A Convolutional Neural Network Model and Method for Improving the Recognition Rate of Crop Diseases and Pests

Abstract

The present invention discloses a convolutional neural network model and method for improving the recognition rate of crop pests and diseases, including: a comprehensive learning module, a reduction module, a residual connection module, a preliminary perception module, a deep learning module, and a summary and induction module connected in sequence; the deep learning module includes four deep learning convolutional units connected in sequence, and the deep learning convolutional unit includes a deep learning convolutional subunit and a deep learning residual connection subunit for connecting with the residual connection module for residual connection. In the technical solution of the present invention, in the design of multiple convolutional kernels, efforts are made to effectively cover features and achieve the most efficient learning. In the overall comprehensive learning, through residual connection, the original image size is restored, reducing the accuracy caused by distortion and deformation. Then, in-depth learning is carried out, continuously increasing the depth and expanding the output to learn more advanced features, thereby improving the recognition and analysis accuracy of pests and diseases and reducing confusion and misjudgment.

Description

Technical Field

[0001] The present invention relates to the technical field of crop pest and disease identification, and particularly relates to a convolutional neural network model and method for improving the identification rate of crop pests and diseases. Background Art

[0002] Crop pests and diseases are important problems in agricultural production. If the pest and disease problems cannot be detected and controlled in time, the pests and diseases will spread, affecting the yield and quality of crops. By improving the identification rate of crop pests and diseases, the types and degrees of pests and diseases can be judged more accurately, so as to take targeted control measures, reduce unnecessary pesticide use, and avoid overuse of pesticides, thereby protecting the ecological environment. During the data collection process, the collected pictures may be blurred, distorted, overexposed or underexposed due to reasons such as light, shooting angle, equipment, etc.; different shooting angles may lead to differences in the color, shape, etc. of pests and diseases; background interference caused by the appearance of other plants or objects; insufficient number of pictures; and a wide variety of pest and disease types. All these problems result in a low accuracy rate of the general neural network structure in analyzing and identifying images, which is extremely likely to cause confusion and misjudgment. The present invention designs a new convolutional neural network model to solve the above-mentioned problems encountered in image recognition. Summary of the Invention

[0003] The main object of the present invention is to provide a convolutional neural network model and method for improving the identification rate of crop pests and diseases, aiming to solve the problems that the existing identification models have low accuracy in identifying and analyzing pests and diseases, and are prone to confusion and misjudgment.

[0004] To achieve the above object, the convolutional neural network model for improving the identification rate of crop pests and diseases proposed by the present invention includes: a comprehensive learning module, a reduction module, a residual connection module, a preliminary perception module, a deep learning module, and a summary and induction module connected in sequence;

[0005] The comprehensive learning module includes a plurality of comprehensive learning convolutional units, the comprehensive learning convolutional units have convolutional kernels of different sizes, and the plurality of comprehensive learning convolutional units are used to learn a plurality of preliminary features;

[0006] The reduction module is used to reduce the size of the preliminary features to the size of the original image;

[0007] The residual connection module is used to connect the original image data and the learned features;

[0008] The preliminary perception module includes a connection convolutional unit and a pooling unit that performs a max pooling operation on the output feature map;

[0009] The deep learning module includes four sequentially connected deep learning convolutional units. The deep learning convolutional unit includes a deep learning convolutional subunit and a deep learning residual connection subunit for connecting with the residual connection module for residual connection. The convolutional kernel sizes of the four deep learning convolutional units are all "3×3", and the stride is "2". The number of output feature maps starts from "64" and doubles layer by layer. The size of the feature maps starts from "56" and halves layer by layer.

[0010] The summary module includes an average pooling layer unit for averaging the feature maps.

[0011] Preferably, the convolutional kernel size range of the comprehensive learning convolutional unit is from "3×3" to " "; the stride is designed as " ", and the number of filters is 1, where width is the width of the image, height is the height of the image, and kernel_size is the size of the convolutional kernel.

