Unmanned aerial vehicle fire target inspection detection method based on YOLOv8

Abstract

The invention discloses a YOLOv8-based unmanned aerial vehicle inspection fire target detection method, and the method comprises the steps: S1, obtaining a public target detection data set, and carrying out the preprocessing of the data set, and obtaining a training set; a data set is divided into a training set and a test set according to the proportion of 8: 2, and a YOLOv8 default data enhancement strategy is reserved; s2, constructing a GMMS-YOLO network which takes a YOLOv8 network as a basic network; wherein a GLMM module is added in a backbone network, a C2fMSDSA module is provided, and the C2fMSDSA module is used for replacing a traditional C2f module of a P5 feature fusion layer of a head network; s3, training a GMMS-YOLO network by using the training set obtained in the step S1 to obtain a GMMS-YOLO model; and S4, inputting an image to be detected into the GMMS-YOLO model for target detection. According to the invention, by introducing a global perception mechanism, the problem of insufficient detection precision caused by the limited receptive field in a complex scene in the prior art is effectively solved, and the target detection performance is improved.

Description

Technical Field

[0001] The present invention relates to the field of target detection, and in particular to a fire target detection method for unmanned aerial vehicle inspection based on YOLOv8. Background Art

[0002] With global climate anomalies and accelerated urbanization, disasters such as forest fires and industrial fires are becoming more frequent, posing a serious threat to ecological security and public property. Traditional fire monitoring relies primarily on satellite remote sensing, ground-based sensor networks, and manual inspections. Satellite remote sensing offers wide coverage, but is limited by low spatial resolution and long revisit cycles, making it difficult to capture early-stage fires. Ground-based sensors require dense deployment and are easily obscured by terrain, making them expensive to deploy in complex environments. Manual inspections, on the other hand, are inefficient and pose a high risk. These methods suffer from significant deficiencies in real-time performance, coverage density, and adaptability, making them difficult to meet the urgent needs of fire prevention and control.

[0003] Drones, with their maneuverability, rapid response, and high-resolution image acquisition capabilities, have become ideal vehicles for fire monitoring. Combined with deep learning algorithms, drones can analyze target features such as fire points and smoke in visible light images in real time. The YOLO series of single-stage detection models excels in target detection due to its efficient end-to-end architecture and real-time inference performance. The YOLOv8 version further improves detection speed and accuracy through a dynamic label allocation strategy and lightweight design. Its multi-scale feature pyramid structure can adapt to the variable scales of fire targets in drone aerial images, such as tiny heat sources, diffuse smoke, and large-scale open flames, providing technical feasibility for fire detection. However, the standard YOLOv8 model still faces bottlenecks in drone fire detection scenarios: First, although its anchor-free mechanism simplifies model design, the paradigm of directly regressing the target bounding box is not sufficiently adaptable to scale fluctuations caused by drastic changes in aerial photography altitude. The shallow semantic information of the feature pyramid is not fully integrated, which easily leads to feature loss of small targets. Second, the receptive field of a single convolution kernel is insufficient for extracting texture features of irregular targets such as smoke and flames. Third, complex background interference, such as clouds and high-temperature equipment, can easily lead to false detections. Summary of the Invention

[0004] In order to make up for the shortcomings of existing traditional technologies, the present invention introduces a global perception mechanism to effectively solve the problem of insufficient detection accuracy caused by limited receptive field in complex scenes in existing technologies, thereby improving target detection performance.

[0005] The present invention makes multi-dimensional improvements to the YOLOv8 model and provides a fire target detection method for drone inspection based on YOLOv8.

[0006] To achieve the above object, the technical solutions adopted by the present invention are as follows:

[0007] The present invention provides a fire target detection method for unmanned aerial vehicle inspection based on YOLOv8, comprising:

[0008] Step S1: Obtain a public object detection dataset and preprocess it to obtain a training set. The dataset is then divided into a training set and a test set in an 8:2 ratio, retaining the default data augmentation strategy of YOLOv8.

[0009] Step S2: Build a GMMS-YOLO network with the YOLOv8 network as the base network. In this process, a GLMM module (Global-Local Mixing Module) is added to the backbone network of the YOLOv8 network. A C2f_MSDSA module (Multi-Scale Dilated Shuffle Attn Module) is proposed and replaced with the traditional C2f module in the P5 feature fusion layer of the head network.

