Soybean salt tolerance degree identification method and system based on SCSFAMNet model
Through the identification method of soybean salt tolerance degree based on the SCSFAM_Net model, the Transformer self-attention mechanism and the feature extraction technology of the SCSFAM module are used to solve the accuracy and efficiency of the identification of salt tolerance of soybean germplasm resources in the existing technology, and achieve a more efficient and accurate identification effect.
Patent Information
- Application Number
- CN202510211471.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-25
- Publication Date
- 2025-06-10
AI Technical Summary
The existing methods for identifying salt tolerance of soybean germplasm resources have problems such as low flux, poor accuracy, time-consuming and labor-intensive and susceptible to subjective factors, making it difficult to effectively screen salt-tolerant soybean germplasm resources.
The soybean salt tolerance degree recognition method and system is adopted based on the SCSFAM_Net model. This method builds a network of Transformer self-attention mechanisms, including the preliminary feature extraction module Stem and four blocks, and uses the SCSFAM module to perform feature extraction of spatial and channel analog self-attention mechanisms, and improves the recognition accuracy through preprocessing, training and testing steps.
It improves the accuracy of identifying the degree of salt tolerance of soybeans, achieves more efficient and accurate salt tolerance identification, reduces the time for redundant feature extraction, and has great prospects in the field of salt tolerance identification of crop seed resources.
Smart Images
Figure CN120126005A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method and system for identifying the salt tolerance degree of soybeans, and in particular to a method and system for identifying the salt tolerance degree of soybeans based on the SCSFAM_Net model, belonging to the technical fields of computer vision image processing and crop level identification. Background Art
[0002] Soybeans (Glycine max L.) originated in China and have a cultivation history of more than 5,000 years. They are the fifth largest food crop after corn, rice, wheat, and potatoes, and are also important oilseed and cash crops. Soybeans are rich in nutrients such as fat, protein, isoflavones, and lecithin, and play an important role in aspects such as food, feed, and processing.
[0003] Saline-alkali land is an important reserve arable land resource. There are a large number of saline-alkali lands globally, and their area is showing a gradually expanding trend. China is rich in soybean germplasm resources, and the salt tolerance of different soybean materials varies. Conducting the identification of the salt tolerance of soybean germplasm resources can provide abundant excellent allelic variations for the cultivation of new salt-tolerant soybean varieties and lay a foundation for cultivating new salt-tolerant soybean varieties.
[0004] There are certain differences in the identification methods, identification indicators, and evaluation methods of salt-tolerant soybean germplasm resources, and there are generally defects such as low throughput, poor accuracy, time-consuming and laborious, and being easily affected by subjective factors, which have seriously hindered the exploration of salt-tolerant soybean germplasm resources. In recent years, phenomics technology characterized by intelligent, high-throughput, and non-destructive precise measurement has developed rapidly, and can achieve the precise identification of crop phenotypes. Establishing an efficient and precise soybean germplasm resource identification technology is of great significance for coping with global food security challenges and promoting sustainable agricultural development. However, the current research on phenotyping technology for identifying the salt tolerance of soybean germplasm resources is still in its infancy. Therefore, studying the identification technology of salt-tolerant soybean germplasm resources based on deep learning technology has important theoretical and practical significance for screening salt-tolerant soybean germplasm resources, cultivating new salt-tolerant soybean varieties, improving soybean yield, and effectively utilizing saline-alkali land. Summary of the Invention
[0005] Object of the Invention: The object of the present invention is to provide a method and system for identifying the salt tolerance degree of soybeans based on the SCSFAM_Net model that can improve the recognition accuracy.
[0006] Technical Solution: A method for identifying the salt tolerance degree of soybeans based on the SCSFAM_Net model according to the present invention includes:
[0007] (1) Collect the pre-acquired data on the salt tolerance level of soybeans and make it into a dataset. Divide the dataset into a training set and a test set according to a preset ratio, and preprocess the training set and the test set respectively.
