-IoU introduces a scale factor r (0 < r < 1) to construct a shrunk auxiliary ground truth bounding box (the side length of the auxiliary box is r times the side length of the original GT box), and uses the intersection over union of the auxiliary box and the predicted box to replace the original IoU for loss calculation, shrinking the focus area of regression optimization to the fine area near the target center, effectively guiding the model to focus on the local fine regression of small targets, accelerating convergence, and significantly improving the adaptability to small automotive targets in remote sensing images. The second component is the dynamic non-monotonic focusing mechanism: In a dense occlusion scenario, low-quality samples (such as severely occluded targets) often produce high loss values but contain a large amount of harmful gradients. If processed with equal weights, it will interfere with the gradient optimization direction of normal samples. This mechanism uses the "outlier degree" of the predicted box (defined as the ratio of the Inner-IoU loss of the current sample to the batch average loss) to replace IoU for quality assessment, and dynamically assigns the gradient gain γ accordingly; the geometric penalty term is based on to calculate , which is used to constrain the geometric alignment degree between the predicted box and the true box, compensating for the defect of gradient disappearance when the predicted box and the GT box have no overlap or partial overlap, and ensuring that an effective gradient direction can still be provided when the target positioning deviation is large; The third component is the geometric penalty term : Based on to calculate , using the area of the minimum enclosing rectangle to constrain the geometric consistency between the predicted box and the true box, and still being able to generate meaningful penalty gradients in the extreme case where the predicted box and the GT box do not overlap at all. The three components work together: Inner-IoU focuses on the fine regression of small targets, the dynamic non-monotonic focusing mechanism shields the harmful gradients of low-quality occluded samples and guides the training focus to medium-quality samples, and the geometric penalty term ensures global positioning convergence. The three complement each other, effectively improving the overall positioning accuracy and robustness of the model in complex remote sensing image scenarios.
[0013] As can be seen from the above, the advantages of this invention are as follows: This invention provides a vehicle target detection method based on an improved YOLOv10 optical remote sensing image. This method performs data augmentation on remote sensing image data to obtain multiple labeled vehicle target images. Subsequently, the augmented image dataset is input into an optical remote sensing image vehicle target detection model, which is trained and optimized using an algorithm based on the improved YOLOv10. Finally, the predicted image is input into the trained optical remote sensing image vehicle target detection model to detect the predicted vehicle target information and obtain labeled detection images. The improved model can significantly improve detection accuracy, thus solving the problem that traditional algorithms cannot accurately detect vehicles in optical remote sensing images with complex backgrounds, small and densely distributed targets, and frequent occlusion in dense scenes. Attached Figure Description
[0014] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0015] Figure 1 A flowchart illustrating a vehicle target detection method based on improved YOLOv10 optical remote sensing images provided by this invention; Figure 2 A schematic diagram of an improved overall network structure provided by the present invention; Figure 3 This is a schematic diagram of the structure of the PSA-INEMA module provided by the present invention; Figure 4 A schematic diagram of the structure of the DIDA dynamic interactive dual attention mechanism module provided by the present invention; Figure 5 This is a schematic diagram of the GMIn-IoU hybrid loss function calculation process provided by the present invention. Detailed Implementation
[0016] The following description sets forth specific details such as particular system architectures and technologies, which are only intended to help better understand the embodiments of this application and are not intended to limit the scope of implementation. However, those skilled in the art should understand that other embodiments of this application can also be implemented without relying on these specific details.
[0017] It should be understood that the term "comprising" as used in this specification and the appended claims is intended to indicate the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the possibility of the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or combinations thereof.
[0018] The technical solutions will now be clearly and completely described with reference to the accompanying drawings of the embodiments of this application. Obviously, the described embodiments are only a part of this application, and not all embodiments. Other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are all within the scope of protection of this application.
[0019] Example Dataset: This invention uses the DOTAv2.0 dataset as the benchmark dataset for car target detection in remote sensing images. This dataset contains 11,268 high-resolution remote sensing images, labeled with 18 target categories, and contains approximately 172,000 target instances. Among them, the car category has the most abundant number and is the key detection target of this invention.
[0020] Data augmentation was performed on the images in the dataset, including adding Gaussian noise and color dithering, which helps improve the robustness of the model. Large remote sensing images were cropped into standard 640×640 sub-images using sliding window cropping, and geometric rotation and scaling transformations were used to increase the diversity of training samples. The LabelImg image annotation tool was used to label the images, recording the category and location information of car targets. Finally, the dataset was divided into training, validation, and test sets in a 7:2:1 ratio for model training, tuning, and evaluation.
