A real-time target detection method and device based on reparameterization large convolution kernel

Abstract

This invention relates to the field of computer vision technology, specifically to a real-time object detection method and apparatus based on reparameterized large convolutional kernels. The method involves inputting the image to be detected into a real-time object detection model. The model extracts hierarchical features from the image based on a backbone network integrating reparameterized large convolutional kernels. A neck network based on a feature pyramid network fuses the features extracted from different levels by the backbone network, generating a feature map fused to a specified feature layer. A head network trained using a triple positive sample allocation strategy outputs the feature map of the specified feature layer from the neck network to generate object classification and bounding box regression results. Compared to existing technologies, this invention achieves a better balance between detection accuracy and inference speed, enabling high real-time inference speed while maintaining high detection accuracy.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a real-time object detection method employing a convolutional neural network architecture with heavily parameterized large convolutional kernels and a sample allocation strategy. Background Technology

[0002] Object detection, a fundamental and crucial task in computer vision, aims to automatically identify all instances of interest from images or video streams and accurately pinpoint their locations (usually in the form of bounding boxes). The introduction of deep learning technology has significantly raised the performance ceiling of object detection, leading to breakthroughs in numerous practical applications. These include vehicle and pedestrian recognition in autonomous driving systems, anomaly detection in intelligent security monitoring, environmental perception in robot navigation and interaction, lesion localization in medical image analysis, product defect detection in industrial automation, and remote sensing image interpretation in aerospace and geographic information systems. It has become a key enabling technology driving the intelligent upgrading of various industries.

[0003] Currently, within the deep learning framework, mainstream object detection methods can be broadly categorized into two main technical approaches based on their core processing flow and design focus:

[0004] The first category comprises multi-stage detection methods that prioritize achieving the highest possible detection accuracy. Examples of this type include Faster R-CNN and Transformer-based DETR algorithms. These methods aim to achieve high-precision detection through multi-stage fine-tuning or iterative correction. However, this high accuracy often comes at the cost of significant computational cost. Their inherent multi-stage processing flow, complex attention calculations, and independent processing of numerous candidate regions or queries result in substantial computational overhead and memory consumption, leading to relatively slow inference speeds and making them unsuitable for applications with strict processing speed constraints, such as real-time video streaming and embedded device deployments.

[0005] The second category comprises single-stage detection methods, such as the YOLO series, which prioritize high efficiency and real-time performance. The core idea of ​​single-stage methods is "one-step": they eliminate the need for independent region proposal generation steps, directly and densely predicting bounding boxes with different spatial locations, preset scales, and aspect ratios, along with their corresponding class confidences, on the network in a single, one-time operation. By highly integrating target localization and classification tasks into a single, end-to-end network structure, the detection process is greatly simplified, significantly reducing computational complexity. This results in single-stage detectors typically possessing very fast inference speeds, easily meeting or even exceeding real-time processing requirements (e.g., reaching 30 FPS or even hundreds of FPS). However, this speed advantage often comes with a certain degree of accuracy loss. Compared to multi-stage methods, single-stage detectors face a more severe imbalance between positive and negative samples during training (background regions far outnumber target regions), and direct prediction on feature maps can lead to relatively low localization accuracy, especially when dealing with small targets, densely overlapping targets, or distinguishing between similar-looking categories, where their performance is usually inferior to top-tier multi-stage methods. Although subsequent techniques such as feature pyramid networks and focus loss were introduced to alleviate these problems, the inherent trade-off between accuracy and speed still exists.