[0008] (2) Based on the Transformer self-attention mechanism, construct a soybean salt tolerance level recognition network based on the SCSFAM_Net model. The soybean salt tolerance level recognition network based on the SCSFAM_Net model includes a preliminary feature extraction module Stem and four block blocks. The Stem module is used for downsampling and increasing the number of channels. The block block contains a backbone module SCSFAM module for feature extraction using spatial and channel analog self-attention mechanisms.
[0009] (3) Use the training set to train the soybean salt tolerance level recognition network based on the SCSFAM_Net model to obtain a trained model, obtain the optimal model weights, test the trained model with the test set, and use the tested model to identify the soybeans to be detected.
[0010] Further, the preprocessing in step (1) includes randomly rotating horizontally, cropping the size, and converting the data format for the training set and the test set respectively.
[0011] Further, among the four block blocks in step (2), the first block block includes one SCSFAM module; the second block block includes one SCSFAM module; the third block block includes 3 SCSFAM modules, and the fourth block block includes 2 SCSFAM modules.
[0012] Further, the feature extraction module Stem in step (2) consists of a 7×7 convolutional kernel, a BatchNorm layer, a RELU activation function, and a max pooling layer.
[0013] Further, the SCSFAM module in step (2) includes: a first branch, where the input features pass through a max pooling and a 1×1 convolutional kernel, and an average pooling layer and a 1×1 convolutional kernel respectively. After multiplying the respective obtained features, they pass through the CHSM module. After the CHSM module passes through the global average pooling and the softmax layer, channel sorting and screening are performed according to the weight size, and the channels are cut according to a ratio and then expanded to the original number of channels. Subsequently, they pass through two 1×1 convolutional kernels respectively. After multiplying the respective obtained features, they pass through the softmax, and then directly multiply with the features obtained through the CHSM module.
[0014] The second branch passes through the global max pooling layer and the global average pooling layer respectively, concatenates the obtained features, then passes through two 3×3 convolutional kernels respectively, multiplies the obtained features, and then passes through softmax to multiply the obtained features with the features that have not passed through the 3×3 convolutional kernel after concatenation;
[0015] The third branch multiplies the input features directly with the result of the residual connection of the features obtained from the first branch and the features obtained from the second branch.
[0016] Based on the same inventive concept, the present invention also provides a soybean salt tolerance degree recognition system based on the SCSFAM_Net model, including:
[0017] A preprocessing module, which is used to collect the pre-obtained soybean salt tolerance degree data and make it into a data set, divide the data set into a training set and a test set according to a preset ratio, and preprocess the training set and the test set respectively;
[0018] A network construction module, which is used to construct a soybean salt tolerance degree recognition network based on the SCSFAM_Net model based on the Transformer self-attention mechanism; the soybean salt tolerance degree recognition network based on the SCSFAM_Net model includes a preliminary feature extraction module Stem and four block blocks; the Stem module is used for downsampling and increasing the number of channels; the block block contains a backbone module SCSFAM module, which is used for feature extraction of spatial and channel analogy self-attention mechanisms;
[0019] A training and recognition module, which is used to train the soybean salt tolerance degree recognition network based on the SCSFAM_Net model using the training set to obtain a trained model, obtain the optimal model weights, test the trained model through the test set, and use the tested model to identify the soybeans to be detected.
[0020] Further, the preprocessing of the preprocessing module includes randomly horizontally rotating, size cropping and data format conversion for the training set and the test set respectively.
[0021] Further, among the four block blocks of the network construction module, the first block block includes one SCSFAM module; the second block block includes one SCSFAM module; the third block block includes 3 SCSFAM modules, and the fourth block block includes 2 SCSFAM modules.
[0022] Further, the feature extraction module Stem of the network construction module is composed of a 7×7 convolutional kernel, a BatchNorm layer, a RELU activation function and a max pooling layer.
