Early pine wood nematode disease identification method based on balance parameters and detection accuracy

Abstract

The present invention discloses an early pine wood nematode disease identification method based on balance parameters and detection accuracy, comprising obtaining a pine wood nematode disease image to be identified, inputting it into a trained pine wood nematode disease identification model based on balance parameters and detection accuracy, and obtaining a corresponding pine wood nematode disease identification result. The pine wood nematode disease identification model includes an input module; a backbone network for extracting features from the pine wood nematode disease image; a neck network for fusing and enhancing multiple output features; an optimized head network for feature representation of multiple fused features; and an output module. The present invention uses a trained pine wood nematode disease identification model based on balance parameters and detection accuracy to identify diseases in pine wood nematode disease images, with strong generalization ability and high recognition accuracy.

Description

Technical Field

[0001] The invention relates to an early pine wood nematode disease identification method based on balance parameters and detection accuracy, and belongs to the technical field of pine wood nematode disease identification. Background Art

[0002] Pine wilt disease (PWD) is a serious threat to pine forests and the timber industry worldwide, which is transmitted from dead trees to healthy trees by vector beetles carrying the pine wood nematode (Bursaphelenchus xylophilus). After pine trees are infected with pine wilt disease, they usually die within a few months. The history of pine wilt disease in Nanjing began in 1982, and the earliest discovery site was located in Purple Mountain in the Nanjing Sun Yat-sen Mausoleum Scenic Area. Since its discovery, the disease has spread rapidly in Nanjing and its surrounding areas, causing serious damage to local forestry resources. The rapid spread of pine wilt disease not only affects the health of forests, but also has a long-term negative impact on the stability of the ecosystem.

[0003] Since pine wilt disease is a fast-spreading disease, once detected, control measures should be taken immediately to prevent it from spreading further to healthy pine trees. Therefore, it is very important to detect pine wilt disease, which can help relevant departments to grasp the current infection situation and take timely measures to avoid further spread of pine wilt disease. Traditional effective measures to control the spread of pine wilt disease are to manually locate infected pine trees and then cut them down, burn them or fumigate them.

[0004] Molecular biology methods are increasingly used in modern research and applications, which improve the speed and accuracy of detection compared to manual identification. However, they are costly, complex to operate, and difficult to directly apply to rapid on-site detection.

[0005] To address these challenges, researchers began to explore the use of remote sensing technology for monitoring. Early remote sensing applications relied mainly on satellite imagery, allowing researchers to conduct assessments without direct contact with infected areas while being able to cover large areas, overcoming the limitations of manual surveys. However, data acquisition from satellite remote sensing is limited by the satellite's orbital period, making it impossible to monitor the rapid spread of pine wood nematode disease in real time. Secondly, weather conditions can significantly affect image acquisition and resolution, limiting the continuity and reliability of monitoring. Finally, the spatial resolution of traditional satellite images is relatively low, making it difficult to distinguish small diseased areas or individual infected trees, affecting the accuracy of monitoring and the ability to identify details.

[0006] In recent years, short-range drone remote sensing technology has shown great potential in monitoring pine wood nematode disease. First, drones can be equipped with optical cameras, hyperspectral sensors, thermal infrared cameras, LiDAR and other equipment for shooting, with high spatial and high temporal resolution. Not only can it reveal subtle changes at the canopy level, including pigment changes and leaf wilting, but it can also provide continuous and repeated monitoring, which is conducive to dynamic tracking of the real-time development of the disease. Secondly, the deployment of drones is not restricted by terrain and can fly in difficult-to-reach areas, which is particularly important for monitoring remote or difficult-to-reach pine forest areas on foot, and has good flexibility.

[0007] Researchers have begun combining drone imagery with traditional machine learning algorithms to identify and analyze trees infected with pine wilt disease. Although traditional machine learning algorithms have made some progress in detecting pine wilt disease, they require manual selection and design of features, which is difficult and time-consuming for complex forest images taken by drones. At the same time, they may face challenges in accuracy and stability due to the dynamic changes in the forest environment, and their generalization ability in different geographical locations or forest types is also limited.

[0008] In recent years, deep learning has developed rapidly. By learning high-level abstract features in data sets, it has shown stronger generalization capabilities than traditional machine learning and has greatly improved the accuracy of target recognition in complex environments. This progress has made deep learning the preferred technology in the field of drone image analysis. However, in practical applications, there are often situations where resources are limited. Summary of the invention

[0009] The purpose of the present invention is to overcome the shortcomings of the prior art and provide an early pine wood nematode disease identification method based on balance parameters and detection accuracy. The pine wood nematode disease images are identified by using a trained pine wood nematode disease identification model based on balance parameters and detection accuracy, which has strong generalization ability and high recognition accuracy.

[0010] To achieve the above object, the present invention is implemented by adopting the following technical solutions:

[0011] The present invention discloses an early pine wood nematode disease identification method based on balance parameters and detection accuracy, comprising the following steps:

[0012] Acquire an image of pine wood nematode disease to be identified;

[0013] Inputting the pine wood nematode disease image into a trained pine wood nematode disease recognition model based on balance parameters and detection accuracy to obtain a corresponding pine wood nematode disease recognition result;

[0014] The pine wood nematode disease recognition model includes an input module, a backbone network based on FasterCSP, a neck network based on GSCSP and RFB-s, a head network based on the CBAM attention mechanism, and an output module connected in sequence;

[0015] The input module is used to input pine wood nematode disease images;

[0016] The backbone network is used to extract features from the pine wood nematode disease image to obtain multiple output features;

[0017] The neck network is used to fuse and enhance multiple output features to obtain multiple fused features;

[0018] The head network is used to optimize the feature representation of multiple fused features to obtain multiple optimized features;

[0019] The output module is used to output pine wood nematode disease identification results based on multiple optimization features.

[0020] Furthermore, the training method of the pine wood nematode disease recognition model is as follows:

[0021] Get training set and validation set;

[0022] According to the training set, the pre-constructed pine wood nematode disease recognition model is trained based on the BCC loss function until a preset iteration termination condition is met and the trained pine wood nematode disease recognition model is output;

[0023] According to the verification set, the performance of the trained pine wood nematode disease identification model is evaluated based on a preset evaluation function that combines the balance parameter and the detection accuracy. In response to the evaluation function values ​​satisfying the corresponding preset evaluation function thresholds, the training is terminated and the trained pine wood nematode disease identification model is output; otherwise, the step of training the pre-constructed pine wood nematode disease identification model based on the BCC loss function is repeated for iterative training.