[0012] Preferably, the reduction module includes a comparison unit, a supplementary convolutional unit, and a reduction convolutional unit. The comparison unit is used to perform integer division comparison with the original image size. The convolutional kernel size of the supplementary convolutional unit is the remainder after integer division, and the stride is the width of the original image size.

[0013] The reduction unit is used to reshape the initially learned aggregated features with the original image size and number of channels as features and restore them to the real image size.

[0014] Preferably, the residual connection module includes a summation unit for summing the original image data and the learned features.

[0015] Preferably, the preliminary perception module includes a convolutional kernel of "7×7" and a connection convolutional unit with a stride of "1", with an output size of "32", and then a pooling unit for performing max pooling on the output feature map. The pooling size of the pooling unit is "3×3", and the stride is "2", and the output size is also "32".

[0016] Preferably, the average pooling layer unit is used to pool the feature maps with an output_size of "7×7" and filters of "512".

[0017] Preferably, the number of learning rounds of the first and fourth deep learning convolutional units of the deep learning module is two rounds, and the number of learning rounds of the second and third deep learning convolutional units of the deep learning module is three rounds.

[0018] Preferably, the deep learning convolutional unit includes two sequentially connected deep learning convolutional subunits for performing two deep convolutions.

[0019] Preferably, the deep learning residual connection sub-unit is used to connect with the residual connection module at the end of each round of learning for residual connection.

[0020] A method for improving the recognition rate of crop diseases and pests includes the following steps:

[0021] Construct the convolutional neural network model for improving the recognition rate of crop diseases and pests as described above, and preset the dynamic connection module and the network structure module;

[0022] Load the data set;

[0023] Train and maintain the model;

[0024] Verify the model;

[0025] Use the verified model for recognizing crop diseases and pests.

[0026] In the technical solution of the present invention, in the design of multiple convolutional kernels, efforts are made to effectively cover features, supplemented by appropriate strides and the selection of dynamic connection methods to achieve maximum efficiency learning. In the overall comprehensive learning, the known features are reshaped, and through residual connection, the original image size is restored, reducing the accuracy caused by distortion and deformation, and then entering in-depth learning, continuously increasing the depth and expanding the output, so as to gradually learn more advanced features and achieve good learning effects at a smaller depth, thereby improving the recognition and analysis accuracy of diseases and pests and reducing confusion and misjudgment. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on the structures shown in these drawings.

[0028] Figure 1 It is a schematic structural diagram of the convolutional neural network model and method for improving the recognition rate of crop diseases and pests of the present invention.

[0029] Figure 2 It is a schematic structural diagram of the residual connection in the convolutional neural network model for improving the recognition rate of crop diseases and pests of the present invention.

[0030] Figure 3 It is a parameter comparison diagram of the convolutional neural network model for improving the recognition rate of crop diseases and pests of the present invention and different models.

[0031] Figure 4These are the training results demonstrated by ResNet-5 in 20 rounds.

[0032] Figure 5 These are the training results demonstrated by ResNet50 in 20 rounds.

[0033] The realization, functional features and advantages of the object of the present invention will be further described in conjunction with the embodiments with reference to the accompanying drawings. Specific Embodiments

[0034] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts belong to the scope of protection of the present invention.

[0035] In addition, in the present invention, descriptions such as "first" and "second" are only for descriptive purposes, and cannot be understood as indicating or implying their relative importance or implicitly indicating the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include at least one of such features. In the description of the present invention, the meaning of "a plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0036] In the present invention, unless otherwise clearly specified and limited, terms such as "connection" and "fixation" should be understood in a broad sense. For example, "fixation" can be a fixed connection, a detachable connection, or integrated; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium, and can be the communication inside two components or the interaction relationship between two components, unless otherwise clearly limited. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific situations.

