Method for detecting infection rate of wheat single-ear gibberellic disease

Abstract

The invention discloses a wheat single-ear gibberellic disease infection rate detection method in the technical field of intelligent agricultural information, and the method comprises the following steps: 1, optimizing a trunk feature extraction network in an AM-YOLOv7 network model; step 2, an activation function in the AM-YOLOv7 network model is optimized; step 3, a feature fusion pyramid network in the AM-YOLOv7 network model is optimized; step 4, the lightweight effect of the optimized AM-YOLOv7 network model is verified; step 5, carrying out transfer learning and model training on the optimized AM-YOLOv7 network model; and step 6, using the optimized AM-YOLOv7 network model to output a detection result. The method ensures that breeders perform real-time detection on the wheat single-ear scab infection rate in an actual scene, is suitable for edge equipment deployment, and solves the problem that the field detection requirement of the wheat single-ear scab infection rate in an actual disease-resistant breeding scene cannot be met due to deep network model hierarchy, large parameter quantity and high requirements on operation memory and computing power.

Description

Technical Field

[0001] The present invention relates to a method for detecting the infection rate of wheat single ear fusarium head blight in the field of smart agricultural information technology. Background Art

[0002] Wheat growth is often accompanied by the threat of various diseases, which directly impact food production security. Among these diseases, wheat fusarium head blight is one of the most threatening. Therefore, in practical disease-resistant breeding scenarios, breeders urgently need a more efficient, accurate, and convenient non-destructive detection method for long-term monitoring of wheat fusarium head blight infection rates within a single ear.

[0003] Computer vision-based deep learning technology is being widely applied by researchers in smart agriculture scenarios, creating the technical foundation for non-destructive diagnosis of crop diseases. Deep learning can rapidly extract feature information from crop images and detect crop diseases. This technology effectively avoids the time-consuming, labor-intensive, and inaccurate nature of manual inspection, while also eliminating the extensive preparatory work required to design feature extractors and classifiers in traditional image recognition techniques. This has led to widespread interest in deep learning for crop disease identification.

[0004] Currently, the YOLO network model, also known as the AM-YOLOv7 network model, is widely used in practical detection. As a groundbreaking object detection model, YOLO can complete the detection and classification tasks of target objects with a single forward propagation, effectively solving the slow inference speed problem of traditional two-stage detection networks and achieving remarkable results in detection accuracy.

[0005] However, traditional deep learning network models face problems in agricultural disease detection applications, such as high network model complexity, large computational effort, long model training time, and high resource consumption. Considering the limited computing power of edge devices, the wheat single ear fusarium infection rate detection model deployed at the edge needs to strictly limit its own parameter quantity and computational effort. Summary of the Invention

[0006] The purpose of the present invention is to provide a method for detecting the infection rate of wheat single ear fusarium wilt. The network model is lightweight designed through model compression methods such as network pruning, group convolution, and depth-separable convolution, so that it can be applied to edge devices and ensure the detection rate in actual scenarios.

[0007] To achieve the above object, the present invention provides a method for detecting the infection rate of wheat single ear scab, comprising the following steps:

[0008] Step 1: Optimize the backbone feature extraction network in the AM-YOLOv7 network model;

[0009] Step 2: Optimize the activation function in the AM-YOLOv7 network model;

[0010] Step 3: Optimize the feature fusion pyramid network in the AM-YOLOv7 network model;

[0011] Step 4: Verify the lightweight effect of the optimized AM-YOLOv7 network model;

[0012] Step 5: Perform transfer learning and model training on the optimized AM-YOLOv7 network model;

[0013] Step 6: Use the optimized AM-YOLOv7 network model to output the detection results.