[0024] Furthermore, the obtaining of the training set and the validation set comprises the following steps:

[0025] Obtain inspection images of pine forests;

[0026] Screening the inspection images to obtain training pine wood nematode disease images;

[0027] After labeling the training pine wood nematode disease images, the labeling box of each pine wood nematode disease target is obtained, and then a data set with labeling boxes is obtained;

[0028] The data set is divided into a training set and a validation set based on a preset division ratio.

[0029] Furthermore, the preset evaluation function combining the balance parameter and the detection accuracy includes an average precision function and an FPS function when the intersection-over-union ratio is 0.5;

[0030] The expression of the average precision function when the intersection-over-union ratio is 0.5 is as follows:

[0031]

[0032] Where mAP@0.5 represents the mean precision function when the intersection-over-union ratio is 0.5; C represents the number of categories of pine wood nematode diseases; AP i represents the average accuracy of the model recognition of the i-th pine wood nematode disease category;

[0033] The FPS function is used to represent the number of frames per second, and the expression is as follows:

[0034]

[0035] In the formula, T represents the average time it takes for the model to complete recognition of all input images.

[0036] Furthermore, the expression of the intersection-over-union ratio is as follows:

[0037]

[0038] Where, IoU represents intersection over union; the predicted box area represents the area of ​​the model predicted box of the pine wood nematode disease target; the true box area represents the area of ​​the labeled box of the pine wood nematode disease target;

[0039] The average precision AP of the model recognition of the i-th pine wood nematode disease category i The expression is as follows:

[0040]

[0041] Where Precision i Represents the precision of the i-th pine wood nematode disease category; Recall i represents the recall rate of the i-th pine wood nematode disease category; Precision i (Recall i ) represents a curve plotted with precision as the y-axis and recall as the x-axis; dRecall i Recall i The partial derivative of

[0042] The expressions of the precision rate Precision and recall rate Recall are as follows:

[0043]

[0044] Where TP represents the number of targets that are actually infected with pine wilt and are identified as such by the model; FP represents the number of targets that are actually not infected with pine wilt and are identified as such by the model; and FN represents the number of targets that are actually infected with pine wilt and are not identified as such by the model.

[0045] Furthermore, the expression of the BCC loss function is as follows:

[0046] Loss=Loss bbox +Loss conf +Loss cls

[0047] In the formula, Loss bbox Represents the bounding box loss function; Loss conf Represents the confidence loss function; Loss cls Represents the classification loss function.

[0048] The bounding box loss function Loss bbox The formula is as follows:

[0049] Loss bbox =1-SIOU

[0050]

[0051] Where, SIOU represents the SIOU similarity measure; A represents the area of ​​the model prediction box of the pine wood nematode disease target; B represents the area of ​​the annotation box of the pine wood nematode disease target; ρ(b A , b a ) represents the center point b of the model prediction box of the pine wood nematode disease target A and the center point b of the annotation box of the pine wood nematode disease target B The Euclidean distance between A h represents the width of the model prediction box of the pine wood nematode disease target; A represents the height of the model prediction box of the pine wood nematode disease target; w B The width of the label box representing the target of pine wood nematode disease; h B The height of the label box representing the pine wood nematode disease target;

[0052] The confidence loss function Loss conf The formula is as follows:

[0053]

[0054] In the formula, y represents the label box of the pine wood nematode disease target, indicating whether there is a pine wood nematode disease target; The model representing the pine wood nematode disease target predicts the probability that the labeled box contains the target;

[0055] The classification loss function Loss cls The formula is as follows:

[0056]

[0057] Where C represents the total number of pine wood nematode disease categories; p c represents the probability distribution of the categories of the labeled pine wood nematode disease targets; Represents the probability distribution of the categories of pine wood nematode disease targets predicted by the model.

[0058] Further, the backbone network includes a first CBLr module, a second CBLr module, a first FasterCSP module, a first MaxPool module, a second FasterCSP module, a second MaxPool module, a third FasterCSP module, a third MaxPool module and a fourth FasterCSP module connected in sequence;

[0059] Wherein, the first CBLr module is used to input the pine wood nematode disease image; the second FasterCSP module is used to output the first output feature; the third FasterCSP module is used to output the second output feature; the fourth FasterCSP module is used to output the third output feature;

[0060] The expression of the first FasterCSP module, the second FasterCSP module, the third FasterCSP module, or the fourth FasterCSP module is as follows:

[0061] FasterCSP out

[0062] =Conv(Concat(FasterNetBlock(Conv(FasterCSP in )), Conv(FasterCSP in )))

[0063] Where, FasterCSP out Represents the output data of any FasterCSP module; Conv represents Conv convolution operation; Concat represents Concat connection operation; FasterNetBlock represents FasterNetBlock operation; FasterCSP in Represents the input data of any FasterCSP module;

[0064] The expression of the FasterNetBlock operation is as follows:

[0065] FasterNetBlock(z1)=z1+(Conv(Conv(Pconv(z1))))

[0066] Pconv(z1)=Concat(Conv(x1),x2)

[0067] x1, x2=Split(z1, [k, Ck], Channel)

[0068] In the formula, z1 represents the input data for FasterNetBlock operation; Pconv represents the Pconv partial convolution operation; Concat represents the Concat connection operation; x1 represents the feature with k channels output by the split operation Figure 1 ; x2 indicates that the number of channels output by the split operation is the feature of Ck Figure 2 ; Split represents the split operation; k represents the number of split operation channels; C represents the number of channel dimensions of z1; Channel represents the channel dimension.

