A YOLO26 target detection acceleration method and system based on reparameterization and dynamic activation

Abstract

The application relates to the field of artificial intelligence and computer vision technology, and provides a YOLO26 target detection acceleration method and system based on reparameterization and dynamic activation, which comprises the following steps: an improved YOLO26 initial network is constructed, a reparameterization module RepC3k2 and RepC2PSA and a lightweight spatial pooling module SPPF-Nano are introduced, and a Dynamic Mish-SiLU dynamic activation function is configured; a target training image is input into the initial network to perform iterative training; the multi-branch weights and biases of the training state network are extracted, structure reparameterization conversion is performed, and a deployment state YOLO26 network is obtained; a to-be-detected image is input into the deployment state network, and a target detection result is output under the bottom constraint of single volume kernel serial calculation and dynamic activation function adaptive weighting. The application completely eliminates branch jump redundancy in the reasoning stage through topological pruning and structure reparameterization, realizes high-speed and accurate real-time detection for resource-constrained edge terminals under the premise of guaranteeing high-precision feature representation.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and computer vision technology, specifically to a method and system for accelerating YOLO26 object detection based on reparameterization and dynamic activation. Background Technology

[0002] Real-time object detection, a key technology in artificial intelligence, plays a crucial supporting role in numerous scenarios such as security monitoring, autonomous driving, smart cities, and industrial automation. With the evolution of deep learning technology, one-stage object detection models, primarily the YOLO (YouOnlyLookOnce) series, have achieved significant breakthroughs in detection accuracy and inference speed. Among them, YOLO26, the latest evolution of this series, is an end-to-end single-stage real-time object detection neural network. Its core architecture essentially relies on a Residual Efficient Layer Aggregation Network to optimize the flow efficiency of deep gradients and deeply collaborates with an efficient region attention mechanism for spatial feature enhancement. This network aims to achieve efficient extraction and accurate capture of multi-scale object features with lower computational power and memory consumption through structured hierarchical aggregation, thus providing a crucial architectural foundation for the lightweight deployment of object detection algorithms on industrial edge devices.

[0003] However, existing YOLO26 and traditional real-time object detection technologies still have significant bottlenecks and shortcomings when deployed in practical applications on resource-constrained edge computing devices (such as embedded development boards and mobile terminals), mainly in the following two key dimensions:

[0004] First, the fundamental challenge of balancing inference speed and feature representation quality remains unresolved. In traditional YOLO models, the non-maximum suppression (NMS) latency in the post-processing stage is high, and there is computational redundancy in the network structure. Existing improvements often attempt to enhance feature representation quality by introducing complex modules such as MBConv or LGAG, but this significantly increases computational demands, directly sacrificing the crucial real-time requirements of edge devices. Furthermore, the static activation functions commonly used in these models (such as SiLU, Mish, or ReLU) have fixed mathematical forms and cannot be dynamically adjusted based on the differentiated needs of nonlinear feature extraction at different network levels and channels. For example, shallow networks often require stronger nonlinear expressions to capture local edges and textures, while deep networks require smooth gradients to preserve high-level semantics. Static activation functions limit the further exploration of the model's overall performance and fitting upper limit.

[0005] Second, there is a severe efficiency gap between the training and inference phases in the core structure of the model. Mainstream high-precision detection architectures typically employ a multi-branch parallel structure during training to maximize feature learning capabilities. However, in actual inference deployment, these parallel branches cannot be effectively merged, resulting in a significant waste of edge computing power on redundant branch jumps and multiple memory reads. Although the "Structural Re-parameterization" (RepVGG) technique, proposed in 2021 by a research team from Tsinghua University and Megvii Technology (Ding Xiaojun, Zhang Xiangyu, et al.) and published in the top computer vision conference CVPR, proposes a "multi-branch during training, single-branch during inference" design, its direct application to YOLO26's core components (such as the C3k2 feature extraction module and the C2PSA position awareness module) lacks targeted structural topology adaptation. Simply applying the re-parameterization module not only fails to effectively retain YOLO26's original unique receptive field and multi-scale feature extraction capabilities but also fails to achieve ultimate inference acceleration in complex backgrounds and small object detection.

[0006] Meanwhile, in existing technologies, standard convolutional layers are often sensitive to isotropic features, while real-world objects such as ship hulls and people have obvious directional elongated features, making it difficult to adapt to training on real-world objects. Summary of the Invention

[0007] This invention addresses the technical problems in existing technologies, such as the difficulty in balancing speed and accuracy in target detection on edge devices, redundant inference in multi-branch structures, rigid representation of static activation functions, and excessive computational load of multi-scale spatial pooling, which is difficult to adapt to training on real objects. It provides a YOLO26 target detection acceleration method and system based on reparameterization and dynamic activation, which achieves high-speed, accurate, and real-time detection of complex scenes and small-sized targets at the edge while ensuring the model's receptive field and feature richness, thus solving the aforementioned problems.