[0010] Step S3: Apply the training set obtained in step S1 to train the GMMS-YOLO network to obtain the GMMS-YOLO model;

[0011] Step S4: Input the image to be detected into the GMMS-YOLO model for target detection.

[0012] Preferably, in step S2, the specific steps of adding the GLMM module include:

[0013] Step S21: Construct a global feature hybrid unit (GMM), initialize the deformable convolution kernel using the height H and width W parameters of the input feature map, establish a vertical-horizontal bidirectional feature recombination mechanism including relative position encoding, use the channel grouping strategy to split the input features into multiple groups of subspaces, and perform cross-region information fusion on the reorganized vertical and horizontal features using adaptively adjusted broadband convolution kernels;

[0014] Step S22: construct a local feature mixture unit (LMM), use asymmetric 3×7 and 7×3 convolution kernels to extract multi-dimensional local features respectively, generate a spatial attention mask through a dynamic weight allocation network, and realize pixel-level weighted fusion of multi-scale features;

[0015] Step S23: Establish a cascade feature enhancement pathway, input the original feature tensor into the GMM module, obtain the feature tensor output by the GMM module, and then input it into the LMM module for secondary refinement. The contribution weights of the global context and local details are automatically adjusted through the gating mechanism, and finally an enhanced representation vector that integrates multi-scale spatial features is output. It is characterized by realizing cross-level feature interaction through the synergy of convolution kernel deformation and reorganization and position-aware coding, and combining the dynamic attention mechanism to improve the model's recognition ability of complex images.

[0016] Preferably, in step S2, a C2f_MSDSA module is proposed and used to replace the traditional C2f module in the P5 feature fusion layer of the head network. Specifically, the C2f module in the P5 feature fusion layer of the head network of the YOLOv8 network is replaced with the C2f_MSDSA module. The C2f_MSDSA module is constructed by replacing the second CBS in the Bottleneck of C2f with MSDSA.

[0017] Preferably, the MSDSA module is constructed, and the specific steps include:

[0018] Step S24: applying a learnable bias parameter to the input features and performing multi-scale feature extraction through a parallel multi-branch quantized convolutional layer, where each branch uses a dilated convolution kernel with a different dilation rate to capture local, mid-range, and global context information respectively;

[0019] Step S25: The three-dimensional feature tensors output by each branch are subjected to dimensional rearrangement and spatial serialization processing, multimodal feature addition and fusion are performed after parameterizing the activation function, and the distribution of the fused features is calibrated using a normalization layer;

[0020] Step S26: reconstruct the dimensions of the calibrated features to restore the spatial topology. When the input and output channel dimensions do not match, cross-channel projection alignment is achieved through lightweight convolution kernels.

[0021] Step S27: Input the channel-aligned features into the attention module based on channel shuffling, and realize adaptive enhancement of feature channels through a dynamic weight allocation mechanism.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] 1. The present invention introduces a deformable convolution kernel and a bidirectional position encoding mechanism through the design of the GLMM module, dynamically adjusts the shape of the convolution kernel and the direction of the receptive field, and realizes cross-level vertical-horizontal feature reorganization. The vertical feature reorganization focuses on the target height change characteristics, and the horizontal reorganization strengthens the spatial continuity. Combined with the channel grouping strategy to separate the multi-scale subspace features, it significantly improves the ability to capture details of the flame edge and irregular targets, thereby compensating for the feature fusion defects of the anchor-free mechanism in scale mutation scenarios. In addition, through the collaboration of row and column bidirectional position encoding and deformable convolution, a spatial long-range dependency model is established, which significantly improves the overall perception ability of complex large scenes, such as forest fire panoramas, and improves the accuracy of target detection.

[0024] 2. The MSDSA module in the present invention captures local to global features based on multi-branch dilated convolution, realizes lightweight multimodal feature fusion through quantization calculation and channel shuffling mechanism, constructs a hierarchical feature pyramid to cover near-view to far-view targets, and uses dynamic weight distribution to enhance key channel responses, effectively suppress background noise interference, and enhance model robustness.