[0014] Compared with the prior art, the beneficial effect of the present invention is that the AM-YOLOv7 network model is optimized, a channel attention mechanism and a new activation function are introduced, and the feature fusion pyramid network is optimized by replacing the backbone feature extraction network of the original network model, thereby reducing the number of parameters and the amount of calculation, and accelerating the network model reasoning speed. The present invention meets the edge device deployment requirements, ensures that breeders can perform real-time detection of wheat single ear fusarium infection rate in actual scenarios, and is suitable for edge device deployment. It solves the problem that the on-site detection requirements of wheat single ear fusarium infection rate in actual disease-resistant breeding scenarios cannot be met due to the deep network model hierarchy, large number of parameters, and high requirements for computing memory and computing power.

[0015] As a further improvement of the present invention, the specific content of step 1 is as follows:

[0016] The MobileNetV3 network model was introduced into the AM-YOLOv7 network model. The original backbone feature extraction network was replaced based on the idea of ​​replacing standard convolution with depthwise separable convolution. The BottleNeck block used in the MobileNetV3 network model can obtain effective feature maps of three different scales: 20×20×160, 40×40×112, and 80×80×40. Each BottleNeck block consists of two pointwise convolution blocks, a depthwise convolution block, and a channel attention mechanism layer SE Layer.

[0017] Replacing the original backbone network with the MobileNetV3 model significantly reduces the number of parameters and computational overhead, thereby accelerating network inference. This lightweight approach effectively addresses the practical application challenges of the original AM-YOLOv7 model in disease resistance breeding scenarios. Furthermore, the key task of the backbone feature extraction network in the AM-YOLOv7 model is to obtain valid feature maps at three scales for single wheat ear samples: 20×20×1024, 40×40×1024, and 80×80×512. MobileNetV3 utilizes numerous BottleNeck blocks to generate valid feature maps at three scales: 20×20×160, 40×40×112, and 80×80×40. Aside from the channel sizes, the sizes of these three feature maps correspond exactly to those in the AM-YOLOv7 model. Therefore, the MobileNetV3 model can directly generate valid feature maps at the same scale as the original model.

[0018] As a further improvement of the present invention, the specific contents in step 2 are as follows:

[0019] The H-swish activation function is used in the BottleNeck block of MobileNetV3 to replace the traditional ReLU function. H-swish approximates the Swish function through ReLU6. The formula is:

[0020]

[0021] Where x represents the pixel value of the feature map.

[0022] In this way, the function can reduce computing resource consumption while improving the nonlinear expression ability of the model.

[0023] As a further improvement of the present invention, the specific contents of step 3 are as follows:

[0024] Taking advantage of the lightweight characteristics of depthwise separable convolution, which occupies fewer parameters and has a lower computational load, the standard convolution blocks (CBS) of the weighted bidirectional feature pyramid network (BiFPN) in the AM-YOLOv7 network model are completely replaced with depthwise separable convolution blocks (DSC). The optimized AM-YOLOv7 network model is named MobileNetV3-YOLOv7 network model.

[0025] Since a single depth-wise separable convolution block DSC consists of a depth-wise convolution block Depthwise Convolution and a point-wise convolution block Pointwise Convolution, and considering that the standard convolution block CBS to be replaced in the weighted bidirectional feature pyramid network BiFPN covers two different convolution step sizes, the step size of the depth-wise separable convolution block DSC can be set according to the specific convolution step size of the standard convolution block CBS to be replaced to complete the group convolution operation, and then the output channel number of the feature map is adjusted through the point-wise convolution block, and the feature is fused with the original input feature map through the residual connection to finally obtain a valid output feature map.

[0026] AM-YOLOv7's original weighted bidirectional feature pyramid network (BiFPN) uses a large number of standard convolutional blocks (CBSs) in its SPC, ELAN-H, and MP2 modules. These blocks perform convolution operations with kernel sizes and strides of three different scales: (1, 1), (3, 1), and (3, 2). The convolution operations of the (3, 1) and (3, 2) convolutional blocks consume significant computational resources due to the kernel size. The computationally complex CBS blocks in the original feature fusion pyramid network are replaced with depthwise separable convolutional blocks (DSCs), achieving a lightweight architecture with fewer parameters and computational overhead.