[0006] Despite significant advancements in all target detection technologies, regardless of their approach, both accuracy- and speed-oriented methods face a common bottleneck when applied to images under specific and challenging non-natural lighting conditions or special imaging modalities, such as remote sensing images, infrared thermal imaging images, and synthetic aperture radar (SAR) images: the difficulty in effectively and efficiently balancing detection accuracy and processing speed. These types of image data often exhibit statistical characteristics and imaging physics mechanisms significantly different from conventional natural light images. For example, remote sensing images cover vast geographical areas with highly variable target scales, complex and diverse backgrounds, and targets may rotate arbitrarily, requiring a large effective receptive field and fine resolution for accurate detection; infrared images lack color texture, have low contrast, poor signal-to-noise ratio, and target recognition relies more on contours and thermal features, making them sensitive to noise; SAR images contain speckle noise and geometric distortion, their grayscale values ​​reflect backscattering characteristics, and their texture structure is expressed in a unique way. When dealing with these special types of data, the inherent design flaws of existing detection frameworks are amplified: if a high-precision model is used, its massive computational load is too slow to process high-resolution remote sensing / SAR images, failing to meet timeliness requirements; if a high-speed model is used, its limited feature representation capabilities and receptive field are often insufficient to cope with the complexity of the aforementioned images, resulting in a significant decrease in detection performance (especially the ability to identify targets in complex backgrounds, small targets, and targets under special imaging conditions), an increase in false negative and false positive rates, and difficulty in achieving the accuracy and robustness required for practical applications.

[0007] Therefore, in the field of target detection, especially in challenging application scenarios such as remote sensing, infrared, and SAR, there is an urgent need to research and develop a new target detection technology paradigm. Summary of the Invention

[0008] The purpose of this invention is to provide a real-time target detection method and apparatus based on a reparameterized large convolutional kernel architecture.

[0009] The objective of this invention can be achieved through the following technical solutions:

[0010] As a first aspect of the present invention, a real-time target detection method based on reparameterized large convolutional kernels is provided, comprising the following steps:

[0011] The image to be detected is input into a real-time target detection model, which includes a backbone network, a neck network, and a head network.

[0012] By using a backbone network based on integrated reparameterized large convolutional kernels, the image to be detected is downsampled and features at different levels are extracted;

[0013] The features extracted from different levels of the backbone network are input into the corresponding feature layers of the neck network based on the feature pyramid network for fusion to generate feature maps of the corresponding feature layers.

[0014] By training the head network using a triple positive sample allocation strategy, and outputting feature maps of specified feature layers based on the neck network, target classification and bounding box regression results are generated.

[0015] As a preferred technical solution, the backbone network includes four reparameterized large convolutional kernel modules. The backbone network sequentially downsamples the image to be detected and extracts features corresponding to the downsampling rate through the reparameterized large convolutional kernel modules. In the latter two reparameterized large convolutional kernel modules, depthwise separable convolution is used, and padding is configured to maintain the spatial resolution of the feature map. At the end of the backbone network, a spatial pyramid pooling module is provided. The spatial pyramid pooling module integrates contextual information of different scales by connecting multiple pooling operations with different kernel sizes in series or parallel.

[0016] As a preferred technical solution, the reparameterized large convolutional kernel module has only one layer in the vertical direction. During training, it adopts a multi-branch structure, which is set up in parallel with: at least one large convolutional kernel branch to capture global and contextual features; at least one small convolutional kernel branch to capture local fine structural information; and an identity mapping branch. During the training phase, the parameters of each branch in the multi-branch structure are learned and updated independently.

[0017] During the inference phase, the parameters of all parallel branches are mathematically equivalently merged into the parameters of a single standard convolutional layer.

[0018] As a preferred technical solution, the neck network uses a feature pyramid network as its basic structure, upsamples and convolves the output features at the ends of the backbone network from bottom to top, and fuses the features from the corresponding downsampling rate level of the backbone network through lateral connections to generate the feature map of the corresponding feature layer.

[0019] As a preferred technical solution, the head network adopts a single-layer head structure trained by a triple positive sample allocation strategy, and uses the feature map of the 1 / 8 feature layer of the neck network as input to perform target classification and bounding box regression prediction.

[0020] As a preferred technical solution, the head network employs a triple positive sample allocation strategy during the training phase, providing dense and high-quality supervision signals to the main detector through auxiliary branches that exist only during training, specifically including:

[0021] The deep-level densely distributed branches include a classification subnetwork and a regression subnetwork. Both the classification subnetwork and the regression subnetwork consist of four feature extraction convolutional layers and one prediction convolutional layer. The dense distribution method treats all regions inside the target as positive samples, generating supervision signals for shallow-level sparse branches.