[0008] The technical solution of this invention is as follows:

[0009] This invention provides a method for accelerating YOLO26 target detection based on reparameterization and dynamic activation, comprising the following steps:

[0010] S1: Initial Network Construction: Construct an improved YOLO26 initial network; wherein, the improved YOLO26 initial network is based on the YOLO26 architecture, introducing a reparameterized feature extraction module RepC3k2 to replace the C3k2 module of RepVGG, a reparameterized position awareness module RepC2PSA to replace the C2PSA module of RepVGG, a lightweight spatial pooling module SPPF-Nano to replace the SPPF module in the backbone network, and replacing the static activation function with the dynamic Mish-SiLU deep learning architecture; the lightweight spatial pooling module SPPF-Nano reduces the number of serial pooling operations of the original SPPF from 3 to 2, and reduces the number of output concatenation paths from 4 to 3;

[0011] S2: Iterative Model Training: Obtain a target training image with labeled information, input the target training image into the improved YOLO26 initial network to perform iterative training, and obtain a training-state YOLO26 network; wherein, during the iterative training process, the main processing routes of the reparameterized feature extraction module RepC3k2 and the reparameterized position awareness module RepC2PSA are both configured as multi-branch parallel structures, and asymmetric convolutional layers are connected in series in the main processing route of the reparameterized feature extraction module RepC3k2;

[0012] S3: Reparameterization Transformation: Extract the weight matrix and bias parameters of the multi-branch structure in the training-state YOLO26 network, and perform a structural reparameterization transformation on the weight matrix and bias parameters to obtain the deployment-state YOLO26 network; wherein, the structural reparameterization transformation is a linear equivalent transformation operation that merges the parameters of the multi-branch parallel structure into a single 3×3 convolutional kernel; at the end node of the main feature enhancement path of the RepC2PSA module in the reparameterized position awareness module, the PSA attention mechanism module is saved;

[0013] S4: Edge-end inference acceleration: Acquire the image to be detected from the edge device and input the image to be detected into the deployed YOLO26 network; under the constraints of the underlying structure of the deployed YOLO26 network with serial computation of a single 3×3 convolution kernel and adaptive weighting of the DynamicMish-SiLU activation function, perform object detection inference operation, output the object detection result of the image to be detected, and complete the object detection acceleration for the edge.

[0014] Preferably, in S1, the lightweight spatial pooling module SPPF-Nano is a lightweight neck network component obtained by performing structural pruning based on the standard SPPF module.

[0015] Preferably, in S1, the lightweight spatial pooling module SPPF-Nano is a topology based on two serial pooling operations and three-way feature concatenation output.

[0016] Under the underlying structural constraints of the two serial pooling operations and the three-way feature concatenation output, the computational load of the network is reduced while preserving the target receptive field.

[0017] Preferably, in S1, the DynamicMish-SiLU activation function is a weighted combination function that incorporates learnable weight parameters;

[0018] The calculation formula for the DynamicMish-SiLU activation function is as follows:

[0019] ,

[0020] In the formula, x represents the feature value of the input feature map, and y represents the activation output value. The learnable weight parameters are... It is a value [0:1] restricted to between 0 and 1 by the Sigmoid mapping;

[0021] Preferably, under the constraints of the iterative training process in S2, the learnable weight parameters are adaptively learned based on gradients obtained through network backpropagation. The value of is used to dynamically adjust the weight ratio of the Mish activation function and the SiLU activation function.

[0022] Preferably, in S2, the multi-branch parallel structure is composed of 3×3 convolution branches, 1×1 convolution branches, and identity mapping branches in parallel;

[0023] Under the constraint of the linear equivalent transformation operation in S3, the weight matrices and bias parameters of the 3×3 convolution branch, the 1×1 convolution branch and the identity mapping branch are linearly superimposed using the additivity of the linear system, and pre-merged and reconstructed into a single 3×3 convolution kernel without branch jumps in the deployed YOLO26 network.

[0024] Preferably, in S4, the edge device includes a resource-constrained embedded development board and a mobile terminal;

[0025] Under the constraints of the serial structure that eliminates multiple memory reads and branch jumps by the single 3×3 convolutional kernel, and the underlying structural constraints that the DynamicMish-SiLU activation function adaptively balances nonlinear expression and computational efficiency, the target detection inference operation is executed to achieve real-time detection of complex scenes and small-sized targets.

[0026] Preferably, in step S2, before inputting the target training image into the improved YOLO26 initial network, a training image scaling step is further included;

[0027] The training image scale normalization step involves performing pixel resampling and dimension alignment preprocessing operations on the dataset images at a uniform resolution scale.

[0028] Under the dimensional alignment constraint of the training image scale normalization step, the obtained training validation set and test set images are uniformly adjusted to a pixel size of 960×960, and then input into the improved YOLO26 initial network to perform the iterative training. This eliminates the interference of different input resolutions on the local receptive field and ensures the scale space consistency of the trained YOLO26 network when facing complex edge scenes and small-sized target detection.

[0029] Preferably, in step S3, after obtaining the deployed YOLO26 network and before performing the target detection inference operation in step S4, a model quantization deployment step is further included;

[0030] The model quantization deployment step is based on the single 3×3 convolution kernel and bias parameters reconstructed by the linear equivalent transformation operation, and converts the floating-point operation precision of the network weights into a low-bit integer precision numerical compression operation.

[0031] Under the synergistic effect of the numerical compression constraint of low bit integer precision and the parameter fusion constraint of the structural reparameterization transformation, the quantized deployment state YOLO26 network is written to the edge device, further reducing the edge memory read / write bandwidth and power consumption load, and improving the edge inference frame rate in high real-time target detection scenarios.