[0023] Furthermore, the SCSFAM module of the network construction module includes: a first branch, where the input features pass through a max pooling layer and a 1×1 convolutional kernel, as well as an average pooling layer and a 1×1 convolutional kernel respectively. After multiplying the features obtained respectively, they pass through the CHSM module. After the CHSM module passes through the global average pooling layer and the softmax layer, channel sorting and screening are performed according to the weight size. After cutting the channels according to a ratio, the channels are expanded to the original number of channels. Subsequently, they pass through two 1×1 convolutional kernels respectively. After multiplying the features obtained respectively, they pass through the softmax, and then directly multiply with the features obtained through the CHSM module;
[0024] A second branch, which passes through a global max pooling layer and a global average pooling layer respectively, concatenates the obtained features, then passes through two 3×3 convolutional kernels respectively, multiplies the obtained features, then passes through the softmax, and multiplies the obtained features with the features that have not passed through the 3×3 convolutional kernel after concatenation;
[0025] A third branch, where the input features directly multiply with the result of the residual connection between the features obtained from the first branch and the features obtained from the second branch.
[0026] Beneficial effects: Compared with the prior art, the present invention has the following remarkable advantages: 1. The present invention proposes a feature channel selection structure, which uses a combination of a convolutional layer and a max pooling layer on each layer structure for channel selection, screens out the important channel features in proportion, and realizes the rapid extraction and recognition of reduced redundant features on each layer structure; 2. The present invention proposes an improved spatial attention mechanism, which enriches the features using a convolutional layer after concatenation through the pooling layer, analogizes the structure of the self-attention mechanism of the transformer to realize feature extraction, and uses softmax for feature selection to optimize the feature information in space; 3. The present invention proposes an improved channel attention mechanism, reduces the convolutional layer, analogizes the structure of the self-attention mechanism of the transformer to realize feature extraction, and then combines it with the features of the improved spatial attention mechanism. Description of the Drawings
[0027] Figure 1 It is the method flow chart of the embodiment of the present invention;
[0028] Figure 2 It is the schematic diagram of the soybean salt tolerance degree recognition network based on the SCSFAM_Net model of the embodiment of the present invention;
[0029] Figure 3 It is the schematic diagram of the SCSFAM module structure of the embodiment of the present invention;
[0030] Figure 4Schematic diagram of the CHSM module according to an embodiment of the present invention. Detailed implementation manners
[0031] In order to enable those skilled in the art to better understand the solution of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application.
[0032] As shown in the attached Figure 1 figure, the method for identifying the degree of salt tolerance of soybeans based on the SCSFAM_Net model in this embodiment includes:
[0033] (1) Collect the pre-acquired data on the degree of salt tolerance of soybeans and make it into a data set, divide the data set into a training set and a test set according to a preset ratio, and preprocess the training set and the test set respectively;
[0034] (2) Based on the Transformer self-attention mechanism, construct a network for identifying the degree of salt tolerance of soybeans based on the SCSFAM_Net model; the network for identifying the degree of salt tolerance of soybeans based on the SCSFAM_Net model includes a preliminary feature extraction module Stem and four block blocks; the Stem module is used for downsampling and increasing the number of channels; the block block contains a backbone module SCSFAM module, which is used for feature extraction of spatial and channel analogy self-attention mechanisms;
[0035] (3) Use the training set to train the network for identifying the degree of salt tolerance of soybeans based on the SCSFAM_Net model to obtain a trained model, obtain the optimal model weights, test the trained model through the test set, and use the tested model to identify the soybeans to be detected.
[0036] Specifically, in step (1), classify the pre-acquired categories of the degree of salt tolerance of soybeans and make them into a data set. Further, the data images can also be divided into 5 categories according to the degree of salt tolerance of soybeans, etc. ((a) Level 1 highly salt-tolerant phenotype, (b) Level 2 salt-tolerant phenotype, (c) Level 3 moderately salt-tolerant phenotype, (d) Level 4 salt-sensitive phenotype, (e) Level 5 highly salt-sensitive phenotype); and divide the data set into a training set and a test set according to a ratio of 8:2, and perform data augmentation on the images including three methods: random horizontal rotation, size cropping, and data format conversion.
