CN116645599A - A lightweight multi-scale CNN model for wheat disease detection - Google Patents

Abstract

The present invention relates to the technical field of wheat disease detection, in particular to a wheat disease detection method of a lightweight multi-scale CNN model; comprising the following steps: S1, using a shooting device to collect wheat disease images, unifying the size of the wheat disease images, and Classify and label each type of disease; S2, expand the wheat disease data by rotating, symmetrically flipping, and improving the contrast, and expand the original data set to 8495 pieces; divide according to the ratio of training set: verification set: test set = 6:2:2 Data set; S3, design the network model of wheat disease image detection; S4, train the wheat disease training samples obtained in S2 in the network model designed by S3 and test on the test set; the present invention has simple structure and small amount of calculation , with wide applicability.

Description

A lightweight multi-scale CNN model for wheat disease detection

technical field

The invention relates to the technical field of wheat disease detection, in particular to a wheat disease detection method of a lightweight multi-scale CNN model.

Background technique

The yield and quality of wheat are affected by many factors. Wheat disease is not only one of the very important factors, but also the main factor restricting the efficient production of high-quality wheat. There are many kinds of wheat diseases, more than 200 kinds of wheat diseases can be found all over the world, and there are about 20 kinds of serious wheat diseases in China, wheat powdery mildew, wheat rust, wheat leaf blight, etc. are very typical and more serious diseases of wheat. In China, due to the impact of diseases, the yield of wheat has been reduced by nearly one-third, which has great harm to grain production.

The wheat disease recognition method based on machine learning relies on the explicit feature extraction of images. The extraction strategies of these image features are formulated by people based on prior knowledge. On the one hand, the steps are complicated and the efficiency is low, and the feature extraction depends on the knowledge and experience of researchers. ; On the other hand, these artificially extracted features often exist in the shallow layer of the image and have narrow application.

The wheat disease identification method based on deep learning requires a lot of resources and time, and the deployment cost is high. It is difficult to apply to resource-constrained scenarios such as mobile devices and embedded devices, and the accuracy is difficult to guarantee.

In rural areas, it is difficult for farmers to hire disease experts or use expensive equipment to identify diseases. They all use experience to judge whether crops are diseased. Therefore, it is of great significance to study a wheat disease identification method with simple structure, small amount of calculation, wide applicability, and can be equipped on the mobile terminal to help farmers identify wheat diseases and improve wheat yield and quality.

Contents of the invention

The purpose of the present invention is to overcome the cumbersome manual detection steps, the low accuracy of machine learning methods for disease identification, and the high deployment and use costs of classic deep learning methods, which are difficult to apply to mobile devices and embedded devices with limited resources. In order to solve the problems of such scenarios, a wheat disease detection method based on a lightweight multi-scale CNN model is provided. The invention has simple structure, small amount of calculation, and wide applicability.

The object of the present invention is achieved by following measures: a kind of wheat disease detection method of lightweight multi-scale CNN model comprises the following steps:

S1. Use the shooting equipment to collect wheat disease images, unify the size of the wheat disease images, and classify and mark each type of disease;

S2, by rotation, symmetric overturning, improving contrast to wheat disease data expansion, original data set is expanded to 8495 pieces; According to training set: verification set: test set=6:2:2 ratio divides data set;

S3, designing a network model for wheat disease image detection;

S4, the wheat disease training samples obtained in S2 are trained in the network model designed by S3 and tested on the test set;

S5. Verify the trained model on the verification set to quickly and accurately identify various wheat diseases.

Preferably, the step S1 is specifically: saving the collected wheat disease images in the same image format, and unifying the size of the images; classifying and labeling the collected images.

Preferably, in the step S2, the collected wheat disease samples are taken from the actual wheat field, and the data set has a total of 3003 pieces. The wheat disease data is expanded by operations such as rotation, symmetrical flipping, and contrast enhancement, and the original data set is expanded to 8495 pieces. ;According to training set: verification set: the ratio division data of test set=6:2:2 obtains 5097 training sets, 1699 verification sets, and 1699 test sets.