[0022] The deep sparse allocation branch includes a classification subnetwork and a regression subnetwork. Both the classification subnetwork and the regression subnetwork consist of four feature extraction convolutional layers and one prediction convolutional layer. A one-to-one matching strategy is used to allocate positive samples as the teacher branch, which generates a set of positive sample targets for the shallow sparse branch.

[0023] The shallow sparse allocation branch includes a classification subnetwork and a regression subnetwork, each of which contains only one convolutional layer as a student branch, and supervised training is performed using the output of the deep sparse allocation branch as a supervision signal.

[0024] During the inference phase, the shallow sparse allocation branch is retained as the detection head for the head network to perform predictions.

[0025] As a preferred technical solution, the loss function of the real-time target detection model is composed of a weighted sum of classification loss and target localization loss; the classification loss adopts the quality focus loss; and the target localization loss adopts the generalized intersection-union loss.

[0026] As a preferred technical solution, for large-sized images to be detected with resolutions far exceeding a predetermined threshold, a sliding window method is used for prediction. The window slides with a fixed step size for the image to be detected, and the image blocks within each window are independently detected using the real-time target detection model. The detection results of all windows are then summarized and processed by non-maximum suppression to obtain the final full-image detection result.

[0027] As a second aspect of the present invention, a real-time target detection device based on reparameterized large convolutional kernels is provided, comprising a memory, a processor, and a program stored in the memory, wherein the processor executes the program to implement the real-time target detection method based on reparameterized large convolutional kernels as described above.

[0028] As a third aspect of the present invention, a storage medium is provided having a program stored thereon, which, when executed, implements the real-time target detection method based on reparameterized large convolutional kernels as described above.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] This invention proposes an innovative solution based on a reparameterized large convolutional kernel architecture.

[0031] 1) This invention provides a novel real-time target detection method that achieves a good balance between high accuracy and high speed in various scenarios, especially challenging ones. By organically combining an efficient architecture with reparameterized large convolutional kernels and a refined training strategy with triple positive sample allocation, the method successfully and significantly improves target detection accuracy while maintaining real-time inference speed. The proposed scheme overcomes the limitations of existing methods, achieving a better balance between detection accuracy and inference speed through innovative network architecture design and training strategies, enabling high-efficiency real-time inference while maintaining high detection accuracy. Especially when processing the aforementioned challenging image data, it provides stable, reliable, and efficient detection performance, thus better serving the practical application needs of key fields such as national defense security, disaster monitoring, resource exploration, and environmental monitoring.

[0032] 2) This invention employs a reparameterized large convolutional kernel architecture. By utilizing large convolutional kernels in the shallow layers of the network to obtain a large receptive field, and combining this with reparameterization techniques to achieve efficient computation during inference, it aims to solve the problem of existing methods struggling to balance speed and receptive field when processing images requiring extensive contextual information. It strives to significantly improve inference speed while maintaining or even enhancing feature representation capabilities. This allows the network to operate in the concise form of standard convolution during inference, while enjoying the broad receptive field provided by large convolutional kernels. This effectively addresses the problem of traditional methods struggling to balance speed and receptive field when processing context-dependent images such as remote sensing, infrared, and SAR.

[0033] 3) This invention also proposes a triple positive sample allocation strategy. By introducing a dedicated allocation branch during the training phase, it provides denser and higher-quality positive sample supervision signals, enabling the network to operate in a concise form of standard convolution during inference while enjoying the broad receptive field brought by large convolutional kernels. This effectively solves the problem of traditional methods struggling to balance speed and receptive field when processing context-dependent images such as remote sensing, infrared, and SAR. It enhances the model's learning ability for difficult samples (such as small targets, low signal-to-noise ratio targets, and targets in complex backgrounds), thereby effectively improving detection accuracy without sacrificing the speed advantage of the heavily parameterized large convolutional kernel architecture, thus addressing the pain point of insufficient accuracy in high-speed models. Attached Figure Description

[0034] Figure 1 This is a flowchart of the real-time target detection method based on reparameterized large convolutional kernels proposed in this invention.