[0025] 3. The present invention uses the deformation adaptability of the deformable convolution kernel and the self-adjustment mechanism of the gating weight to enable the model to dynamically adapt to the geometric transformation and texture differences of different fire targets: the broadband convolution kernel of the GLMM module captures cross-regional contextual associations through adaptive expansion rate adjustment, enhancing the ability to reason about unknown fire spread patterns; the cascaded feature enhancement pathway automatically balances the contribution ratio of global semantics and local details through a gating mechanism, constructing a multi-scale feature adaptive fusion network to avoid overfitting a single data distribution and enhance the generalization of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0027] Figure 1 A flow chart of a method provided by an embodiment of the present invention;

[0028] Figure 2(a) is a schematic diagram of the overall network structure of GLMS-YOLO of the present invention;

[0029] FIG2( b ) is a schematic diagram of the GLMS-YOLO network structure components of the present invention;

[0030] Figure 3 Schematic diagram of the GMM module structure in the present invention;

[0031] Figure 4 Schematic diagram of the LMM module structure in the present invention;

[0032] Figure 5 Schematic diagram of the C2f_MSDSA module structure in the present invention;

[0033] Figure 6 Schematic diagram of the MSDSA module structure in the present invention. DETAILED DESCRIPTION

[0034] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0035] The purpose of this paper is to provide a fire target detection method for drone inspections based on YOLOv8, which can address the shortcomings of existing traditional technologies. By introducing a global perception mechanism, it effectively solves the problem of insufficient detection accuracy caused by the limited receptive field of existing technologies in complex scenes, thereby improving target detection performance.

[0036] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0037] Figure 1 A flow chart of the method provided in the embodiment of the present invention is shown in FIG. Figure 1 As shown, the present invention provides a fire target detection method for drone inspection based on YOLOv8, comprising:

[0038] Step S1: Obtain a public object detection dataset and preprocess it to obtain a training set.

[0039] Annotated images of fires were obtained from a publicly available open-source dataset. Preprocessing included data normalization, scaling the input images to a uniform resolution of 640×640 pixels, using bilinear interpolation to preserve geometric features, and normalizing the HSV color space to simulate varying lighting conditions. The dataset was partitioned using stratified sampling, with a training set and a test set constructed in an 8:2 ratio to ensure that the two subsets had identical distributions across feature dimensions such as scene type, fire morphology, and meteorological conditions.

[0040] Step S2: Construct a GMMS-YOLO network with the YOLOv8 network as the base network; wherein, add the GLMM module to the backbone network, propose the C2f_MSDSA module, and replace it with the traditional C2f module in the P5 feature fusion layer of the head network.

[0041] As shown in Figure 2(a), the GLMM module is added to the tenth layer of the backbone network, that is, after the SPPF layer as the last layer of the backbone network. This module inputs the feature tensor output by the GMM into the LMM for secondary refinement. Through a gating mechanism, it automatically adjusts the contribution weights of global context and local details, and ultimately outputs an enhanced representation vector that integrates multi-scale spatial features. Its characteristics are characterized by the synergy of convolutional kernel deformation and reorganization and position-aware encoding to achieve cross-level feature interaction, and combined with a dynamic attention mechanism to improve the model's recognition ability for complex images.

[0042] Step S21: Figure 3 As shown in the figure, the implementation details of the GMM (Global Mixing Module) involved in the GLMM module are as follows:

[0043] 1. Input feature decomposition and spatial reorganization: Assume that the input feature is a tensor of size B×C×H×W (B is the batch size, C is the number of channels, H and W are the spatial dimensions). The module splits the channel dimension into Through tensor deformation operations, the original features are reorganized into two complementary representations:

[0044] Column feature block: The size is N×C1×H×(p×W), that is, the p channels of each group are spliced ​​along the width direction to form an extended column dimension.

[0045] Row feature block: The size is N×C1×(p×H)×W, that is, the p channels of each group are spliced ​​along the height direction to form an extended row dimension.