[0037] In step (2), construct a network for identifying the degree of salt tolerance of soybeans based on the SCSFAM_Net model, as Figure 2As shown, first construct the preliminary feature extraction module, and then build four blocks as the backbone network. Specifically, it initially includes the feature extraction Stem module and six feature extraction SCSFAM modules. The Stem module performs simple downsampling and increases the number of channels. The four blocks stack the SCSFAM modules in a ratio of 1:1:3:1 for deep feature extraction; the first block includes one SCSFAM module; the second block includes one SCSFAM module; the third block includes three SCSFAM modules; the fourth block includes two SCSFAM modules. Among them, the SCSFAM module extracts spatial and channel features respectively. Among the channel features, the CHSM module filters the channel weights and then uses element-wise multiplication to expand the dimensional structure features. The spatial features are expanded by the spatial dimensional structure features through element-wise multiplication. The structure of the CHSM module is as Figure 4 shown.
[0038] The preliminary feature extraction Stem module consists of a 7×7 convolutional kernel. First, in the preprocessing stage of the training sample image input to the model, the feature vector S0 is obtained through image preprocessing; S0 is input into the preliminary feature extraction module, and S0 undergoes a convolutional operation with a stride of 1, padding of 3, and kernel_size = 7, and then passes through the BatchNorm2d, Relu activation function, and max pooling layer to obtain the feature vector S1.
[0039] The first block structure of the backbone module network consists of one SCSFAM module, as Figure 3As shown, input S1 into the SCSFAM module. After S1 is input into the first branch, it passes through a convolutional kernel with a maximum pooling of kernel_size = 3, a stride of 1, and a padding of 1, followed by a convolutional kernel with kernel_size = 1, an average pooling layer with a convolutional kernel of kernel_size = 3, a stride of 1, and a padding of 1, and a convolutional kernel with kernel_size = 1. The features obtained from the two branches are combined, and then passed through the CHSM module. In the CHSM module, after global average pooling and the softmax layer, channel sorting and screening are performed according to the weight size. After cutting the channels in proportion, the channels are expanded again to the original number of channels. Then, the obtained features are multiplied by a convolutional kernel with kernel_size = 1, and after softmax, they are multiplied by the feature vector that has not passed through the convolutional kernel with kernel_size = 1. The features obtained from the second branch after global max pooling and global average pooling are concatenated, and then the obtained features are multiplied by a convolutional kernel with kernel_size = 3, a stride of 1, and a padding of 1. The features obtained after softmax are multiplied by the features that have not passed through the convolutional kernel with kernel_size = 3. After the feature vectors of the first branch and the second branch are connected by residual connection, the feature vector S11 is obtained through the activation function. The obtained feature vector is downsampled by a factor of two to obtain the feature vector Z1.
[0040] Enter the second block. Input S11 into the SCSFAM module. After S11 is input into the first branch, it passes through a convolutional kernel with a maximum pooling of kernel_size = 3, a stride of 1, and a padding of 1, followed by a convolutional kernel with kernel_size = 1, an average pooling layer with a convolutional kernel of kernel_size = 3, a stride of 1, and a padding of 1, and a convolutional kernel with kernel_size = 1. The features obtained from the two branches are then processed. After that, it passes through the CHSM module. In the CHSM module, after global average pooling and the softmax layer, channel sorting and screening are performed according to the weight size. After cutting the channels in proportion, the channels are expanded again to the original number of channels. Then, the obtained features are multiplied by a convolutional kernel with kernel_size = 1. After softmax, the result is multiplied by the result without passing through the convolutional kernel with kernel_size = 1. The features obtained from the second branch through global max pooling and global average pooling layers are concatenated. Then, the obtained features are multiplied by a convolutional kernel with kernel_size = 3, a stride of 1, and a padding of 1. The features obtained after softmax are multiplied by the features without passing through the convolutional kernel with kernel_size = 3. After the feature vectors obtained from the first branch and the second branch are subjected to residual connection, they pass through an activation function. The obtained feature vectors are downsampled by a factor of two and multiplied by Z1, and pass through a maximum pooling layer with a convolutional kernel of kernel_size = 2 and the RELU activation function. Then, the obtained feature vectors are downsampled by a factor of two to obtain the feature vectors and S22.