[0069] Further, the neck network includes an RFB-s module, a third GSConv module, a first Upsampled module, a first Concat module, a first GSCSP module, a fourth GSConv module, a second Upsampled module, a second Concat module, a second GSCSP module, a fifth GSConv module, a third Concat module, a third GSCSP module, a sixth GSConv module, a fourth Concat module and a fourth GSCSP module connected in sequence, and the first GSCSP module is connected to the third Concat module, and the RFB-s module is connected to the fourth Concat module;

[0070] Among them, the RFB-s module is used to input the first output feature; the first Concat module is also connected to the first GSConv module, and the first GSConv module is used to input the second output feature; the second Concat module is also connected to the second GSConv module, and the second GSConv module is used to input the third output feature; the second GSCSP module is used to output the first fusion feature; the third GSCSP module is used to output the second fusion feature; the fourth GSCSP module is used to output the third fusion feature;

[0071] The expression of the RFB-s module is as follows:

[0072] RFB out =ReLU(ConvLinear(Concat(branchi (RFB in )))×scale+shortcut(RFB in ))

[0073] In the formula, RFB out Represents the output data of the RFB-s module; ReLU represents the ReLU activation function; scale represents the adjustment of ConvLinear(Concat(branch i (RFB in )))'s impact ratio; shortcut represents a skip connection operation; ConvLinear represents a convolution and full connection operation; Concat represents a Concat connection operation; branch i For RFB in The operation of the i-th branch of RFB, the value of i is 0, 1, 2; branch0 represents the operation of RFB in Convolution with a kernel size of 1×1, convolution with a kernel size of 3×3, and convolution with a kernel size of 3×3 and a dilation rate of 1 are performed in sequence; branch1 represents RFB in Convolution with a kernel size of 1×1, convolution with a kernel size of 3×3, and convolution with a kernel size of 3×3 and a dilation rate of 3 are performed in sequence; branch2 represents RFB in Convolution with a kernel size of 1×1, convolution with a kernel size of 3×3, and convolution with a kernel size of 3×3 and a dilation rate of 5 are performed in sequence; RFB in Represents the input data of the RFB-s module;

[0074] The expression of the first GSCSP module or the second GSCSP module or the third GSCSP module or the fourth GSCSP module is as follows:

[0075] GSCSP out =Conv(Concat(GSBottleneck(Conv(GSCSP in )), Conv(GSCSP in )))

[0076] In the formula, GSCSP out Represents the output data of any GSCSP module; Conv represents the Conv convolution operation; GSCSP in Represents the input data of any GSCSP module; GSBottleneck represents the GSBottleneck operation, and the expression is as follows:

[0077] GSBottleneck(z2)=z2+GSConv(GSConv(z2))

[0078] Where GSBottleneck(z2) represents the output data of the GSBottleneck operation on the input data z2; z2 represents the input data of the GSBottleneck operation; GSConv represents the GSConv operation. The expression is as follows:

[0079] GSConv(z3)=Shuffle(Concat(Conv(z3),DWConv(Conv(z3)))

[0080] Wherein, GSConv(z3) represents the output data of GSConv operation on input data z3; z3 represents the input data of GSConv operation; Shuffle represents the shuffle operation on the channel; DWConv represents the separation channel convolution operation.

[0081] Furthermore, the head network includes:

[0082] A first CBAM module is used to optimize the feature representation of the input first fusion feature based on the CBAM attention mechanism, and output an initial first optimized feature;

[0083] a seventh GSConv module, used to perform a global sparse convolution operation on the initial first optimized feature, and output a final first optimized feature;

[0084] The second CBAM module optimizes the feature representation of the input second fusion feature based on the CBAM attention mechanism and outputs the initial second optimized feature;

[0085] an eighth GSConv module, used to perform a global sparse convolution operation on the initial second optimized feature and output a final second optimized feature;

[0086] The third CBAM module optimizes the feature representation of the input third fusion feature based on the CBAM attention mechanism and outputs the initial third optimized feature;

[0087] a ninth GSConv module, configured to perform a global sparse convolution operation on the initial third optimized feature and output a final third optimized feature;

[0088] The CBAM attention mechanism combines the CAM channel attention mechanism and the SAM spatial attention mechanism, and the expression is as follows:

[0089] CBAM out =CBAM in +MC (CBAM in )+M s (CBAM in +M C (CBAM in ))

[0090] Where CBAM out Represents the output data of any CBAM module; CBAM in Represents the input data of any CBAM module; M C represents the CAM channel attention mechanism; M s Represents the SAM spatial attention mechanism.

[0091] Furthermore, the expression of the CAM channel attention mechanism is as follows:

[0092] M C (CBAM in )=σ(MLP(AvgPool(CBAM in ))+MLP(MaxPool(CBAM in )))

[0093] Where M C (CBAM in ) represents the output data after being processed by the CAM channel attention mechanism; σ represents the Sigmoid activation function; MLP represents the fully connected operation; AvgPool represents the average pooling operation; MaxPool represents the maximum pooling operation;

[0094] M s (CBAM in )=σ(f 7×7 ([AvgPool(CBAM in ); MaxPool(CBAM in )]))

[0095] Where M s (CBAM in ) represents the output data after being processed by the SAM spatial attention mechanism; f 7×7 Represents a convolution operation with a convolution kernel size of 7×7.

[0096] Compared with the prior art, the present invention has the following beneficial effects:

[0097] The early pine wood nematode disease recognition method based on balanced parameters and detection accuracy of the present invention uses a trained pine wood nematode disease recognition model based on balanced parameters and detection accuracy to identify the pine wood nematode disease image, with strong generalization ability and high recognition accuracy. Among them, the backbone network based on the model uses FasterCSP to balance the lightness and real-time performance of the model. The neck network based on the model uses GSCSP to reduce the complexity of the model and improve the detection accuracy. The receptive field of the model is improved by RFB-s, and the long-distance information collection ability of the model is enhanced. The head network based on the model uses the CBAM attention mechanism, and by analyzing the global spatial information of each channel, the importance of different channels is learned, and the attention to spatial details is strengthened, so that the model can better focus on small target pine wood nematode diseases, and further improve the detection performance of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0098] Figure 1 is a schematic flow chart of the method for identifying early pine wood nematode disease based on balance parameters and detection accuracy provided in Example 1;

[0099] Figure 2 is a schematic diagram of the structure of the backbone network provided in Example 1;

[0100] Figure 3 is a schematic diagram of the structure of the neck network provided in Example 1;