[0037] In addition, the technical solutions between the various embodiments of the present invention can be combined with each other, but it must be based on the fact that those of ordinary skill in the art can implement it. When the combination of technical solutions appears to be contradictory or unable to be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0038] Please refer to Figures 1 - 5 , the present invention proposes a convolutional neural network model for improving the recognition rate of crop pests and diseases, including; a comprehensive learning module, a reduction module, a residual connection module, a preliminary perception module, a deep learning module and a summary and induction module connected in sequence;

[0039] The comprehensive learning module includes multiple comprehensive learning convolutional units, and the comprehensive learning convolutional units have convolutional kernels of different sizes. The multiple comprehensive learning convolutional units are used to learn multiple preliminary features;

[0040] The reduction module is used to reduce the size of the preliminary features to the size of the original image;

[0041] The residual connection module is used to connect the original image data and the learned features;

[0042] The preliminary perception module includes a connection convolutional unit and a pooling unit that performs a max pooling operation on the output feature map;

[0043] The deep learning module includes four sequentially connected deep learning convolutional units. The deep learning convolutional unit includes a deep learning convolutional subunit and a deep learning residual connection subunit used to connect with the residual connection module for residual connection; the convolutional kernel sizes of the four deep learning convolutional units are all "3×3", and the stride is "2"; the number of output feature maps starts from "64" and doubles layer by layer; while the size of the feature map starts from "56" and halves layer by layer;

[0044] The summary module includes an average pooling layer unit for averaging the feature map.

[0045] In the technical solution of the present invention, in the design of multiple convolutional kernels, efforts are made to effectively cover features, supplemented by appropriate strides and selection of dynamic connection methods to achieve learning with the highest efficiency. In the overall comprehensive learning, the learned features are reshaped, and through residual connection, the size of the original image is restored, reducing the accuracy loss caused by distortion and deformation. Then, it enters in-depth learning, continuously increasing the depth and expanding the output, so as to gradually learn more advanced features and achieve good learning effects at a relatively small depth, thereby improving the accuracy of identifying and analyzing pests and diseases and reducing confusion and misjudgment.

[0046] Specifically, the convolutional neural network structure simulates the learning process of the human brain and reflects a complete learning process.

[0047] It includes a total of six layers. The first layer is the comprehensive learning layer, which realizes an overall understanding through a reasonably designed learning block;

[0048] The second layer is the reduction layer, which tries to fit the knowledge learned preliminarily with the original data as much as possible;

[0049] The third layer is the residual connection layer, which combines the original data for a general review to reduce omissions and errors;

[0050] The fourth layer is the preliminary perception layer, which performs preliminary processing on visual information, such as detecting edges, corners, etc., extracts basic image features, and lays a foundation for subsequent in-depth understanding.

[0051] The fifth layer is the deep learning layer. As learning progresses, the breadth of knowledge will continuously shrink, while the understanding of knowledge will continuously increase. That is, the output size of each layer of the network will decrease as the network depth increases, and the number of filters will increase as the network depth increases, enabling the extraction of richer and more complex features.

[0052] The sixth layer is the summarization layer. After accumulating a large amount of information, it gradually summarizes and generalizes more abstract and high-level knowledge and rules.

[0053] In another embodiment of the present invention, the convolution kernel size range of the comprehensive learning convolution unit is from "3×3" to " "; the stride is designed as " ", and the number of filters is 1, where width is the width of the image, height is the height of the image, and kernel_size is the size of the convolution kernel.

[0054] Specifically, if a full-size convolution kernel is designed, that is, in the size range from "1×1, 2×2, 3×3, 4×4,..., w×w", although the resulting effect is good, the economy cannot bear it. As the convolution kernel increases, the computational complexity of the network also increases, which will lead to slower training and inference times; it will increase the number of network parameters and model size, which may lead to an increase in the cost of model storage and transmission; it will increase the risk of overfitting in the network because each convolution kernel can learn different features, and if not restricted, the network may over-rely on certain specific features; at the same time, more training data is required, otherwise the network may have an underfitting problem.