[0041] Enter the third block. Input S22 into the first SCSFAM module in sequence. After S22 is input into the first branch, it passes through a convolutional kernel with a maximum pooling of kernel_size = 3, a stride of 1, and a padding of 1, followed by a convolutional kernel with kernel_size = 1, an average pooling layer with a convolutional kernel of kernel_size = 3, a stride of 1, and a padding of 1, and a convolutional kernel with kernel_size = 1. The features obtained from the two branches are combined. Then, it passes through the CHSM module. In the CHSM module, after global average pooling and the softmax layer, channel sorting and screening are performed according to the weight size. After cutting the channels according to a ratio, the channels are expanded again to the original number of channels. Then, the obtained features are multiplied by a convolutional kernel with kernel_size = 1. After passing through the softmax, the result is multiplied by the feature without passing through the convolutional kernel with kernel_size = 1. The features obtained from the second branch by passing through the global max pooling and the global average pooling layer are concatenated. Then, the obtained features are multiplied by a convolutional kernel with kernel_size = 3, a stride of 1, and a padding of 1. The feature obtained after passing through the softmax is multiplied by the feature without passing through the convolutional kernel with kernel_size = 3. After the feature vectors obtained from the first branch and the second branch are subjected to residual connection, the feature vector S33 is obtained through the activation function. Input S33 into the second SCSFAM module. After S33 is input into the first branch, it passes through a convolutional kernel with a maximum pooling of kernel_size = 3, a stride of 1, and a padding of 1, followed by a convolutional kernel with kernel_size = 1, an average pooling layer with a convolutional kernel of kernel_size = 3, a stride of 1, and a padding of 1, and a convolutional kernel with kernel_size = 1. The features obtained from the two branches are combined. Then, it passes through the CHSM module. In the CHSM module, after global average pooling and the softmax layer, channel sorting and screening are performed according to the weight size. After cutting the channels according to a ratio, the channels are expanded again to the original number of channels. Then, the obtained features are multiplied by a convolutional kernel with kernel_size = 1. After passing through the softmax, the result is multiplied by the feature without passing through the convolutional kernel with kernel_size = 1. The features obtained from the second branch by passing through the global max pooling and the global average pooling layer are concatenated. Then, the obtained features are multiplied by a convolutional kernel with kernel_size = 3, a stride of 1, and a padding of 1. The feature obtained after passing through the softmax is multiplied by the feature without passing through the convolutional kernel with kernel_size = 3.After the first branch and the second branch obtain the feature vectors and perform residual connection, they pass through the activation function. The obtained feature vectors are downsampled by a factor of two, multiplied by S22, and passed through a max-pooling layer with a kernel size of kernel_size = 2 and the RELU activation function. Then, the obtained feature vectors are downsampled by a factor of two to obtain the feature vectors and S44. S44 is input into the second SCSFAM module. After S44 is input into the first branch, it passes through a convolutional kernel with a kernel size of kernel_size = 3, a stride of 1, and a padding of 1, then passes through a convolutional kernel with a kernel size of kernel_size = 1, an average pooling layer with a kernel size of kernel_size = 3, a stride of 1, and a padding of 1, and a convolutional kernel with a kernel size of kernel_size = 1. The features obtained from the two branches are combined, and then passed through the CHSM module. In the CHSM module, after global average pooling and the softmax layer, the channels are sorted and filtered according to the weight size. After cutting the channels according to the ratio, the channels are expanded again to the original number of channels. Then, the obtained features are multiplied by a convolutional kernel with a kernel size of kernel_size = 1, and after passing through the softmax, they are multiplied by the features that have not passed through the convolutional kernel with a kernel size of kernel_size = 1. The features obtained from the global max-pooling and global average pooling layers of the second branch are concatenated. Then, the obtained features are multiplied by a convolutional kernel with a kernel size of kernel_size = 3, a stride of 1, and a padding of 1, and the features obtained after passing through the softmax are multiplied by the features that have not passed through the convolutional kernel with a kernel size of kernel_size = 3. After the first branch and the second branch obtain the feature vectors and perform residual connection, they pass through the activation function. The obtained feature vectors are downsampled by a factor of two, multiplied by S44, and passed through a max-pooling layer with a kernel size of kernel_size = 2 and the RELU activation function. Then, the obtained feature vectors are downsampled by a factor of two and multiplied by S22, and passed through a max-pooling layer with a kernel size of kernel_size = 2 and the RELU activation function to obtain the feature vector S55. Then, the obtained feature vectors are downsampled by a factor of two to obtain the feature vectors and Z2.