[0032] This invention also provides a YOLO26 target detection acceleration system based on reparameterization and dynamic activation, comprising:

[0033] An initial network building module is used to build an improved YOLO26 initial network. The improved YOLO26 initial network is based on the YOLO26 architecture and introduces a reparameterized feature extraction module RepC3k2, a reparameterized position awareness module RepC2PSA, a lightweight spatial pooling module SPPF-Nano, and replaces the static activation function with the DynamicMish-SiLU deep learning architecture.

[0034] A network training and optimization module, connected to the initial network construction module, is used to acquire target training images with labeled information, input the target training images into the improved YOLO26 initial network to perform iterative training, and obtain a training-state YOLO26 network; wherein, during the iterative training process, the reparameterized feature extraction module RepC3k2 and the reparameterized position awareness module RepC2PSA are both configured as multi-branch parallel structures, and asymmetric convolutional layers are connected in series in the reparameterized feature extraction module RepC3k2;

[0035] The structure reparameterization transformation module, connected to the network training and optimization module, is used to extract the weight matrix and bias parameters of the multi-branch structure in the training state YOLO26 network, and perform structure reparameterization transformation on the weight matrix and bias parameters to obtain the deployment state YOLO26 network; wherein, the structure reparameterization transformation is a linear equivalent transformation operation that merges the parameters of the multi-branch parallel structure into a single 3×3 convolution kernel;

[0036] The edge-end inference acceleration module, connected to the structure reparameterization conversion module, is used to acquire the image to be detected from the edge device and input the image to be detected into the deployed YOLO26 network. Under the constraints of the underlying structure of the deployed YOLO26 network, which uses a single 3×3 convolutional kernel for serial computation and the DynamicMish-SiLU dynamic activation function for adaptive weighting, the module performs object detection inference operations and outputs the object detection result of the image to be detected, thus completing the object detection acceleration for the edge.

[0037] The beneficial effects of this invention are as follows:

[0038] This invention provides a YOLO26 target detection acceleration method and system based on reparameterization and dynamic activation. In the iterative training phases S1 to S2, the initial network introduces multi-branch parallel structures "RepC3k2" and "RepC2PSA" modules, replacing the original non-parallel C3k2 and C2PSA modules. This allows the high-capacity network topology of the multi-branch parallel structure to greatly expand the fitting upper limit of local details, channel mixing, and semantic extraction, significantly increasing the diversity of the receptive field. This allows the network to learn local details, channel mixing, and original features simultaneously, significantly improving the model's fitting upper limit and ensuring that the model has excellent feature representation and high-precision recognition capabilities in complex backgrounds and small-sized targets. In the inference and deployment phases S3 to S4, the massive multi-branch parameters are linearly equivalently merged into a single "3×3" convolutional kernel through structural reparameterization, effectively "collapsed" into a compact and efficient serial structure. This approach retains the high-precision representation of complex networks while achieving lightweight operation at the inference end. It solves the problems of frame rate drops caused by blindly adding complex modules for refinement, or missed detections of small targets and complex scenes due to rough pruning for speed. The cascaded asymmetric convolutional layers in the RepC3k2 reparameterized feature extraction module are designed for real-world objects that are not as sensitive to isotropic features as standard symmetric convolutional layers. Real-world objects such as people and ships have distinct directional elongated features. Therefore, cascading asymmetric convolutional layers allows for the use of parameters in asymmetric directions to specifically capture texture features and edge information in those directions, improving the model's robustness to shape-sensitive targets. The dynamic activation function introduced in S1 addresses the rigidity of static activation function representations. Specifically, in S3 and S4, S3 performs a linear equivalent transformation on the weight matrix and bias parameters, while S4 enforces inference execution under serial computation with a single 3×3 convolutional kernel, achieving zero-branch jumps and smooth continuous memory access. This significantly reduces edge memory read / write bandwidth consumption and system latency, improving the real-time inference frame rate of the model on resource-constrained embedded development boards and mobile terminals. It also completely eliminates branch jumps and multiple memory read operations in the inference graph, transforming the originally complex parallel computation into highly efficient serial convolution operations. This not only significantly reduces computational latency but also optimizes memory access continuity, saving computational resources, reducing inference speed, and improving accuracy. At the final node of the RepC2PSA module's main feature enhancement path, a PSA attention mechanism module is stored. This terminal PSA attention mechanism module, by introducing positional encoding and self-attention operations, can adaptively suppress background noise generated by the preceding modules, accurately locking target regions with specific positional sensitivity. Simultaneously, thanks to the efficient feature preprocessing of the preceding RepVGG module, the overall load on the model is reduced, achieving a balance between computational efficiency and detection accuracy at the edge.

[0039] Meanwhile, the lightweight spatial pooling module "SPPF-Nano" introduced in S1 directly reduces the computational load of the network through structured pruning, while strictly preserving the multi-scale receptive field coverage of the core target. This effectively solves the pain point of excessive computational load in multi-scale spatial pooling and further reduces the memory usage pressure of the edge system.