[0055] The minimum size range of the convolution kernel is designed as "3×3", and the maximum is designed as from "3×3" to " "; the stride is designed as " ", and the best effect is achieved when the number of filters is 1. Here, width is the width of the image, height is the height of the image, kernel_size is the size of the convolution kernel, and filters is the number of filters.

[0056] After each convolution kernel has completed learning, the learned features need to be aggregated. At this time, only simple appending and accumulation are required, without the need for complex processing.

[0057] In yet another embodiment of the present invention, the reduction module includes a comparison unit, a supplementary convolution unit, and a reduction convolution unit. The comparison unit is used to perform an integer division comparison with the original image size. The convolution kernel size of the supplementary convolution unit is the remainder after integer division, and the stride is the width of the original image size;

[0058] The reduction unit is used to reshape the preliminarily learned aggregated features with the original image size and number of channels and restore them to the real image size.

[0059] Specifically, the preliminary features learned by convolution kernels of various different sizes have been aggregated, and a preliminary understanding of the knowledge has been obtained. At this time, there is already a certain difference from the size of the original image, so it is necessary to restore the size in order to consolidate learning on the original knowledge basis. Among them, reshaping the preliminarily learned aggregated features and restoring them to the real image size is reshape, which restores the feature points after convolution to the original image size.

[0060] First, flatten the preliminary learning features and compare them with the original image size "width×height". If it cannot be divided evenly, use the remainder as the size of the convolution kernel, the number of filters is 1, and the stride is the width of the image for supplementary learning, which is equivalent to a process of checking for omissions and making up deficiencies, and summarize the features of the supplementary part into the first-layer features.

[0061] Then, with the original image size and number of channels as features, reshape the preliminarily learned aggregated features and restore them to the real image size. This feature will be saved for a long time and used in subsequent deep learning. This is similar to human learning. When encountering problems in in-depth learning, one has to constantly go back to the content learned before.

[0062] In another embodiment of the present invention, the residual connection module includes a summation unit that sums the original image data and the learned features.

[0063] Specifically, after restoring the preliminarily learned features, it is necessary to combine the original image information for residual connection. This can directly transfer the gradient to the original image layer, alleviate the problem of gradient disappearance; accelerate the convergence speed of the network, reduce the training time, and improve the accuracy of the model; and can make the network easier to learn and capture the non-linear relationships in the data. The residual connection is relatively simple because the second reduction layer has been reshaped and has the same dimension as the original input data, and at this time, direct summation is only required.

[0064] In another embodiment of the present invention, the preliminary perception module includes a convolution kernel of "7×7" and a connecting convolution unit with a stride of "1", the output size is "32", and then a pooling unit that performs a max-pooling operation on the output feature map. The pooling size of the pooling unit is "3×3", the stride is "2", and the output size is also "32".

[0065] Specifically, during the training process, this setting can significantly improve the performance and accuracy of the network. Because a larger convolutional kernel can cover a larger local area, thereby increasing the receptive field of each convolutional layer, enabling the network to learn more global feature information; using a convolutional operation with a stride of "2" can reduce the image size by half, which can reduce the number of model parameters, and also reduce the computational amount and memory occupancy; a larger convolutional kernel can avoid the problem of gradient disappearance, and can better initialize the network weights, thereby improving the performance ability of the model.

[0066] In another embodiment of the present invention, the average pooling layer unit is used to perform pooling on the feature map with an output size of "7×7" and the number of filters of "512".

[0067] Specifically, after the completion of deep learning, at this time, pooling is performed on the feature map with an output size of "7×7" and the number of filters of "512", which can effectively reduce the number of parameters and reduce the risk of overfitting; and performing an average operation on the feature map can retain more feature information, enabling the classifier to better understand and recognize images; the average pooling layer has a certain invariance, and for image translation, rotation, scaling and other transformations, its output result will not change, which can also enhance the robustness of the network.

[0068] In another embodiment of the present invention, the number of learning rounds of the first and fourth deep learning convolutional units of the deep learning module is two rounds, and the number of learning rounds of the second and third deep learning convolutional units of the deep learning module is three rounds.