[0035] Figure 2 This is a schematic diagram of the structure of the real-time target detection model of the present invention.

[0036] Figure 3 This is a schematic diagram illustrating the principle of structural transformation of the reparameterized large convolutional kernel during the training and inference phases of this invention. Detailed Implementation

[0037] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are implemented based on the technical solution of the present invention, providing detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0038] Example 1

[0039] This invention proposes an innovative solution based on a reparameterized large convolutional kernel architecture. It addresses the common problem in existing target detection methods that struggle to balance accuracy and speed when processing non-natural light images such as remote sensing, infrared, and SAR. By using a reparameterized large convolutional kernel architecture, a large receptive field is obtained through large convolutional kernels in the shallow layers of the network. Combined with reparameterization techniques, efficient computation during inference is achieved. This aims to solve the problem of existing methods struggling to balance speed and receptive field when processing images requiring extensive contextual information (such as remote sensing and SAR images), striving to significantly improve inference speed while maintaining or even enhancing feature representation capabilities. Simultaneously, a triple positive sample allocation strategy is proposed. By introducing a dedicated allocation branch during the training phase, more dense and high-quality positive sample supervision signals are provided, aiming to improve the model's learning ability for difficult samples (such as small targets, low signal-to-noise ratio targets, and targets in complex backgrounds). Thus, without sacrificing the speed advantage brought by the reparameterized large convolutional kernel architecture, detection accuracy is effectively improved, addressing the pain point of insufficient accuracy in high-speed models. By combining these two core technologies, this invention aims to provide a new real-time target detection method that can achieve a good balance between high accuracy and high speed in various scenarios, especially challenging scenarios.

[0040] like Figure 1 As shown, the specific implementation steps of the real-time target detection method based on reparameterized large convolutional kernels proposed in this invention are as follows:

[0041] S1. Construct a backbone network integrating heavily parameterized large convolutional kernels as the basic module for feature extraction.

[0042] The backbone network forms the foundation of the entire detection framework, responsible for extracting hierarchical features from the input image. To quickly obtain a large effective receptive field in a shallow network, especially when processing remote sensing and SAR images with strong contextual dependence, this step employs the following design:

[0043] 1.1) In the feature extraction of the backbone network, re-parameterized large kernel convolutional networks (RepLKNet) are used to replace traditional small convolutional kernels such as 3x3. The size of the large convolutional kernel can be selected according to the specific task and computational resources, including but not limited to one or more of 11x11, 9x9, or 7x7, such as... Figure 2As shown, H is the image height, W is the image width, C is the number of image channels, k is the kernel size, and s is the stride. The backbone network contains four reparameterized large convolutional kernel modules. In the first two modules, normal convolution (Conv) is used, while in the latter two modules, due to the increased number of channels, depthwise separable convolution is employed to reduce computational cost. Simultaneously, appropriate padding is configured to maintain the spatial resolution of the feature maps and prevent premature information loss. This design aims to rapidly expand the network's effective receptive field, enabling the backbone network to capture a wide range of contextual information early on, thereby reducing the number of layers in the backbone network.

[0044] 1.2) As Figure 3 As shown, the reparameterized large convolutional kernel module employs a multi-branch structure during training. This structure internally sets up at least one large convolutional kernel Conv branch (e.g., 11x11) and at least one small convolutional kernel Conv branch (e.g., 3x3 or 1x1) in parallel. The large kernel branch captures global and contextual features, while the small kernel branch focuses on capturing fine-grained local structural information. Additionally, an identity mapping branch can be included to stabilize the training process. This multi-branch design enriches the diversity of features and enhances the model's learning ability.

[0045] 1.3) During the training phase, the parameters of each branch in the multi-branch structure are learned and updated independently. However, during the inference phase, to achieve efficient deployment, a structural reparameterization technique is used to mathematically and equivalently fuse the parameters of all parallel branches (large kernel, small kernel, identity mapping, etc.) into the parameters of a single standard convolutional layer. This way, only one efficient large-kernel convolutional layer performs forward propagation during inference, significantly improving computational speed and achieving both high performance during training and fast speed during inference.

