CN116403056B - Ginseng grading system and method - Google Patents

Abstract

A ginseng grading system and a ginseng grading method belong to the technical field of classification and identification, and particularly relate to the technical field of ginseng classification and identification. It solves the problem of low efficiency and accuracy of ginseng classification. The system comprises: the input module sequentially comprises a two-dimensional convolution layer and a batch normalization layer; the trunk module sequentially comprises 4 layers of improved ConvNeXt networks, which are named as a first improved ConvNeXt network, a second improved ConvNeXt network, a third improved ConvNeXt network and a fourth improved ConvNeXt network respectively; and the output module sequentially comprises an average pooling layer, a batch normalization layer and a full connection layer. The system and the method can be applied to the field of classification of ginseng.

Description

Ginseng grading system and method

Technical Field

The invention belongs to the technical field of classification and identification, and particularly relates to the technical field of ginseng classification and identification.

Background

Classification of ginseng is mainly developed from both chemical and physical viewpoints, classification of ginseng is performed from chemical viewpoints, classification of ginseng is mainly performed by detecting the content of active ingredients in ginseng, for example:

the saponin content is one of the main active ingredients in the ginseng, and has obvious effects of immunoregulation, anti-tumor, antioxidation and the like, so that in general, the higher the saponin content is, the better the quality of the ginseng is.

Iron content, which is one of essential trace elements of the human body, plays an important role in the generation and transportation of blood, and thus, iron content is also an important index for evaluating the quality of ginseng.

The volatile oil content is an important component in ginseng, and has the effects of calming, resisting bacteria, resisting inflammation and the like, so that the volatile oil content is also one of important indexes of the quality of ginseng.

Polysaccharide content, which is a complex high molecular compound with the functions of enhancing human immunity, resisting tumor, resisting oxidization and the like, is also an important index for evaluating the quality of ginseng.

Ginseng grading is performed from a physical point of view, and the grading of ginseng is mainly performed by artificial evaluation of physical parameters or appearance of ginseng, for example:

humidity, the humidity of ginseng refers to its water content, and high humidity can cause ginseng to be contaminated with bacteria and mold, thereby affecting its quality.

Density, the density of ginseng is also an important index for evaluating its quality, and generally, the higher density ginseng is better in quality because its components are more concentrated.

Elasticity, the elasticity of ginseng refers to the ability of ginseng to recover its original shape after being deformed under stress, and generally, ginseng with better elasticity has higher quality.

Color, ginseng color is also an important index for evaluating its quality. Generally, the quality of the ginseng with bright and uniform color is better.

Whether ginseng is classified by a chemical or physical method, the classification is mainly performed by a manual assay or identification mode, and the main defects are that: firstly, a great deal of manpower is wasted, professional staff is seriously relied on, the grading speed is low, and the grading accuracy is low; secondly, grading standards are difficult to unify, and modern management means are lacked, so that the quality of ginseng products is difficult to ensure.

Disclosure of Invention

The invention provides a ginseng classification system and a ginseng classification method, which aim to solve the problem of low ginseng classification efficiency and accuracy.

In one aspect, a ginseng grading system, the system comprising:

an input module: sequentially comprising a two-dimensional convolution layer and a batch normalization layer;

a backbone module: the ConvNeXt network sequentially comprises 4 layers of improvement, which are named as a first improvement ConvNeXt network, a second improvement ConvNeXt network, a third improvement ConvNeXt network and a fourth improvement ConvNeXt network respectively:

the first modified ConvNeXt network: the system comprises a channel shuffling module and an optimized ConvNeXt module in sequence, wherein the number of channels of the optimized ConvNeXt module is 128, and the operation of the ConvNeXt module is repeated for 3 times;

the second modified ConvNeXt network: the system comprises a downsampling module, a channel shuffling module and an optimized ConvNeXt module in sequence, wherein the number of channels of the optimized ConvNeXt module is 256, and the operation of the ConvNeXt module is repeated for 3 times;

the third modified ConvNeXt network: the system comprises a downsampling module, a channel shuffling module and an optimized ConvNeXt module in sequence, wherein the number of channels of the optimized ConvNeXt module is 512, and the operation of the ConvNeXt module is repeated 27 times;

the fourth modified ConvNeXt network: the system comprises a downsampling module, a channel shuffling module and an optimized ConvNeXt module in sequence, wherein the number of channels of the optimized ConvNeXt module is 1024, and the operation of the ConvNeXt module is repeated for 3 times;

and an output module: the method comprises an average pooling layer, a batch normalization layer and a full connection layer in sequence.