Preferably, in the step S2, CBAM and NAM attention modules are added to each layer of residual modules; the overall attention process of CBAM can be summarized as:

Among them, Y∈C×W×H is the input feature map, C represents the number of channels of the feature map, W represents the width of the feature map, and H represents the height of the feature map; Mc represents the channel attention mechanism; Represents element-wise multiplication; Ms represents the spatial attention mechanism; Y' and Y" are the output feature maps;

The overall attention process of NAM can be summarized as:

Channel attention module:

x _c ＝GlobalAvgPool9x)∈R ^C

x _s =GlobalStdPool(x)∈R ^C

s=ScaleFactor(x)∈R

w _c ＝s·(x _c +x _s )∈R ^C

w _c ＝Threshold(w _c )∈R ^C

x'=x⊙w _c

Among them, x is the input feature map, C is the number of channels, GlobalAvgPool and GlobalStdPool are the global average pooling and global standard deviation pooling respectively, ScaleFactor is the scaling factor, w _c is the channel weight vector, Threshold is the threshold function, ⊙ is the element-wise phase Multiply, x′ is the output feature map;

Spatial attention module:

x _f ＝Conv ₁ (x)∈R ^H×W

x _f ＝Conv ₂ (x _f )∈R ^H×W

s=ScaleFactor(x _f )∈R

w _s =s x _f ∈R ^H×W

w _s ＝Threshold(w _s )∈R ^H×W

x'=x⊙w _s

Among them, x is the input feature map, H and W are the height and width, respectively, Conv ₁ and Conv ₂ are two convolutional layers, ScaleFactor is the scaling factor, w _s is the spatial weight matrix, Threshold is the threshold function, ⊙ is the step-by-step Element-wise multiplication, x' is the output feature map.

Preferably, in the step S3, the network model adopted is Inception-Resnet-CN, and the specific process is as follows:

The input of the model is a 224×224×3 RGB image. The model architecture includes three different Inception structures, two maximum pooling layers, six Residual-CN structures, an average pooling layer, and a fully connected layer; the Inception module can reduce With complex matrix dimensions and visual information aggregated in different sizes, the Residual-CN module strengthens the ability to represent disease features and reduces the impact of complex backgrounds in images on model recognition performance.

Preferably, the training method of the network model in the step S4 is specifically: during training, the image is scaled to 224×224, and the learning rate adjustment method is specifically adopted: Adam optimizes the gradient descent, and the number of pictures for each round of iteration batch processing is set to 64 , set the number of iterations to 70, and the learning rate to 0.001.

Preferably, in the step S4, cross-entropy is used to calculate the classification loss, and L2 regularization is added to punish weight parameters. The mathematical formula is: J(θ)=-∑p(x)log ₂ q(x)+λ|| θ|| ² where: p(x) is the target probability distribution; q(x) is the forecast distribution; θ is the weight parameter; λ is the regularization coefficient; ||θ|| ² is the regularization term.

Preferably, the collecting of wheat disease images in the step S1 specifically includes: taking images of farmland wheat in batches by a collection device, and then performing image preprocessing.

Beneficial effects of the present invention: Compared with the traditional machine learning wheat disease detection method, the inventive method has high speed, high accuracy and good robustness; compared with the existing deep learning wheat disease detection method, the inventive method can be used in complex It can efficiently and quickly identify wheat diseases in the background, has low resource requirements, saves costs, and can be deployed on mobile devices; compared with the detection of wheat diseases based on spectral images, the present invention reduces the cost of experimental equipment; the present invention has a reasonable design, The detection of wheat diseases in the real environment is realized, and the accurate identification of disease categories is realized, and it has the advantage of fast speed.

Description of drawings

Fig. 1 is the flow chart of wheat disease detection;

Fig. 2 is a wheat disease category diagram;

Figure 3 is a structural diagram of the Inception-Resnet-CN model;

Figure 4 shows the training results of the seven models on the dataset.

