CN114241309A - Rice sheath blight identification method and system based on ShuffleNet V2-Unet - Google Patents

Rice sheath blight identification method and system based on ShuffleNet V2-Unet Download PDF

Info

Publication number
CN114241309A
CN114241309A CN202111554870.XA CN202111554870A CN114241309A CN 114241309 A CN114241309 A CN 114241309A CN 202111554870 A CN202111554870 A CN 202111554870A CN 114241309 A CN114241309 A CN 114241309A
Authority
CN
China
Prior art keywords
shufflenet
layer
module
model
rice
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202111554870.XA
Other languages
Chinese (zh)
Inventor
李志忠
秦俊豪
李优新
程昱
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Guangdong University of Technology
Original Assignee
Guangdong University of Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Guangdong University of Technology filed Critical Guangdong University of Technology
Priority to CN202111554870.XA priority Critical patent/CN114241309A/en
Publication of CN114241309A publication Critical patent/CN114241309A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computational Linguistics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Evolutionary Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a rice sheath blight identification method and system based on ShuffleNet V2-Unet, the method comprises the following steps: s1: collecting a rice disease data set, wherein the rice disease data set comprises an image with rice sheath blight disease streaks; s2: establishing a ShuffleNet V2-Unet model, wherein the ShuffleNet V2-Unet model is used for identifying the stringy and stringy striae of rice; s3: training the ShuffleNet V2-Unet model by using the rice disease data set in the step S1 to obtain a trained ShuffleNet V2-Unet model; s4: and (4) identifying the rice sheath blight disease of the input picture by using the trained ShuffleNet V2-Unet model. According to the method, rice sheath blight is identified through the ShuffleNet V2-Unet model, rice sheath blight lines can be effectively identified in a complex environment, the detection speed can be kept while high detection precision is kept, meanwhile, the method is low in requirement on the detection environment for detecting the rice sheath blight, and operation for detecting the rice sheath blight is convenient.