[0101] Figure 4 is a schematic diagram of the structure of the head network provided in Example 1;

[0102] Figure 5 This is an example diagram of the model output image provided in Example 1. DETAILED DESCRIPTION

[0103] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0104] Example 1

[0105] This embodiment 1 provides an early pine wood nematode disease identification method based on balance parameters and detection accuracy, comprising the following steps:

[0106] Acquire an image of pine wood nematode disease to be identified;

[0107] The pine wood nematode disease image is input into the trained pine wood nematode disease recognition model based on the balance parameter and the detection accuracy to obtain the corresponding pine wood nematode disease recognition result;

[0108] Among them, the pine wood nematode disease recognition model includes an input module connected in sequence, a backbone network based on FasterCSP, a neck network based on GSCSP and RFB-s, a head network based on the CBAM attention mechanism, and an output module;

[0109] An input module, used to input pine wood nematode disease images;

[0110] The backbone network is used to extract features from pine wood nematode disease images and obtain multiple output features;

[0111] The neck network is used to fuse and enhance multiple output features to obtain multiple fused features;

[0112] The head network is used to optimize the feature representation of multiple fused features to obtain multiple optimized features;

[0113] The output module is used to output the pine wood nematode disease identification results based on multiple optimization features.

[0114] The technical concept of the present invention is: using a trained pine wood nematode disease recognition model based on balanced parameters and detection accuracy to identify pine wood nematode disease images, with strong generalization ability and high recognition accuracy. Among them, based on the backbone network of the model, FasterCSP is used to balance the lightness and real-time performance of the model. Based on the neck network of the model, GSCSP is used to reduce the complexity of the model and improve the detection accuracy. RFB-s is used to improve the receptive field of the model and enhance the long-distance information collection ability of the model. Based on the head network of the model, the CBAM attention mechanism is used to analyze the global spatial information of each channel, learn the importance of different channels, strengthen the attention to spatial details, enable the model to better focus on small target pine wood nematode diseases, and further improve the detection performance of the model.

[0115] like Figure 1 As shown in the figure, the specific steps of the early pine wood nematode disease identification method based on balance parameters and detection accuracy are as follows:

[0116] Step 1: Pre-construction of the model.

[0117] A pre-built pine wood nematode disease recognition model is constructed, including a sequentially connected input module, a FasterCSP-based backbone network, a GSCSP and RFB-s-based neck network, a CBAM attention mechanism-based head network, and an output module.

[0118] 1.1. Input module, used to input pine wood nematode disease images.

[0119] 1.2. The backbone network is used to extract features from pine wood nematode disease images and obtain multiple output features.

[0120] like Figure 2 As shown, the backbone network includes a first CBLr module, a second CBLr module, a first FasterCSP module, a first MaxPool module, a second FasterCSP module, a second MaxPool module, a third FasterCSP module, a third MaxPool module and a fourth FasterCSP module which are connected in sequence;

[0121] Among them, the first CBLr module is used to input the pine wood nematode disease image; the second FasterCSP module is used to output the first output feature; the third FasterCSP module is used to output the second output feature; the fourth FasterCSP module is used to output the third output feature;

[0122] The expression of the first FasterCSP module or the second FasterCSP module or the third FasterCSP module or the fourth FasterCSP module is as follows:

[0123] FasterCSP out

[0124] =Conv(Concat(FasterNetBlock(Conv(FasterCSP in )), Conv(FasterCSP in )))

[0125] Where, FasterCSP out Represents the output data of any FasterCSP module; Conv represents Conv convolution operation; Concat represents Concat connection operation; FasterNetBlock represents FasterNetBlock operation; FasterCSP in Represents the input data of any FasterCSP module;

[0126] The expression of FasterNetBlock operation is as follows:

[0127] FasterNetBlock(z1)=z1+(Conv(Conv(Pconv(z1))))

[0128] Pconv(z1)=Concat(Conv(x1),x2)

[0129] x1, x2=Split(z1, [k, Ck], Channel)

[0130] In the formula, z1 represents the input data for FasterNetBlock operation; Pconv represents the Pconv partial convolution operation; Concat represents the Concat connection operation; x1 represents the feature with k channels output by the split operation Figure 1 ; x2 indicates that the number of channels output by the split operation is the feature of Ck Figure 2 ; Split represents the split operation; k represents the number of split operation channels; C represents the number of channel dimensions of z1; Channel represents the channel dimension.

[0131] The innovation of the backbone network in this embodiment mainly lies in the application of the FasterCSP module, which consists of a partial convolution PConv operation and a FasterNetBlock operation. The partial convolution PConv operation only performs a conventional convolution operation on part of the input channels, thereby balancing the lightness and real-time performance of the model.

[0132] 1.3. Neck network, used to fuse and enhance multiple output features to obtain multiple fused features;

[0133] like Figure 3 As shown, the neck network includes an RFB module, a third GSConv module, a first Upsampled module, a first Concat module, a first GSCSP module, a fourth GSConv module, a second Upsampled module, a second Concat module, a second GSCSP module, a fifth GSConv module, a third Concat module, a third GSCSP module, a sixth GSConv module, a fourth Concat module and a fourth GSCSP module connected in sequence, and the first GSCSP module is connected to the third Concat module, and the RFB-s module is connected to the fourth Concat module;

[0134] Among them, the RFB module is used to input the first output feature; the first Concat module is also connected to the first GSConv module, and the first GSConv module is used to input the second output feature; the second Concat module is also connected to the second GSConv module, and the second GSConv module is used to input the third output feature; the second GSCSP module is used to output the first fusion feature; the third GSCSP module is used to output the second fusion feature; the fourth GSCSP module is used to output the third fusion feature;

[0135] The expression of the RFB-s module is as follows:

[0136] RFB out =ReLU(ConvLinear(Concat(branch i (RFB in)))×scale+shortcut(RFB in ))