[0021] In this embodiment, YOLOv10 is used as the baseline model, and its structure is optimized and improved in three aspects to enhance the detection capability of small car targets in remote sensing images. An improved target detection network is constructed, and the network structure is as follows: Figure 2 As shown.
[0022] Backbone section: An improved INEMA (Improved Efficient Multi-Scale Attention) unit is embedded in the PSA (Partial Self-Attention) module of the YOLOv10 backbone network to form a PSA-INEMA module, replacing the original multi-head self-attention (MHSA) module. The module structure is as follows. Figure 3As shown, the INEMA unit is an improvement on the EMA (Efficient Multi-Scale Attention) module. Through feature grouping and the design of three parallel sub-networks, it effectively avoids the information loss caused by channel dimensionality reduction: Branch 1 performs one-dimensional global average pooling in the X direction to capture horizontal spatial context information, and Branch 2 performs one-dimensional global average pooling in the Y direction to extract vertical spatial dependencies. The spatial features of the two branches are activated by Sigmoid to generate a spatial attention weight map; Branch 3 captures local spatial patterns through 3×3 convolution and then performs two-dimensional global average pooling to generate a channel attention weight map. To address the problem of reduced cross-channel correlation capture capability of group normalization (GN) in EMA when processing remote sensing images with complex backgrounds, INEMA replaces GN in the three parallel branches with layer normalization (LN) and adds a parallel 1×1 convolutional residual branch to provide a more direct propagation path for gradients, effectively suppressing the risk of deep gradient fluctuations and improving the ability to preserve small target features and integrate cross-channel information.
[0023] Neck section: A DIDA (Dynamic Interactive Dual Attention) module is embedded in the key feature fusion nodes at the Neck end. Specifically, it is integrated after the upsampling structure and after each C2f and Upsampling module in the downsampling stage. The module structure is as follows: Figure 4 As shown. The DIDA module is an improvement on DA-Net (DualAttention Network). It interactively improves the spatial attention module and the channel attention module by inputting the feature map output from the spatial attention module's adder to the multiplication calculation position of the channel attention module, and simultaneously inputting the feature map output from the channel attention module to the corresponding position of the spatial attention module. This achieves bidirectional dynamic interaction and cross-modulation between spatial and channel attention, guiding the network to focus on weak vehicle target features while suppressing redundant background information. In the final fusion stage, the channel branch output... and spatial branch output The corresponding dynamic weight maps are generated through lightweight 1×1 convolutional layers. and After performing Softmax normalization along the branch dimension, the formula is applied. It achieves adaptive weighted fusion, effectively solves the feature aliasing problem, significantly improves the detail response in the road area and suppresses noise in the empty area.
[0024] Head section: A hybrid GMIn-IoU loss function is introduced to replace the original CIoU loss function to optimize the bounding box regression localization loss. The calculation process is as follows: Figure 5As shown in the figure. GMIn-IoU takes Inner-IoU as the core framework, introduces a scale factor r (0 < r < 1) to construct a shrunk auxiliary ground truth bounding box (the side length of the auxiliary box is r times the side length of the original GT box), and uses the intersection over union of the scaled auxiliary GT box and the predicted box to replace the original IoU for loss calculation, effectively focusing on the fine regression region of small targets, accelerating model convergence and improving the adaptability to small car targets in remote sensing images. On this basis, GMIn-IoU incorporates a dynamic non-monotonic focusing mechanism, defines the "outlier degree" of the predicted box , and uses this to evaluate the quality of the predicted box, and dynamically assigns gradient gains according to the outlier degree . Finally, the loss function is expressed as ; where the geometric penalty term is based on for calculation , measuring the geometric alignment degree between the predicted box and the ground truth box, and compensating for the defect of gradient disappearance of pure IoU in non-overlapping cases. The three work together to effectively reduce the harmful gradients generated by low-quality occluded samples, make the model focus on the optimization of medium-quality predicted boxes, and improve the overall localization accuracy and robustness.
[0025] The hardware of the experimental environment for this experiment is as follows: CPU processor Intel Core i9-13900HX; memory is 64GB DDR5; GPU graphics card is NVIDIA GeForce RTX 4090; storage hard disk is 2TB PCIe 4.0 NVMe solid state drive. The software of the experimental environment is as follows: operating system Ubuntu 22.04; programming language Python 3.10; model framework PyTorch 2.1.0; CUDA version 12.1.