[0027] As a further improvement of the present invention, the specific content of step 4 is as follows:

[0028] Ablation experiments were used to replace the backbone feature extraction network BackBone of the AM-YOLOv7 network model with the MobileNetV3 network model, and the standard convolution block CBS in the weighted bidirectional feature pyramid network BiFPN of the AM-YOLOv7 network model was replaced with the depthwise separable convolution block DSC. The effects of each operation step on the number of parameters and computational complexity of the AM-YOLOv7 network model were analyzed.

[0029] This can more comprehensively verify the lightweight effect of the MobileNetV3-YOLOv7 network model.

[0030] As a further improvement of the present invention, the specific contents of step 5 are as follows:

[0031] A total of 100 epochs are set. For the training of the wheat single ear scab dataset, a total of 300 epochs are set. The first 20 epochs are frozen training, and the last 280 epochs are non-frozen training.

[0032] In this way, the loss curves of the MobileNetV1-YOLOv7, MobileNetV2-YOLOv7, MobileNetV3-YOLOv7, DenseNet-YOLOv7 and GhostNet-YOLOv7 network models obtained after using lightweight networks to replace the backbone feature extraction network in the original AM-YOLOv7 model all maintain similar convergence trends. The loss functions of the training of each lightweight network model gradually converge and tend to be flat after nearly 70 epochs. Among them, the MobileNetV3-YOLOv7 network model benefits from the fact that the MobileNetV3 network not only retains the depthwise separable convolution module in the MobileNetV1 network and the inverse residual structure and linear bottleneck structure in the MobileNetV2 network, but also pioneered the introduction of the channel attention mechanism and the new nonlinear activation function H-swish. Compared with other lightweight network models, the MobileNetV3-YOLOv7 network model shows better convergence effect and higher detection performance during training.

[0033] As a further improvement of the present invention, the specific content of step 6 is as follows:

[0034] In step 6.1, the input of the MobileNetV3-YOLOv7 network model preprocesses the wheat ear sample image, including adjusting the input image size to 640×640 pixels to meet the network input requirements;

[0035] In step 6.2, the preprocessed wheat ear sample image is fed into the MobileNetV3 network model. The BottleNeck block generates feature maps of three different scales: 20×20×160, 40×40×112, and 80×80×40, corresponding to the detection requirements of large, medium, and small objects, respectively.

[0036] In step 6.3, after the feature map is input into the weighted bidirectional feature pyramid network (BiFPN), it first passes through the point-by-point convolution block to complete the channel dimension increase operation, then extracts feature information through the depth convolution block and the channel attention layer, and then passes through the point-by-point convolution block again to complete the channel dimension reduction operation. The feature is fused with the original input feature map through the residual connection to finally obtain the effective feature map;

[0037] In step 6.4, the detection head in the MobileNetV3-YOLOv7 network model uses the fused effective feature map to complete positioning and category classification. Since the feature map contains the semantic information of healthy and diseased grains in a single wheat ear, the detection results can be output and the detected target can be classified to determine whether it is a diseased grain or a healthy grain.

[0038] In this way, the channel attention mechanism is introduced to improve the model's detection accuracy of wheat fusarium infection rate; by introducing the new H-Swish nonlinear activation function, the computational overhead required by the model is reduced and computing resources are optimized; introducing the MobileNetV3 network model into the backbone feature extraction network can significantly accelerate the model inference speed; the standard convolution block CBS is completely replaced by the depthwise separable convolution block DSC, and the grouping operation is set according to the specific convolution step size, effectively reducing the number of parameters and computational complexity of the network model. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is the MobileNetV3 network model structure;

[0040] Figure 2 is the number of network model parameters and the amount of calculation;

[0041] Figure 3 It is the MobileNet3-YOLOv7 network model architecture;

[0042] Figure 4 This is the BottleNeck block structure diagram;

[0043] Figure 5 It is a new nonlinear activation function of H-swish;

[0044] Figure 6 This is a schematic diagram of the ELAN-H module network structure;