Further, the channel shuffling module comprises a channel reorganization layer, a channel transposition layer and a channel flattening layer in sequence.

Further, the downsampling module sequentially comprises a batch normalization layer and a two-dimensional sampling layer.

Further, the specific improvement of the optimized ConvNeXt module is as follows: adding an improved structure re-parameterization module and replacing the GELU activation function with the PReLU activation function, wherein the improved structure re-parameterization module is placed before the depth separable convolutional layer.

Further, in the improved structural reparameterization module, the main network sequentially comprises a depth separable convolution layer and a batch standardization layer, the main network is added with a branch network, the branch network sequentially comprises the depth separable convolution layer and the batch standardization layer, the convolution kernel size of the main network is 13×13, and the convolution kernel size of the branch network is 5×5.

Scheme II, a ginseng grading method:

the method is carried out by using the ginseng grading system according to the scheme one, and the method specifically comprises the following steps:

s1, arranging shot ginseng high-definition color pictures into a data set, and performing data expansion;

s2, dividing the data set into a training set and a verification set;

s3, inputting a training set into the ginseng grading system;

s4, inputting the verification set into the ginseng grading system, and calculating the accuracy of the system;

s5, inputting the ginseng atlas to be classified into the ginseng classification system for classification.

The beneficial effects of the invention are as follows:

(1) Firstly, extracting a characteristic map of ginseng by adopting a grading extraction method, firstly extracting shallow features of the ginseng by an input module, and then further extracting the features by a main module on the basis, so that the feature expression capability of a system can be improved; when the system is constructed, the important and main characteristics of ginseng grading are considered, so that a multi-layer network is independently designed to extract the characteristics, and the full extraction of the details of the ginseng root is ensured;

(2) The channel shuffling module is embedded after the trunk module is downsampled, so that channel characteristics are fully fused to improve classification accuracy, the input of the next adopted grouping convolution is ensured to come from different groups, information can flow among different groups, and the operation accuracy is improved by 3.70% compared with that of an original model;

(3) According to the invention, an improved structural re-parameterization module is newly added in an optimized ConvNeXt module, a relatively small (5 multiplied by 5) kernel is added into a large (13 multiplied by 13) kernel through linear transformation, so that a model similar to VGG is generated, the calculation efficiency of an original network is improved, and compared with the original model, the operation accuracy is improved by 4.93%;

(4) According to the invention, the GELU activation function in the original ConvNeXt module is replaced by the PReLU activation function, and the batch standardization layer is combined in the optimized structural heavy parameterization module, so that gradient disappearance and gradient explosion are prevented, and the operation accuracy is improved by 1.23% compared with that of the original model. Compared with different activation functions, the method has the highest precision and the shortest time consumption.

The system and the method can be applied to the field of classification of ginseng.

Drawings

FIG. 1 is a diagram of a ginseng classification system according to an embodiment of the present invention;

FIG. 2 is a block diagram of a channel shuffling module according to an embodiment of the present invention;

FIG. 3 is a block diagram of a downsampling module according to an embodiment of the present invention;

FIG. 4 is a block diagram of an optimized ConvNeXt module according to an embodiment of the present invention;

FIG. 5 is a block diagram of an improved structural reparameterization module according to an embodiment of the present invention.

Detailed Description

The following description of the embodiments of the present invention will be made apparent and fully in view of the accompanying drawings, in which some, but not all embodiments of the invention are shown. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to fall within the scope of the invention.

Example 1,

This embodiment provides a ginseng grading system, as shown in fig. 1, comprising:

an input module: the shallow feature information is used for extracting the shallow feature information of the image to obtain a feature map;

the method comprises a two-dimensional convolution Layer (Conv 2 d) and a batch normalization Layer (Layer Norm) in sequence, wherein the convolution kernel size K=4 of the two-dimensional convolution Layer and the step length S=4 of the two-dimensional convolution Layer; the size of the ginseng atlas to be classified is 256 multiplied by 256, the number of channels is 3, and the characteristic atlas with the size of 56 multiplied by 56 and the number of channels of 128 is obtained after the ginseng atlas is output by the input module.