Description

Rice sheath blight identification method and system based on ShuffleNet V2-Unet
Technical Field
The invention relates to the field of image processing and pattern recognition, in particular to a rice sheath blight recognition method and system based on ShuffleNet V2-Unet.
Background
The rice sheath blight disease is one of main diseases of rice planting, the disease mainly occurs on leaf sheaths and leaves of rice, stripes of the disease are usually in irregular moire shapes, and the yellow and withered leaves are caused during the disease attack, so that the planting and production of the rice are seriously affected. The traditional method for diagnosing rice diseases is to identify the diseases of rice manually, farmers determine the types of the diseases by means of self-accumulated experience, book reference or professional consultation and the like, and the method is low in efficiency and difficult to monitor and treat the diseases in real time. Therefore, the method for detecting the rice diseases in real time and determining the types of the rice diseases is very urgent.
At present, the practical application of machine vision to rice sheath blight detection is less, and the main reasons are as follows: 1) the detection environment is complex, the disease lines of the sheath blight are irregular, the leaf sheaths and the leaves have different degrees of cloud lines, the color and the disease lines of the leaf sheath are relatively identified with dead leaves, and the leaf sheath are easy to confuse. 2) The early stage of rice sheath blight disease is difficult to be identified quickly. 3) The rice sheath blight and lesion is complex, single fixed characteristic is not available, and accurate category identification is difficult. 4) The traditional rice sheath blight identification algorithm has high detection environment requirement and is difficult to be practically applied.
The existing schemes applied to rice sheath blight identification mainly comprise:
(1) identification method based on convolution neural network (Liuting Ting, Wang Ting, Hulin. identification of rice sheath blight image based on convolution neural network [ J ]. China Rice science, 2019,33(01):90-94.DOI: 10.16819/j.1001-7216.2019.8051.): in the method, an author builds a convolutional neural network by using an Alexnet model, cuts the picture size into 227 x 227, and inputs the 227 x 227 into the model to identify the rice sheath blight disease.
(2) Identifying rice sheath blight disease (plum shijgh, lotus root cause channel, dingh as min, poplar red soldier, shenjuqing, trejalin based on hyperspectral imaging techniques [ J ] identified based on hyperspectral imaging techniques, proceedings of south china agricultural university, 2018,39(06): 97-103): taking healthy and banded sclerotial blight infected rice seedlings as research objects, collecting hyperspectral images of leaves and canopy, removing obvious noise parts, and respectively preprocessing by different methods to obtain the spectral curves of the rice leaves. And modeling the spectrum of different pretreatments by adopting partial least square-discriminant analysis. And extracting characteristic information of the original spectral data of the canopy by adopting an MNF algorithm, and establishing a linear discriminant analysis model and an error back propagation neural network discriminant model based on the characteristic information. .
(3) Research on a rice leaf disease rapid identification method based on image processing (research on a rice leaf disease rapid identification method based on image processing [ D ]. northeast agriculture university, 2018 ]): selecting a mobile phone as disease image acquisition equipment; establishing a rice leaf lesion database by image preprocessing methods such as graying, image denoising, lesion segmentation and the like; analyzing the characteristics of the three disease spots, and extracting characteristic parameters from three aspects of color, shape and texture; and identifying by combining the comprehensive characteristics of the color, the shape and the texture of the picture by utilizing a BP neural network model.
(4) Detection and identification based on hyperspectral images and chlorophyll content (early detection and identification of rice sheath blight disease [ J ] of hyperspectral images and chlorophyll content; spectroscopy and spectral analysis, 2019,39(06):1898 + 1904): and (3) performing early identification on the rice sheath blight disease by using a hyperspectral imaging technology and combining the chlorophyll content.
(5) Methods of classification using a Support Vector Machine (SVM) classifier (Yuan, Chenley, Wuna, et al. Rice sheath blight disease image identification processing methods research [ J ] agricultural mechanization research, 2016,38(06):84-87+ 92.): the rice sheath blight disease was classified using a Support Vector Machine (SVM) combining color and texture.