[0046] 2. Use two-way depthwise separable convolution to process column and row feature blocks separately to achieve feature fusion in the spatial dimension. The column-wise convolution includes the following steps: input the reorganized column feature block xh into the depthwise separable convolution layer projh (convolution kernel 3×3, grouping number C1); the convolution operation slides along the expanded width dimension to capture the long-range dependency in the vertical direction, and the output feature size remains N×C1×H×(p×W); the result is restored to the original spatial size N×C×H×W through tensor deformation, and fused with the original input through a residual connection, the formula is: xh′=Conv(xh)+x. Similarly, the row-wise convolution includes the following steps: input the reorganized row feature block xw into the depthwise separable convolution layer projw (convolution kernel 3×3, grouping number C1); the convolution kernel slides along the expanded height dimension to capture the contextual association in the horizontal direction, and the output feature size remains N×C1×(p×H)×W; the result is restored to N×C×H×W, and cross-path fused with the output of the column processing path. The row-column separation convolution in this step reduces the O(k 2) The computational complexity is reduced to O(k), and the number of parameters is reduced to the traditional convolution At the same time, the global modeling capability is enhanced through dual-path complementary features.

[0047] 3. To eliminate the translation invariance deviation of the convolution operation, the module introduces dual-path relative position encoding. The height direction encoding poS is generated based on the row index difference h , reflecting the relative position relationship of the feature points in the vertical direction; generating the width direction code poS based on the column index difference w , reflecting the relative position relationship of the feature points in the horizontal direction; the generated position code is injected into the row and column processing paths in an additive form:

[0048] xh=x+pos h

[0049] xW=xh′+pos w

[0050] 4. Integrate row and column features through a two-stage fusion strategy and output them. The column path fusion is to splice the column processing result xh′ with the original input, compress the channel through 1×1 convolution, and then output the intermediate features through batch normalization and GELU activation function. The row path fusion is to splice the row processing result xw with the intermediate features and generate the final output through 1×1 convolution. The formula is as follows:

[0051] y h =GELU(BN(Conv([x h ′,x])))

[0052] y out =Conv([y h , x w ])

[0053] Step S22: Figure 4 As shown in Figure 2, the implementation steps of the LMM (Local Mixing Module) involved in the GLMM module are as follows:

[0054] 1. Construct asymmetric convolution groups of 3×7 and 7×3. The input feature map is processed by two heterogeneous convolution layers, one horizontal and one vertical. The 3×7 convolution kernel slides along the width of the image, effectively capturing long-range correlations between pixels in the same row; the 7×3 convolution kernel moves vertically to establish feature correlations between pixels across rows. The two heterogeneous convolution layers are represented as follows:

[0055] X w =Conv 3×7 (X)

[0056] X h =Conv 7×3 (X)

[0057] The two-way convolution output maintains the same spatial dimension as the original features.

[0058] 2. Fuse the original features with the two convolution output features. Figure 4 The AvgPool shown in is to obtain channel statistics through global average pooling:

[0059] A base =AvgPool(X+X W +X h )

[0060] 3. Generate spatial attention weights based on the three-layer fully connected network MLP, and then input the channel statistics into the spatial attention weights. After the statistics are input into the feature reweighted network, they are processed by two levels of 1×1 convolution and nonlinear activation to output three sets of weight coefficients with dimensions [N, C]. Through the dimensional expansion operation, the weight coefficients are converted into a three-dimensional attention map that matches the spatial dimensions of the feature map. The MLP is a multi-layer perception unit that contains two levels of 1×1 convolution layers. The first level of convolution compresses the channel dimension to 1 / 2, and the second level restores the original dimension and outputs three sets of weight parameters to achieve intelligent screening of feature channels. The implementation process formula is shown below:

[0061] w=MLP(A base )=W2(GELU(W1(A base ))

[0062] in GAP stands for global average pooling.

[0063] 4. Use the Swish activation function to perform nonlinear calibration on the attention weights, and finally complete the feature fusion through weighted summation. The Swish function involved in this step introduces a linear multiplier term based on the Sigmoid response to enhance the gradient propagation capability. Its expression is:

[0064] Swish(x)=x·σ(x)

[0065] Where σ represents the Sigmoid function.

[0066] Multiply the three calibrated weights by the original features, horizontal features, and vertical features channel by channel:

[0067] X out =A[0]·Xh+A[1]·Xw+A[2]·X

[0068] Finally, feature fusion is completed through weighted summation. This process retains the original features as the reference signal to ensure the stability of network training.