A backbone module: the method is used for improving the characteristic representation capability of the model and optimizing the characteristic communication among channels so as to improve the calculation efficiency of the original network and the network operation precision; the method comprises the steps of sequentially comprising a 4-layer improved ConvNeXt network, which are named as a first improved ConvNeXt network, a second improved ConvNeXt network, a third improved ConvNeXt network and a fourth improved ConvNeXt network respectively;

the first modified ConvNeXt network: the system sequentially comprises a Channel shuffling module (channel_shuffle) and an optimized ConvNeXt module (ConvNeXt Block), wherein the number of channels of the optimized ConvNeXt module is 128 (dim=128), the operation of the module is repeated for 3 times, and the size of an output characteristic diagram is 56 multiplied by 56;

the second modified ConvNeXt network: the device sequentially comprises a downsampling module (downsampling), a Channel shuffling module (channel_shuffle) and an optimized ConvNeXt module, wherein the number of channels of the optimized ConvNeXt module is 256 (dim=256), the operation of the module is repeated for 3 times, and the size of an output characteristic diagram is 28 multiplied by 28;

the third modified ConvNeXt network: the device sequentially comprises a downsampling module (downsampling), a Channel shuffling module (channel_shuffle) and an optimized ConvNeXt module, wherein the number of channels of the optimized ConvNeXt module is 512 (dim=512), the operation of the module is repeated 27 times, and the size of an output characteristic diagram is 14 multiplied by 14;

the fourth modified ConvNeXt network: the device comprises a downsampling module (downsampling), a Channel shuffling module (channel_shuffle) and an optimized ConvNeXt module in sequence, wherein the number of channels of the optimized ConvNeXt module is 1024 (dim=1024), the operation of the module is repeated for 3 times, and the size of an output characteristic diagram is 7 multiplied by 7;

and an output module: the method is used for obtaining probabilities of the ginseng belonging to each grade through full connection, wherein the probability is the result of a predicted sample; the method sequentially comprises an average pooling Layer (Global Avg Pooling), a batch normalization Layer (Layer Norm) and a full connection Layer (Linear).

EXAMPLE 2,

This embodiment is further defined in embodiment 1, as shown in fig. 2, the Channel shuffling module sequentially includes a Channel reorganizing layer (channel_view), a Channel transpose layer (channel_transit), and a Channel flattening layer (channel_flat), and the system embeds the Channel shuffling module after the main network is downsampled, so that the Channel features are fully fused to improve the classification accuracy, ensure that the next adopted packet convolution has inputs from different groups, and the information can flow between different groups, and through this operation accuracy is improved by 3.70% compared with the original model.

EXAMPLE 3,

This embodiment is further defined in embodiment 1, and as shown in fig. 3, the downsampling module sequentially includes a batch normalization Layer (Layer Norm) and a two-dimensional convolution Layer (Conv 2 d), where the convolution kernel size k=2 of the two-dimensional convolution Layer and the step size s=2 of the two-dimensional convolution Layer.

EXAMPLE 4,

This example is a further limitation of example 1, and as shown in fig. 4, the specific modification of the optimized ConvNeXt module is as follows: adding a modified structure Re-parameterization module (Re-parameter) and replacing the GELU activation function with the PReLU activation function, wherein the modified structure Re-parameterization module is placed before the depth separable convolutional layer (Depthwick Conv2 d).

As shown in fig. 5, in the improved structural reparameterization module, a backbone network sequentially comprises a depth separable convolution layer (dw) and a batch normalization layer (BN), a branch network is added to the backbone network, the branch network sequentially comprises the depth separable convolution layer (dw) and the batch normalization layer (BN), the convolution kernel of the backbone network is 13×13, and the convolution kernel of the branch network is 5×5; a relatively small (5×5) kernel is added to a large 13×13 kernel through linear transformation, so that a model similar to VGG is generated, the calculation efficiency of the original network is improved, and the operation accuracy is improved by 4.93% compared with that of the original model.

EXAMPLE 5,

This example illustrates the beneficial effects of the system of the present invention by a series of comparisons. The system test configuration environment is a GPU parallel computing workstation, the processor is Xeon (R) CPU E5-2680v4, the display card is GeForce GTX 1080Ti, the Ubuntu16.04LTS operation system is adopted, the Anaconda3-5.2.0-Linux version is arranged in a soft part configuration mode, and a Pytorch deep learning framework is built based on Python 3.7.0 programming language.