The existing method realizes the identification of the rice sheath blight disease to different degrees, but the rice sheath blight disease is mainly characterized in that the leaf sheath and the leaves have different degrees of cloud streaks, and the cloud streaks are densely distributed in an actual scene. Therefore, although the conventional method based on the conventional image processing achieves certain results, the accuracy and the recognition efficiency have a large improvement space.
Disclosure of Invention
The invention mainly aims to provide a rice sheath blight identification method based on ShuffleNet V2-Unet, which can accurately identify complex rice sheath blight and accurately and quickly identify the rice sheath blight.
The invention further aims to provide a rice sheath blight identification system based on ShuffleNet V2-Unet.
In order to solve the technical problems, the technical scheme of the invention is as follows:
a rice sheath blight identification method based on ShuffleNet V2-Unet comprises the following steps:
s1: collecting a rice disease data set, wherein the rice disease data set comprises an image with rice sheath blight disease streaks;
s2: establishing a ShuffleNet V2-Unet model, wherein the ShuffleNet V2-Unet model is used for identifying the stringy and stringy striae of rice;
s3: training the ShuffleNet V2-Unet model by using the rice disease data set in the step S1 to obtain a trained ShuffleNet V2-Unet model;
s4: and (4) identifying the rice sheath blight disease of the input picture by using the trained ShuffleNet V2-Unet model.
Preferably, in step S2, the shefflenetv 2-Unet model includes a trunk extraction network and a feature enhancement network, where the trunk extraction network is a pruned shefflenetv 2 model, the trunk extraction network extracts features of different sizes of the input image and outputs the features to the feature enhancement network for feature enhancement, and the feature enhancement network outputs the recognition result.
Preferably, the pruned shefflenetv 2 model is a network obtained by removing the last convolutional layer, the global pooling layer and the full-link layer of the original shefflenetv 2 model.
Preferably, the ShuffleNet V2 model of pruning specifically comprises:
the method comprises the steps of obtaining a first extraction feature after the input image is subjected to convolutional layer stacking, obtaining a second extraction feature after the first extraction feature is subjected to maximum pooling after down sampling, obtaining a third extraction feature after the second extraction feature is subjected to down sampling and then passes through a third extraction module, obtaining a fourth extraction feature after the third extraction feature is subjected to down sampling and then passes through a fourth extraction module, and obtaining a fifth extraction feature after the fourth extraction feature is subjected to down sampling and then passes through a fifth extraction module.
Preferably, the third extraction module, the fourth extraction module and the fifth extraction module are all formed by stacking a random channel separation module and a down-sampling module.
Preferably, the random channel separation module specifically comprises:
the random channel separation module randomly divides an input channel into two parts, wherein one part is directly output without any convolution operation, and the other part continuously processes the input through three convolutions; and then, the channel outputs of the two parts are added in parallel, and finally, the obtained feature diagram is subjected to channel random mixing operation, wherein the input features and the output features of the random channel separation module have the same size.
Preferably, the down-sampling module specifically is:
the down-sampling module respectively conveys the input features to the two branches for down-sampling processing, and carries out channel splicing on the processing results of the two branches, the sizes of the input features and the output features of the down-sampling module are different, and the down-sampling module compresses the size of the input features.
Preferably, the feature enhancement network is composed of 4 upsampling layers, specifically:
sampling the fifth extraction feature output by the pruned ShuffleNet V2 model by a bilinear interpolation method to expand the size of the feature layer, connecting the upsampled fifth extraction feature with the fourth extraction feature output by the pruned ShuffleNet V2 model by using a concat function to obtain a first combined feature, and performing two times of 3 × 3 convolution operation and ReLU activation function operation on the first combined feature to obtain the output of a first layer upsampling layer; the output of the first layer of up-sampling is up-sampled by a bilinear interpolation method to expand the size of a characteristic layer, then a concat function is used for connecting the output of the up-sampled first layer of up-sampling layer with a third extraction characteristic output by the pruned ShuffleNet V2 model to obtain a