Automatic detection method and equipment for invasive plant solidago canadensis and storage medium

Abstract

The invention provides an invasive plant solidago canadensis automatic detection method and device, and a storage medium, and the method comprises the steps: building a training data set, obtaining a target species image through the remote sensing of an unmanned plane, and generating a training sample through image processing and data enhancement; constructing a detection network, including replacing an original module in a target detection network with an improved feature extraction module to form a multi-stage cascade backbone network; the improved feature extraction module sequentially comprises a redundant feature generation unit, a gradient enhancement unit and a channel adjustment unit; and training the target detection network, and then executing detection, including inputting the to-be-analyzed image to the trained target detection network to obtain an automatic detection result of the invasive species. The invention provides a lightweight feature extraction module which is used for resource-limited scenes such as unmanned aerial vehicle remote sensing image invasive species detection.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and target detection technology, specifically relating to an automatic detection technology for the invasive plant Goldenrod in Canada. Background Technology

[0002] Invasive alien species (IAS) are alien species that have been introduced and established, posing a threat or harm to ecosystems, habitats, and species, impacting my country's ecological environment, and damaging the sustainable development of agriculture, forestry, animal husbandry, and fisheries, as well as biodiversity. Solidago canadensis L., a highly damaging invasive alien plant (IAP), has caused significant negative impacts on my country's ecological environment, biodiversity, and economy. In 2022, it was included in my country's "List of Key Invasive Alien Species under Management."

[0003] Currently, the removal of Canadian goldenrod is mostly achieved through manual location followed by physical, chemical, or biological control methods. Manual location is time-consuming, labor-intensive, and ineffective. Utilizing advanced computer vision technology for efficient and accurate location and monitoring of Canadian goldenrod would help outdoor workers pinpoint its location and eradicate it according to the actual situation. This is of great significance for maintaining ecosystem health and biodiversity, and for providing fundamental data for agricultural, forestry, and grassland management and ecological environmental protection.

[0004] Early object detection methods relied on manually designed features (such as HOG and SIFT) and classifiers like support vector machines to achieve object recognition. These methods suffer from insufficient feature generalization ability in complex scenarios and require feature design for specific tasks, making them difficult to adapt to the needs of multi-scale and multi-morphological object detection.

[0005] Deep learning-based object detection technologies (such as the YOLO series and Faster R-CNN) automatically learn features through convolutional neural networks, significantly improving detection accuracy. However, existing models still face the following challenges in practical applications: 1) Traditional bottleneck structures (BottleNeck) suffer information loss when reducing the number of channels, affecting the ability to detect small objects; 2) Multi-branch structures (such as Inception and ResNet) enhance feature representation but significantly increase computational overhead; 3) Traditional Feature Pyramid Networks (FPN) rely on unidirectional feature flow, resulting in insufficient feature fusion for small objects.

[0006] In recent years, lightweight methods have made progress. For example, GhostNet generates redundant features through inexpensive operations to reduce computation, but lacks nonlinear expressive power; RepConv technology enhances feature extraction through reparameterization, but it has not been effectively combined with lightweight design. Existing methods still have shortcomings in balancing accuracy and efficiency, making it difficult to meet the needs of practical applications. Summary of the Invention

[0007] To address the aforementioned issues, this invention proposes a lightweight feature extraction module, C2f-RepG, which reduces computational complexity while maintaining detection accuracy.

[0008] This invention provides an automatic detection method for the invasive species *Solidago canadensis*, comprising the following steps: Construct a training dataset, including acquiring images of target species through UAV remote sensing, and generating training samples through image processing and data augmentation; Constructing a detection network involves replacing the original modules in the target detection network with improved feature extraction modules to form a multi-level cascaded backbone network; the improved feature extraction module sequentially includes a redundant feature generation unit, a gradient enhancement unit, and a channel adjustment unit. The redundant feature generation unit divides the input feature map into blocks through convolution operations, generates redundant feature maps through linear transformations, and concatenates them with the original features to form intermediate features. The gradient enhancement unit adopts a reparameterized convolutional structure. During the training phase, features are extracted and fused through multi-branch convolution, and during the inference phase, it is equivalent to a single convolutional layer. The channel adjustment unit dynamically controls the output channel dimension through a scaling factor. The target detection network is trained and then the detection is performed, including inputting the image to be analyzed into the trained target detection network to obtain the automatic detection results of invasive species.