[0042] Enter the fourth block. After entering the SCSFAM module, input S55 into the SCSFAM module. After S55 enters the first branch, it passes through a convolutional kernel with a maximum pooling of kernel_size = 3, a padding of 1, and a stride of 1, then passes through a convolutional kernel with kernel_size = 1, an average pooling layer with a convolutional kernel of kernel_size = 3, a padding of 1, and a stride of 1, and a convolutional kernel with kernel_size = 1. The features obtained from the two branches are combined. Then, it passes through the CHSM module. In the CHSM module, after global average pooling and the softmax layer, channel sorting and screening are performed according to the weight size. After cutting the channels according to the ratio, the channels are expanded again to the original number of channels. Then, the obtained features are multiplied through convolutional kernels with kernel_size = 1, and after softmax, they are multiplied with the features that have not passed through the convolutional kernel with kernel_size = 1. The features obtained from the global max pooling and global average pooling layers of the second branch are concatenated. Then, the obtained features are multiplied through convolutional kernels with kernel_size = 3, a padding of 1, and a stride of 1, and the features obtained after softmax are multiplied with the features that have not passed through the convolutional kernel with kernel_size = 3. After the feature vectors obtained from the first branch and the second branch are subjected to residual connection, the feature vector S66 is obtained through the activation function.
[0043] Step (3) includes inputting the feature vector obtained in the above steps into an adaptive pooling layer, a convolutional layer with kernel_size = 2, a Relu activation function, and a fully connected layer, and finally outputting the classification result F. Calculate the loss between the obtained classification result and the true value. By continuously iterating the forward and backward propagation processes, and then updating the model parameters through the momentum gradient descent algorithm to optimize the network parameters until the best weights are obtained, and the training is completed.
[0044] Use the best training model weights obtained as the pre-trained model to verify the data pictures in the test set, and evaluate the performance of the model through four indicators: accuracy (AccuracS), precision, recall, and F1-Measure.
[0045] The calculation formulas for each indicator are as follows:
[0046]
[0047] Among them, TP (True Positives): True positive examples, which are predicted as positive examples and are actually positive examples; FP (False Positives): False positive examples, which are predicted as positive examples but are actually negative examples; FN (false Negatives): False negative examples, which are predicted as negative examples but are actually positive examples; TN (True Negatives): True negative examples, which are predicted as negative examples and are actually negative examples.
[0048] Table 1 Comparison table of experimental results of this embodiment with the performance of traditional models and lightweight models
[0049]
[0050]
[0051] The comparison table of the experimental results of the SCSFAM_Net model proposed in this embodiment with the performance of traditional models and lightweight models is shown in Table 2. It can be seen that the recognition efficiency of the model proposed by the present invention reaches 96.50%, has high robustness and practical application value, and has great prospects in the field of identifying the salt tolerance of crop seed resources at different levels.
[0052] The variable descriptions involved in this embodiment are shown in Table 2.