second combination characteristic, and the second combination characteristic is subjected to two times of 3 multiplied by 3 convolution operation and ReLU activation function operation to obtain the output of the second layer of up-sampling layer; the output of the second layer upsampling is upsampled by a bilinear interpolation method to expand the size of the characteristic layer, then a concat function is used for connecting the output of the second layer upsampled layer after the upsampling with the second extraction characteristic output by the pruned ShuffleNet V2 model to obtain a third combined characteristic, and the third combined characteristic is subjected to two times of 3 x 3 convolution operation and ReLU activation function operation to obtain the output of the third layer upsampling layer; the output of the third layer of up-sampling is up-sampled by a bilinear interpolation method to expand the size of the feature layer, then a concat function is used for connecting the output of the up-sampled third layer of up-sampling layer with the first extraction feature output by the pruned ShuffleNet V2 model to obtain a fourth combination feature, and the fourth combination feature is subjected to two times of 3 × 3 convolution operation and ReLU activation function operation to obtain the output of the fourth layer of up-sampling layer; and obtaining an output characteristic diagram after the output of the sampling layer on the fourth layer is subjected to 1 × 1 convolution.
Preferably, the shufflenet v 2-uet model further includes a CBAM attention mechanism module, an input of the CBAM attention mechanism module is a first extracted feature of the pruned shufflenet v2 model output, an input of the CBAM attention mechanism module is connected to an output of the third upsampling layer, and the CBAM attention mechanism module specifically is:
the CBAM attention mechanism module comprises a channel attention module and a space attention module, wherein the input features are multiplied by the input features after passing through the channel attention module to obtain intermediate features, and the intermediate features are multiplied by the intermediate features after passing through the space attention module to obtain output features of the CBAM attention mechanism module.
A rice sheath blight identification system based on ShuffleNet V2-Unet comprises:
the system comprises a data acquisition module, a data processing module and a data processing module, wherein the data acquisition module acquires a rice disease data set, and the rice disease data set comprises an image with rice sheath blight and rice streak;
the model establishing module is used for establishing a ShuffleNet V2-Unet model, and the ShuffleNet V2-Unet model is used for identifying stringy blight and striae of rice;
a model training module, which trains the ShuffleNet V2-Unet model by using the rice disease data set in the step S1 to obtain a trained ShuffleNet V2-Unet model;
and the recognition module is used for recognizing the rice sheath blight disease of the input picture by using the trained ShuffleNet V2-Unet model.
Compared with the prior art, the technical scheme of the invention has the beneficial effects that:
according to the method, rice sheath blight is identified through the ShuffleNet V2-Unet model, rice sheath blight lines can be effectively identified in a complex environment, the detection speed can be kept while high detection precision is kept, meanwhile, the method is low in requirement on the detection environment for detecting the rice sheath blight, and operation for detecting the rice sheath blight is convenient.
Drawings
FIG. 1 is a schematic flow chart of the method of the present invention.
FIG. 2 is a schematic structural diagram of the ShuffLeNet V2-Unet model provided by the embodiment.
Fig. 3 is a schematic structural diagram of a CBAM attention mechanism module according to an embodiment.
Fig. 4 is a schematic operation flow chart provided by the embodiment.
Fig. 5 is a schematic diagram illustrating comparison between different models for identifying rice sheath blight disease according to an embodiment, where fig. 5(a) is an original diagram, fig. 5(b) is an Unet model identification result, fig. 5(c) is a PSPnet model identification result, fig. 5(d) is a deplab v3 model identification result with Xception as a backbone network, fig. 5(e) is a deplab v3 model identification result with Mobilenetv2 as a backbone network, and fig. 5(f) is a ShuffleNetV2-Unet model identification result.
FIG. 6 is a block diagram of the system of the present invention.
Detailed Description
The drawings are for illustrative purposes only and are not to be construed as limiting the patent;
for the purpose of better illustrating the embodiments, certain features of the drawings may be omitted, enlarged or reduced, and do not represent the size of an actual product;
it will be understood by those skilled in the art that certain well-known structures in the drawings and descriptions thereof may be omitted.