[0069] Specifically, the first round of learning is not deep enough and the learning difficulty is not great, so the number of learning rounds is small. As the learning progresses, the complexity of the middle two rounds has increased, so there should be more learning rounds. In the last round, the learned features have been basically mastered and do not require too much repeated learning, so the number of learning rounds has also decreased.

[0070] In another embodiment of the present invention, the deep learning convolutional unit includes two sequentially connected deep learning convolutional sub-units to perform two deep convolutions.

[0071] Specifically, each round of learning for each small layer includes 2 deep convolutions.

[0072] In another embodiment of the present invention, the deep learning residual connection sub-unit is used to connect with the residual connection module for residual connection at the end of each round of learning.

[0073] Specifically, after the first round of learning for each small layer, a residual connection needs to be made with the previous second-layer restoration layer to prevent the disappearance of gradients caused by the increase in the learning depth. This is like in the process of human brain learning, some important information may be forgotten or fade away, resulting in an insufficiently in-depth and comprehensive understanding and mastery of knowledge. Similarly, in the process of neural network training, the disappearance of gradients will also cause some key information to gradually be lost during transmission, affecting the learning effect and performance of the network.

[0074] A method for improving the recognition rate of crop pests and diseases includes the following steps:

[0075] Construct the convolutional neural network model for high crop pest and disease recognition rate as described in any of the above embodiments, and preset a dynamic connection module and a network structure module;

[0076] Load the data set;

[0077] Train and maintain the model;

[0078] Verify the model;

[0079] Use the verified model for crop pest and disease recognition.

[0080] In another embodiment of the present invention, taking the crop pest and disease picture library as an example, the specific process of using this neural network structure is described.

[0081] The first step: Design of the dynamic connection module. In order to manually calculate the shape of the output tensor and complete the weight initialization during network compilation.

[0082] The role of selecting the dynamic connection method is to optimize the model complexity and computational efficiency of the convolutional neural network. A custom layer class needs to be inherited from tf.keras.layers.Layer, and functions such as __init__, build, call, and compute_output_shape need to be implemented in the code.

[0083] For example, in the build function, the operations executed during network compilation are to complete the weight initialization, and the input parameter input_shape is the shape of the first input parameter in call.

[0084] def build(self, input_shape):

[0085] kernel_shape = tf.TensorShape((self.kernel_size[0], self.kernel_size[1],

[0086] input_shape[-1], self.output_dim))

[0087] self.choice = self.add_weight(name='kernel', shape=kernel_shape,

[0088] initializer=tf.initializers.he_normal(), trainable=True)

[0089] Step 2: Design the custom network structure module. It includes basic structures such as basic convolution blocks, reduction blocks, full convolution blocks, depth convolution blocks, residual reduction blocks, and residual depth blocks. The code for the basic convolution block structure is as follows:

[0090] def surface_conv(self):

[0091] y = None

[0092] filters = 1

[0093] end = math.floor(math.sqrt(self.width))

[0094] for kernel_size in range(3, end):

[0095] stride = math.floor(math.sqrt(kernel_size))

[0096] if stride <= 1:

[0097] stride = 2

[0098] x = self.basic_conv_block(self.input_source, filters, kernel_size, stride)

[0099] if y is not None:

[0100] y = tf.concat([y, Flatten()(x)], 1)

[0101] else:

[0102] y = Flatten()(x)

[0103] if y.shape[-1] % np.square(self.width) != 0:

[0104] x = self.basic_conv_block(self.input_source,

[0105] np.square(self.width) - y.shape[-1] % np.square(self.width), 1, self.width)

[0106] y = tf.concat([y, Flatten()(x)], 1)

[0107] The code for constructing the model structure is as follows:

[0108] def make(self):

[0109] # Comprehensive summary learning

[0110] x = self.surface_conv()

[0111] # Summary after learning

[0112] y1 = self.reshape_block(x)

[0113] # Final review after learning

[0114] x = self.reshape_residual_block(x)