[0053] Table 2 Variable description table
[0054] Variable Variable Description SCSFAM Feature Extraction Backbone Module UpCFP Multi-Scale Feature Extraction Module S0 Feature Vector Output by Image Preprocessing S1 Feature Vector of the Preliminary Feature Extraction Module S11 Feature Vector Output by the SCSFAM Module in the First Block Z1 Feature Vector with S11 Downsampled by a Factor of Two S22 Feature Vector Output by the SCSFAM Module in the Second Block S33 Feature Vector Output by the First SCSFAM Module in the Third Block S44 Feature Vector Output by the Second SCSFAM Module in the Third Block S55 Feature Vector Output by the Third SCSFAM Module in the Third Block Z2 Feature Vector Obtained by Downsampling by a Factor of Two after the Third Block S66 Feature Vector Output by the SCSFAM Module in the Fourth Block F Output Classification Result
[0055] Based on the same inventive concept, this embodiment also provides a soybean salt tolerance degree recognition system based on the SCSFAM_Net model, including:
[0056] A preprocessing module for collecting the soybean salt tolerance degree data obtained in advance and making it into a data set, dividing the data set into a training set and a test set according to a preset ratio, and preprocessing the training set and the test set respectively;
[0057] A network construction module for constructing a soybean salt tolerance degree recognition network based on the SCSFAM_Net model based on the Transformer self-attention mechanism; the soybean salt tolerance degree recognition network based on the SCSFAM_Net model includes a preliminary feature extraction module Stem and four block blocks; the Stem module is used for downsampling and increasing the number of channels; the block block includes a backbone module SCSFAM module for performing feature extraction of spatial and channel analogy self-attention mechanisms;
[0058] The training and recognition module is used to train the soybean salt tolerance degree recognition network based on the SCSFAM_Net model using the training set, obtain the trained model, acquire the optimal model weights, test the trained model with the test set, and use the tested model to recognize the soybeans to be detected.
[0059] Furthermore, the preprocessing in the preprocessing module includes performing random horizontal rotation, size cropping, and data format conversion on the training set and the test set respectively.
[0060] Furthermore, in the four block blocks of the network construction module, the first block block includes one SCSFAM module; the second block block includes one SCSFAM module; the third block block includes 3 SCSFAM modules, and the fourth block block includes 2 SCSFAM modules.
[0061] Furthermore, the feature extraction module Stem in the network construction module consists of a 7×7 convolutional kernel, a BatchNorm layer, a RELU activation function, and a max pooling layer.
[0062] Furthermore, the SCSFAM module in the network construction module includes: The first branch, the input features respectively pass through a max pooling and a 1×1 convolutional kernel, as well as an average pooling layer and a 1×1 convolutional kernel. After multiplying the respective obtained features, they pass through the CHSM module. After the CHSM module passes through the global average pooling and the softmax layer, it performs channel sorting and screening according to the weight size, cuts the channels according to a ratio, and then expands the channels to the original number of channels. Subsequently, they respectively pass through two 1×1 convolutional kernels, multiply the respective obtained features, pass through the softmax, and then directly multiply the features obtained after passing through the CHSM module;
[0063] The second branch respectively passes through the global max pooling and the global average pooling layer, splices the obtained features, then respectively passes through two 3×3 convolutional kernels, multiplies the obtained features, then passes through the softmax, and multiplies the obtained features by the features that have not passed through the 3×3 convolutional kernel after splicing;
[0064] The third branch directly multiplies the input features by the result of the residual connection of the features obtained from the first branch and the features obtained from the second branch.
Claims
1. A method for identifying the degree of salt tolerance of soybean based on the SCSFAM_Net model, characterized in that: include: (1) Collecting the previously acquired soybean salt tolerance data and making it into a data set, dividing the data set into a training set and a test set according to a preset ratio, and preprocessing the training set and the test set respectively; (2) Based on the Transformer self-attention mechanism, a soybean salt tolerance degree recognition network based on the SCSFAM_Net model is constructed; the soybean salt tolerance degree recognition network based on the SCSFAM_Net model includes a preliminary feature extraction module Stem and four blocks; the Stem module is used for downsampling and increasing the number of channels; the block block includes a backbone module SCSFAM module for feature extraction of spatial and channel analog self-attention mechanisms; (3) The training set is used to train the soybean salt tolerance recognition network based on the SCSFAM_Net model to obtain the trained model and the optimal model weight. The trained model is tested on the test set and the tested model is used to identify the soybean to be tested.
2. The method for identifying the degree of soybean salt tolerance based on the SCSFAM_Net model according to claim 1, characterized in that: The preprocessing in step (1) includes random horizontal rotation, size cropping and data format conversion of the training set and the test set respectively.
3. The method for identifying the degree of soybean salt tolerance based on the SCSFAM_Net model according to claim 1, characterized in that: Among the four blocks in step (2), the first block includes one SCSFAM module; the second block includes one SCSFAM module; the third block includes three SCSFAM modules, and the fourth block includes two SCSFAM modules.