[0026] In summary, this application proposes an improved method for detecting car targets in optical remote sensing images based on YOLOv10. Aiming at the problems of misdetection, missed detection of targets and insufficient regression loss accuracy caused by factors such as complex backgrounds, small sizes of car targets, dense distributions, and frequent occlusions in dense scenes in optical remote sensing images, through three collaborative improvements of introducing PSA-INEMA in the PSA module of the backbone network, embedding the DIDA module at the key fusion node of the Neck, and adopting the GMIn-IoU loss function at the Head, a new improved detection model is designed. This model significantly improves the detection accuracy and efficiency of car targets in optical remote sensing images, significantly reduces the missed detection rate and false detection rate in complex backgrounds, and at the same time maintains the real-time detection ability, has high practical application value, and is especially suitable for scenarios such as traffic monitoring, intelligent traffic management, and urban planning in satellite images.
[0027] The above embodiments illustrate only one specific implementation of the present invention. Although described in detail, they do not limit the scope of patent protection of the present invention. It should be particularly noted that those skilled in the art can make various modifications and improvements without departing from the basic concept of the present invention, and all such modifications and improvements should be considered within the scope of protection of the present invention.
Claims
1. A method for detecting vehicle targets in optical remote sensing images based on an improved version of YOLOv10, characterized in that, include: S1: Obtain the optical remote sensing image dataset of vehicle targets, preprocess and accurately label the dataset, and after image cropping, denoising and data augmentation, divide it into training set, validation set and test set in a 7:2:1 ratio; S2: Using YOLOv10 as a baseline, an improved detection network is constructed: the original multi-head self-attention module is replaced with the PSA-INEMA module in the backbone PSA module; the DIDA module is introduced in the key feature fusion node at the Neck end; and the CIoU loss function is replaced with the GMIn-IoU hybrid loss function at the Head end. S3: Input the training set into the detection network for training, and iteratively adjust the model parameters based on the validation set results until convergence; S4: Input the test set into the trained detection network and output the category and location detection results of car targets in the optical remote sensing image.
2. The method for detecting vehicle targets in optical remote sensing images as described in claim 1, characterized in that, The PSA-INEMA module replaces the multi-head self-attention (MHSA) module in the PSA structure of the YOLOv10 backbone network with INEMA units; the INEMA unit replaces group normalization (GN) with layer normalization (LN) in the three parallel branches of the EMA module, and adds parallel 1×1 convolutional residual branches.
3. The method for detecting vehicle targets in optical remote sensing images as described in claim 2, characterized in that, The INEMA unit contains three parallel branches: branch 1 performs one-dimensional global average pooling in the X direction; branch 2 performs one-dimensional global average pooling in the Y direction; the outputs of branch 1 and branch 2 are fused after being activated by Sigmoid to generate a spatial attention weight map. Branch 3 undergoes a 3×3 convolution followed by a two-dimensional global average pooling process to generate a channel attention weight map; the outputs of the three branches are then weighted and fused to serve as the final output of the module.
4. The method for detecting vehicle targets in optical remote sensing images as described in claim 1, characterized in that, The DIDA module is an improvement on DA-Net: it introduces the feature map output by the adder of the spatial attention module into the multiplication operation position of the channel attention module, and at the same time introduces the output feature map of the channel attention module into the corresponding multiplication position of the spatial attention module, thereby realizing bidirectional cross-modulation of spatial attention and channel attention.
5. The method for detecting vehicle targets in optical remote sensing images as described in claim 4, characterized in that, During the fusion phase, the DIDA module outputs channel branches. and spatial branch output Weight maps are generated by 1×1 convolutional layers respectively. and After Softmax normalization, according to the formula Achieve adaptive weighted fusion.
6. The method for detecting vehicle targets in optical remote sensing images as described in claim 4, characterized in that, The DIDA module is embedded after each Upsample upsampling structure in the Neck of the YOLOv10 network, and before the convolution operation after each C2f module in the downsampling stage.
7. The method for detecting vehicle targets in optical remote sensing images as described in claim 1, characterized in that, The GMIn-IoU hybrid loss function takes Inner-IoU as the core framework: introducing a scale factor r (0 < r < 1) to construct an auxiliary true bounding box (the side length of the auxiliary box is r times the side length of the original GT box) for calculating the intersection over union; integrating a dynamic non-monotonic focusing mechanism based on the outlier degree of the predicted box to dynamically allocate gradient gains ; where is a geometric penalty term based on ; finally, the loss function is expressed as . 8. The method for detecting vehicle targets in optical remote sensing images as described in claim 1, characterized in that, The training in S3 includes inputting the data-enhanced training set into the network model described in this application for training, and iteratively optimizing the model parameters based on the validation set results until convergence.