[0040] Under the underlying structural constraints of S4 object detection inference operations, the model can adaptively weight the layers based on the different characteristics of shallow networks (strong nonlinear representations to capture local texture edges) and deep networks (smooth gradients to preserve high-level semantics). Closed-loop regulation endows neurons with the ability to dynamically perceive feature differences, accelerating network convergence and significantly raising the upper limit of overall accuracy without significantly increasing the inference load.

[0041] In some embodiments of the present invention, the dual-parallel topology of "RepC3k2" and "RepC2PSA" ensures lossless gradient backpropagation in deep networks and improves the convergence stability of difficult-to-detect small targets. Among them, "RepC3k2" significantly enhances the robustness of recognizing long strip textures and vertical edges with extremely low computational cost by cascading low-parameter asymmetric convolutional layers; "RepC2PSA" utilizes the high-frequency information source of the front-end "RepVGG" in conjunction with the end-end "PSA" mechanism to accurately locate key regions and effectively suppress background noise.

[0042] The "SPPF-Nano" module reduces the computational load by approximately one-third by reducing serial pooling to two iterations and simplifying output concatenation to three paths, while maintaining strict lossless multi-scale receptive field coverage. This significantly reduces system memory usage and computational pressure. The Dynamic Mish-SiLU activation function utilizes learnable parameters... A closed loop is established with the Sigmoid mapping, enabling the network to adaptively update weight ratios based on the true gradient. This mechanism effectively prevents the activation function from diverging or becoming distorted during iteration and ensures the smoothness of differentiation with rigorous mathematical constraints. Before training, scale normalization at a uniform 960×960 resolution eliminates scale noise interference caused by dataset inconsistencies, ensuring high scale-space adaptability during model deployment. During deployment, the structural advantages of the "branchless topology" combined with low-bit quantization further compress the model size, maximizing power consumption and memory read / write bandwidth savings for edge devices, achieving ultimate hardware acceleration in high frame rate scenarios. Attached Figure Description

[0043] Figure 1 This is a flowchart illustrating a YOLO26 target detection acceleration method based on reparameterization and dynamic activation in an embodiment of the present invention.

[0044] Figure 2 This is a schematic diagram of the improved YOLO26 overall network structure in an embodiment of the present invention.

[0045] Figure 3 This is a schematic diagram of the design architecture of the basic RepVGG module in an embodiment of the present invention.

[0046] Figure 4 This is a schematic diagram comparing the structural equivalence transformations of the reparameterized feature extraction module (RepC3k2) in the training and inference phases in an embodiment of the present invention.

[0047] Figure 5 This is a schematic diagram comparing the structural equivalence transformations of the reparameterized position awareness module (RepC2PSA) during the training and inference phases in an embodiment of the present invention.

[0048] Figure 6 This is a schematic diagram of the network topology of the lightweight space pooling module (SPPF-Nano) in an embodiment of the present invention.

[0049] Figure 7 This is a schematic diagram illustrating the model training and experimental verification process of a YOLO26 target detection acceleration method based on reparameterization and dynamic activation, provided in an embodiment of the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the following detailed description, in conjunction with the accompanying drawings and specific embodiments, provides a further detailed explanation of the YOLO26 target detection acceleration method and system based on reparameterization and dynamic activation proposed in this invention. The advantages and features of this invention will become clearer from the following description. It should be noted that the accompanying drawings are in a very simplified form and use non-precise proportions, used only to facilitate and clearly illustrate the purpose of the embodiments of this invention. Please refer to the accompanying drawings to make the objectives, features, and advantages of this invention more apparent and understandable. It should be understood that the structures, proportions, sizes, etc., depicted in the accompanying drawings are only used to complement the content disclosed in the specification, for those skilled in the art to understand and read, and are not intended to limit the implementation conditions of this invention. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportional relationships, or adjustments to the size, without affecting the effects and objectives achieved by this invention, should still fall within the scope of the technical content disclosed in this invention.

[0051] Figure 1 This diagram illustrates a flowchart of an accelerated YOLO26 object detection method based on reparameterization and dynamic activation, provided by an embodiment of the present invention. As shown in the diagram, this acceleration method mainly consists of four stages at the system architecture level: initial network construction, iterative model training, reparameterization transformation, and edge inference acceleration. Specifically, it includes the following steps:

[0052] S1: Initial Network Construction: Construct an improved YOLO26 initial network; wherein, the improved YOLO26 initial network is based on the YOLO26 architecture, introducing a reparameterized feature extraction module RepC3k2, a reparameterized position-aware module RepC2PSA, and a lightweight spatial pooling module SPPF-Nano, and replacing the static activation function with a dynamic Mish-SiLU deep learning architecture; In some embodiments of the present invention, the macroscopic topology of the improved YOLO26 initial network is as follows Figure 2 As shown. Compared with the traditional YOLO26 architecture, this invention has... Figure 2 The backbone and neck networks shown have undergone structural modifications at their feature interaction nodes, with RepC3k2, RepC2PSA, and SPPF-Nano modules explicitly introduced, and the DynamicMish-SiLU activation function fully deployed in the underlying operators.

[0053] In some embodiments of the present invention, the introduced reparameterized feature extraction module RepC3k2 is the C3k2 module that replaces RepVGG, the reparameterized position awareness module RepC2PSA is the C2PSA module that replaces RepVGG, and the lightweight spatial pooling module SPPF-Nano is the SPPF module that replaces the backbone network.