4. The method for identifying the degree of soybean salt tolerance based on the SCSFAM_Net model according to claim 1, characterized in that: The feature extraction module Stem in step (2) is composed of a 7×7 convolution kernel, a BatchNorm layer, a RELU activation function and a maximum pooling layer.
5. The method for identifying the degree of soybean salt tolerance based on the SCSFAM_Net model according to claim 1, characterized in that: The SCSFAM module in step (2) comprises: a first branch, wherein input features are respectively subjected to maximum pooling and a 1×1 convolution kernel, and an average pooling layer and a 1×1 convolution kernel, and the features obtained are multiplied and then passed through a CHSM module. The CHSM module sorts and screens channels according to weights after passing through global average pooling and a softmax layer, and then expands the channels to the original number of channels after cutting the channels in proportion, and then respectively passes through two 1×1 convolution kernels, and the features obtained are multiplied and then passed through softmax, and then directly multiplied with the features obtained by the CHSM module; In the second branch, the features are concatenated after passing through the global maximum pooling layer and the global average pooling layer, and then they are multiplied by two 3×3 convolution kernels, and then they are multiplied by softmax and the concatenated features without passing through the 3×3 convolution kernel. In the third branch, the input feature is directly multiplied by the result of residual connection of the feature obtained by the first branch and the feature obtained by the second branch.
6. A soybean salt tolerance identification system based on SCSFAM_Net model, characterized in that: include: A preprocessing module is used to collect the pre-acquired soybean salt tolerance data and make it into a data set, divide the data set into a training set and a test set according to a preset ratio, and preprocess the training set and the test set respectively; A network construction module is used to construct a soybean salt tolerance degree recognition network based on the SCSFAM_Net model based on the Transformer self-attention mechanism; the soybean salt tolerance degree recognition network based on the SCSFAM_Net model includes a preliminary feature extraction module Stem and four blocks; the Stem module is used for downsampling and increasing the number of channels; the block includes a backbone module SCSFAM module for feature extraction of spatial and channel analog self-attention mechanisms; The training and recognition module is used to train the soybean salt tolerance recognition network based on the SCSFAM_Net model using the training set, obtain the trained model, obtain the optimal model weight, test the trained model through the test set, and use the tested model to identify the soybean to be detected.
7. The soybean salt tolerance identification system based on the SCSFAM_Net model according to claim 6, characterized in that: Preprocessing module The preprocessing includes random horizontal rotation, size cropping and data format conversion for the training set and the test set respectively.
8. The soybean salt tolerance identification system based on the SCSFAM_Net model according to claim 6, characterized in that: Among the four blocks of the network construction module, the first block includes one SCSFAM module; the second block includes one SCSFAM module; the third block includes three SCSFAM modules, and the fourth block includes two SCSFAM modules.
9. The soybean salt tolerance identification system based on the SCSFAM_Net model according to claim 6, characterized in that: The feature extraction module Stem of the network construction module is composed of a 7×7 convolution kernel, a BatchNorm layer, a RELU activation function and a maximum pooling layer.
10. The soybean salt tolerance identification system based on the SCSFAM_Net model according to claim 6, characterized in that: The SCSFAM module of the network construction module includes: a first branch, the input features are respectively subjected to maximum pooling and 1×1 convolution kernel, and average pooling layer and 1×1 convolution kernel, and the features obtained are multiplied, and then passed through the CHSM module, and the CHSM module sorts and screens the channels according to the weight size after the global average pooling and softmax layer, and expands the channels to the original number of channels after cutting the channels according to the proportion, and then passes through two 1×1 convolution kernels, respectively, and the features obtained are multiplied, and then passed through softmax, and then directly multiplied with the features obtained by the CHSM module; In the second branch, the features are concatenated after passing through the global maximum pooling layer and the global average pooling layer, and then they are multiplied by two 3×3 convolution kernels, and then they are multiplied by softmax and the concatenated features without passing through the 3×3 convolution kernel. In the third branch, the input feature is directly multiplied by the result of residual connection of the feature obtained by the first branch and the feature obtained by the second branch.