[0046] 1.4) At the end of the backbone network or near the output, a Spatial Pyramid Pooling – Fast (SPPF) module is introduced. This module effectively integrates contextual information at different scales by cascading or paralleling multiple pooling operations with different kernel sizes (e.g., 5x5), further enhancing the contextual information integration capability of the output features and the robustness and expressiveness to changes in the target scale.

[0047] 1.5) In order to serve the goal of real-time detection, the backbone network adopts a relatively shallow layer design. Compared with previous modules such as C2f in YOLOv8 and C3f2 in YOLO11, which stack two and three layers of convolution and have multiple horizontal convolutions respectively, the large convolution kernel reparameter module we use has only one layer in the vertical direction and only one layer in the horizontal direction during the inference stage, thereby directly reducing the amount of computation and network latency.

[0048] S2. Construct a neck network that also integrates heavily parameterized large convolutional kernels to fuse features from different levels.

[0049] The role of the neck network is to fuse features extracted from different levels by the backbone network to generate feature maps that have good representation capabilities for targets at different scales. In this embodiment, the construction of the neck network includes:

[0050] A Feature Pyramid Network (FPN) is used as the basic structure. The output features of the backbone network are convolved and upsampled from bottom to top, and corresponding scale features from the backbone network are fused through lateral connections. Specifically, this process involves fusing features from a feature layer with a downsampling rate of 1 / 32 (P5), a feature layer with a downsampling rate of 1 / 16 (P4), and a feature layer with a downsampling rate of 1 / 8 (P3) from the backbone network. In the bottom-up fusion path, when fusing to the 1 / 8 feature layer P3, which has a high resolution and is crucial for small object detection, a reparameterized large convolutional kernel module, RepLKNet, is specifically employed. The purpose of this design is to further expand the receptive field on the fused high-resolution feature map, making it not only rich in detailed information but also rich in global contextual information, providing higher-quality input for subsequent detection heads.

[0051] S3. Design a head network that employs a triple positive sample allocation strategy to optimize target prediction and classification.

[0052] The head network performs final object classification and bounding box regression based on the feature maps output by the neck network. To balance accuracy and speed, this invention designs a unique head network:

[0053] First, based on the feature maps of the P3 feature layer with a large receptive field obtained from the neck network, an efficient single-layer head structure was designed. This structure uses only the P3 feature layer as input for prediction, avoiding the computational overhead of prediction on multiple feature layers and ensuring detection speed.

[0054] Secondly, to compensate for the potential accuracy loss caused by single-layer heads and high-speed models, an innovative triple positive sample allocation strategy is introduced during the training phase. This strategy provides dense and high-quality supervision signals to the main detector by designing additional auxiliary branches that exist only during training. Specifically, this strategy includes:

[0055] 1. Deep Dense Allocation Branch: The classification and regression sub-networks of this branch each consist of four feature extraction convolutional layers and one prediction convolutional layer. It employs a dense allocation method, treating all regions within the target as positive samples. This design aims to provide rich supervision signals, effectively compensating for the gradient sparsity problem caused by the extremely small number of positive samples in deep sparse branches. This avoids convergence instability or getting stuck in local optima during the early stages of training, ensuring that the neck and backbone networks can quickly and stably capture the key appearance and geometric features of the target.

[0056] 2. Deep Sparse Allocation Branch: This branch adopts the same network structure as the deep dense allocation branch, but uses a strict one-to-one matching strategy when allocating positive samples. Specifically, this strategy constructs a prediction-aware cost matrix based on the classification confidence and regression localization accuracy of the prediction results, and uses the Hungarian Algorithm for global bipartite graph matching. This dynamically selects the predicted bounding boxes that achieve a globally optimal match with the real target in both category and location from among numerous prediction candidates as positive samples. Its core purpose is not to directly supervise the head network, but rather to act as a teacher branch, generating a set of high-quality, easy-to-learn positive sample targets for the shallow sparse branch.