The system replaces the GELU activation function in the original ConvNeXt module with the PReLU activation function and combines the batch standardization layer in the optimized structural heavy parameterization module, so that gradient disappearance and gradient explosion are prevented, and the operation accuracy is improved by 1.23% compared with that of the original model. The comparison of the different activation functions is shown in table 1, which not only has the highest precision but also has the shortest time consumption.

Table 1:

in order to comprehensively measure the effectiveness of the proposed system, the Accuracy (Accuracy), recall (Recall), precision (Precision) and Specificity (Specificity) are adopted as evaluation indexes in the experiment, wherein TP is the number of samples correctly predicted as positive samples, namely the number of accurately identified ginseng samples, TN is the number of samples correctly predicted as negative samples, namely the number of samples of accurately identified other ginseng samples, FP is the number of samples incorrectly predicted as positive samples, namely the number of samples of ginseng identified as incorrect, FN is the number of samples incorrectly predicted as negative samples, namely the number of samples of ginseng identified as other varieties, and the calculation formula is as follows:

，

，

，

；

as shown in Table 2, in general, convNeXt+Rep (method 3), convNeXt+channel shuffle+Rep (method 5), convNeXt+channel shuffle+PReLU (method 6) and ConvNeXt+channel shuffle+Rep+PReLU (method 8) have identification accuracy of 90.12%, 92.59%, 91.36% and 94.44% respectively, and compared with other 4 types of models, the method 8 model can more accurately position and identify ginseng features, effectively improve the precision of the ConvNeXt model, better balance the grading results of various types of ginseng, and realize accurate identification of ginseng grades.

The details in Table 2 show that the addition of the channel_buffer module, the Re-parameter module and the PReLU can alleviate the problems of lower accuracy and precision and the like of the ConvNeXt model to a certain extent, so that the model is more suitable for the grade classification of ginseng. Under the action of a single module, the Re-parameter module is introduced to have the best overall identification effect on the ConvNeXt model, and the accuracy, recall rate and specificity of the model are respectively 90.12%, 85.09% and 85.23%, so that compared with other functional modules, the identification performance of the ConvNeXt model is better improved by introducing the Re-parameter module in the aspect of module fusion. Under the condition of fusion of two modules, the combined action effect of the channel_buffer module and the Re-parameter module is best, compared with the original module, the combined action effect of the PReLU and the Re-parameter module on the accuracy is improved by 7.4%, the combined action effect of the PReLU and the Re-parameter module on the accuracy is the lowest, the competitive advantage of the channel_buffer module is verified, and the channel_buffer module has stronger robustness and generalization capability.

Table 2:

the comparison test is carried out by respectively selecting a classical network and a ginseng data set, and the four types of performance indexes in table 3 are combined, so that the improved ConvNeXt module has good classification performance on the ginseng data set. Overall, compared with ResNet-50, resNet-101, denseNet-121 and InceptionV3, the original ConvNeXt network is respectively improved by 2.47 percentage points, 1.86 percentage points, 0.62 percentage point and 4.94 percentage points, and the recall rate is 78.06 percent, which indicates that the original ConvNeXt network has more stable results and better performance compared with the traditional network. Compared with the original basic ConvNeXt network model, the improved ConvNeXt module has the advantages that the accuracy, precision, recall rate and specificity are respectively improved by 9.25%, 13.86%, 12.98% and 7.02%, and a plurality of indexes are better than those of Vision Transformer and Swim converter networks, the average accuracy reaches 94.44%, and the following three reasons are that the improved ConvNeXt module has better effects:

1) A channel shuffling module is added after the downsampling module, so that channel characteristics are fully fused, and network operation precision is improved;

2) An improved structural re-parameterization module is added into the optimized ConvNeXt module, so that the high precision of the network is maintained;

3) By using the PReLU activation function, the nonlinear variability of the neural network model is increased, the network operation rate is improved, and the network operation precision is improved.

Table 3:

EXAMPLE 6,

The present embodiment provides a ginseng grading method using the ginseng grading system of any one of embodiments 1-4, specifically:

s1, arranging shot ginseng high-definition color pictures into a data set, and performing data expansion;

s2, dividing the data set into a training set and a verification set;

s3, inputting a training set into the ginseng grading system;

s4, inputting the verification set into the ginseng grading system, and calculating the accuracy of the system;

s5, inputting the ginseng atlas to be classified into the ginseng classification system for classification.