[0137] In the formula, RFB out Represents the output data of the RFB-s module; ReLU represents the ReLU activation function; scale represents the adjustment of ConvLinear(Concat(branch i (RFB in )))'s impact ratio; shortcut represents a skip connection operation; ConvLinear represents a convolution and full connection operation; Concat represents a Concat connection operation; branch i For RFB in The operation of the i-th branch of RFB, the value of i is 0, 1, 2; branch0 represents the operation of RFB in Convolution with a kernel size of 1×1, convolution with a kernel size of 3×3, and convolution with a kernel size of 3×3 and a dilation rate of 1 are performed in sequence; branch1 represents RFB in Convolution with a kernel size of 1×1, convolution with a kernel size of 5×5, and convolution with a kernel size of 3×3 and a dilation rate of 3 are performed in sequence; branch2 represents RFB in Convolution with a kernel size of 1×1, convolution with a kernel size of 3×3, and convolution with a kernel size of 3×3 and a dilation rate of 5 are performed in sequence; RFB in Represents the input data of the RFB-s module;

[0138] The expression of the first GSCSP module or the second GSCSP module or the third GSCSP module or the fourth GSCSP module is as follows:

[0139] GSCSP out =Conv(Concat(GSBottleneck(Conv(GSCSP in )), Conv(GSCSP in )))

[0140] In the formula, GSCSP out Represents the output data of any GSCSP module; Conv represents the Conv convolution operation; GSCSP in Represents the input data of any GSCSP module; GSBottleneck represents the GSBottleneck operation, and the expression is as follows:

[0141] GSBottleneck(z2)=z2+GSConv(GSConv(z2))

[0142] Where GSBottleneck(z2) represents the output data of the GSBottleneck operation on the input data z2; z2 represents the input data of the GSBottleneck operation; GSConv represents the GSConv operation. The expression is as follows:

[0143] GSConv(z3)=Shuffle(Concat(Conv(z3),DWConv(Conv(z3)))

[0144] Wherein, GSConv(z3) represents the output data of GSConv operation on input data z3; z3 represents the input data of GSConv operation; Shuffle represents the shuffle operation on the channel; DWConv represents the separation channel convolution operation.

[0145] The innovation of the neck network in this embodiment mainly lies in the application of the RFB-s module and the GSCSP module. The RFB-s module mainly uses dilated convolution to improve the receptive field of the model and improve the model's ability to obtain contextual information. Any GSCSP module is a lightweight convolution module built on GSConv, which further reduces the number of model parameters while maintaining a high detection accuracy.

[0146] 1.4. Head network, used to optimize the feature representation of multiple fusion features to obtain multiple optimized features;

[0147] like Figure 4 As shown, the head network includes:

[0148] A first CBAM module is used to optimize the feature representation of the input first fusion feature based on the CBAM attention mechanism, and output an initial first optimized feature;

[0149] A seventh GSConv module, used for performing a global sparse convolution operation on the initial first optimized feature, and outputting a final first optimized feature;

[0150] The second CBAM module optimizes the feature representation of the input second fusion feature based on the CBAM attention mechanism and outputs the initial second optimized feature;

[0151] An eighth GSConv module, used to perform a global sparse convolution operation on the initial second optimized features and output a final second optimized feature;

[0152] The third CBAM module optimizes the feature representation of the input third fusion feature based on the CBAM attention mechanism and outputs the initial third optimized feature;

[0153] A ninth GSConv module, used for performing a global sparse convolution operation on the initial third optimized feature, and outputting a final third optimized feature;

[0154] The CBAM attention mechanism combines the CAM channel attention mechanism and the SAM spatial attention mechanism, and the expression is as follows:

[0155] CBAM out =CBAM in +M C (CBAM in )+M s (CBAM in +M C (CBAM in ))

[0156] Where CBAM out Represents the output data of any CBAM module; CBAM in Represents the input data of any CBAM module; M C represents the CAM channel attention mechanism; M s Represents the SAM spatial attention mechanism.

[0157] The expression of CAM channel attention mechanism is as follows:

[0158] M C (CBAM in )=σ(MLP(AvgPool(CBAM in ))+MLP(MaxPool(CBAM in )))

[0159] Where M C (CBAM in ) represents the output data after being processed by the CAM channel attention mechanism; σ represents the Sigmoid activation function; MLP represents the fully connected operation; AvgPool represents the average pooling operation; MaxPool represents the maximum pooling operation;

[0160] M s (CBAM in )=σ(f 7×7 ([AvgPool(CBAM in ); MaxPool(CBAM in )]))

[0161] Where M s (CBAM in ) represents the output data after being processed by the SAM spatial attention mechanism; f 7×7 Represents a convolution operation with a convolution kernel size of 7×7.

[0162] The innovation of the head network in this embodiment mainly lies in the CBAM module based on the CBAM attention mechanism, which includes CAM (Channel Attention Module, CAM) channel attention and SAM (SpatialAttention Module, SAM) spatial attention mechanism. Among them, the CAM channel attention mechanism can learn the importance of different channels by analyzing the global spatial information of each channel; the spatial attention module mechanism can strengthen the attention to spatial details and further improve the detection performance of the model by focusing on important spatial areas in the image.

[0163] 1.5. Output module, used to output pine wood nematode disease recognition results based on multiple optimized features. The recognition results are the results of the model detecting early pine wood nematode disease and late pine wood nematode disease on the input recognition image. The example is shown in the figure below. Figure 5 shown.

[0164] Step 2: Model training.

[0165] 2.1. Preprocessing of dataset.

[0166] Use the visible light camera on the drone to obtain inspection images of the pine forest in a certain area.

[0167] According to the inspection images, screening is performed to obtain training pine wood nematode disease images.

[0168] After labeling the training pine wood nematode disease images, the labeling box of each pine wood nematode disease target is obtained, and then a data set with labeling boxes is obtained.

[0169] The dataset with annotated boxes is divided into training set, validation set and test set in a ratio of 8:1:1.

[0170] The latitude and longitude of the specific area are 31°55'38.70"-31°55'54.65"N, 118°46'16.56"-118°46'40.03"E. The area has typical northern subtropical humid climate characteristics, with an annual precipitation of 819.8-971.8mm and an annual average temperature of 16.9-18.0℃. The main tree species in the forest in the study area is Pinus massoniana.

[0171] 2.2. Set the training parameters of the model, and train the pre-built pine wood nematode disease recognition model based on the training set obtained in step 2.1 based on the BBC loss function until the preset iteration termination condition is met and the trained pine wood nematode disease recognition model is output.