The technical solution of the present invention is further described below with reference to the accompanying drawings and examples.
Example 1
The embodiment provides a rice sheath blight identification method based on ShuffleNet V2-Unet, as shown in FIG. 1, comprising the following steps:
s1: collecting a rice disease data set, wherein the rice disease data set comprises an image with rice sheath blight disease streaks;
s2: establishing a ShuffleNet V2-Unet model, wherein the ShuffleNet V2-Unet model is used for identifying the stringy and stringy striae of rice;
s3: training the ShuffleNet V2-Unet model by using the rice disease data set in the step S1 to obtain a trained ShuffleNet V2-Unet model;
s4: and (4) identifying the rice sheath blight disease of the input picture by using the trained ShuffleNet V2-Unet model.
The shuffle netv2-Unet model in step S2 is shown in fig. 2, and includes a trunk extraction network and a feature enhancement network, where the trunk extraction network is a pruned shuffle netv2 model, the trunk extraction network extracts features of different sizes of an input image and outputs the features to the feature enhancement network for feature enhancement, and the feature enhancement network outputs a recognition result.
The pruned ShuffleNet V2 model is a network obtained by removing the last convolutional layer, the global pooling layer and the full-link layer of the original ShuffleNet V2 model.
The ShuffLeNet V2 model of pruning specifically comprises the following steps:
the method comprises the steps of obtaining a first extraction feature after the input image is subjected to convolutional layer stacking, obtaining a second extraction feature after the first extraction feature is subjected to maximum pooling after down sampling, obtaining a third extraction feature after the second extraction feature is subjected to down sampling and then passes through a third extraction module, obtaining a fourth extraction feature after the third extraction feature is subjected to down sampling and then passes through a fourth extraction module, and obtaining a fifth extraction feature after the fourth extraction feature is subjected to down sampling and then passes through a fifth extraction module.
The third extraction module, the fourth extraction module and the fifth extraction module are all formed by stacking a random channel separation module and a down-sampling module.
The random channel separation module specifically comprises:
the random channel separation module randomly divides an input channel into two parts, wherein one part is directly output without any convolution operation, and the other part continuously processes the input through three convolutions; and then, the channel outputs of the two parts are added in parallel, and finally, the obtained feature diagram is subjected to channel random mixing operation, wherein the input features and the output features of the random channel separation module have the same size.
The down-sampling module is specifically as follows:
the down-sampling module respectively conveys the input features to the two branches for down-sampling processing, and carries out channel splicing on the processing results of the two branches, the sizes of the input features and the output features of the down-sampling module are different, and the down-sampling module compresses the size of the input features.
The characteristic enhancement network is composed of 4 upper sampling layers, and specifically comprises the following steps:
sampling the fifth extraction feature output by the pruned ShuffleNet V2 model by a bilinear interpolation method to expand the size of the feature layer, connecting the upsampled fifth extraction feature with the fourth extraction feature output by the pruned ShuffleNet V2 model by using a concat function to obtain a first combined feature, and performing two times of 3 × 3 convolution operation and ReLU activation function operation on the first combined feature to obtain the output of a first layer upsampling layer; the output of the first layer of up-sampling is up-sampled by a bilinear interpolation method to expand the size of a characteristic layer, then a concat function is used for connecting the output of the up-sampled first layer of up-sampling layer with a third extraction characteristic output by the pruned ShuffleNet V2 model to obtain a second combination characteristic, and the second combination characteristic is subjected to two times of 3 multiplied by 3 convolution operation and ReLU activation function operation to obtain the output of the second layer of up-sampling layer; the output of the second layer upsampling is upsampled by a bilinear interpolation method to expand the size of the characteristic layer, then a concat function is used for connecting the output of the second layer upsampled layer after the upsampling with the second extraction characteristic output by the pruned ShuffleNet V2 model to obtain a third combined characteristic, and the third combined