[0045] Figure 7 It is a schematic diagram of the SPPCSPC module structure;

[0046] Figure 8 This is a schematic diagram of the MP2 module structure;

[0047] Figure 9 This is the training curve for the wheat single ear scab data set;

[0048] Figure 10 These are the results of wheat single ear fusarium infection rate detection. DETAILED DESCRIPTION

[0049] The present invention will be further described below in conjunction with the accompanying drawings:

[0050] like Figure 1-10 The method for detecting the infection rate of wheat single ear fusarium head blight shown in the present invention comprises the following steps:

[0051] Step 1: Optimize the backbone feature extraction network in the AM-YOLOv7 network model;

[0052] The MobileNetV3 network model was introduced into the AM-YOLOv7 network model. The original backbone feature extraction network was replaced based on the idea of ​​replacing standard convolution with depthwise separable convolution. The BottleNeck block used in the MobileNetV3 network model can obtain effective feature maps of three different scales: 20×20×160, 40×40×112, and 80×80×40. Each BottleNeck block consists of two pointwise convolution blocks, a depthwise convolution block, and a channel attention mechanism layer SE Layer.

[0053] Step 2: Optimize the activation function in the AM-YOLOv7 network model;

[0054] The H-swish activation function is used in the BottleNeck block of MobileNetV3 to replace the traditional ReLU function. H-swish approximates the Swish function through ReLU6. The formula is:

[0055]

[0056] Where x represents the pixel value of the feature map.

[0057] Step 3: Optimize the feature fusion pyramid network in the AM-YOLOv7 network model;

[0058] Taking advantage of the lightweight characteristics of depthwise separable convolution, which occupies fewer parameters and has a lower computational load, the standard convolution blocks (CBS) of the weighted bidirectional feature pyramid network (BiFPN) in the AM-YOLOv7 network model are completely replaced with depthwise separable convolution blocks (DSC). The optimized AM-YOLOv7 network model is named MobileNetV3-YOLOv7 network model.

[0059] Since a single depth-wise separable convolution block DSC consists of a depth-wise convolution block Depthwise Convolution and a point-wise convolution block Pointwise Convolution, and considering that the standard convolution block CBS to be replaced in the weighted bidirectional feature pyramid network BiFPN covers two different convolution step sizes, the step size of the depth-wise separable convolution block DSC can be set according to the specific convolution step size of the standard convolution block CBS to be replaced to complete the group convolution operation, and then the output channel number of the feature map is adjusted through the point-wise convolution block, and the feature is fused with the original input feature map through the residual connection to finally obtain a valid output feature map.

[0060] Step 4: Verify the lightweight effect of the optimized AM-YOLOv7 network model;

[0061] Ablation experiments were used to replace the backbone feature extraction network BackBone of the AM-YOLOv7 network model with the MobileNetV3 network model, and the standard convolution block CBS in the weighted bidirectional feature pyramid network BiFPN of the AM-YOLOv7 network model was replaced with the depthwise separable convolution block DSC. The effects of each operation step on the number of parameters and computational complexity of the AM-YOLOv7 network model were analyzed.

[0062] Step 5: Perform transfer learning and model training on the optimized AM-YOLOv7 network model;

[0063] A total of 100 epochs are set. For the training of the wheat single ear scab dataset, a total of 300 epochs are set. The first 20 epochs are frozen training, and the last 280 epochs are non-frozen training.

[0064] Step 6: Use the optimized AM-YOLOv7 network model to output the detection results.

[0065] In step 6.1, the input of the MobileNetV3-YOLOv7 network model preprocesses the wheat ear sample image, including adjusting the input image size to 640×640 pixels to meet the network input requirements;

[0066] In step 6.2, the preprocessed wheat ear sample image is fed into the MobileNetV3 network model. The BottleNeck block generates feature maps of three different scales: 20×20×160, 40×40×112, and 80×80×40, corresponding to the detection requirements of large, medium, and small objects, respectively.