[0009] Moreover, the redundant feature map is generated by performing linear transformation operations on the segmented feature sub-maps and then summing them up.

[0010] Moreover, the gradient enhancement unit fuses multi-branch convolution parameters into a single convolutional layer parameter during the inference phase.

[0011] Furthermore, the scaling factor is set to a value less than one.

[0012] Furthermore, the detection results are output through image slicing and segmentation.

[0013] Moreover, the data augmentation strategies include a variety of combinations of rotation, translation, flipping, brightness adjustment, random occlusion, and noise addition.

[0014] Moreover, the bounding box loss function simultaneously considers the overlap between the predicted box and the ground truth box, the distance between the center points, and the consistency of the aspect ratio.

[0015] On the other hand, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the automatic detection method for the invasive species goldenrod of Canada as described above.

[0016] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the automatic detection method for the invasive species Goldenrod in Canada as described above.

[0017] On the other hand, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the automatic detection method for the invasive species Goldenrod in Canada as described above.

[0018] This invention proposes a lightweight feature extraction module, C2f-RepG, which effectively improves the feature extraction capability of target detection models and reduces computational complexity by integrating GhostNet ideas and RepConv reparameterization technology. It is suitable for resource-constrained scenarios such as invasive species detection in UAV remote sensing images.

[0019] Compared with the prior art, the key improvement of the present invention is as follows: (1) Design a new module C2f-RepG to replace the original C2f module in YOLOv8 and build a new Backbone network; (2) Compared with the original Backbone network, C2f-RepG adopts a lightweight design, utilizes the GhostNet principle, uses inexpensive operations to generate a portion of redundant feature maps and uses split to divide the feature maps into blocks; (3) The introduction of reparameterized convolution RepConv compensates for the performance loss caused by removing the bottleneck module BottleNeck from the original C2f module. In addition, RepConv can increase the nonlinearity of the model without increasing the computational burden. Finally, the scaling factor can be controlled to accommodate models of different sizes. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the C2f-RepG module structure according to an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of the RepConv reparameterization process according to an embodiment of the present invention.

[0022] Figure 3This is a network structure diagram of an embodiment of the present invention. Detailed Implementation

[0023] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0024] To address the shortcomings of existing object detection models, such as insufficient feature extraction capabilities, high computational complexity, and poor lightweight performance, this invention achieves the following technical solutions: First, a training sample set is constructed by collecting image data of the invasive species *Gynostemma pentaphyllum* (Canada goldenrod) in complex natural environments using drones. Target bounding boxes and category labels are annotated, and diverse training datasets are generated using data augmentation strategies such as rotation, translation, flipping, brightness adjustment, and noise addition. Second, a C2f-RepG module is constructed, comprising: 1) a feature generation unit based on GhostNet principles, which generates redundant feature maps through inexpensive operations and performs block processing; 2) a gradient flow enhancement unit, which uses RepConv reparameterization technology to replace the traditional bottleneck structure, enhancing feature representation capabilities through a multi-branch structure during training and fusing it into a single-branch convolution during inference to reduce computational load; 3) a channel dimension control mechanism, which adjusts the number of feature map channels through scaling factors to balance model capacity and computational efficiency; and 4) inputting drone images to be detected, achieving efficient and accurate detection in large-scale scenes using slice-assisted hyper-inference (SAHI) technology. The C2f-RepG module is then applied to the object detection network to achieve efficient detection of multi-scale targets in complex scenes. Experimental results show that the method of the present invention maintains high detection accuracy (97.0% mAP@0.5) while reducing the number of model parameters to 2.123M and the floating-point operation to 6.2GFLOPs, which significantly improves the lightweight performance and inference speed of the model and is suitable for resource-constrained scenarios such as invasive species detection in UAV remote sensing images.

[0025] This invention provides an automatic detection method for Canadian goldenrod based on an improved C2f module, comprising the following steps: Step 1: Build a high-quality dataset.

[0026] The preferred embodiment proposes that, based on image data acquired through UAV remote sensing technology, and combined with image cropping, annotation, and data augmentation strategies, the problem of insufficient sample data is solved.