Detailed ways

The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some, not all, embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

Embodiment 1: as shown in Fig. 1-Fig. 4, a kind of wheat disease detection method of lightweight multi-scale CNN model comprises the following steps:

S1. Use the shooting equipment to collect wheat disease images, unify the size of the wheat disease images, and classify and mark each type of disease;

S2. Expand the wheat disease data by rotating, symmetrically flipping, and improving the contrast, and expand the original data set to 8495 pieces; divide the data set according to the ratio of training set: verification set: test set = 6:2:2;

S3, designing a network model for wheat disease image detection;

S4, the wheat disease training samples obtained in S2 are trained in the network model designed by S3 and tested on the test set;

S5. Verify the trained model on the verification set to quickly and accurately identify various wheat diseases.

Step S1 is specifically as follows: save the collected wheat disease images in the same image format, and unify the size of the images; classify and label the collected images; collect wheat disease images in step S1 specifically: take batches of farmland wheat images through the collection equipment, Then image preprocessing is performed.

As shown in Figure 2, in step S2, the collected wheat disease samples were photographed in the actual wheat field. The data set has a total of 3003 pieces. The wheat disease data was expanded by operations such as rotation, symmetrical flip, and contrast enhancement, and the original data set was expanded to 8495 Zhang, (leaf rust 1156, powdery mildew 1380, wheat smut 1342, root rot 1096, scab 1161, tar spot 1272, healthy 1086); according to training set: validation set: test Set = 6:2:2 The ratio of data is divided to obtain 5097 training sets, 1699 verification sets, and 1699 test sets. Here, each category is allocated according to 6:2:2.

In the step S2, CBAM and NAM attention modules are added to each layer of residual modules; the overall attention process of CBAM can be summarized as:

Among them, Y∈C×W×H is the input feature map C represents the number of channels of the feature map, W represents the width of the feature map, and H represents the height of the feature map; Mc represents the channel attention mechanism; Represents element-wise multiplication; Ms represents the spatial attention mechanism; Y' and Y" are the output feature maps;

The overall attention process of NAM can be summarized as:

Channel attention module:

x _c ＝GlobalAvgPool(x)∈R ^C

x _s =GlobalStdPool(x)∈R ^C

s=ScaleFactor(x)∈R

w _c ＝s×(x _c +x _s )∈R ^C

w _c ＝Threshold(w _c )∈R ^C

x'=x⊙w _c

Among them, x is the input feature map, C is the number of channels, GlobalAvgPool and GlobalStdPool are the global average pooling and global standard deviation pooling respectively, ScaleFactor is the scaling factor, w _c is the channel weight vector, Threshold is the threshold function, ⊙ is the element-wise phase Multiply, x′ is the output feature map.

Spatial attention module:

x _f ＝Conv ₁ (x)∈R ^H×W

x _f ＝Conv ₂ (x _f )∈R ^H×W

s=ScaleFactor(x _f )∈R

w _s =s×x _f ∈R ^H×W

w _s ＝Threshold(w _s )∈R ^H×W

x'=x⊙w _s

Among them, x is the input feature map, H and W are the height and width, respectively, Conv ₁ and Conv ₂ are two convolutional layers, ScaleFactor is the scaling factor, w _s is the spatial weight matrix, Threshold is the threshold function, ⊙ is the step-by-step Element-wise multiplication, x' is the output feature map.

As shown in Figure 3, in step S3, the network model adopted is Inception-Resnet-CN, and the specific process is as follows:

The input of the model is a 224×224×3 RGB image. The model architecture includes three different Inception structures, two maximum pooling layers, six Residual-CN structures, an average pooling layer, and a fully connected layer; the Inception module can reduce Complex matrix dimensions and aggregating visual information at different dimensions. The Residual-CN module strengthens the ability to represent disease features and reduces the impact of complex backgrounds in images on model recognition performance. Among them, the Inception-A structure uses 1×1 convolution and 3×3 convolution, and the 5×5 convolution replaced by two 3×3 convolutions is combined in parallel. The number of branch1～branch4 convolution kernels is 8, 12, respectively. 24, 8, 12, 24, 24. The Residual-CN-A structure includes two convolutional layers with a convolution kernel of 3×3, a CBAM module, a NAM module, and an identity map. The number of convolution kernels of Residual-CN-A1 to A4 is 64, 128, 128, and 256, respectively. On the basis of the Residual-CN-A structure, the Residual-CN-B structure realizes the matching of the number of channels of the two channels through the 1×1 convolution at the identity map. The number of convolution kernels of Residual-CN-B1～B2 is 128 and 256. The Inception-B structure is combined using 1×1 convolution, asymmetric 1×7 convolution, and 7×1 convolution in series. The number of convolution kernels in branch1~branch4 is 32, 32, 64, 64, 32, 64, 64, 64, 64, and 32, respectively. The Inception-C structure uses a series-parallel combination of 1×1 convolution, asymmetric 1×3 convolution, and 3×1 convolution. The number of convolution kernels of branch1~branch4 are 64, 64, 128, 96, 96, 256, 256, 256, 96, 96 respectively.

The training method of the network model in step S4 is as follows: during training, the image is scaled to 224×224, and the learning rate adjustment method is specifically adopted: Adam optimizes the gradient descent, sets the number of pictures for each round of iteration batch processing to 64, and sets the number of iterations to 70 with a learning rate of 0.001.

In step S4, cross-entropy is used to calculate the classification loss, and L2 regularization is added to punish the weight parameters. The mathematical formula is: J(θ)=-∑p(x)log ₂ q(x)+λ‖θ‖ ² where: p(x) is the target probability distribution; q(x) is the predictive distribution; θ is the weight parameter; λ is the regularization coefficient; ‖θ‖ ² is the regularization term.

Model Performance Evaluation Metrics

Precision refers to the proportion of correct predictions in the samples predicted to be positive, Accuracy accuracy rate, Recall refers to the proportion of correct prediction results in the actual positive samples, F1 value is the weighted average of precision rate and recall rate. Among them, TP is the real situation is true, the prediction result is the number of positives samples, TN is the real situation is true, the prediction result is the number of negatives samples, similarly, FP is the real situation is false, the prediction result is the number of positives samples, FN is The true situation is false, and the predicted result is the number of negatives samples.

Referring to Figure 4, the obtained data is trained in various models, and the performance of various algorithms is compared.

The InceptionResnet-CN model has the highest accuracy, reaching 98.76%, and the model performance is good. After the evaluation, the model detects wheat disease images, outputs disease categories, and achieves a high detection rate, which can make important contributions to food security and agricultural development. .

Finally, it should be noted that: the above is only a preferred embodiment of the present invention, and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, for those skilled in the art, it still It is possible to modify the technical solutions recorded in the foregoing embodiments, or to perform equivalent replacements on some of the technical features. Any modifications, equivalent replacements, improvements, etc. within the spirit and principles of the present invention shall be included in the within the protection scope of the present invention.

Claims (8)

1. A wheat disease detection method of a lightweight multi-scale CNN model is characterized by comprising the following steps of: the method comprises the following steps:

s1, acquiring wheat disease images by using shooting equipment, unifying the sizes of the wheat disease images, and classifying and labeling each type of disease;

s2, expanding the wheat disease data by rotation, symmetrical overturning and contrast improvement, and expanding an original data set to 8495 pieces; dividing the data set according to the proportion of training set to verification set to test set=6:2:2;

s3, designing a network model for detecting wheat disease images;

s4, training the wheat disease training sample obtained in the S2 in a network model designed in the S3 and testing the wheat disease training sample on a test set;

and S5, verifying the trained model on a verification set, and rapidly and accurately identifying various wheat diseases.

2. The method for detecting wheat diseases by using a lightweight multi-scale CNN model according to claim 1, which is characterized in that: the step S1 is to store the collected wheat disease images in the same picture format and unify the sizes of the pictures; and classifying and labeling the acquired images.