For step S1, the ginseng dataset includes 367 ginseng pieces, and the northeast ginseng pieces are photographed at a fixed height (40 cm) in multiple directions by a small professional studio (Sutefoto, guangdong, china). The classification is based on the ginseng classification standard in Jilin provincial and regional crude drug ginseng release edition as shown in table 4, and the classification confirmation is carried out on ginseng by a Chinese medicinal herb university expert teacher, the classification comprises 1668 pictures of sun-dried ginseng shot at multiple angles, the classification is divided into 3 categories, and classification is carried out according to the rules in table 4, wherein the classification is carried out on 549 special ginseng, 830 first-stage ginseng and 289 second-stage ginseng.

Table 4:

for step S1, data expansion is performed on the data set in 3 ways, namely random rotation and overturn; the relief effect is combined by horizontal overturning; horizontal flipping incorporates gaussian blur.

For step S2, the training set and validation set before and after expansion are shown in table 5:

table 5:

for step S3, the specific test parameter information used in training with the training set is shown in table 6.

Table 6:

for step S5, under the condition that other parameters are the same, a comparison test is performed on the data set before expansion (1680 sheets) and the data set after expansion (5116 sheets), and the training set and the verification set pictures are divided according to the ratio of 8:2 as shown in table 5. The accuracy of the test results is 78.67%, 85.97% respectively, and the improvement is 7.3%.

Claims (4)

1. A ginseng grading system, wherein a ginseng image to be graded is input to the system, the system outputs a ginseng grading result, the system comprising:

an input module: sequentially comprising a two-dimensional convolution layer and a batch normalization layer;

a backbone module: the method comprises the steps of sequentially comprising a 4-layer improved ConvNeXt network, which are named as a first improved ConvNeXt network, a second improved ConvNeXt network, a third improved ConvNeXt network and a fourth improved ConvNeXt network respectively;

the first modified ConvNeXt network: the system comprises a channel shuffling module and an optimized ConvNeXt module in sequence, wherein the number of channels of the optimized ConvNeXt module is 128, and the operation of the module is repeated for 3 times;

the second modified ConvNeXt network: the system comprises a downsampling module, a channel shuffling module and an optimized ConvNeXt module in sequence, wherein the number of channels of the optimized ConvNeXt module is 256, and the operation of the module is repeated for 3 times;

the third modified ConvNeXt network: the system comprises a downsampling module, a channel shuffling module and an optimized ConvNeXt module in sequence, wherein the number of channels of the optimized ConvNeXt module is 512, and the operation of the module is repeated 27 times;

the fourth modified ConvNeXt network: the system comprises a downsampling module, a channel shuffling module and an optimized ConvNeXt module in sequence, wherein the number of channels of the optimized ConvNeXt module is 1024, and the operation of the module is repeated for 3 times;

and an output module: the device comprises an average pooling layer, a batch normalization layer and a full connection layer in sequence;

the specific improvement of the optimized ConvNeXt module is as follows: adding an improved structure re-parameterization module and replacing the GELU activation function with the PReLU activation function, wherein the improved structure re-parameterization module is arranged in front of the depth separable convolution layer;

in the improved structural reparameterization module, a main network sequentially comprises a depth separable convolution layer and a batch standardization layer, a branch network is added to the main network, the branch network sequentially comprises the depth separable convolution layer and the batch standardization layer, the convolution kernel size of the main network is 13 multiplied by 13, and the convolution kernel size of the branch network is 5 multiplied by 5.

2. The ginseng grading system of claim 1, wherein the channel shuffling module comprises a channel reorganization layer, a channel transpose layer, and a channel flattening layer in that order.

3. The ginseng grading system of claim 2, wherein the downsampling module comprises a batch normalization layer and a two-dimensional sampling layer in sequence.

4. A method for classifying ginseng, characterized in that the method is carried out using the ginseng classification system according to any one of claims 1-3, in particular:

s1, arranging shot ginseng high-definition color pictures into a data set, and performing data expansion;

s2, dividing the data set into a training set and a verification set;

s3, inputting a training set into the ginseng grading system;

s4, inputting the verification set into the ginseng grading system, and calculating the accuracy of the system;

s5, inputting the ginseng atlas to be classified into the ginseng classification system for classification.