[0027] In the embodiments, the preferred recommendations are implemented by including the following steps: Step 1.1: Acquire raw imagery of the target location using a quadcopter drone. In practice, existing quadcopter drones can be used, such as the DJI Phantom 4 RTK version.

[0028] Step 1.2: Select UAV image processing software (such as Pix4Dmapper) to perform orthophoto stitching on the acquired raw images.

[0029] Step 1.3: Construct the original sample dataset, crop the stitched images according to the corresponding size and overlap, and filter out invalid images such as blurry or uncertain images to obtain preliminary images of Canadian goldenrod.

[0030] Step 1.4 involves labeling the image of a single goldenrod from Canada, using a standardized data storage structure. In practice, it is recommended to use existing software tools such as LabelImg for labeling.

[0031] Step 1.5 involves performing data augmentation operations on the original sample dataset, including rotation, translation, flipping, brightness adjustment, cutout, and adding noise to expand the dataset. Specifically, Python can be used for data augmentation, and the augmented dataset can be randomly divided into training, validation, and test sets in a 7:2:1 ratio for subsequent training and validation.

[0032] Step 2: Set up an automatic detection network for Canadian goldenrod based on the improved C2f module. Specifically, replace the original C2f module in YOLOv8 with the new module C2f-RepG to construct a new backbone network. This invention designs the C2f-RepG module to reduce computational load and parameter count, forming a new network structure.

[0033] The C2f-RepG module is an improved C2f module based on reparameterized convolution and phantom features. The module proposed in this invention cleverly references the ideas of GhostNet, generates partially redundant feature maps through inexpensive operations, and uses RepConv on the gradient flow branch to enhance feature extraction capabilities in order to compensate for the performance loss caused by discarding BottleNeck residual blocks. The new network increases the nonlinearity of the model without increasing the computational burden.

[0034] See Figure 1 The C2f-RepG module provided in this embodiment has the following structure: Redundant feature generation unit: generates the input feature map Split is divided into main branches and redundant branches using 1×1 convolutions. The main branch directly retains the original features and inputs them into the channel adjustment unit, while the redundant branch generates redundant feature maps. The feature is then fed into the gradient enhancement unit and concatenated with the original feature by channel to form an intermediate feature. .

[0035] "Spilt" refers to a channel splitting operation of a feature map tensor. It optimizes model performance by splitting the feature map into multiple independent branches along the channel dimension, assigning different computational paths and functional roles to each branch. Preferably, this invention involves equally dividing the input feature map along the channel dimension at a ratio of 0.5, resulting in a main branch and redundant branches, with each branch having half the number of channels as the original feature map.

[0036] The formula for generating redundant feature maps is: (1) in, For linear transformation operations, Let i be the sub-feature map after block division, where i is the block number and k is the number of blocks. This is an operation for generating redundant feature maps.

[0037] Gradient enhancement unit: It adopts a RepConv (reparameterized convolution) structure set in sequence, a cascaded structure of n-1 3×3 convolutional layers Conv2d (which can be denoted as Conv3×3×(n-1), where n is the number of convolutional blocks in the unit, and it is recommended to take a value of 2), and a 1×1 convolutional layer to unify the number of channels of the feature map.

[0038] RepConv is a dynamically structured convolutional module. During training, it includes three branches: 3×3 convolution + BN, 1×1 convolution + BN, and an identity mapping branch + BN. During inference, these branches are fused into a single convolutional layer. The formula is as follows: (2) in, , For convolution kernels of different branches, Let be the identity mapping matrix, and This refers to the convolution kernel weights after fusion during the inference stage.

[0039] BN stands for normalization layer. Each branch has independent convolutional kernel weights and BN parameters. During the inference phase, in order to reduce the amount of computation, these branches are fused through parameter fusion to form an equivalent 3×3 convolutional layer.

[0040] See Figure 2 The input feature map is processed simultaneously through three parallel paths during the training phase: Branch 1: The main branch, consisting of 3×3 convolutions and batch normalization (BN), is used to extract spatial features. Branch 2: 1×1 convolution + BN improves channel interaction and enhances expressive power. Branch 3: Identity Mapping + BN is used for information preservation. Features will only flow into this branch when the input and output channels are equal and Stride=1. Stride is the stride of the convolution kernel moving on the input data during the convolution operation.

[0041] The outputs of the three paths are added together using the ReLU activation function to form an enhanced feature.

[0042] During the inference phase, all branches are fused into an equivalent 3×3 convolution. This fusion operation is completely equivalent to the original multi-branch structure output. Before inference, each branch needs to be fused into an equivalent convolution operation. Specifically, for the 3×3 convolution + BN branch, the BN is fused into the convolution kernel using mathematical formula (3); for the 1×1 + BN branch, the convolution needs to be zero-padding to expand it to 3×3; for the identity mapping + BN branch, a 3×3 convolution kernel can be added as an identity matrix, i.e., the middle is 1 and the rest are 0.

[0043] The batch normalization layer (BN) is set up, which includes statistical parameters (training set mean). ,variance ) and learnable parameters (scaling factor) Offset During training, each convolutional branch is followed by its own independent Batch Normalization (BN) layer, meaning that parameters are not shared between branches. Adding BN after the identity mapping ensures that all branches participate uniformly in reparameterization merging. The formula for calculating BN is as follows: (3) in, The input feature values ​​for the batch normalization layer, These are the output feature values ​​after batch normalization. It is a very small constant to prevent the denominator from being zero.

[0044] make (4) in, The convolution weights are the result of incorporating the parameters of the BN layer. This is the convolution bias after incorporating the parameters of the BN layer.

[0045] During the inference phase, Batch Normalization (BN) is incorporated into the convolution weights and biases, and the convolution results are fed into the system according to the idea of ​​structure reparameterization. This allows us to obtain the merged convolution parameters.

[0046] (5) in, x The feature map is input to the convolutional layer. These are the original weight parameters of the convolutional layer. These are the original bias parameters of the convolutional layer. The weights of the merged convolution. This is the bias of the convolution after fusion.

[0047] Channel adjustment unit: via scaling factor s The formula for dynamically controlling the number of output channels is: (6) in, The number of channels in the feature map input to the channel adjustment unit. This refers to the number of channels in the output feature map after processing by the channel adjustment unit.

[0048] In the C2f-RepG module, after the redundant feature generation unit generates redundant feature maps, the channel adjustment unit concatenates the original features of the main branch with the feature maps generated by the redundant branches through the gradient enhancement unit by channel using "Concat". Finally, in the channel adjustment unit, the number of channels of the feature maps is unified by a 1×1 conv2d convolution, and the final feature map is output.

[0049] The working process of the above C2f-RepG module network structure is as follows: Step 2.1: After dividing the input feature map into blocks using the redundant feature generation unit, redundant features are generated through linear transformation and then concatenated with the original features by channel to form intermediate features. .

[0050] Step 2.2, intermediate features After the initial features are obtained through three branches and ReLU activation summation in the RepConv module, the gradient enhancement unit further processes these features: first, they are cascaded through n-1 3×3 convolutional layers (n is preferably 2, i.e., one 3×3 convolutional layer) to further deepen spatial feature extraction and enhance the nonlinear expressive power of the features; then, a 1×1 convolutional layer is used to uniformly adjust the number of channels in the feature map to obtain the enhanced features. Ultimately The feature maps obtained from each layer in this unit are identically mapped and superimposed onto the channel adjustment unit for Concat stitching.

[0051] Specifically, the block feature maps are input into RepConv to enhance the gradient. This part contains three parallel branches: 3×3 convolution, 1×1 convolution, and identity mapping. During the training phase, the input feature maps are processed through these three branches respectively, and then the outputs of the three branches are summed and passed through the ReLU activation function to obtain the enhanced features. During the inference phase, in order to reduce computation and memory access, the parameters of the three branches from the training phase are fused to obtain an equivalent single convolutional layer.

[0052] Step 2.3: Adjust the number of channels using the last 1×1 Conv2d convolution in the channel adjustment process to output the final features. .

[0053] See Figure 3 The embodiment further proposes replacing the C2f module in the YOLOv8 network with a C2f-RepG module to construct a 4-level cascaded backbone network. A preferred recommendation is that, except for operations involving changing the number of feature map channels using 1×1, all other operations involving feature extraction use a 3×3 convolutional kernel size.

[0054] In practice, the size of the input feature map can be denoted as H×W×C. In this embodiment, we take an 80×80×256 feature map size as an example of the input C2f-RepG in the Backbone. This means that the input image has 80 pixels in both height and width, and 256 channels.