[0067] In step 6.3, after the feature map is input into the weighted bidirectional feature pyramid network (BiFPN), it first passes through the point-by-point convolution block to complete the channel dimension increase operation, then extracts feature information through the depth convolution block and the channel attention layer, and then passes through the point-by-point convolution block again to complete the channel dimension reduction operation. The feature is fused with the original input feature map through the residual connection to finally obtain the effective feature map;

[0068] In step 6.4, the detection head in the MobileNetV3-YOLOv7 network model uses the fused effective feature map to complete positioning and category classification. Since the feature map contains the semantic information of healthy and diseased grains in a single wheat ear, the detection results can be output and the detected target can be classified to determine whether it is a diseased grain or a healthy grain.

[0069] In the present invention, Figure 3 The figure shows the overall architecture of the MobileNetV3-YOLOv7 network model. First, the input of the MobileNetV3-YOLOv7 network preprocesses a sample image of a single wheat ear, resizing it to 640×640 pixels to accommodate the network's input requirements. The image data then enters the backbone feature extraction network, built on MobileNetV3. This network leverages its efficient depthwise separable convolution operations, channel-wise attention mechanism (SELayer), and H-swish activation function to extract multi-scale features from the image.

[0070] In the MobileNetV3 network model, BottleNeck blocks are used extensively to implement image feature extraction. Each BottleNeck block consists of two pointwise convolution blocks, a depthwise convolution block, and a channel attention mechanism layer. The detailed structure is as follows: Figure 4 After the feature map is input to the BottleNeck module, it first passes through the point-by-point convolution block to complete the channel dimension increase operation, then extracts feature information through the depth convolution block and channel attention layer, and then passes through the point-by-point convolution block again to complete the channel dimension reduction operation. The feature is fused with the original input feature map through the residual connection, and finally a valid output feature map is obtained.

[0071] In addition, in order to obtain higher detection accuracy, the present invention also introduces a new activation function called H-swish in each BottleNeck module to replace the original ReLU activation function, as shown in Figure 5 The image and formula for this activation function are defined to ensure effective extraction of semantic information from feature maps while maintaining low computational complexity. This module outputs feature maps of three scales (80×80×40, 40×40×112, and 20×20×160), corresponding to the detection requirements of small, medium, and large objects, respectively. The output feature maps are passed to a feature fusion pyramid network, which is responsible for fusing feature maps of different scales to enhance the model's detection capabilities.

[0072] In order to further optimize computational efficiency and model performance, the traditional standard convolution block (CBS) is replaced by the depthwise separable convolution block (DSC). Specifically, Figure 6 、 Figure 7 、 Figure 8As shown, in the present invention, all standard convolution blocks CBS are replaced by depthwise separable convolution blocks DSC. Each depthwise separable convolution block DSC consists of a depthwise convolution block and a pointwise convolution block. The step size in the depthwise convolution block can be set according to the specific convolution step size of the standard convolution block to be replaced to complete the grouped convolution operation. Subsequently, the number of output channels of the feature map is adjusted by the pointwise convolution block, and the feature fusion process after the depthwise separable convolution operation is optimized by residual connection, and finally an effective output feature map is obtained. This improvement not only significantly reduces the amount of computation and parameters, but also further improves the efficiency of feature extraction through cross-scale connections and weighted feature fusion. BiFPN introduces learnable weights in the fusion process, enabling the network to adaptively adjust the contribution of feature maps of different scales, thereby more effectively extracting and transmitting rich semantic information. The feature maps after feature fusion are passed to the detection head part, which uses these feature maps to locate and classify the target, and finally outputs the detection results of healthy grains and diseased grains in a single wheat ear.

[0073] like Figure 1 The figure shows the overall structure of the MobileNetV3 network model, where Input represents the size of the input feature map of the network layer, OperatorType represents the operation type of the network layer, Conv2D represents the convolution layer, BottleNeck represents the depth-separable convolution block, KernelSize represents the size of the convolution kernel, Output represents the size of the output feature map of the network layer, Stride represents the step size of the convolution operation, SE represents whether the network layer uses the channel attention mechanism, and AF represents the type of activation function used by the network layer, including ReLU and H-swish activation functions.