characteristic is subjected to two times of 3 x 3 convolution operation and ReLU activation function operation to obtain the output of the third layer upsampling layer; the output of the third layer of up-sampling is up-sampled by a bilinear interpolation method to expand the size of the feature layer, then a concat function is used for connecting the output of the up-sampled third layer of up-sampling layer with the first extraction feature output by the pruned ShuffleNet V2 model to obtain a fourth combination feature, and the fourth combination feature is subjected to two times of 3 × 3 convolution operation and ReLU activation function operation to obtain the output of the fourth layer of up-sampling layer; and obtaining an output characteristic diagram after the output of the sampling layer on the fourth layer is subjected to 1 × 1 convolution.
As shown in fig. 4, the original feature extraction network of Unet is replaced by the lightweight network ShuffleNetV2, the global pooling layer and the full connection layer on the structure of ShuffleNetV2 are pruned, the outputs of the first four stages of the extraction model are connected with the up-sampling structure of the original Unet model, the deconvolution of the up-sampling of Unet is replaced by the up-sampling structure of the bilinear interpolation method, and finally the first layer feature layer of the model is connected with the up-sampling layer by the attention mechanism module, so that the model construction is completed. Using rice sheath blight dataTraining the model to obtain the model prediction weight, and using the rice sheath blight model prediction weight by the ShuffleNet V2-Unet model, the rice sheath blight can be recognized. Table 1 is a parameter table of ShuffleNet V2 after pruning, wherein Output indicates Output size, Stride indicates step length, Repeat indicates module repetition times, and Oi indicates Output channel number. Using output O1.5The number of output channels as the ShuffLeNet V2-Unet model.
TABLE 1
Figure BDA0003418276790000081
Example 2
In this embodiment, on the basis of embodiment 1, the ShuffleNetV 2-uet model further includes a CBAM attention mechanism module, as shown in fig. 3, an input of the CBAM attention mechanism module is a first extracted feature output by the pruned ShuffleNetV2 model, an input of the CBAM attention mechanism module is connected to an output of the sampling layer on the third layer, and the CBAM attention mechanism module specifically includes:
the CBAM attention mechanism module comprises a channel attention module and a space attention module, wherein the input features are multiplied by the input features after passing through the channel attention module to obtain intermediate features, and the intermediate features are multiplied by the intermediate features after passing through the space attention module to obtain output features of the CBAM attention mechanism module.
The performance of the semantic segmentation model selected in the experiment is tested by using a public field provided by the Chinese academy of sciences as a disease data set image, 850 pictures of rice sheath blight are selected as a training set, 150 pictures are selected as a verification set, and 150 pictures are selected as a test set.
As shown in fig. 5, a comparison of the recognition results of different models is shown, and table 2 shows the performance evaluation of different models.
TABLE 2
Figure BDA0003418276790000091
Example 3
A rice sheath blight identification system based on ShuffleNet V2-Unet, as shown in FIG. 5, comprises:
the system comprises a data acquisition module, a data processing module and a data processing module, wherein the data acquisition module acquires a rice disease data set, and the rice disease data set comprises an image with rice sheath blight and rice streak;
the model establishing module is used for establishing a ShuffleNet V2-Unet model, and the ShuffleNet V2-Unet model is used for identifying stringy blight and striae of rice;
a model training module, which trains the ShuffleNet V2-Unet model by using the rice disease data set in the step S1 to obtain a trained ShuffleNet V2-Unet model;
and the recognition module is used for recognizing the rice sheath blight disease of the input picture by using the trained ShuffleNet V2-Unet model.
The same or similar reference numerals correspond to the same or similar parts;
the terms describing positional relationships in the drawings are for illustrative purposes only and are not to be construed as limiting the patent;
it should be understood that the above-described embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. Other variations and modifications will be apparent to persons skilled in the art in light of the above description. And are neither required nor exhaustive of all embodiments. Any modification, equivalent replacement, and improvement made within the spirit and principle of the present invention should be included in the protection scope of the claims of the present invention.