[0172] The training parameters of the model include input image size, number of training rounds, number of batch images, optimizer, initial learning rate, and mosaic data enhancement probability. Specifically, the input image size is 640×640 pixels; the number of training rounds is 200 rounds; the number of batch images is 32; the optimizer used is SGD; the initial learning rate is 0.001; and the mosaic data enhancement probability is 100%. The preset iteration termination condition is to reach the preset number of training rounds.

[0173] The expression of the BBC loss function used in this embodiment is as follows:

[0174] Loss=Loss bbox +Loss conf +LoSS cls

[0175] In the formula, Loss bbox Represents the bounding box loss function; Loss conf Represents the confidence loss function; Loss cls Represents the classification loss function.

[0176] Bounding Box Loss bbox The formula is as follows:

[0177] Loss bbos =1-SIOU

[0178]

[0179] Where, SIOU represents the SIOU similarity measure; A represents the area of ​​the model prediction box of the pine wood nematode disease target; B represents the area of ​​the annotation box of the pine wood nematode disease target; ρ(b A ,b a ) represents the center point b of the model prediction box of the pine wood nematode disease target A and the center point b of the annotation box of the pine wood nematode disease target B The Euclidean distance between A h represents the width of the model prediction box of the pine wood nematode disease target; A represents the height of the model prediction box of the pine wood nematode disease target; w B The width of the label box representing the target of pine wood nematode disease; h B The height of the annotation box representing the pine wood nematode disease target.

[0180] Confidence loss function Loss conf The formula is as follows:

[0181]

[0182] In the formula, y represents the label box of the pine wood nematode disease target, indicating whether there is a pine wood nematode disease target; The model representing the pine wood nematode disease target predicts the probability that the labeled box contains the target.

[0183] The classification loss function Loss cls The formula is as follows:

[0184]

[0185] Where C represents the total number of pine wood nematode disease categories; p c represents the probability distribution of the categories of the labeled pine wood nematode disease targets; Represents the probability distribution of the categories of pine wood nematode disease targets predicted by the model.

[0186] 2.3. According to the validation set obtained in step 2.1, the performance of the trained pine wood nematode disease recognition model is evaluated based on the preset evaluation function that combines the balance parameter and the detection accuracy. If the evaluation function values ​​meet the corresponding preset evaluation function thresholds, the training is terminated and the trained pine wood nematode disease recognition model is output; otherwise, the training is repeated back to step 2.2 to iteratively train the pre-built pine wood nematode disease recognition model based on the BBC loss function.

[0187] Among them, the preset evaluation function combining the balance parameter and the detection accuracy includes the average precision function and the FPS function when the intersection-over-union ratio is 0.5, as follows:

[0188] 2.3.1. The expression of the average precision function when the intersection-over-union ratio is 0.5 is as follows:

[0189]

[0190] Where mAP@0.5 represents the mean precision function when the intersection-over-union ratio is 0.5; C represents the number of categories of pine wood nematode diseases; AP i Represents the average accuracy of the model recognition of the i-th pine wood nematode disease category.

[0191] The expression of intersection-over-union ratio is as follows:

[0192]

[0193] Where IoU represents intersection over union; the predicted box area represents the area of ​​the model predicted box of the pine wood nematode disease target; and the true box area represents the area of ​​the labeled box of the pine wood nematode disease target.

[0194] In this embodiment, the intersection-over-union ratio of 0.5 is used in the average precision function, which refers to the evaluation standard of the PASCAL VOC data set.

[0195] Average precision AP of model recognition of the i-th pine wood nematode disease category i The expression is as follows:

[0196]

[0197] Where Precision i Represents the precision of the i-th pine wood nematode disease category; Recall i represents the recall rate of the i-th pine wood nematode disease category; Precision i (Recall i ) represents a curve plotted with precision as the y-axis and recall as the x-axis; dRecall i Recall i The partial derivative of ; i represents the pine wood nematode disease category and can only be 0 or 1; i = 0 represents early pine wood nematode disease (PWD_P); i = 1 represents late pine wood nematode disease (PWD_A).

[0198] The expressions of precision and recall are as follows:

[0199]

[0200] Where TP represents the number of targets that are actually infected with pine wilt and are identified as such by the model; FP represents the number of targets that are actually not infected with pine wilt and are identified as such by the model; and FN represents the number of targets that are actually infected with pine wilt and are not identified as such by the model.

[0201] 2.3.2, FPS function is used to represent the number of frames per second, the expression is as follows:

[0202]

[0203] Where T represents the average time it takes for the model to complete recognition of all input images divided by the number of images.

[0204] 2.3.3. The preset evaluation function threshold in this embodiment is:

[0205] The average accuracy function reaches 75% and the FPS function reaches 100 to pass.

[0206] Step 3: Practical application of the model.

[0207] Acquire an image of pine wood nematode disease to be identified;

[0208] The pine wood nematode disease image is input into the trained pine wood nematode disease recognition model based on balance parameters and detection accuracy to obtain the corresponding pine wood nematode disease recognition result.

[0209] Example 2

[0210] This embodiment 2 detects and identifies the captured pine wood nematode disease images based on the early pine wood nematode disease identification method based on balance parameters and detection accuracy provided in embodiment 1.

[0211] The model proposed in the present invention and other widely used models, including YOLOv5, YOLOv7, Faster R-CNN and SSD, are respectively trained using the data set constructed by the present invention, and then the trained models are used to predict the pine wood nematode disease images in the test set constructed by the present invention.

[0212] Model mAP@0.5 <![CDATA[AP0]]> <![CDATA[AP1]]> FPS YOLOv5 78.5％ 71.9％ 85.1％ 141 YOLOv7 78.9％ 72.4％ 85.4％ 145 Faster R-CNN 70.2％ 64.2％ 76.2％ 55 SSD 69.3％ 61.4％ 77.2％ 72 Model of the present invention 80.9％ 73.9％ 87.9％ 133

[0213] It can be seen that compared with other widely used recognition models, the pine wood nematode disease recognition model based on balanced parameters and detection accuracy trained by the present invention is used to recognize pine wood nematode disease images, which has strong generalization ability and high recognition accuracy.

[0214] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0215] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0216] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0217] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0218] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. An early pine wood nematode disease identification method based on balance parameters and detection accuracy, characterized in that: The steps include: Acquire an image of pine wood nematode disease to be identified; Inputting the pine wood nematode disease image into a trained pine wood nematode disease recognition model based on balance parameters and detection accuracy to obtain a corresponding pine wood nematode disease recognition result; The pine wood nematode disease recognition model includes an input module, a backbone network based on FasterCSP, a neck network based on GSCSP and RFB-s, a head network based on the CBAM attention mechanism, and an output module connected in sequence; The input module is used to input pine wood nematode disease images; The backbone network is used to extract features from the pine wood nematode disease image to obtain multiple output features; The neck network is used to fuse and enhance multiple output features to obtain multiple fused features; The head network is used to optimize the feature representation of multiple fused features to obtain multiple optimized features; The output module is used to output the pine wood nematode disease identification result according to the multiple optimization features; The principle expression of FasterCSP is as follows: FasterCSP out =Conv(Concat(FasterNetBlock(Conv(FasterCSP in )),Conv(FasterCSP in ))); Where, FasterCSP out Represents the output data of any FasterCSP module; Conv represents Conv convolution operation; Concat represents Concat connection operation; FasterNetBlock represents FasterNetBlock operation; FasterCSP in Represents the input data of any FasterCSP module; The expression of the FasterNetBlock operation is as follows: FasterNetBlock(z1)=z1+(Conv(Conv(Pconv(z1)))); Pconv(z1)=Concat(Conv(x1),x2); x1, x2=Split(z1, [k, Ck], Channel); In the formula, z1 represents the input data for FasterNetBlock operation; Pconv represents the Pconv partial convolution operation; Concat represents the Concat connection operation; x1 represents the feature map 1 with the number of channels k output by the split operation; x2 represents the feature map 2 with the number of channels Ck output by the split operation; Split represents the split operation; k represents the number of channels for the split operation; C represents the number of channel dimensions of z1; Channel represents the channel dimension; The principle expression of the GSCSP is as follows: GSCSP out =Conv(Concat(GSBottleneck(Conv(GSCSP in )),Conv(GSCSP in ))); In the formula, GSCSP out Represents the output data of any GSCSP module; Conv represents the Conv convolution operation; GSCSP in Represents the input data of any GSCSP module; GSBottleneck represents the GSBottleneck operation, and the expression is as follows: GSBottleneck(z2)=z2+GSConv(GSConv(z2)); Where GSBottleneck(z2) represents the output data of the GSBottleneck operation on the input data z2; z2 represents the input data of the GSBottleneck operation; GSConv represents the GSConv operation. The expression is as follows: GSConv(z3)=Shuffle(Concat(Conv(z3),DWConv(Conv(z3)))); Where, GSConv(z3) represents the output data of GSConv operation on input data z3; z3 represents the input data of GSConv operation; Shuffle represents shuffling operation on channels; DWConv represents separation channel convolution operation; The principle expression of the RFB-s is as follows: RFB out =ReLU(ConvLinear(Concat(branch i (RFB in )))×scale+shortcut(RFB in )); In the formula, RFB out Represents the output data of the RFB-s module; ReLU represents the ReLU activation function; scale represents the adjustment of ConvLinear(Concat(branch i (RFB in )))'s impact ratio; shortcut represents a skip connection operation; ConvLinear represents a convolution and full connection operation; Concat represents a Concat connection operation; branch i For RFB in The operation of the i-th branch of RFB, the value of i is 0, 1, 2; branch0 represents the operation of RFB in Convolution with a kernel size of 1×1, convolution with a kernel size of 3×3, and convolution with a kernel size of 3×3 and a dilation rate of 1 are performed in sequence; branch1 represents RFB in Convolution with a kernel size of 1×1, convolution with a kernel size of 3×3, and convolution with a kernel size of 3×3 and a dilation rate of 3 are performed in sequence; branch2 represents RFB in Convolution with a kernel size of 1×1, convolution with a kernel size of 3×3, and convolution with a kernel size of 3×3 and a dilation rate of 5 are performed in sequence; RFB in Represents the input data of the RFB-s module.

2. The early pine wood nematode disease identification method based on balance parameters and detection accuracy according to claim 1 is characterized in that: The training method of the pine wood nematode disease recognition model is as follows: Get training set and validation set; According to the training set, the pre-constructed pine wood nematode disease recognition model is trained based on the BCC loss function until a preset iteration termination condition is met and the trained pine wood nematode disease recognition model is output; According to the validation set, the performance of the trained pine wood nematode disease recognition model is evaluated based on a preset evaluation function combining the balance parameter and the detection accuracy. In response to the evaluation function values ​​satisfying the corresponding preset evaluation function thresholds, the training is terminated and the trained pine wood nematode disease recognition model is output; otherwise, the training is repeated to return to the step of training the pre-constructed pine wood nematode disease recognition model based on the BCC loss function for iterative training; The BCC loss function includes a bounding box loss function, a confidence loss function and a classification loss function.

3. The early pine wood nematode disease identification method based on balance parameters and detection accuracy according to claim 2 is characterized in that: The step of obtaining a training set and a validation set comprises the following steps: Obtain inspection images of pine forests; Screening the inspection images to obtain training pine wood nematode disease images; After labeling the training pine wood nematode disease images, the labeling box of each pine wood nematode disease target is obtained, and then a data set with labeling boxes is obtained; The data set is divided into a training set and a validation set based on a preset division ratio.

4. The early pine wood nematode disease identification method based on balance parameters and detection accuracy according to claim 2 is characterized in that: The preset evaluation function combining the balance parameter and the detection accuracy includes an average precision function and an FPS function when the intersection-over-union ratio is 0.5; The expression of the average precision function when the intersection-over-union ratio is 0.5 is as follows: Where mAP@0.5 represents the mean precision function when the intersection-over-union ratio is 0.5; C represents the number of categories of pine wood nematode diseases; AP i represents the average accuracy of the model recognition of the i-th pine wood nematode disease category; The FPS function is used to represent the number of frames per second, and the expression is as follows: In the formula, T represents the average time it takes for the model to complete recognition of all input images.