[0069] Step S23: Establish a cascaded feature enhancement pathway, input the original feature tensor into the GMM module, obtain the feature tensor output by the GMM module, and then input it into the LMM module for secondary refinement. The contribution weights of the global context and local details are automatically adjusted through the gating mechanism, and finally an enhanced representation vector that integrates multi-scale spatial features is output.

[0070] In step S2, the C2f module in the feature fusion layer of the YOLOv8 network head network is replaced with the C2f_MSDSA module, as shown in Figure 2(a). Specifically, the C2f module in the P5 feature fusion layer of the YOLOv8 network head network is replaced with the C2f_MSDSA module. Figure 5 As shown in the figure, the construction steps are: replace the second CBS module of Bottleneck in C2f with the MSDSA module.

[0071] Step S24: The MSDSA module involved in the C2f_MSDSA module is as follows: Figure 6 The implementation details are as follows:

[0072] 1. Input feature map X∈R B×C×H×W , passing it through a learnable bias layer whose core parameter is β∈R 1×C×1×1 , dynamically learning the offset of the feature distribution through back propagation. This operation is formulated as:

[0073] X biased =X+β·1 H×W

[0074] Among them 1 H×W The matrix is ​​all 1s, and the spatial dimension is expanded through the broadcast mechanism. This step enhances the model's adaptability to feature distribution by introducing a trainable bias.

[0075] 2. Use three sets of parallel dilated convolution branches with the following parameter configurations:

[0076] Branch 1: 3×3 convolution kernel, dilation rate d=1, number of output channels C.

[0077] Branch 2: 3×3 convolution kernel, dilation rate d=3, number of output channels C.

[0078] Branch 3: 3×3 convolution kernel, dilation rate d=5, number of output channels C.

[0079] The calculation process of each branch is:

[0080] X1=Conv2D(X biased , kernel=3, dialation=1)

[0081] X2=Conv2D(X biased , kernel=3, dialation=3)

[0082] X3=Conv2D(X biased , kernel=3, dialation=5)

[0083] Step S25:

[0084] 1. Each set of convolution is followed by an RPreLU activation function, whose parameters μ and υ are optimized through end-to-end training, and the formula is:

[0085] RPReLU(x)=RPReLU(x-μ)+υ

[0086] This activation function enhances nonlinear expression capabilities through shift operations.

[0087] 2. Fuse the outputs of the three branches along the channel dimension and normalize the fusion features along the channel dimension. The formula is expressed as:

[0088] X sum =X1+X2+X3

[0089] Step S26: Reshape the feature from (B, H×W, C) to (B, C, H, W) to restore its spatial structure. The formula is:

[0090] X norm =LayerNorm(X sum )

[0091] Step S27:

[0092] 1. Divide the channels into G groups, each group contains Channels disrupt the inherent order between channels and force information between different groups to interact. The formula is:

[0093]

[0094] 2. Compress the spatial dimension of the output of step 4 above to 1×1 through adaptive average pooling, and then use G groups of 1×1 convolutions to generate attention weights, which are expressed as follows:

[0095] X pool =AvgPool2D(X shuffle )

[0096] W attn =σ(Conv2D(X pool ,groups=G))

[0097] Where σ is the Sigmoid function, and the output weight matrix W attn ∈R B×C×1×1 .

[0098] The number of output channels of the accumulation layer remains C, and each group is calculated independently The weight of each channel.

[0099] 3. Restore the weight matrix to the original channel order to ensure that the final weights strictly correspond to the channel arrangement of the input feature map to avoid semantic confusion.

[0100]

[0101] Finally, the reorganized weights are applied to the original input features, and through the weight recovery mechanism, the original channel semantic structure is retained after cross-group interaction. The formula is as follows:

[0102]

[0103] in represents channel-by-channel multiplication, that is:

[0104] Output[b,c,h,w]=X[b,c,h,w]×W rearranged [b,c,0,0]

[0105] S3: Apply the training set obtained in step S1 to train the GMMS-YOLO network to obtain the GMMS-YOLO model;

[0106] S4: Input the image to be detected into the GMMS-YOLO model for target detection.

[0107] To verify the effectiveness of the method described in this solution, we conducted ablation experiments on the two improved modules using YOLOv8 as the baseline model. The following table shows the ablation test results. Model a represents the YOLOv8n network, model b adds a GLMM module to the backbone, model c replaces the C2f module in the head with a C2f_MSDSA module, and model d simultaneously adds a GLMM module to the backbone and replaces the C2f module in the head with a C2f_MSDSA module. Model d was used to verify the final detection performance of the present invention, and its test results are shown in Table 1.