3. The method for detecting wheat diseases by using a lightweight multi-scale CNN model according to claim 1, which is characterized in that: in the step S2, the collected wheat disease samples are photographed in real wheat fields, the total data sets are 3003, the wheat disease data are expanded through operations such as rotation, symmetrical overturning, contrast improvement and the like, and the original data sets are expanded to 8495; the data were divided according to the training set: validation set: test set = 6:2:2 ratio to obtain 5097 training sets, 1699 validation sets, 1699 test sets.

4. A method for detecting wheat diseases by using a lightweight multi-scale CNN model according to claim 3, wherein: in the step S2, adding a CBAM and NAM attention module into each layer of residual error module; the overall attention process of CBAM can be summarized as:

wherein Y epsilon C X W X H is an input feature map, C represents the channel number of the feature map, W represents the width of the feature map, and H represents the height of the feature map; mc represents a channel attention mechanism;representing element-by-element multiplication; ms represents spatial attention mechanisms; y 'and Y' are output feature graphs;

the overall attention process of NAM can be summarized as:

channel attention module:

x _c ＝GlobalAvgPoolx∈R ^C

x _s ＝GlobalStdPoolx∈R ^C

s＝ScaleFactorx∈R

w _c ＝s·(x _c +x _s )∈R ^C

w _c ＝Threshold(w _c )∈R ^C

x′＝x⊙w _c

wherein x is the input feature map, C is the channel number, globalAvgPool and GlobalStdPool are global average pooling and global standard deviation pooling, respectively, scaleFactor is the scaling factor, w _c Is the channel weight vector, threshold is the Threshold function, as is element-wise multiplication, x' is the output feature map;

spatial attention module:

x _f ＝Conv ₁ (x)∈R ^H×W

x _f ＝Conv ₂ (x _f )∈R ^H×W

s＝ScaleFactor(x _f )∈R

w _s ＝s·x _f ∈R ^H×W

w _s ＝Threshold(w _s )∈R ^H×W

x′＝x⊙w _s

wherein x is an input feature map, H and W are height and width, respectively, conv ₁ And Conv ₂ Two convolution layers, scaleFactor is the scaling factor, w _s Is a spatial weight matrix, threshold is a Threshold function, +..

5. The method for detecting wheat diseases by using a lightweight multi-scale CNN model according to claim 1, which is characterized in that: in the step S3, the adopted network model is an admission-Resnet-CN, and the specific process is as follows:

the input of the model is a 224×224×3 RGB image, and the model architecture comprises three different acceptance structures, two maximum pooling layers, six Residual-CN structures, an average pooling layer and a full connection layer; the acceptance module can reduce the dimension of a complex matrix and aggregate visual information on different sizes, and the Residual-CN module enhances the characterization capability of disease features and reduces the influence of a complex background in an image on the recognition performance of the model.

6. The method for detecting wheat diseases by using a lightweight multi-scale CNN model according to claim 1, which is characterized in that: the training method of the network model in the step S4 specifically comprises the following steps: during training, the images are scaled to 224 multiplied by 224, and the learning rate adjustment mode specifically adopts: adam optimizes gradient descent, sets the number of pictures in each iteration batch to be 64, sets the iteration number to be 70 and sets the learning rate to be 0.001.

7. The method for detecting wheat diseases by using a lightweight multi-scale CNN model according to claim 1, which is characterized in that: in the step S4, the cross entropy is adopted to calculate the classification loss, and the L2 regularization is added to punish the weight parameter, and the mathematical formula is as follows: j (θ) = - Σp (x) log ₂ q(x)+λ‖θ‖ ² Wherein: p (x) is a target probability distribution; q (x) is the prediction distribution;θ is a weight parameter; lambda is the regularization coefficient; II theta II ² Is a regularization term.

8. The method for detecting wheat diseases by using a lightweight multi-scale CNN model according to claim 1, which is characterized in that: the step S1 is to collect wheat disease images, specifically, the farmland wheat images are shot in batches through collecting equipment, and then image preprocessing is carried out.