Claims (10)

1. A rice sheath blight identification method based on ShuffleNet V2-Unet is characterized by comprising the following steps:
s1: collecting a rice disease data set, wherein the rice disease data set comprises an image with rice sheath blight disease streaks;
s2: establishing a ShuffleNet V2-Unet model, wherein the ShuffleNet V2-Unet model is used for identifying the stringy and stringy striae of rice;
s3: training the ShuffleNet V2-Unet model by using the rice disease data set in the step S1 to obtain a trained ShuffleNet V2-Unet model;
s4: and (4) identifying the rice sheath blight disease of the input picture by using the trained ShuffleNet V2-Unet model.
2. The method for identifying sheath blight of rice based on ShuffleNet V2-Unet of claim 1, wherein the ShuffleNet V2-Unet model in step S2 comprises a trunk extraction network and a feature enhancement network, wherein the trunk extraction network is a pruning ShuffleNet V2 model, the trunk extraction network extracts features of different sizes of input images and outputs the features to the feature enhancement network for feature enhancement, and the feature enhancement network outputs the identification result.
3. The method for identifying the rice sheath blight disease based on ShuffleNet V2-Unet of claim 2, wherein the ShuffleNet V2 model is a network obtained by removing the last convolutional layer, the global pooling layer and the full connection layer of the original ShuffleNet V2 model.
4. The method for identifying rice sheath blight disease based on ShuffleNet V2-Unet of claim 3, wherein the ShffleNet V2 model of pruning is specifically as follows:
the method comprises the steps that a first extraction feature layer is obtained after an input image is subjected to convolutional layer stacking, a second extraction feature layer is obtained after the first extraction feature layer is subjected to maximum pooling after down-sampling, a third extraction feature is obtained after the second extraction feature layer is subjected to third extraction module after down-sampling, a fourth extraction feature is obtained after the third extraction feature layer is subjected to fourth extraction module after down-sampling, and a fifth extraction feature is obtained after the fourth extraction feature layer is subjected to fifth extraction module after down-sampling.
5. The method for identifying sheath blight of rice based on ShuffleNet V2-Unet of claim 4, wherein the third extraction module, the fourth extraction module and the fifth extraction module are all stacked by a random channel separation module and a down-sampling module.
6. The rice sheath blight identification method based on ShuffleNet V2-Unet of claim 5, wherein the random channel separation module specifically comprises:
the random channel separation module randomly divides an input channel into two parts, wherein one part is directly output without any convolution operation, and the other part continuously processes the input through three convolutions; and then, the channel outputs of the two parts are added in parallel, and finally, the obtained feature diagram is subjected to channel random mixing operation, wherein the input features and the output features of the random channel separation module have the same size.
7. The rice sheath blight disease identification method based on ShuffleNet V2-Unet of claim 6, wherein the down-sampling module is specifically:
the down-sampling module respectively conveys the input features to the two branches for down-sampling processing, and carries out channel splicing on the processing results of the two branches, the sizes of the input features and the output features of the down-sampling module are different, and the down-sampling module compresses the size of the input features.
8. The rice sheath blight disease identification method based on ShuffleNet V2-Unet of claim 7, wherein the feature enhancement network consists of 4 upsampling layers, and specifically comprises:
sampling the fifth extraction feature output by the pruned ShuffleNet V2 model by a bilinear interpolation method to expand the size of the feature layer, connecting the upsampled fifth extraction feature with the fourth extraction feature output by the pruned ShuffleNet V2 model by using a concat function to obtain a first combined feature, and performing two times of 3 × 3 convolution operation and ReLU activation function operation on the first combined feature to obtain the output of a first layer upsampling layer; the output of the first layer of up-sampling is up-sampled by a bilinear interpolation method to expand the size of a characteristic layer, then a concat function is used for connecting the output of the up-sampled first layer of up-sampling layer with a third extraction characteristic output by the pruned ShuffleNet V2 model to obtain a second combination characteristic, and the second combination characteristic is subjected to two times of 3 multiplied by 3 convolution operation and ReLU activation function operation to obtain the output of the second layer of up-sampling layer; the output of the second layer upsampling is upsampled by a bilinear interpolation method to expand the size of the characteristic layer, then a concat function is used for connecting the output of the second layer upsampled layer after the upsampling with the second extraction characteristic output by the pruned ShuffleNet V2 model to obtain a third combined characteristic, and the third combined characteristic is subjected to two times of 3 x 3 convolution operation and ReLU activation function operation to obtain the output of the third layer upsampling layer; the output of the third layer of up-sampling is up-sampled by a bilinear interpolation method to expand the size of the feature layer, then a concat function is used for connecting the output of the up-sampled third layer of up-sampling layer with the first extraction feature output by the pruned ShuffleNet V2 model to obtain a fourth combination feature, and the fourth combination feature is subjected to two times of 3 × 3 convolution operation and ReLU activation function operation to obtain the output of the fourth layer of up-sampling layer; and obtaining an output characteristic diagram after the output of the sampling layer on the fourth layer is subjected to 1 × 1 convolution.
9. The method for identifying sheath blight of rice based on ShuffleNet V2-Unet of claim 8, wherein the ShffleNet V2-Unet model further comprises a CBAM attention mechanism module, an input of the CBAM attention mechanism module is the first extracted feature output by the pruned ShffleNet V2 model, an input of the CBAM attention mechanism module is connected with an output of the sampling layer on the third layer, and the CBAM attention mechanism module is specifically:
the CBAM attention mechanism module comprises a channel attention module and a space attention module, wherein the input features are multiplied by the input features after passing through the channel attention module to obtain intermediate features, and the intermediate features are multiplied by the intermediate features after passing through the space attention module to obtain output features of the CBAM attention mechanism module.
10. A rice sheath blight identification system based on ShuffleNet V2-Unet, comprising:
the system comprises a data acquisition module, a data processing module and a data processing module, wherein the data acquisition module acquires a rice disease data set, and the rice disease data set comprises an image with rice sheath blight and rice streak;
the model establishing module is used for establishing a ShuffleNet V2-Unet model, and the ShuffleNet V2-Unet model is used for identifying stringy blight and striae of rice;
a model training module, which trains the ShuffleNet V2-Unet model by using the rice disease data set in the step S1 to obtain a trained ShuffleNet V2-Unet model;
and the recognition module is used for recognizing the rice sheath blight disease of the input picture by using the trained ShuffleNet V2-Unet model.
CN202111554870.XA 2021-12-17 2021-12-17 Rice sheath blight identification method and system based on ShuffleNet V2-Unet Pending CN114241309A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111554870.XA CN114241309A (en) 2021-12-17 2021-12-17 Rice sheath blight identification method and system based on ShuffleNet V2-Unet