[0055] This invention sets the scaling factor to a value less than one. In this embodiment, the scaling factor is set to 0.5, which means that the input feature map is divided into two blocks using a 1×1 convolution. That is, the channel dimension of the input feature map is divided into two equal parts, resulting in two sub-feature maps, each with a size of 80×80×128.

[0056] Step 3: Train the network model based on the training sample set.

[0057] The embodiments further provide preferred implementation schemes as follows: Step 3.1: Use the stochastic gradient descent (SGD) optimizer with an initial learning rate of 0.01, momentum of 0.937, and weight decay of 0.0005.

[0058] Stochastic gradient descent is an existing technique, and will not be described further in this invention.

[0059] Step 3.2 employs a cosine annealing learning rate strategy with a period of 200 epochs. During training, the learning rate decreases with a cosine curve as the number of training epochs increases, which helps the model finely adjust parameters in the later stages of training.

[0060] Step 3.3: Use the bounding box loss CIOU Loss as the loss function. The bounding box loss function simultaneously considers the overlap between the predicted box and the ground truth box, the distance between the center points, and the consistency of the aspect ratio. (7) in, This represents the Euclidean distance between the centroids of the grounding truth (GT) and the predicted bounding box. h , w These represent the height and width of the prediction box, respectively. h gt and w gt This indicates the height and width of GT. c h and c w These represent the height and width of the smallest rectangular box formed by the predicted bounding box and the ground truth (GT), respectively. It is calculated by dividing the area of ​​the intersection of the predicted bounding box and the ground truth bounding box by the area of ​​their union. CIOU Loss for bounding box The parameters for the prediction box include the coordinates of the center point of the prediction box. x,y ) and width and height w,h Information such as ) The parameters of the ground truth bounding box include the coordinates of the center point of the ground truth bounding box. x gt , y gt ) and width and height ( w gt , h gt Information such as these serve as a reference benchmark during model training.

[0061] Step 3.4 sets the number of images per batch for training to 32, and trains for a total of 200 epochs. Simultaneously, mixed precision training is employed, utilizing PyTorch's automatic mixed precision (AMP) function to reduce memory usage and accelerate the training process.

[0062] Step 4: Test the dataset used for training and test the generalization on the dataset not used for training.

[0063] The embodiments further provide preferred implementation schemes as follows: Step 4.1: Input the dataset used for testing into the trained network model to generate the location information of the target in the image.

[0064] Step 4.2: Evaluate the final target detection results using various evaluation metrics, including Mean Average Precision (mAP). 0.5 and mAP 0.5-0.95 Indicators such as the number of parameters and training time are used to evaluate the detection accuracy and efficiency of the model.

[0065] In practice, generalization ability tests can be conducted on image test datasets that were not used for training to evaluate the model's detection performance in unknown scenarios.

[0066] The preferred approach is to use Slice-Assisted Hyper-Inference (SAHI) to divide a large image into several regions by internal image slicing during the model inference process, perform inference on the entire image, and output the inference result.

[0067] The results of intelligent detection of Canadian goldenrod data collected in complex real-world environments using this invention demonstrate that it can achieve rapid location and accurate monitoring of invasive species, providing technical support for intelligent detection of invasive alien species in ecological environment monitoring and a scientific basis for controlling the spread of alien species and protecting ecosystem diversity.

[0068] In the process of accuracy verification and performance evaluation using this invention, the evaluation indicators involved in step 4.2 are used to compare it with other mainstream models. This also verifies the rationality of selecting YOLOv8n as the benchmark and the effectiveness of its improvement. The comparison models include YOLOX, Faster-RCNN, Cascade-RCNN, TOOD, and RTMDet. The quantitative comparison results of each model are shown in Table 1.

[0069] Table 1 Comparison of detection results from different models

[0070] YOLOX focuses on a balance between accuracy and speed, but its detection accuracy was low in this experiment. Faster-RCNN, Cascade-RCNN, and TOOD use complex ResNet50 structures, resulting in a large number of parameters and floating-point operations. Although these three methods can provide better accuracy than YOLOX in complex scenes, they require more computational resources. Introducing C2f-RepG into mAP... 0.5 and mAP 0.5-0.95The improvements achieved are 1.0% and 5.2% compared to the original model. These improvements are mainly attributed to the idea of ​​generating more feature maps with low-cost operations borrowed from GhostNet in C2f-RepG, and the use of RepConv to compensate for the performance loss caused by removing the Bottleneck module, making feature propagation smoother. The C2f-RepG module maintains the model's expressiveness while reducing computational cost, resulting in a 0.747M reduction in the number of model parameters.