[0108] Table 1 Comparison table of experimental results of each model

[0109] GLMM C2f_MSDSA Precision(%) Recall (%) mAP50 (%) mAP50-90 (%) a 69.9 61.8 69.1 37.8 b √ 71.8 62.4 70.3 37.8 C √ 73.3 62.1 71.3 37.9 d √ √ 75.5 62.0 72.7 38.0

[0110] As shown in Table 1, model d of the present invention achieves significant improvements in Precision% (75.5) and mAP50% (72.7), respectively, compared to baseline model a by 5.6% and 3.6%, and outperforms models b or c, which only add either module. This demonstrates the synergistic optimization effect of the GLMM and C2f_MSDSA modules, achieving improved precision and validating the effectiveness of the joint global-local feature optimization.

[0111] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0112] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A fire target detection method for drone inspection based on YOLOv8, characterized in that: include: S1: Obtain a public object detection dataset and preprocess it to obtain a training set; The dataset is divided into training set and test set in a ratio of 8:2, and the default data enhancement strategy of YOLOv8 is retained; S2: Build a GMMS-YOLO network based on the YOLOv8 network. In this network, add a GLMM module to the backbone network and propose a C2f_MSDSA module, which replaces the traditional C2f module in the P5 feature fusion layer of the head network. S3: Apply the training set obtained in step S1 to train the GMMS-YOLO network to obtain the GMMS-YOLO model; S4: Input the image to be detected into the GMMS-YOLO model for target detection.

2. The method according to claim 1, characterized in that In step S2, adding the GLMM module includes the following steps: Step S21: construct a global feature mixing unit (GMM), initialize the deformable convolution kernel using the height H and width W parameters of the input feature map, establish a vertical-horizontal bidirectional feature recombination mechanism including relative position encoding, use the channel grouping strategy to split the input features into multiple groups of subspaces, and perform cross-region information fusion on the reorganized vertical and horizontal features using adaptively adjusted broadband convolution kernels; Step S22: construct a local feature mixing unit LMM, use asymmetric 3×7 and 7×3 convolution kernels to extract multi-dimensional local features respectively, generate a spatial attention mask through a dynamic weight allocation network, and realize pixel-level weighted fusion of multi-scale features; Step S23, establish a cascade feature enhancement path, input the original feature tensor into the GMM module, obtain the feature tensor output by the GMM module, and then input it into the LMM module for secondary refinement. The contribution weights of the global context and local details are automatically adjusted through the gating mechanism, and finally an enhanced representation vector that integrates multi-scale spatial features is output. It is characterized by realizing cross-level feature interaction through the synergy of convolution kernel deformation and reorganization and position-aware coding, and combining the dynamic attention mechanism to improve the model's recognition ability of complex images.

3. The method according to claim 1, characterized in that In step S2, replace the C2f module in the feature fusion layer of the YOLOv8 network's head network with the C2f_MSDSA module. Specifically, replace the C2f module in the P5 feature fusion layer of the YOLOv8 network's head network with the C2f_MSDSA module. To construct the C2f_MSDSA module: Replace the second CBS in the Bottleneck of C2f with MSDSA.

4. The MSDSA module according to claim 3 is constructed, comprising the following steps: Step S24: applying a learnable bias parameter to the input features and performing multi-scale feature extraction through a parallel multi-branch quantized convolutional layer, where each branch uses a dilated convolution kernel with a different dilation rate to capture local, mid-range, and global context information respectively; Step S25: The three-dimensional feature tensors output by each branch are subjected to dimensional rearrangement and spatial serialization processing, multimodal feature addition and fusion are performed after parameterizing the activation function, and the distribution of the fused features is calibrated using a normalization layer; Step S26: reconstruct the dimensions of the calibrated features to restore the spatial topology. When the input and output channel dimensions do not match, cross-channel projection alignment is achieved through lightweight convolution kernels. Step S27: Input the channel-aligned features into the attention module based on channel shuffling, and realize adaptive enhancement of feature channels through a dynamic weight allocation mechanism.

Images