[0074] like Figure 2 It is the impact of each operation step on the parameter amount and computational cost of the original AM-YOLOv7 network model. The ablation experiment results show that after completing the lightweight operations of the original backbone feature extraction network and the original feature fusion pyramid network, the parameter amount decreased by 13.017M and 18.470M respectively, and the computational cost decreased by 65.581G and 46.913G respectively. The parameter amount and computational cost of the lightweight MobileNetV3-YOLOv7 network model decreased by 28.103M and 89.336G respectively. Compared with the original AM-YOLOv7 network model, the overall parameter amount decreased by 72.642%, and the overall computational cost decreased by 84.306%.

[0075] Figure 9This is the training curve for the wheat single-ear scab dataset of the present invention. The loss curve and mAP curve intuitively demonstrate the training performance of different lightweight network models on the wheat single-ear scab dataset. The MobileNetV3-YOLOv7 model exhibited rapid convergence and high detection accuracy during training, showing significant advantages over other lightweight models. These results demonstrate that the MobileNetV3-YOLOv7 model can effectively meet the requirements for wheat single-ear scab infection rate detection, providing reliable technical support for on-site testing in actual disease-resistant breeding scenarios.

[0076] Figure 10 In order to further analyze and verify the detection performance of the MobileNetV3-YOLOv7 network model, the present invention uses the gradient weighted class activation mapping technology Grad-CAM to visualize the main focus of the network model in the task of detecting wheat single ear fusarium wilt grains in the experiment, and completes the detection of wheat single ear fusarium wilt infection rate.

[0077] Through the above implementation, the present invention significantly reduces the complexity of the model while ensuring detection accuracy, is suitable for edge device deployment, and realizes real-time and non-destructive detection of wheat fusarium infection rate.

[0078] The present invention designs a wheat single ear fusarium infection rate detection model based on the optimization of lightweight convolutional neural networks. It is suitable for edge device deployment and solves the problem that the network model has a deep hierarchy, a large number of parameters, and high requirements for computing memory and computing power, which cannot meet the needs of on-site detection of wheat single ear fusarium infection rate in actual disease-resistant breeding scenarios. The present invention combines the MobileNet network model with the AM-YOLOv7 network model, introduces a channel attention mechanism and a new activation function, and replaces the backbone feature extraction network of the original network model by replacing the standard convolution block in the feature fusion pyramid network with a depth-separable convolution block, thereby achieving the purpose of reducing the number of parameters and the amount of calculation, and accelerating the network model reasoning speed. The present invention meets the edge device deployment requirements and ensures that breeders can perform real-time detection of wheat single ear fusarium infection rate in actual scenarios.

[0079] The present invention is not limited to the above-mentioned embodiments. On the basis of the technical solution disclosed herein, those skilled in the art can make some substitutions and modifications to some of the technical features therein according to the disclosed technical content without creative labor, and these substitutions and modifications are all within the protection scope of the present invention.

Claims

1. A method for detecting the infection rate of wheat single ear fusarium head blight, characterized in that: The following steps are included: Step 1: Optimize the backbone feature extraction network in the AM-YOLOv7 network model; Step 2: Optimize the activation function in the AM-YOLOv7 network model; Step 3: Optimize the feature fusion pyramid network in the AM-YOLOv7 network model; Step 4: Verify the lightweight effect of the optimized AM-YOLOv7 network model; Step 5: Perform transfer learning and model training on the optimized AM-YOLOv7 network model; Step 6: Use the optimized AM-YOLOv7 network model to output the detection results.