[0054] In some embodiments of the present invention, both the reparameterized feature extraction module RepC3k2 and the reparameterized position awareness module RepC2PSA include a dual-path parallel topology; wherein, the dual-path parallel topology is a feature extraction architecture composed of a main processing path and a bypass direct path;

[0055] In some embodiments of the present invention, the lightweight spatial pooling module SPPF-Nano is a lightweight neck network component obtained by performing structural pruning based on a standard SPPF module. For some embodiments of the present invention, please refer to... Figure 6 ,picture 6 The network topology of the lightweight spatial pooling module SPPF-Nano is shown. Combined with... Figure 6 It can be seen that the lightweight spatial pooling module SPPF-Nano is based on a topology structure with two serial pooling operations and three-way feature concatenation output. Under the underlying structure constraints of the two serial pooling operations and three-way feature concatenation output (i.e., the number of serial pooling operations of the original SPPF is reduced to two, and the number of concatenation outputs is reduced from four to three), the computational load of the network is reduced while preserving the target receptive field.

[0056] In some embodiments of the present invention, the calculation formula for the DynamicMish-SiLU activation function is as follows:

[0057] In the formula, x represents the feature value of the input feature map, and y represents the activation output value. The learnable weight parameters are... It is a value [0:1] restricted to between 0 and 1 by the Sigmoid mapping;

[0058] The innovation in activation functions is based on improvements to the shortcomings of traditional YOLO series activation functions. Traditional YOLO series activation functions often employ static or fixed combinations of activation functions. However, in computer vision networks, different network layers and channels have different requirements for nonlinear feature extraction. Shallow networks may require stronger nonlinearity to capture edges, while deep networks may require smoother gradients to preserve semantics. Therefore, this invention proposes an innovative dynamic activation function, DynamicMish-SiLU, to meet the different needs of nonlinear feature extraction in network structures. Its core innovation lies in the use of learnable weight parameters. This parameter is used to weight and combine the high-precision but computationally complex Mish activation function with the high-efficiency but inaccurate SiLU activation function.

[0059] During training, the network structure will automatically learn based on gradients. The value of is used to dynamically adjust the weight ratio of the Mish and SiLU activation functions:

[0060] when DynamicMish-SiLU behaves like the Mish activation function, aiming for high accuracy.

[0061] when DynamicMish-SiLU behaves like the SiLU activation function, aiming for high efficiency.

[0062] The formula for DynamicMish-SiLU is as follows:

[0063]

[0064]

[0065] Among them, learnable parameters Using a Sigmoid map to learnable parameters The value of is restricted to between 0 and 1, thus ensuring the stability of the model. The formulas for the Mish and SiLU functions in this invention are as follows:

[0066]

[0067] .

[0068] S2: Iterative Model Training: Obtain target training images with labeled information, input the target training images into the improved YOLO26 initial network to perform iterative training, and obtain the training-state YOLO26 network; wherein, during the iterative training process, combined with Figure 4 and Figure 5 As can be seen from the structural equivalence transformation comparison diagram in the upper part (training phase), the main processing routes of the reparameterized feature extraction module RepC3k2 and the reparameterized position awareness module RepC2PSA are both configured as multi-branch parallel structures, and asymmetric convolutional layers are connected in series in the main processing route of the reparameterized feature extraction module RepC3k2.

[0069] In some embodiments of the present invention, before inputting the target training image into the improved YOLO26 initial network, a training image scale normalization step is further included; wherein, the training image scale normalization step performs pixel resampling and dimension alignment preprocessing operations on the dataset images at a uniform resolution scale; under the dimension alignment constraint, the obtained training validation set and test set images are uniformly adjusted to a pixel size of 960×960 to eliminate the interference of different input resolutions on the local receptive field and ensure the scale space consistency of the trained YOLO26 network when facing complex edge scenes;

[0070] In some embodiments of the present invention, please refer to Figure 4 The basic RepVGG module design architecture shown here employs a RepVGG multi-branch architecture for the main processing route to perform mixed extraction of local details and channels. The multi-branch parallel structure consists of parallel 3×3 convolutional branches, 1×1 convolutional branches, and identity mapping branches.

[0071] In some embodiments of the present invention, an asymmetric convolutional layer with one-third the number of parameters of a standard convolution is cascaded after the RepVGG multi-branch architecture, specifically designed to capture vertical texture features and edge information for objects (such as ship hulls and people) with obvious directional elongated features in reality.

[0072] In some embodiments of the present invention, under the constraints of the iterative training process of S2, the learnable weight parameters are adaptively learned based on the true gradients of network backpropagation. The value of is dynamically adjusted to change the weight ratio of the Mish activation function and the SiLU activation function, preventing the activation functions from diverging or becoming distorted during iteration.