[0057] 3. Shallow Sparse Allocation Branch: This branch features a minimalist, lightweight structure, with its classification and regression subnetworks each containing only one convolutional layer. As the student branch, it does not directly learn the original labels but instead uses the output of the deep sparse allocation branch as its supervision signal. This branch is the final detection head retained and used for prediction during the inference stage, and its concise structure is key to achieving high-speed inference.

[0058] Through triple supervision provided by these three dedicated branches, complex "teacher-student" knowledge distillation is achieved. The model learns sufficiently during training, effectively improving its ability to detect difficult samples. During inference, the two complex deep teacher branches are completely removed, retaining only the fully trained lightweight student branch for final prediction. This significantly improves inference speed without sacrificing accuracy, ensuring the model's real-time performance.

[0059] S4. Integrate the improved backbone, neck and head networks mentioned above to form a complete real-time target detection architecture.

[0060] The heavily parameterized large-kernel backbone network designed in step S1, the enhanced neck network designed in step S2, and the single-layer efficient head network designed in step S3 are integrated to form an end-to-end, complete real-time object detection model architecture. This architecture is designed to optimize the entire process from feature extraction to final prediction, aiming to achieve the best balance between speed and accuracy.

[0061] S5. Train the object detection network using the training image set, optimize the network parameters, and obtain the final model.

[0062] During model training, the dataset is first divided into training and testing sets. A variety of data augmentation techniques are employed to enhance the model's generalization ability, including but not limited to random translation, scaling, mosaic stitching, and color space transformation. The resolution of the input images is dynamically adjusted; for example, the shorter side of the image is adjusted to 800 pixels while ensuring the longer side does not exceed 1333 pixels. In terms of training strategies, Exponential Moving Average (EMA) is used to update the model weights, which helps stabilize the training process and improve the performance of the final model.

[0063] The loss function for network training consists of a weighted sum of classification loss and target localization loss.

[0064] 5.1 Classification Loss: Quality Focal Loss (QFL) is used, and its expression is as follows:

[0065]

[0066] in, It is the globally optimal solution. Acting as a moderating factor when the predicted quality estimate is inaccurate and deviates from the label. At this time, the adjustment factor is relatively large, so it will pay more attention to learning this difficult example. Parameters The weight reduction can be smoothly controlled in this embodiment, with parameters... Set to 2.

[0067] 5.2 Target localization loss: GIoU loss is selected, and its expression is as follows:

[0068]

[0069] in, A and B These are the prediction bounding box and the corresponding label bounding box, C It can simultaneously wrap the prediction box. A and annotation box B The minimum convex set bounding box, where IoU is the predicted box. A and annotation box B The intersection-union ratio between the two bounding boxes. Compared to the traditional IoU loss, the GIoU loss can provide an effective gradient signal even when the two bounding boxes do not intersect, which helps to optimize the localization process.

[0070] S6. Apply the trained model to the image to be detected, perform target prediction, and output the detection results.

[0071] During the inference or application phase, the trained and optimized model is deployed to the real-world scenario. For an image to be detected, preprocessing is first performed: the shorter side of the image is adjusted to a preset size, such as 800 pixels, while ensuring that the longer side does not exceed an upper limit, such as 1333 pixels, to match the input requirements of the model.

[0072] For large images (such as large-format remote sensing images) with resolutions far exceeding the predetermined threshold, a sliding window method is used for prediction. Specifically, the window is slid across the image at a fixed step size, such as 512 pixels, and each image patch within the window is detected independently. Finally, the detection results from all windows are aggregated and processed using non-maximum suppression (NMS) to obtain the final full-image detection result.

[0073] The proposed solution overcomes the limitations of existing methods by achieving a better balance between detection accuracy and inference speed through innovative network architecture design and training strategies. It enables high-efficiency real-time inference while maintaining high detection accuracy. Especially when processing challenging image data, it provides stable, reliable, and efficient detection performance, thus better serving the practical application needs of key fields such as national defense security, disaster monitoring, resource exploration, and environmental monitoring.

[0074] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0075] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A real-time target detection method based on reparameterized large convolutional kernels, characterized by the following steps: include: The image to be detected is input into a real-time target detection model, which includes a backbone network, a neck network, and a head network. By using a backbone network based on integrated reparameterized large convolutional kernels, the image to be detected is downsampled and features at different levels are extracted; The features extracted from different levels of the backbone network are input into the corresponding feature layers of the neck network based on the feature pyramid network for fusion to generate feature maps of the corresponding feature layers. By training the head network using a triple positive sample allocation strategy, and outputting feature maps of specified feature layers based on the neck network, target classification and bounding box regression results are generated. The head network is a single-layer head structure trained using a triple positive sample allocation strategy. It uses the feature map of 1 / 8 of the feature layer of the neck network as input to perform target classification and bounding box regression prediction. During the training phase, the head network provides dense and high-quality supervision signals to the main detector through auxiliary branches that exist only during training. Specifically, this includes: The deep-level densely distributed branches include a classification subnetwork and a regression subnetwork. Both the classification subnetwork and the regression subnetwork consist of four feature extraction convolutional layers and one prediction convolutional layer. The dense distribution method treats all regions inside the target as positive samples, generating supervision signals for shallow-level sparse branches. The deep sparse allocation branch includes a classification subnetwork and a regression subnetwork. Both the classification subnetwork and the regression subnetwork consist of four feature extraction convolutional layers and one prediction convolutional layer. A one-to-one matching strategy is used to allocate positive samples as the teacher branch, which generates a set of positive sample targets for the shallow sparse branch. The shallow sparse allocation branch includes a classification subnetwork and a regression subnetwork, each of which contains only one convolutional layer as a student branch, and supervised training is performed using the output of the deep sparse allocation branch as a supervision signal. During the inference phase, the shallow sparse allocation branch is retained as the detection head for the head network to perform predictions.

2. The real-time target detection method based on reparameterized large convolutional kernels according to claim 1, characterized in that, The backbone network includes four reparameterized large convolutional kernel modules. The backbone network downsamples the image to be detected sequentially and extracts features corresponding to the downsampling rate through the reparameterized large convolutional kernel modules. In the latter two reparameterized large convolutional kernel modules, depthwise separable convolution is used and padding is configured to maintain the spatial resolution of the feature map. At the end of the backbone network, a spatial pyramid pooling module is set. The spatial pyramid pooling module integrates contextual information of different scales by cascading or paralleling multiple pooling operations with different kernel sizes.

3. The real-time target detection method based on reparameterized large convolutional kernels according to claim 2, characterized in that, The reparameterized large convolutional kernel module has only one layer in the vertical direction. During training, it adopts a multi-branch structure, which has at least one large convolutional kernel branch set in parallel to capture global and contextual features. At least one small convolutional kernel branch is used to capture fine local structural information; And the identity mapping branch; During the training phase, the parameters of each branch in the multi-branch structure are learned and updated independently. During the inference phase, the parameters of all parallel branches are mathematically equivalently merged into the parameters of a single standard convolutional layer.

4. The real-time target detection method based on reparameterized large convolutional kernels according to claim 1, characterized in that, The neck network uses a feature pyramid network as its basic structure. It upsamples and convolves the output features at the ends of the backbone network from bottom to top, and then fuses the features from the corresponding downsampling rate level of the backbone network through lateral connections to generate the feature map of the corresponding feature layer.

5. The real-time target detection method based on reparameterized large convolutional kernels according to claim 1, characterized in that, The loss function of the real-time target detection model consists of a weighted sum of classification loss and target localization loss; the classification loss uses quality focus loss; and the target localization loss uses generalized intersection-union loss.

6. The real-time target detection method based on reparameterized large convolutional kernels according to claim 1, characterized in that, For large-sized images to be detected with resolutions far exceeding a predetermined threshold, a sliding window method is used for prediction. The window slides with a fixed step size for the image to be detected, and the image blocks within each window are detected independently using the real-time target detection model. The detection results of all windows are then aggregated and processed by non-maximum suppression to obtain the final full-image detection result.

7. A real-time target detection device based on reparameterized large convolutional kernels, comprising a memory, a processor, and a program stored in the memory, characterized in that, When the processor executes the program, it implements the real-time target detection method based on reparameterized large convolutional kernels as described in any one of claims 1-6.

Images