2. A wheat single ear fusarium infection rate detection method according to claim 1, characterized in that: The specific contents of step 1 are as follows: The MobileNetV3 network model was introduced into the AM-YOLOv7 network model. The original backbone feature extraction network was replaced based on the idea of ​​replacing standard convolution with depthwise separable convolution. The BottleNeck block used in the MobileNetV3 network model can obtain effective feature maps of three different scales: 20×20×160, 40×40×112, and 80×80×40. Each BottleNeck block consists of two pointwise convolution blocks, a depthwise convolution block, and a channel attention mechanism layer SE Layer.

3. A wheat single ear fusarium infection rate detection method according to claim 2, characterized in that: The specific contents of step 2 are as follows: The H-swish activation function is used in the BottleNeck block of MobileNetV3 to replace the traditional ReLU function. H-swish approximates the Swish function through ReLU6. The formula is: Where x represents the pixel value of the feature map.

4. A wheat single ear fusarium infection rate detection method according to claim 3, characterized in that: The specific contents of step 3 are as follows: Taking advantage of the lightweight characteristics of depthwise separable convolution, which occupies fewer parameters and has a lower computational load, the standard convolution blocks (CBS) of the weighted bidirectional feature pyramid network (BiFPN) in the AM-YOLOv7 network model are completely replaced with depthwise separable convolution blocks (DSC). The optimized AM-YOLOv7 network model is named MobileNetV3-YOLOv7 network model. Since a single depth-wise separable convolution block DSC consists of a depth-wise convolution block Depthwise Convolution and a point-wise convolution block Pointwise Convolution, and considering that the standard convolution block CBS to be replaced in the weighted bidirectional feature pyramid network BiFPN covers two different convolution step sizes, the step size of the depth-wise separable convolution block DSC can be set according to the specific convolution step size of the standard convolution block CBS to be replaced to complete the group convolution operation, and then the output channel number of the feature map is adjusted through the point-wise convolution block, and the feature is fused with the original input feature map through the residual connection to finally obtain a valid output feature map.

5. A wheat single ear fusarium infection rate detection method according to claim 4, characterized in that: The specific contents of step 4 are as follows: Ablation experiments were used to replace the backbone feature extraction network BackBone of the AM-YOLOv7 network model with the MobileNetV3 network model, and the standard convolution block CBS in the weighted bidirectional feature pyramid network BiFPN of the AM-YOLOv7 network model was replaced with the depthwise separable convolution block DSC. The effects of each operation step on the number of parameters and computational complexity of the AM-YOLOv7 network model were analyzed.

6. A wheat single ear fusarium infection rate detection method according to claim 5, characterized in that: The specific contents of step 5 are as follows: A total of 100 epochs are set. For the training of the wheat single ear scab dataset, a total of 300 epochs are set. The first 20 epochs are frozen training, and the last 280 epochs are non-frozen training.

7. A wheat single ear fusarium infection rate detection method according to claim 6, characterized in that: The specific contents of step 6 are as follows: In step 6.1, the input of the MobileNetV3-YOLOv7 network model preprocesses the wheat ear sample image, including adjusting the input image size to 640×640 pixels to meet the network input requirements; In step 6.2, the preprocessed wheat ear sample image is fed into the MobileNetV3 network model. The BottleNeck block generates feature maps of three different scales: 20×20×160, 40×40×112, and 80×80×40, corresponding to the detection requirements of large, medium, and small objects, respectively. In step 6.3, after the feature map is input into the weighted bidirectional feature pyramid network (BiFPN), it first passes through the point-by-point convolution block to complete the channel dimension increase operation, then extracts feature information through the depth convolution block and the channel attention layer, and then passes through the point-by-point convolution block again to complete the channel dimension reduction operation. The feature is fused with the original input feature map through the residual connection to finally obtain the effective feature map; In step 6.4, the detection head in the MobileNetV3-YOLOv7 network model uses the fused effective feature map to complete positioning and category classification. Since the feature map contains the semantic information of healthy and diseased grains in a single wheat ear, the detection results can be output and the detected target can be classified to determine whether it is a diseased grain or a healthy grain.

Images