[0073] S3: Reparameterization Transformation: Extract the weight matrix and bias parameters of the multi-branch structure in the training-state YOLO26 network, and perform a structural reparameterization transformation on the weight matrix and bias parameters to obtain the deployment-state YOLO26 network; wherein, the structural reparameterization transformation is a linear equivalent transformation operation that merges the parameters of the multi-branch parallel structure into a single 3×3 convolutional kernel; at the end node of the main feature enhancement path of the RepC2PSA module in the reparameterized position awareness module, the PSA attention mechanism module is saved;

[0074] In some embodiments of the present invention, combined with Figure 4 and Figure 5 As can be seen from the structural diagram in the lower half, under the constraint of the linear equivalent transformation operation of S3, the weight matrices and bias parameters of the 3×3 convolution branch, the 1×1 convolution branch and the identity mapping branch are linearly superimposed using the additivity of the linear system, and pre-merged and reconstructed into a single 3×3 convolution kernel without branch jump in the deployed YOLO26 network.

[0075] In some embodiments of the present invention, under the constraints of the underlying structure output by splicing the main feature enhancement route and the bypass information direct route, the position-sensitive attention PSA mechanism of the end node is used in combination with the pre-processing features to adaptively suppress background noise and accurately lock the position-sensitive region.

[0076] In some embodiments of the present invention, after obtaining the deployed YOLO26 network and before performing the inference operation in S4, a model quantization deployment step is further included; wherein, the model quantization deployment step is based on the single 3×3 convolution kernel and bias parameters reconstructed by the linear equivalent transformation operation, and converts the floating-point operation precision of the network weights into a low-bit integer precision numerical compression operation.

[0077] S4: Edge-end inference acceleration: Acquire the image to be detected from the edge device and input the image to be detected into the deployed YOLO26 network; under the constraints of the underlying structure of the deployed YOLO26 network with serial computation of a single 3×3 convolution kernel and adaptive weighting of the DynamicMish-SiLU activation function, perform object detection inference operation, output the object detection result of the image to be detected, and complete the object detection acceleration for the edge.

[0078] In some embodiments of the present invention, the edge device includes a resource-constrained embedded development board and a mobile terminal;

[0079] In some embodiments of the present invention, under the synergistic effect of the numerical compression constraint of low bit integer precision and the parameter fusion constraint of the structural reparameterization transformation, the quantized deployment state YOLO26 network is written to the edge device, further reducing the edge memory read / write bandwidth and power consumption load, and improving the edge inference frame rate in high real-time target detection scenarios.

[0080] In some embodiments of the present invention, under the constraints of the serial structure that completely eliminates multiple memory reads and branch jumps by the single 3×3 convolutional kernel, and under the constraints of the underlying structure that the DynamicMish-SiLU activation function adaptively balances nonlinear expression and computational efficiency according to the different characteristics of shallow and deep networks, the target detection inference operation is performed to achieve real-time, high-speed and accurate detection of complex scenes and small targets at the edge.

[0081] This invention provides a YOLO26 target detection acceleration system based on reparameterization and dynamic activation, the overall execution flow of which is as follows: Figure 1 As shown, it includes:

[0082] An initial network building module is used to build an improved YOLO26 initial network. The improved YOLO26 initial network is based on the YOLO26 architecture and introduces a reparameterized feature extraction module RepC3k2, a reparameterized position awareness module RepC2PSA, a lightweight spatial pooling module SPPF-Nano, and replaces the static activation function with the DynamicMish-SiLU deep learning architecture.

[0083] A network training and optimization module, connected to the initial network construction module, is used to acquire target training images with labeled information, input the target training images into the improved YOLO26 initial network to perform iterative training, and obtain a training-state YOLO26 network; wherein, during the iterative training process, the reparameterized feature extraction module RepC3k2 and the reparameterized position awareness module RepC2PSA are both configured as multi-branch parallel structures, and asymmetric convolutional layers are connected in series in the reparameterized feature extraction module RepC3k2;

[0084] The structure reparameterization transformation module, connected to the network training and optimization module, is used to extract the weight matrix and bias parameters of the multi-branch structure in the training state YOLO26 network, and perform structure reparameterization transformation on the weight matrix and bias parameters to obtain the deployment state YOLO26 network; wherein, the structure reparameterization transformation is a linear equivalent transformation operation that merges the parameters of the multi-branch parallel structure into a single 3×3 convolution kernel;

[0085] The edge-end inference acceleration module, connected to the structure reparameterization conversion module, is used to acquire the image to be detected from the edge device and input the image to be detected into the deployed YOLO26 network. Under the constraints of the underlying structure of the deployed YOLO26 network, which uses a single 3×3 convolutional kernel for serial computation and the DynamicMish-SiLU dynamic activation function for adaptive weighting, the module performs object detection inference operations and outputs the object detection result of the image to be detected, thus completing the object detection acceleration for the edge.

[0086] In some embodiments of the present invention Figure 7 This invention provides a model training and experimental verification process for an accelerated YOLO26 target detection method based on reparameterization and dynamic activation, as described in an embodiment of the present invention.