[0115] # In-depth learning + summary review

[0116] x = self.deep_conv(x, y1)

[0117] x = Flatten()(x)

[0118] x = Dropout(0.25)(x)

[0119] y = Dense(self.classes)(x)

[0120] y = Softmax(axis=-1)(y)

[0121] model = Model([self.input_source], [y])

[0122] Step 3: Dataset Loading. Collect crop pest and disease images, and after processing, put them into the image library. The processing includes converting images in different formats into the specified format, which is png here; processing the images into the specified size, which is 224*224 in this example; performing deformation operations such as uniformly scaling, cutting, and rotating the images; and dividing the dataset into a training set and a validation set, with a ratio of 1:4 here.

[0123] Step 4: Model Training and Saving.

[0124] def training(model, save_path, train_dir, model_file):

[0125] train_datagen = ImageDataGenerator(rescale = 1. / 255, shear_range = 0.2, zoom_range = 0.2,

[0126] validation_split = 0.25, horizontal_flip = True)

[0127] train_generator = train_datagen.flow_from_directory(train_dir,

[0128] target_size=(width, width), batch_size = batchSize)

[0129] val_generator = train_datagen.flow_from_directory(train_dir,

[0130] target_size=(width, width), batch_size = batchSize)

[0131] model.compile(optimizer = SGD(learning_rate = 0.001, momentum = 0.9),

[0132] loss='categorical_crossentropy', metrics=['accuracy'])

[0133] train = model.fit(train_generator, steps_per_epoch = len(train_generator), epochs = epochs,

[0134] validation_data = val_generator, validation_steps = len(val_generator))

[0135] Step 5: Model validation. Mainly select several key indicators such as Epoch, Training Accuracy, Validation Accuracy, Training loss, and Validation loss.

[0136] The main code is as follows, and the verification results:

[0137] acc = train.history['accuracy']

[0138] val_acc = train.history['val_accuracy']

[0139] loss = train.history['loss']

[0140] val_loss = train.history['val_loss']

[0141] ep = np.arange(0, epochs)

[0142] plt.figure()

[0143] plt.plot(ep, acc, label='Training Accuracy')

[0144] plt.plot(ep, val_acc, label='Validation Accuracy')

[0145] plt.plot(ep, loss, label='Training loss')

[0146] plt.plot(ep, val_loss, label='Validation loss')

[0147] plt.title('Training and Validation Accuracy / Loss')

[0148] plt.xlabel("Epoch#")

[0149] plt.ylabel("Loss / Accuracy")

[0150] plt.legend().

[0151] The verification results can be referred to Figures 3 - 5 :

[0152] The custom convolutional neural network structure of the present invention has achieved good results, mainly including the following aspects

[0153] 1. With the same number of network layers, the number of parameters and computational complexity of the neural network are greatly reduced.

[0154] For details, see Figure 3 , the rightmost ResNet-5 layer is the convolutional neural network structure of the present invention. Obviously, it only has 25 layers, and the number of parameters and computational complexity are also the smallest. Under the same dataset and parameter conditions, its training time per round saves 7% to 38% of the time compared to ResNet50.

[0155] 2. For the recognition of crop pests and diseases, compared with the traditional network structure, the recognition rate has been significantly improved (for details, see Figure 4 and Figure 5 ).

[0156] Figure 4 is the training result shown by ResNet-5 at 20 rounds, Figure 5 is the training result shown by ResNet50 at 20 rounds.

[0157] Obviously, Figure 4 the learning efficiency of the training set is much higher, the accuracy rises extremely fast, and the loss rate gradually drops to a lower level; the accuracy of the validation set is also gradually rising, and the loss rate rises instead of falling after the 10th round, showing an overfitting phenomenon.

[0158] Figure 5 the learning efficiency of the training set of Figure 4 is relatively much lower, and the loss rate drops relatively gently; the accuracy of the validation set rises steadily, and the loss rate is better than that of Figure 4

[0159] The above is only the preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structural transformation made under the concept of the present invention by using the content of the specification and drawings of the present invention, or directly / indirectly applied in other related technical fields, is included in the patent protection scope of the present invention.