5. The early pine wood nematode disease identification method based on balance parameters and detection accuracy according to claim 4 is characterized in that: The expression of the intersection-over-union ratio is as follows: Where, IoU represents intersection over union; the predicted box area represents the area of ​​the model predicted box of the pine wood nematode disease target; the true box area represents the area of ​​the labeled box of the pine wood nematode disease target; The average precision AP of the model recognition of the i-th pine wood nematode disease category i The expression is as follows: Where Precision i Represents the precision of the i-th pine wood nematode disease category; Recall i represents the recall rate of the i-th pine wood nematode disease category; Precision i (Recall i ) represents a curve plotted with precision as the y-axis and recall as the x-axis; dRecall i Recall i The partial derivative of The expressions of the precision rate Precision and recall rate Recall are as follows: Where TP represents the number of targets that are actually infected with pine wilt and are identified as such by the model; FP represents the number of targets that are actually not infected with pine wilt and are identified as such by the model; and FN represents the number of targets that are actually infected with pine wilt and are not identified as such by the model.

6. The early pine wood nematode disease identification method based on balance parameters and detection accuracy according to claim 2 is characterized in that: The expression of the BCC loss function is as follows: Loss=Loss bbox +Loss conf +Loss cls ; In the formula, Loss bbox Represents the bounding box loss function; Loss conf Represents the confidence loss function; Loss cls represents the classification loss function; The bounding box loss function Loss bbox The formula is as follows: Loss bbox =1-SIOU; Where, SIOU represents the SIOU similarity measure; A represents the area of ​​the model prediction box of the pine wood nematode disease target; B represents the area of ​​the annotation box of the pine wood nematode disease target; ρ(b A , b a ) represents the center point b of the model prediction box of the pine wood nematode disease target A and the center point b of the annotation box of the pine wood nematode disease target B The Euclidean distance between A h represents the width of the model prediction box of the pine wood nematode disease target; A represents the height of the model prediction box of the pine wood nematode disease target; w B The width of the label box representing the target of pine wood nematode disease; h B The height of the label box representing the pine wood nematode disease target; The confidence loss function Loss conf The formula is as follows: In the formula, y represents the label box of the pine wood nematode disease target, indicating whether there is a pine wood nematode disease target; The model representing the pine wood nematode disease target predicts the probability that the labeled box contains the target; The classification loss function Loss cls The formula is as follows: Where C represents the total number of pine wood nematode disease categories; p c represents the probability distribution of the categories of the labeled pine wood nematode disease targets; Represents the probability distribution of the categories of pine wood nematode disease targets predicted by the model.

7. The early pine wood nematode disease identification method based on balance parameters and detection accuracy according to claim 1 is characterized in that: The backbone network includes a first CBLr module, a second CBLr module, a first FasterCSP module, a first MaxPool module, a second FasterCSP module, a second MaxPool module, a third FasterCSP module, a third MaxPool module and a fourth FasterCSP module connected in sequence; Among them, the first CBLr module is used to input pine wood nematode disease images; the second FasterCSP module is used to output first output features; the third FasterCSP module is used to output second output features; and the fourth FasterCSP module is used to output third output features.

8. The method for identifying early pine wood nematode disease based on balance parameters and detection accuracy according to claim 7, characterized in that: The neck network includes an RFB-s module, a third GSConv module, a first Upsampled module, a first Concat module, a first GSCSP module, a fourth GSConv module, a second Upsampled module, a second Concat module, a second GSCSP module, a fifth GSConv module, a third Concat module, a third GSCSP module, a sixth GSConv module, a fourth Concat module and a fourth GSCSP module connected in sequence, and the first GSCSP module is connected to the third Concat module, and the RFB-s module is connected to the fourth Concat module; Among them, the RFB-s module is used to input the first output feature; the first Concat module is also connected to the first GSConv module, and the first GSConv module is used to input the second output feature; the second Concat module is also connected to the second GSConv module, and the second GSConv module is used to input the third output feature; the second GSCSP module is used to output the first fusion feature; the third GSCSP module is used to output the second fusion feature; the fourth GSCSP module is used to output the third fusion feature.

9. The method for identifying early pine wood nematode diseases based on balance parameters and detection accuracy according to claim 8, characterized in that: The head network includes: A first CBAM module is used to optimize the feature representation of the input first fusion feature based on the CBAM attention mechanism, and output an initial first optimized feature; a seventh GSConv module, used to perform a global sparse convolution operation on the initial first optimized feature, and output a final first optimized feature; The second CBAM module optimizes the feature representation of the input second fusion feature based on the CBAM attention mechanism and outputs the initial second optimized feature; an eighth GSConv module, used to perform a global sparse convolution operation on the initial second optimized feature and output a final second optimized feature; The third CBAM module optimizes the feature representation of the input third fusion feature based on the CBAM attention mechanism and outputs the initial third optimized feature; a ninth GSConv module, configured to perform a global sparse convolution operation on the initial third optimized feature and output a final third optimized feature; The CBAM attention mechanism combines the CAM channel attention mechanism and the SAM spatial attention mechanism, and the expression is as follows: CBAM out =CBAM in +M C (CBAM in )+M s (CBAM in +M C (CBAM in )); Where CBAM out Represents the output data of any CBAM module; CBAM in Represents the input data of any CBAM module; M C represents the CAM channel attention mechanism; M s Represents the SAM spatial attention mechanism.

10. The early pine wood nematode disease identification method based on balance parameters and detection accuracy according to claim 9, characterized in that: The expression of the CAM channel attention mechanism is as follows: M C (CBAM in )=σ(MLP(AvgPool(CBAM in ))+MLP(MaxPool(CBAM in ))); Where M C (CBAM in ) represents the output data after being processed by the CAM channel attention mechanism; σ represents the Sigmoid activation function; MLP represents the fully connected operation; AvgPool represents the average pooling operation; MaxPool represents the maximum pooling operation; M s (CBAM in )=σ(f 7×7 ([AvgPool(CBAM in );MaxPool(CBAM in )])) Where M s (CBAM in ) represents the output data after being processed by the SAM spatial attention mechanism; f 7×7 Represents a convolution operation with a convolution kernel size of 7×7.

Images