[0087] In some embodiments of the present invention, the self-built dataset is first divided into a training / validation set and a test set at a 9:1 ratio. Secondly, all images in both the training / validation set and the test set are uniformly adjusted to a size of 960×960 to ensure fairness in the testing. Subsequently, the training set is fed into the improved YOLO26 algorithm for training. The improvements include: replacing the original C3k2 module with a RepC3k2 module, replacing the original C2PSA module with a RepC2PSA module, replacing the SPPF module in the backbone network with an SPPF-Nano module, and replacing the original activation function with a DynamicMish-SiLU activation function. Finally, the algorithm is experimentally analyzed and its performance evaluated using the test set. The evaluation results show that, compared with existing technologies, the improved YOLO26 object detection scheme based on reparameterization and dynamic adaptation provided by the present invention has the following significant beneficial effects:

[0088] (1) Significantly improves reasoning speed and achieves efficient edge acceleration:

[0089] This invention innovatively designs a reparameterized core module, replacing the C3k2 and C2PSA modules with REPC3k2 and RepC2PSA modules, transforming the original multi-branch structure into an architecture of "multi-branch enhancement during training and single-branch fusion during inference." Through structural reparameterization technology, multiple convolutional branches are seamlessly fused into a single convolutional operation during the inference phase, effectively eliminating the overhead of multi-branch computation. Experimental results show that the overall model's inference speed is significantly improved, and the acceleration effect is even more pronounced after combining quantization technology, greatly meeting the needs of high real-time detection scenarios.

[0090] (2) While ensuring high speed, the detection accuracy is effectively improved:

[0091] By introducing a dynamic activation mechanism and employing the Dynamic Mish-SiLU activation function, the model can automatically learn and adjust the form of the activation function based on the characteristics of different parts of the network. This enhances the model's nonlinear expressiveness and feature adaptability, accelerates model convergence, and optimizes the final accuracy. Real-world data shows that while inference speed is significantly improved, the model's overall detection accuracy is also markedly enhanced, achieving a breakthrough in both speed and accuracy.

[0092] (3) Significantly reduce the number of model parameters and computational load:

[0093] The optimized spatial pooling module SPPF-Nano proposed in this invention effectively reduces the computational load of related modules by simplifying computational branches while maintaining receptive field coverage. Combined with the parameter fusion effect brought about by the reparameterization module, the number of parameters in the overall model is significantly reduced, which not only reduces the memory footprint and power consumption of edge devices, but also further reduces deployment costs.

[0094] (4) Enhanced adaptability to small targets and complex scenarios:

[0095] The improved module cleverly combines reparameterized convolution with an attention mechanism, preserving the attention mechanism's ability to focus on target regions while leveraging reparameterization to improve computational efficiency. This makes the improved model more robust in small target detection and complex backgrounds, effectively overcoming the common problems of missed or false detections of small targets in traditional acceleration schemes.

[0096] Finally, it should be noted that the above embodiments are only used to help further understand the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that various substitutions and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the content disclosed in the embodiments.

Claims

1. A method for accelerating target detection using YOLO26 based on reparameterization and dynamic activation, characterized in that, Includes the following steps: S1: Initial Network Construction: Construct an improved YOLO26 initial network; wherein, the improved YOLO26 initial network is based on the YOLO26 architecture, introducing a reparameterized feature extraction module RepC3k2 to replace the C3k2 module of RepVGG, a reparameterized position awareness module RepC2PSA to replace the C2PSA module of RepVGG, a lightweight spatial pooling module SPPF-Nano to replace the SPPF module in the backbone network, and replacing the static activation function with the dynamic Mish-SiLU deep learning architecture; the lightweight spatial pooling module SPPF-Nano reduces the number of serial pooling operations of the original SPPF from 3 to 2, and reduces the number of output concatenation paths from 4 to 3; S2: Iterative Model Training: Obtain a target training image with labeled information, input the target training image into the improved YOLO26 initial network to perform iterative training, and obtain a training-state YOLO26 network; wherein, during the iterative training process, the main processing routes of the reparameterized feature extraction module RepC3k2 and the reparameterized position awareness module RepC2PSA are both configured as multi-branch parallel structures, and asymmetric convolutional layers are connected in series in the main processing route of the reparameterized feature extraction module RepC3k2; S3: Reparameterization Transformation: Extract the weight matrix and bias parameters of the multi-branch structure in the training-state YOLO26 network, and perform a structural reparameterization transformation on the weight matrix and bias parameters to obtain the deployment-state YOLO26 network; wherein, the structural reparameterization transformation is a linear equivalent transformation operation that merges the parameters of the multi-branch parallel structure into a single 3×3 convolutional kernel; at the end node of the main feature enhancement path of the RepC2PSA module in the reparameterized position awareness module, the PSA attention mechanism module is saved; S4: Edge-end inference acceleration: Acquire the image to be detected from the edge device and input the image to be detected into the deployed YOLO26 network; under the constraints of the underlying structure of the deployed YOLO26 network with serial computation of a single 3×3 convolution kernel and adaptive weighting of the DynamicMish-SiLU activation function, perform object detection inference operation, output the object detection result of the image to be detected, and complete the object detection acceleration for the edge.

2. The method for accelerating YOLO26 target detection based on reparameterization and dynamic activation according to claim 1, characterized in that, In S1, the lightweight spatial pooling module SPPF-Nano is a lightweight neck network component obtained by performing structural pruning based on the standard SPPF module.

3. The YOLO26 target detection acceleration method based on reparameterization and dynamic activation according to claim 2, characterized in that, In S1, the lightweight spatial pooling module SPPF-Nano is based on a topology structure with two serial pooling operations and three-way feature concatenation output. Under the underlying structural constraints of the two serial pooling operations and the three-way feature concatenation output, the computational load of the network is reduced while preserving the target receptive field.