Claims

1. A convolutional neural network model for improving the recognition rate of crop diseases and insect pests, characterized in that: include: A comprehensive learning module, a restoration module, a residual connection module, a preliminary perception module, a deep learning module and a summary module connected in sequence; The comprehensive learning module includes a plurality of comprehensive learning convolution units, wherein the comprehensive learning convolution units have convolution kernels of different sizes, and the plurality of comprehensive learning convolution units are used to learn a plurality of preliminary features; The restoration module is used to restore the size of the preliminary feature to the size of the original image; The restoration module includes a comparison unit, a supplementary convolution unit and a restoration convolution unit, wherein the comparison unit is used to perform integer division comparison with the original image size, the convolution kernel size of the supplementary convolution unit is the remainder after integer division, and the step length is the width of the image size and the step length is the width of the original image size; The restoration convolution unit is used to reshape the initially learned summary features based on the original image size and the number of channels, and restore them to the real image size; The residual connection module is used to connect the original image data and the learning features; The preliminary perception module includes a connection convolution unit and a pooling unit for performing a maximum pooling operation on the output feature map; The deep learning module includes four deep learning convolution units connected in sequence, and the deep learning convolution unit includes a deep learning convolution subunit and a deep learning residual connection subunit for connecting to the residual connection module for residual connection; the convolution kernel size of the four deep learning convolution units is "3×3" and the step size is "2"; the number of feature maps output starts from "64" and doubles layer by layer; and the size of the feature map starts from "56" and decreases layer by layer; The summarization module includes an average pooling layer unit for performing an average operation on the feature map.

2. The convolutional neural network model for improving the recognition rate of crop diseases and insect pests as claimed in claim 1, characterized in that: The convolution kernel size of the fully learned convolution unit ranges from "3×3" to " "; step length is designed as" ", the number of filters is 1, where width is the width of the image, height is the height of the image, and kernel_size is the size of the convolution kernel.

3. The convolutional neural network model for improving the recognition rate of crop diseases and insect pests as claimed in claim 1, characterized in that: The residual connection module includes a summing unit for summing the original image data and the learned features.

4. The convolutional neural network model for improving the recognition rate of crop diseases and insect pests as claimed in claim 1, characterized in that: The preliminary perception module includes a convolution kernel of "7×7" and a connected convolution unit with a stride of "1", and an output size of "32". Then, a pooling unit that performs a maximum pooling operation on the output feature map has a pooling size of "3×3", a stride of "2", and an output size of "32".

5. The convolutional neural network model for improving the recognition rate of crop diseases and insect pests as claimed in claim 1, characterized in that: The average pooling layer unit is used to pool the feature map with output_size of "7×7" and filters of "512".

6. The convolutional neural network model for improving the recognition rate of crop pests and diseases as described in any one of claims 1 to 5, characterized in that: The number of learning rounds of the first and fourth deep learning convolution units of the deep learning module is two rounds, and the number of learning rounds of the second and third deep learning convolution units of the deep learning module is three rounds.

7. The convolutional neural network model for improving the recognition rate of crop diseases and insect pests as claimed in claim 6, characterized in that: The deep learning convolution unit includes two deep learning convolution sub-units connected in sequence to perform two deep convolutions.

8. The convolutional neural network model for improving the recognition rate of crop diseases and insect pests as claimed in claim 6, characterized in that: The deep learning residual connection subunit is used to connect with the residual connection module to perform residual connection at the end of each round of learning.

9. A method for improving the recognition rate of crop diseases and insect pests, characterized in that: The steps include: Constructing a convolutional neural network model with a high crop disease and insect pest recognition rate as described in any one of claims 1 to 8, and presetting a dynamic connection module and a network structure module; Load the dataset; Train and maintain the model; Validate the model; The validated model is used to identify crop diseases and pests.

Images