[0071] In specific implementation, the method proposed in the technical solution of this invention can be automatically executed by those skilled in the art using computer software technology. System devices for implementing the method, such as computer-readable storage media storing the corresponding computer program of the technical solution of this invention and computer equipment including the computer program running the corresponding computer program, should also be within the protection scope of this invention.

[0072] The following embodiments describe the electronic device provided by the present invention. The electronic device described below can be referred to in correspondence with the automatic detection method for the invasive species Goldenrod in Canada described above.

[0073] The electronic device may include a processor, a communications interface, memory, and a communication bus. The processor, communications interface, and memory communicate with each other via the communication bus. The processor can invoke logical instructions from the memory to execute an automatic detection method for the invasive species *Goldenrod decanterii*, primarily including the software processing steps described above.

[0074] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present 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 the present 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] On the other hand, embodiments of the present invention also provide a computer program product, the computer program product including a computer program, the computer program being stored on a non-transitory computer-readable storage medium, and when the computer program is executed by a processor, the computer is able to execute the software processing portion of the automatic detection method for the invasive species Goldenrod in Canada provided by the above methods.

[0076] In another aspect, embodiments of the present invention also provide a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the software processing portion of the automatic detection method for the invasive species Goldenrod in Canada provided by the above methods.

[0077] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0078] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0079] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for automatic detection of the invasive species Solidago canadensis, characterized in that, The method comprises the following steps: Constructing a training data set, including obtaining target species images through unmanned aerial vehicle remote sensing, generating training samples through image processing and data enhancement; Constructing a detection network, including replacing the original module in the target detection network with an improved feature extraction module to form a multi-level cascade backbone network; the improved feature extraction module comprises a redundant feature generation unit, a gradient enhancement unit and a channel adjustment unit in sequence, The redundant feature generation unit divides the input feature map into blocks through convolution operation, generates a redundant feature map through linear transformation, and splices the original feature to form an intermediate feature; The gradient enhancement unit adopts a reparameterization convolution structure, and in the training stage, the features extracted through multi-branch convolution are fused, and in the inference stage, it is equivalent to a single convolution layer; The channel adjustment unit dynamically controls the output channel dimension through a scaling factor. Training the target detection network and then performing detection, including inputting the image to be analyzed into the trained target detection network to obtain the automatic detection result of the invasive species.

2. The method of claim 1, wherein the invasive species of Solidago canadensis is automatically detected, characterized by, The redundant feature map is obtained by linearly transforming the feature sub-graphs after block division and then collecting them.

3. The method of claim 1, wherein the invasive species of Solidago canadensis is detected automatically. The gradient enhancement unit fuses the multi-branch convolution parameters into single convolution layer parameters in the inference stage.

4. The method of claim 1, wherein the invasive species of Solidago canadensis is detected automatically. The scaling factor is set to a value less than one.

5. The method of claim 1, wherein the invasive species of Solidago canadensis is detected automatically. The image slices are divided and processed to output the detection result.

6. The method of claim 1, wherein the invasive species of Solidago canadensis is detected automatically. The data enhancement strategy includes multiple combinations of rotation, translation, flipping, brightness adjustment, random occlusion and noise addition.

7. The method of claim 1, wherein the invasive species of Solidago canadensis is detected automatically. The bounding box loss function simultaneously considers the overlap, center point distance and aspect ratio consistency of the predicted box and the real box.

8. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that: The processor executes the program to realize the automatic detection method of the invasive species Solidago canadensis as claimed in any one of claims 1 to 7. 9.A non-transitory computer-readable storage medium having stored thereon a computer program. The computer program is executed by the processor to realize the automatic detection method of the invasive species Solidago canadensis as claimed in any one of claims 1 to 7.

10. A computer program product comprising a computer program, characterized in that: The computer program is executed by the processor to realize the automatic detection method of the invasive species Solidago canadensis as claimed in any one of claims 1 to 7.

Images