4. The YOLO26 target detection acceleration method based on reparameterization and dynamic activation according to claim 3, characterized in that, In S1, the DynamicMish-SiLU activation function is a weighted combination function that incorporates learnable weight parameters; The calculation formula for the DynamicMish-SiLU activation function is as follows: ， In the formula, x represents the feature value of the input feature map, and y represents the activation output value. The learnable weight parameters are... It is a value [0:1] that is restricted to between 0 and 1 by the Sigmoid mapping.

5. The YOLO26 target detection acceleration method based on reparameterization and dynamic activation according to claim 3, characterized in that, Under the constraints of the iterative training process of S2, the learnable weight parameters are adaptively learned based on gradients through network backpropagation. The value of is used to dynamically adjust the weight ratio of the Mish activation function and the SiLU activation function.

6. The YOLO26 target detection acceleration method based on reparameterization and dynamic activation according to claim 1, characterized in that, In S2, the multi-branch parallel structure is composed of 3×3 convolution branches, 1×1 convolution branches, and identity mapping branches in parallel. Under the constraint of the linear equivalent transformation operation in S3, the weight matrices and bias parameters of the 3×3 convolution branch, the 1×1 convolution branch and the identity mapping branch are linearly superimposed using the additivity of the linear system, and pre-merged and reconstructed into a single 3×3 convolution kernel without branch jumps in the deployed YOLO26 network.

7. The method for accelerating YOLO26 target detection based on reparameterization and dynamic activation according to claim 1, characterized in that, In S4, the edge device includes a resource-constrained embedded development board and a mobile terminal; Under the constraints of the serial structure that eliminates multiple memory reads and branch jumps by the single 3×3 convolutional kernel, and the underlying structural constraints that the DynamicMish-SiLU activation function adaptively balances nonlinear expression and computational efficiency, the target detection inference operation is executed to achieve real-time detection of complex scenes and small-sized targets.

8. The YOLO26 target detection acceleration method based on reparameterization and dynamic activation according to claim 1, characterized in that, In S2, before inputting the target training image into the improved YOLO26 initial network, a training image scale normalization step is also included; The training image scale normalization step involves performing pixel resampling and dimension alignment preprocessing operations on the dataset images at a uniform resolution scale. Under the dimensional alignment constraint of the training image scale normalization step, the obtained training validation set and test set images are uniformly adjusted to a pixel size of 960×960, and then input into the improved YOLO26 initial network to perform the iterative training. This eliminates the interference of different input resolutions on the local receptive field and ensures the scale space consistency of the trained YOLO26 network when facing complex edge scenes and small-sized target detection.

9. The YOLO26 target detection acceleration method based on reparameterization and dynamic activation according to claim 1, characterized in that, In step S3, after obtaining the deployed YOLO26 network and before performing the target detection inference operation in step S4, a model quantization deployment step is also included. The model quantization deployment step is based on the single 3×3 convolution kernel and bias parameters reconstructed by the linear equivalent transformation operation, and converts the floating-point operation precision of the network weights into a low-bit integer precision numerical compression operation. Under the synergistic effect of the numerical compression constraint of low bit integer precision and the parameter fusion constraint of the structural reparameterization transformation, the quantized deployment state YOLO26 network is written to the edge device, further reducing the edge memory read / write bandwidth and power consumption load, and improving the edge inference frame rate in high real-time target detection scenarios.

10. A YOLO26 target detection acceleration system based on reparameterization and dynamic activation, characterized in that, include: An initial network building module is used to build an improved YOLO26 initial network. The improved YOLO26 initial network is based on the YOLO26 architecture and introduces a reparameterized feature extraction module RepC3k2, a reparameterized position awareness module RepC2PSA, a lightweight spatial pooling module SPPF-Nano, and replaces the static activation function with the DynamicMish-SiLU deep learning architecture. A network training and optimization module, connected to the initial network construction module, is used to acquire target training images with labeled information, input the target training images into the improved YOLO26 initial network to perform iterative training, and obtain a training-state YOLO26 network; wherein, during the iterative training process, the reparameterized feature extraction module RepC3k2 and the reparameterized position awareness module RepC2PSA are both configured as multi-branch parallel structures, and asymmetric convolutional layers are connected in series in the reparameterized feature extraction module RepC3k2; The structure reparameterization transformation module, connected to the network training and optimization module, is used to extract the weight matrix and bias parameters of the multi-branch structure in the training state YOLO26 network, and perform structure reparameterization transformation on the weight matrix and bias parameters to obtain the deployment state YOLO26 network; wherein, the structure reparameterization transformation is a linear equivalent transformation operation that merges the parameters of the multi-branch parallel structure into a single 3×3 convolution kernel; The edge-end inference acceleration module, connected to the structure reparameterization conversion module, is used to acquire the image to be detected from the edge device and input the image to be detected into the deployed YOLO26 network. Under the constraints of the underlying structure of the deployed YOLO26 network, which uses a single 3×3 convolutional kernel for serial computation and the DynamicMish-SiLU dynamic activation function for adaptive weighting, the module performs object detection inference operations and outputs the object detection result of the image to be detected, thus completing the object detection acceleration for the edge.

Images