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111554870.XA CN114241309A (en) 2021-12-17 2021-12-17 Rice sheath blight identification method and system based on ShuffleNet V2-Unet

Publications (1)

Publication Number Publication Date
CN114241309A true CN114241309A (en) 2022-03-25

Family

ID=80758432

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111554870.XA Pending CN114241309A (en) 2021-12-17 2021-12-17 Rice sheath blight identification method and system based on ShuffleNet V2-Unet

Country Status (1)

Country Link
CN (1) CN114241309A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109993228A (en) * 2019-04-02 2019-07-09 南通科技职业学院 Plant protection drone rice sheath blight disease recognition methods based on machine vision
CN115359411A (en) * 2022-10-21 2022-11-18 成都工业学院 Transformer substation environment understanding method based on improved deep Lab V3+ network

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109993228A (en) * 2019-04-02 2019-07-09 南通科技职业学院 Plant protection drone rice sheath blight disease recognition methods based on machine vision
CN115359411A (en) * 2022-10-21 2022-11-18 成都工业学院 Transformer substation environment understanding method based on improved deep Lab V3+ network

Similar Documents

Publication Publication Date Title
CN107247971B (en) Intelligent analysis method and system for ultrasonic thyroid nodule risk index
CN106909784A (en) Epileptic electroencephalogram (eeg) recognition methods based on two-dimentional time-frequency image depth convolutional neural networks
CN112241762B (en) Fine-grained identification method for pest and disease damage image classification
CN110348357B (en) Rapid target detection method based on deep convolutional neural network
CN114241309A (en) Rice sheath blight identification method and system based on ShuffleNet V2-Unet
CN106934418B (en) Insulator infrared diagnosis method based on convolution recursive network
CN114038037B (en) Expression label correction and identification method based on separable residual error attention network
CN109753996B (en) Hyperspectral image classification method based on three-dimensional lightweight depth network
CN114332572B (en) Method for extracting breast lesion ultrasonic image multi-scale fusion characteristic parameters based on saliency map-guided hierarchical dense characteristic fusion network
CN113112498B (en) Grape leaf spot identification method based on fine-grained countermeasure generation network
CN111597920A (en) Full convolution single-stage human body example segmentation method in natural scene
CN109472287A (en) Three-dimensional fluorescence spectrum feature extracting method based on Two-Dimensional Gabor Wavelets
CN115294075A (en) OCTA image retinal vessel segmentation method based on attention mechanism
CN116129405A (en) Method for identifying anger emotion of driver based on multi-mode hybrid fusion
Suo et al. Casm-amfmnet: A network based on coordinate attention shuffle mechanism and asymmetric multi-scale fusion module for classification of grape leaf diseases
Song et al. Red blood cell classification based on attention residual feature pyramid network
WO2024104035A1 (en) Long short-term memory self-attention model-based three-dimensional medical image segmentation method and system
CN112966698A (en) Freshwater fish image real-time identification method based on lightweight convolutional network
CN112132137A (en) FCN-SPP-Focal Net-based method for identifying correct direction of abstract picture image
CN113192076B (en) MRI brain tumor image segmentation method combining classification prediction and multi-scale feature extraction
CN112633400B (en) Shellfish classification and identification method and device based on computer vision
CN115471675A (en) Disguised object detection method based on frequency domain enhancement
CN115471724A (en) Fine-grained fish epidemic disease identification fusion algorithm based on self-adaptive normalization
CN115063700A (en) Detection method based on small-scale pine wood nematode disease tree
CN115170987A (en) Method for detecting diseases of grapes based on image segmentation and registration fusion

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination