CN116883735A - Domain self-adaptive wheat seed classification method based on public features and private features - Google Patents
Domain self-adaptive wheat seed classification method based on public features and private features Download PDFInfo
- Publication number
- CN116883735A CN116883735A CN202310825251.2A CN202310825251A CN116883735A CN 116883735 A CN116883735 A CN 116883735A CN 202310825251 A CN202310825251 A CN 202310825251A CN 116883735 A CN116883735 A CN 116883735A
- Authority
- CN
- China
- Prior art keywords
- domain
- image
- wheat seed
- class
- sample
- 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.)
- Granted
Links
- 241000209140 Triticum Species 0.000 title claims abstract description 110
- 235000021307 Triticum Nutrition 0.000 title claims abstract description 110
- 238000000034 method Methods 0.000 title claims abstract description 33
- 238000013145 classification model Methods 0.000 claims abstract description 36
- 238000012549 training Methods 0.000 claims abstract description 19
- 238000000605 extraction Methods 0.000 claims description 85
- 238000012937 correction Methods 0.000 claims description 22
- 230000003595 spectral effect Effects 0.000 claims description 22
- 238000002310 reflectometry Methods 0.000 claims description 7
- 238000005520 cutting process Methods 0.000 claims description 3
- 230000003044 adaptive effect Effects 0.000 claims 8
- 238000013508 migration Methods 0.000 abstract description 8
- 230000005012 migration Effects 0.000 abstract description 8
- 230000006870 function Effects 0.000 description 23
- 230000004913 activation Effects 0.000 description 7
- 238000009826 distribution Methods 0.000 description 6
- 238000001514 detection method Methods 0.000 description 5
- 238000005516 engineering process Methods 0.000 description 5
- 238000010606 normalization Methods 0.000 description 5
- 238000011176 pooling Methods 0.000 description 5
- 230000000694 effects Effects 0.000 description 4
- 238000004364 calculation method Methods 0.000 description 3
- 230000008859 change Effects 0.000 description 3
- 239000000284 extract Substances 0.000 description 3
- ISWSIDIOOBJBQZ-UHFFFAOYSA-N Phenol Chemical compound OC1=CC=CC=C1 ISWSIDIOOBJBQZ-UHFFFAOYSA-N 0.000 description 2
- 238000012271 agricultural production Methods 0.000 description 2
- 238000000701 chemical imaging Methods 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 238000010186 staining Methods 0.000 description 2
- NNJPGOLRFBJNIW-HNNXBMFYSA-N (-)-demecolcine Chemical compound C1=C(OC)C(=O)C=C2[C@@H](NC)CCC3=CC(OC)=C(OC)C(OC)=C3C2=C1 NNJPGOLRFBJNIW-HNNXBMFYSA-N 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000012742 biochemical analysis Methods 0.000 description 1
- 238000009395 breeding Methods 0.000 description 1
- 230000001488 breeding effect Effects 0.000 description 1
- 235000013339 cereals Nutrition 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 230000001066 destructive effect Effects 0.000 description 1
- 238000010191 image analysis Methods 0.000 description 1
- 238000002329 infrared spectrum Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000000877 morphologic effect Effects 0.000 description 1
- 238000011206 morphological examination Methods 0.000 description 1
- 238000009659 non-destructive testing Methods 0.000 description 1
- 239000002245 particle Substances 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000001228 spectrum Methods 0.000 description 1
- 238000005728 strengthening Methods 0.000 description 1
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A40/00—Adaptation technologies in agriculture, forestry, livestock or agroalimentary production
- Y02A40/10—Adaptation technologies in agriculture, forestry, livestock or agroalimentary production in agriculture
Abstract
The application discloses a domain self-adaptive wheat seed classification method based on public features and private features, which relates to the technical field of hyperspectrum, wherein a wheat seed classification model built by the method can learn the public features, class-sensitive private features and domain-sensitive private features of a sample image in a source domain and a target domain, and is guided by different constraints during training, so that the model can carry out targeted migration, only the public features and the class-sensitive private features which are helpful for classification learning are migrated, and the wheat seed classification model can be quickly adapted to a new environment of the target domain by using a small number of samples through a training strategy based on the public features and the private features, thereby ensuring that the classification accuracy of the model in the target domain is higher, the generalization capability is stronger, avoiding negative migration of the model and remarkably improving the classification accuracy of the model on the target domain after the migration.
Description
Technical Field
The application relates to the technical field of hyperspectrum, in particular to a domain self-adaptive wheat seed classification method based on public characteristics and private characteristics.
Background
Wheat is one of main grain crops in China, and is planted throughout China, seeds of the wheat have a large influence on yield, and the wheat is also important for strengthening and improving agricultural economy. With the rapid development of breeding research, various high-quality seeds are introduced into the market, and the seed variety detection problem is complicated while the agricultural production ecology is enriched.
The traditional seed variety detection methods include morphology, phenol staining, DNA molecular technology and the like. Morphological examination requires that the observer identify the seed based on its appearance. This process is highly dependent on the experience of the observer, is prone to error and requires a lot of manpower. Biochemical analysis methods such as phenol staining and DNA molecular techniques have high accuracy, but these methods are time consuming and can cause seed damage, failing to meet the rapid, non-destructive demands of modern agricultural production.
With the development of hyperspectral imaging technology and computer technology, methods utilizing visible light near infrared, near infrared and short wave infrared spectrum image analysis technology have been widely applied in the field of nondestructive testing of agricultural products. The hyperspectral image simultaneously comprises image information (external morphological characteristics) and spectrum information (internal characteristics) of a sample to be detected, and then the model for variety identification of wheat seeds can be obtained through training by combining the internal data rule of deep learning hyperspectral image, so that nondestructive detection based on hyperspectral imaging technology is realized. However, in practical application, due to differences among wheat seeds in different years and interference of external environments, there is a domain deviation in hyperspectral images of different wheat seeds, so that the trained model often has a limited application range, for example, the model obtained by training the hyperspectral images of wheat seeds harvested in 2019 has poor classification effect on the wheat seeds harvested in 2020, for example, the model obtained by training the hyperspectral images of wheat seeds harvested in instrument A has poor classification effect on the hyperspectral images of wheat seeds harvested in instrument B. On one hand, the accuracy of nondestructive detection by using the model obtained by training is low, the detection accuracy is difficult to guarantee, and on the other hand, the generalization capability of the model obtained by training is not high.
Disclosure of Invention
Aiming at the problems and the technical requirements, the inventor provides a domain self-adaptive wheat seed classification method based on public characteristics and private characteristics, and the technical scheme of the application is as follows:
a domain-adaptive wheat seed classification method based on public and private features, the method comprising:
the method for constructing the network model of the wheat seed classification model comprises the following steps: for extracting common features F n Public feature extraction module for extracting sensitive-like private feature F c Class sensitive private feature extraction module for extracting domain sensitive private feature F d Domain sensitive private feature extraction module, class domain classifier, class classifier and domain classifier, public feature F n And class sensitive private feature F c Inputting the spliced input classes into a classifier with common characteristics F n Sum domain sensitive private feature F d Input domain classifier after splicing, public feature F n Class sensitive private feature F c Sum domain sensitive private feature F d Inputting the spliced class domain classifier;
acquiring a source domain sample image of a wheat seed sample in a source domain and a target domain sample image of the wheat seed sample in a target domain, wherein each sample image is a hyperspectral image with a domain label and a class label, the domain label of the sample image is used for indicating the domain to which the sample image belongs, the class label is used for indicating the variety of the wheat seed sample in the sample image, and the target domain is the same as the variety of the wheat seed sample in the source domain;
training a wheat seed classification model by using the source domain sample image and the target domain sample image;
obtaining a hyperspectral image to be identified of the wheat seeds to be identified in the target domain, inputting the hyperspectral image to be identified into a wheat seed classification model obtained through training, and obtaining a variety identification result of the wheat seeds to be identified, which is output by a class classifier of the wheat seed classification model;
wherein the common feature F of each hyperspectral image n Is the common characteristic and similar sensitive private characteristic F of hyperspectral images in different domains c Is the unique characteristic of hyperspectral image on the belonging domain and is used for distinguishing different varieties and domain sensitive private characteristic F d Is a feature of the hyperspectral image being different over different domains.
The network model of the wheat seed classification model further comprises an initial feature extraction module, wherein the initial feature extraction module is used for extracting initial low-dimensional features F from hyperspectral images input into the wheat seed classification model;
the public feature extraction module performs feature extraction on the initial low-dimensional feature F to obtain public feature F n The class-sensitive private feature extraction module performs feature extraction on the initial low-dimensional feature F to obtain a class-sensitive private feature F c The domain sensitive private feature extraction module performs feature extraction on the initial low-dimensional feature F to obtain domain sensitive private feature F d 。
The further technical scheme is that the training of the wheat seed classification model by using the source domain sample image and the target domain sample image comprises the following steps:
randomly matching the target domain sample image and the source domain sample image to form a sample image pair, respectively inputting the sample images into a wheat seed classification model, and adjusting network parameters of an initial feature extraction module and a public feature extraction module until the public features of the source domain sample image and the public features of the target domain sample image are aligned;
inputting any one sample image in the source domain sample image and the target domain sample image into a wheat seed classification model, and adjusting network parameters of a class-sensitive private feature extraction module and a domain-sensitive private feature extraction module based on class labels and domain labels of the input sample image and by combining an output result of a class classifier, an output result of a domain classifier and an output result of the class domain classifier.
The method further comprises the steps of adjusting network parameters of an initial feature extraction module and a public feature extraction module, and comprising the following steps:
common features F based on source domain sample images in the same sample image pair n And common features F of the sample image of the target domain n Calculating loss function loss dist According to loss function loss dist And adjusting network parameters of the initial feature extraction module and the public feature extraction module.
The further technical proposal is that the loss function loss is calculated dist Comprising the following steps:
common features F of source-domain sample images in the same sample image pair n And common features F of the sample image of the target domain n Euclidean distance between as a loss function loss dist 。
The method comprises the further technical scheme that the adjusting of network parameters of the class-sensitive private feature extraction module and the domain-sensitive private feature extraction module comprises the steps of for any input sample image:
class label calculation loss function loss based on output result of class classifier and sample image class Calculating a loss function loss based on an output result of the domain classifier and a domain label of the sample image domain Class labels based on output result of class domain classifier and sample image and domain labels to calculate loss function loss class-domain ；
According to loss function loss class 、loss domain、 loss class-domain And adjusting network parameters of the class-sensitive private feature extraction module and the domain-sensitive private feature extraction module.
The further technical scheme is that, for any one sample image of the wheat seed sample in the source domain sample image and the target domain sample image, the obtaining the sample image of the wheat seed sample comprises:
acquiring an original hyperspectral image of a wheat seed sample;
performing black-and-white plate correction on the original hyperspectral image;
and extracting the region of interest from the original hyperspectral image after the black-and-white plate correction is completed to obtain a sample image of the wheat seed sample.
The further technical scheme is that extracting the interested region from the original hyperspectral image after the black-white plate correction to obtain the sample image of the wheat seed sample comprises the following steps:
spectral reflectivities at three preset wavelengths in an original hyperspectral image after black-and-white plate correction are extracted and used as RGB channel components to form a pseudo RGB image;
by passing throughConverting the pseudo RGB image into a YCbCr color space, and selecting a Cb channel component as a gray scale image;
and extracting an interested region in the original hyperspectral image corrected by the black-and-white plate according to the gray value, filling a zero value to a preset image size, and obtaining a sample image of the wheat seed sample.
According to the further technical scheme, the region of interest in the original hyperspectral image after black-and-white plate correction is extracted according to the gray value comprises the following steps:
dividing the gray image into a foreground area and a background area by using an Ojin algorithm, and performing binarization treatment to obtain a binary image, wherein the foreground area is an area where a wheat seed sample is located;
performing open operation and close operation on the binary image by using a circular structure variable with the diameter of 15 to obtain a binary mask of a wheat seed sample;
and carrying out 8-connection on the binary mask to find the connection region of the binary mask, returning the mask of each connection region, and obtaining the region of interest in the original hyperspectral image after the correction of the black-and-white plate according to the original hyperspectral image after the correction of the black-and-white plate is completed by cutting the mask.
The further technical scheme is that the spectral reflectances at three preset wavelengths in the original hyperspectral image after the black-and-white plate correction are extracted and respectively used as RGB channel components, and the method comprises the following steps:
spectral reflectance at 460nm was extracted as a B-channel component, spectral reflectance at 525nm was extracted as a G-channel component, and spectral reflectance at 621nm was extracted as an R-channel component.
The beneficial technical effects of the application are as follows:
the application discloses a domain self-adaptive wheat seed classification method based on public features and private features, which is characterized in that the method respectively extracts public features, class-sensitive private features and domain-sensitive private features of hyperspectral images, and guides the hyperspectral images by giving different constraints during training, so that the models can be purposefully migrated, only the public features and the class-sensitive private features which are helpful for classification learning are migrated, and the wheat seed classification model can be quickly adapted to new environments by using a small number of samples through a training strategy based on the public features and the private features, thereby ensuring that the classification accuracy of the model is higher, the generalization capability is stronger, and the model can be prevented from negative migration due to no migration of the domain-sensitive private features, and the classification accuracy of the model on a target domain after migration is remarkably improved.
Drawings
FIG. 1 is a network model diagram of a wheat seed classification model in one embodiment of the application.
FIG. 2 is a schematic representation of the calculation of four classes of loss functions during training of a wheat seed classification model in one embodiment of the application.
Detailed Description
The following describes the embodiments of the present application further with reference to the drawings.
The application discloses a domain self-adaptive wheat seed classification method based on public characteristics and private characteristics, which comprises the following steps:
and 1, constructing a network model of the wheat seed classification model.
Referring to a network structure diagram shown in fig. 1, the wheat seed classification model constructed by the application comprises a public feature extraction module, a class sensitive private feature extraction module, a domain sensitive private feature extraction module, a class domain classifier, a class classifier and a domain classifier.
The public feature extraction module is used for extracting public features F of the hyperspectral image input into the wheat seed classification model n Common feature F of each hyperspectral image n The characteristic shared by the hyperspectral images in different domains is that the distribution drift is not generated due to the change of the domain to which the hyperspectral image belongs.
The class-sensitive private feature extraction module and the domain-sensitive private feature extraction module are both used for extracting private features of hyperspectral images input into the wheat seed classification model, and correspond to public features F n The private feature of each hyperspectral image is a unique feature of the hyperspectral image in the belonging domain, and the private feature shifts in distribution along with the change of the belonging domain.
The application further subdivides the private feature into sensitive-like private feature F c Sum domain sensitive private feature F d : class-sensitive private feature F of each hyperspectral image c Is unique in the domain to which the hyperspectral image belongs and is used for distinguishing the characteristics of different varieties, and is similar to sensitive private characteristics F c A drift in distribution occurs with a change in domain, but at the same time contains knowledge of the domain to which it belongs that aids in classification of the variety. Domain sensitive private feature F of each hyperspectral image d Is the characteristic of the difference of hyperspectral images in different domains, namely domain sensitive private characteristic F d Features are included that reflect differences between different domains, but do not assist in classification learning over the target domain. It follows that even though it is also a private feature, the class sensitive private feature F c Sum domain sensitive private feature F d Also has great difference, and is similar to sensitive private characteristic F c Helpful to classification learning on target domain, domain sensitive private feature F d The classification learning on the target domain is not helpful, but knowledge with negative effects generated by the classification task of the target domain is learned on the source domain to generate negative migration, and the two types of private features are further refined and respectively extracted and processed in a targeted manner. Correspondingly, the class-sensitive private feature extraction module is used for extracting hyperspectral images input into the wheat seed classification modelClass sensitive private feature F c The domain sensitive private feature extraction module is used for extracting domain sensitive private features F of hyperspectral images input into the wheat seed classification model d 。
Extracted public feature F n And class sensitive private feature F c Inputting the spliced class classifier, and outputting the result R of the class classifier c And the variety classification result is used for indicating the wheat seeds in the input hyperspectral image. Extracted public feature F n Sum domain sensitive private feature F d After splicing, inputting the input domain classifier, and outputting the result R of the domain classifier d And the device is used for indicating the domain of the input hyperspectral image. Extracted public feature F n Class sensitive private feature F c Sum domain sensitive private feature F d After being spliced, the input class domain classifier is used for outputting the result R of the class domain classifier c-d And the device is used for indicating the domain to which the input hyperspectral image belongs and the variety classification result.
In another embodiment, the network model of the wheat seed classification model further comprises an initial feature extraction module for extracting an initial low-dimensional feature F from the hyperspectral image input to the wheat seed classification model and inputting the initial low-dimensional feature F to the public feature extraction module, the domain-sensitive private feature extraction module and the domain-sensitive private feature extraction module, respectively, considering that part of feature extraction operations of the three modules, i.e., the public feature extraction module, the class-sensitive private feature extraction module and the domain-sensitive private feature extraction module, are similar. The public feature extraction module further performs feature extraction on the initial low-dimensional feature F to obtain public feature F n The class-sensitive private feature extraction module further performs feature extraction on the initial low-dimensional feature F to obtain a class-sensitive private feature F c The domain sensitive private feature extraction module further performs feature extraction on the initial low-dimensional feature F to obtain domain sensitive private feature F d . The initial feature extraction module is used for extracting the initial low-dimensional feature F in a unified mode, and then the initial low-dimensional feature F is input into the three modules to further extract corresponding features respectively, so that the three modules can share the initial feature extraction module, and the complexity of the model is reduced.
In practical application, the initial feature extraction module, the public feature extraction module, the sensitive private feature extraction module and the domain sensitive private feature extraction module are respectively realized by a convolutional neural network. In one embodiment, (1) the initial feature extraction module comprises a first convolution layer and a second convolution layer, the first convolution layer comprising, in order, 128 convolution kernel normalization layers of 3*3, a ReLu activation function, and a 2 x 2 max pooling layer. The second convolution layer comprises, in order, 32 convolution kernel normalization layers of 3*3, a ReLu activation function, and a 2 x 2 max pooling layer. (2) The common feature extraction module comprises 16 convolution kernel normalization layers of 3*3, a ReLu activation function and a 2 x 2 maximum pooling layer in sequence. (3) The sensitive private feature extraction module sequentially comprises 8 convolution kernel normalization layers 3*3, a ReLu activation function and a 2 x 2 maximum pooling layer. (4) The domain sensitive private feature extraction module sequentially comprises 8 convolution kernel normalization layers of 3*3, a ReLu activation function and a 2 x 2 maximum pooling layer. (5) The class classifier, the domain classifier and the class domain classifier are all composed of full connection layers and have the same structure, and each classifier sequentially comprises a full connection layer with 120 nodes, a ReLu activation function, a full connection layer with 60 nodes, a ReLu activation function and a softmax classifier.
And 2, acquiring a source domain sample image of the wheat seed sample in the source domain and a target domain sample image of the wheat seed sample in the target domain, wherein the target domain is the same as the variety of the wheat seed sample in the source domain.
Each sample image is a hyperspectral image with a domain label and a class label, the domain label of the sample image is used for indicating the domain to which the sample image belongs, and the class label is used for indicating the variety of the wheat seed sample in the sample image. In the application, the spectral data distribution of the source domain sample image of the wheat seed sample of the same variety in the source domain and the target domain sample image in the target domain has drift. In practical application, at least one of the source domain and the target domain of the wheat seed sample of the same variety is different from each other in the year, the place of production and the used hyperspectral image acquisition equipment.
In one embodiment, for a sample image of a wheat seed sample in any one of a source domain sample image and a target domain sample image, a method of obtaining a sample image of a wheat seed sample includes:
(1) And acquiring an original hyperspectral image of the wheat seed sample by using hyperspectral image acquisition equipment. In one embodiment, a wheat seed sample is placed in a fluted black plastic particle counter plate, and a linear push-broom hyperspectral image acquisition device is used to acquire an original hyperspectral image, wherein the acquired original hyperspectral image comprises an image of a region of interest and an image of a background region. In one embodiment, the raw hyperspectral image acquired covers 94 bands (spectral resolution 6.4 nm) in the spectral range 400-1000 nm.
(2) Black and white plate correction is performed on the original hyperspectral image. Comprising following the steps ofProcessing the spectral reflectivity of each pixel point in the original hyperspectral image, wherein the spectral reflectivity of any pixel point in the original hyperspectral image is I, and the spectral reflectivity after the correction of the black-and-white plate is I cablibration ，I Black Is the spectral reflectivity of the pixel point in the blackboard image, I White The spectral reflectance of the pixel in the whiteboard image.
(3) And extracting the region of interest from the original hyperspectral image after the black-and-white plate correction is completed to obtain a sample image of the wheat seed sample. Comprising the following steps:
a. spectral reflectivities at three preset wavelengths in the original hyperspectral image after the black-and-white plate correction are extracted and used as RGB channel components to form a pseudo RGB image. In one embodiment, spectral reflectance at 460nm is extracted as the B-channel component, spectral reflectance at 525nm is extracted as the G-channel component, and spectral reflectance at 621nm is extracted as the R-channel component.
b. By passing throughThe pseudo RGB image is converted into YCbCr color space and Cb channel components are selected as gray scale images.
c. And extracting an interested region in the original hyperspectral image corrected by the black-and-white plate according to the gray value, filling a zero value to a preset image size, and obtaining a sample image of the wheat seed sample. In one embodiment, an oxford algorithm is utilized to divide the gray level image into a foreground area and a background area, and binarization processing is carried out to obtain a binary image, wherein the foreground area is an area where a wheat seed sample is located. And performing open operation and close operation treatment on the binary image by using a circular structure variable with the diameter of 15 to obtain a binary mask of the wheat seed sample. And carrying out 8-connection on the binary mask to find the connection region of the binary mask, returning the mask of each connection region, and obtaining the region of interest in the original hyperspectral image after the correction of the black-and-white plate according to the original hyperspectral image after the correction of the black-and-white plate is completed by cutting the mask.
And step 3, training a wheat seed classification model by using the source domain sample image and the target domain sample image.
Firstly, randomly matching a target domain sample image and a source domain sample image to form a sample image pair, respectively inputting the sample images into a wheat seed classification model, and adjusting network parameters of an initial feature extraction module and a public feature extraction module until the public features of the source domain sample image and the public features of the target domain sample image are aligned. Comprising common features F of source-domain sample images in the same sample image pair as shown in FIG. 2 n And common features F of the sample image of the target domain n Calculating loss function loss dist According to loss function loss dist And adjusting network parameters of the initial feature extraction module and the public feature extraction module. In one embodiment, the common features F of the source domain sample images in the same sample image pair n And common features F of the sample image of the target domain n Euclidean distance between as a loss function loss dist 。
Then inputting any sample image in the source domain sample image and the target domain sample image into a wheat seed classification model, and combining an output result of a class classifier, an output result of a domain classifier and an output of the class domain classifier based on class labels and domain labels of the input sample imageAnd finally, adjusting network parameters of the class-sensitive private feature extraction module and the domain-sensitive private feature extraction module. Class label calculation loss function loss including class classifier based output result and sample image class Calculating a loss function loss based on an output result of the domain classifier and a domain label of the sample image domain Class labels based on output result of class domain classifier and sample image and domain labels to calculate loss function loss class-domain . Then according to loss function loss class 、lossd omain 、loss class-domain And adjusting network parameters of the class-sensitive private feature extraction module and the domain-sensitive private feature extraction module until model training is completed. Therefore, the wheat seed classification model of the application can learn how to distribute the characteristics of Ji Yuanyu and target domains, and can also notice the difference between the characteristic distribution of the source domain and the characteristic distribution of the target domain.
And 4, acquiring a hyperspectral image to be identified of the wheat seeds to be identified in the target domain, inputting the hyperspectral image to be identified into the wheat seed classification model obtained through training, and obtaining the variety identification result of the wheat seeds to be identified, which is output by a class classifier of the wheat seed classification model. In this step, the step of obtaining the hyperspectral image to be identified of the wheat seed to be identified is the same as the step of obtaining the sample image, and it is also necessary to sequentially correct and extract the region of interest through a black-and-white plate after the original hyperspectral image is obtained. It can be seen that the classification learning on the target domain does not help the domain sensitive private feature F which is instead prone to negative migration d And does not participate in the prediction of the classification result.
The above is only a preferred embodiment of the present application, and the present application is not limited to the above examples. It is to be understood that other modifications and variations which may be directly derived or contemplated by those skilled in the art without departing from the spirit and concepts of the present application are deemed to be included within the scope of the present application.
Claims (10)
1. A domain-adaptive wheat seed classification method based on public and private features, the method comprising:
the method for constructing the network model of the wheat seed classification model comprises the following steps: for extracting common features F n Public feature extraction module for extracting sensitive-like private feature F c Class sensitive private feature extraction module for extracting domain sensitive private feature F d Domain sensitive private feature extraction module, class domain classifier, class classifier and domain classifier, public feature F n And class sensitive private feature F c Inputting the spliced input classes into a classifier with common characteristics F n Sum domain sensitive private feature F d Input domain classifier after splicing, public feature F n Class sensitive private feature F c Sum domain sensitive private feature F d Inputting the spliced class domain classifier;
acquiring a source domain sample image of a wheat seed sample in a source domain and a target domain sample image of the wheat seed sample in a target domain, wherein each sample image is a hyperspectral image with a domain label and a class label, the domain label of the sample image is used for indicating the domain to which the sample image belongs, the class label is used for indicating the variety of the wheat seed sample in the sample image, and the target domain is the same as the variety of the wheat seed sample in the source domain;
training the wheat seed classification model by using a source domain sample image and a target domain sample image;
acquiring a hyperspectral image to be identified of wheat seeds to be identified in a target domain, inputting a wheat seed classification model obtained by training, and obtaining a variety identification result of the wheat seeds to be identified, which is output by a class classifier of the wheat seed classification model;
wherein the common feature F of each hyperspectral image n Is the common characteristic and similar sensitive private characteristic F of the hyperspectral image in different domains c Is the characteristic of the hyperspectral image which is unique on the belonging domain and is used for distinguishing different varieties, and the domain sensitive private characteristic F d Is a feature of the hyperspectral image being different over different domains.
2. The domain adaptive wheat seed classification method of claim 1, wherein the network model of the wheat seed classification model further comprises an initial feature extraction module for extracting initial low-dimensional features F for hyperspectral images input to the wheat seed classification model;
the public feature extraction module performs feature extraction on the initial low-dimensional feature F to obtain a public feature F n The class-sensitive private feature extraction module performs feature extraction on the initial low-dimensional feature F to obtain a class-sensitive private feature F c The domain sensitive private feature extraction module performs feature extraction on the initial low-dimensional feature F to obtain domain sensitive private feature F d 。
3. The domain adaptive wheat seed classification method of claim 2, wherein the training the wheat seed classification model using the source domain sample image and the target domain sample image comprises:
randomly matching a target domain sample image and a source domain sample image to form a sample image pair, respectively inputting the sample images into the wheat seed classification model, and adjusting network parameters of the initial feature extraction module and the public feature extraction module until the public features of the source domain sample image and the public features of the target domain sample image are aligned;
inputting any one sample image in the source domain sample image and the target domain sample image into the wheat seed classification model, and adjusting network parameters of the class-sensitive private feature extraction module and the domain-sensitive private feature extraction module based on class labels and domain labels of the input sample image and by combining an output result of a class classifier, an output result of a domain classifier and an output result of the class domain classifier.
4. A domain adaptive wheat seed classification method according to claim 3, wherein said adjusting network parameters of said initial feature extraction module and said common feature extraction module comprises:
common features F based on source domain sample images in the same sample image pair n And common features F of the sample image of the target domain n Calculating loss function loss dist According to loss function loss dist And adjusting network parameters of the initial feature extraction module and the public feature extraction module.
5. The domain adaptive wheat seed classification method of claim 4, wherein the calculating a loss function loss dist Comprising the following steps:
common features F of source-domain sample images in the same sample image pair n And common features F of the sample image of the target domain n Euclidean distance between as a loss function loss dist 。
6. A domain adaptive wheat seed classification method as claimed in claim 3, wherein said adjusting network parameters of the class sensitive private feature extraction module and the domain sensitive private feature extraction module comprises for any one sample image input:
calculating a loss function loss based on the output result of the class classifier and the class label of the sample image class Calculating a loss function loss based on the output result of the domain classifier and the domain label of the sample image domain Calculating a loss function loss based on the output result of the class domain classifier, the class label of the sample image and the domain label class-domain ；
According to loss function loss class 、loss domain 、loss class-domain And adjusting network parameters of the class-sensitive private feature extraction module and the domain-sensitive private feature extraction module.
7. The domain adaptive wheat seed classification method of claim 1, wherein, for a sample image of a wheat seed sample of any one of a source domain sample image and a target domain sample image, acquiring a sample image of the wheat seed sample comprises:
acquiring an original hyperspectral image of a wheat seed sample;
performing black-and-white plate correction on the original hyperspectral image;
and extracting an interested region from the original hyperspectral image after the black-and-white plate correction to obtain a sample image of the wheat seed sample.
8. The method of claim 7, wherein extracting the region of interest from the corrected original hyperspectral image to obtain the sample image of the wheat seed sample comprises:
spectral reflectivities at three preset wavelengths in an original hyperspectral image after black-and-white plate correction are extracted and used as RGB channel components to form a pseudo RGB image;
Y＝0.257×R+0.504×G+0.098×B+16
converting the pseudo RGB image to cr=0.439 xr-0.368 xg-0.071 xb+128 by cb= -0.148 xr-0.291 xg+0.439xb+128
In the YCbCr color space, selecting a Cb channel component as a gray scale image;
and extracting an interested region in the original hyperspectral image subjected to black-and-white plate correction according to the gray value, filling a zero value to a preset image size, and obtaining a sample image of the wheat seed sample.
9. The domain adaptive wheat seed classification method according to claim 8, wherein the extracting the region of interest in the original hyperspectral image after the black-and-white plate correction according to the gray value comprises:
dividing the gray image into a foreground area and a background area by using an Ojin algorithm, and performing binarization treatment to obtain a binary image, wherein the foreground area is an area where a wheat seed sample is located;
performing open operation and close operation on the binary image by using a circular structure variable with the diameter of 15 to obtain a binary mask of a wheat seed sample;
and carrying out 8-connection on the binary mask to find the connection region of the binary mask, returning the mask of each connection region, and obtaining the region of interest in the original hyperspectral image after the correction of the black-and-white plate according to the original hyperspectral image after the correction of the black-and-white plate by cutting the mask.
10. The domain adaptive wheat seed classification method according to claim 8, wherein the extracting the spectral reflectances at three predetermined wavelengths in the original hyperspectral image after the black-and-white plate correction is performed as RGB channel components, respectively, comprises:
spectral reflectance at 460nm was extracted as a B-channel component, spectral reflectance at 525nm was extracted as a G-channel component, and spectral reflectance at 621nm was extracted as an R-channel component.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310825251.2A CN116883735B (en) | 2023-07-05 | 2023-07-05 | Domain self-adaptive wheat seed classification method based on public features and private features |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310825251.2A CN116883735B (en) | 2023-07-05 | 2023-07-05 | Domain self-adaptive wheat seed classification method based on public features and private features |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN116883735A true CN116883735A (en) | 2023-10-13 |
| CN116883735B CN116883735B (en) | 2024-03-08 |
Family
ID=88265599
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310825251.2A Active CN116883735B (en) | 2023-07-05 | 2023-07-05 | Domain self-adaptive wheat seed classification method based on public features and private features |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116883735B (en) |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180330205A1 (en) * | 2017-05-15 | 2018-11-15 | Siemens Aktiengesellschaft | Domain adaptation and fusion using weakly supervised target-irrelevant data |
| CN112308158A (en) * | 2020-11-05 | 2021-02-02 | 电子科技大学 | Multi-source field self-adaptive model and method based on partial feature alignment |
| CN112949725A (en) * | 2021-03-09 | 2021-06-11 | 江南大学 | Wheat seed classification method based on multi-scale feature extraction |
| WO2021120752A1 (en) * | 2020-07-28 | 2021-06-24 | 平安科技（深圳）有限公司 | Region-based self-adaptive model training method and device, image detection method and device, and apparatus and medium |
| CN113920376A (en) * | 2021-11-02 | 2022-01-11 | 江南大学 | Method for identifying wheat seed varieties based on light-weight convolutional neural network |
| CN114863175A (en) * | 2022-05-10 | 2022-08-05 | 南京信息工程大学 | Unsupervised multi-source partial domain adaptive image classification method |
| CN115049627A (en) * | 2022-06-21 | 2022-09-13 | 江南大学 | Steel surface defect detection method and system based on domain self-adaptive deep migration network |
| CN115620160A (en) * | 2022-10-21 | 2023-01-17 | 中国地质大学（武汉） | Remote sensing image classification method based on multi-classifier active transfer learning resistance |
-
2023
- 2023-07-05 CN CN202310825251.2A patent/CN116883735B/en active Active
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180330205A1 (en) * | 2017-05-15 | 2018-11-15 | Siemens Aktiengesellschaft | Domain adaptation and fusion using weakly supervised target-irrelevant data |
| WO2021120752A1 (en) * | 2020-07-28 | 2021-06-24 | 平安科技（深圳）有限公司 | Region-based self-adaptive model training method and device, image detection method and device, and apparatus and medium |
| CN112308158A (en) * | 2020-11-05 | 2021-02-02 | 电子科技大学 | Multi-source field self-adaptive model and method based on partial feature alignment |
| US20220138495A1 (en) * | 2020-11-05 | 2022-05-05 | University Of Electronic Science And Technology Of China | Model and method for multi-source domain adaptation by aligning partial features |
| CN112949725A (en) * | 2021-03-09 | 2021-06-11 | 江南大学 | Wheat seed classification method based on multi-scale feature extraction |
| CN113920376A (en) * | 2021-11-02 | 2022-01-11 | 江南大学 | Method for identifying wheat seed varieties based on light-weight convolutional neural network |
| CN114863175A (en) * | 2022-05-10 | 2022-08-05 | 南京信息工程大学 | Unsupervised multi-source partial domain adaptive image classification method |
| CN115049627A (en) * | 2022-06-21 | 2022-09-13 | 江南大学 | Steel surface defect detection method and system based on domain self-adaptive deep migration network |
| CN115620160A (en) * | 2022-10-21 | 2023-01-17 | 中国地质大学（武汉） | Remote sensing image classification method based on multi-classifier active transfer learning resistance |
Non-Patent Citations (7)
| Title |
|---|
| ERIC TZENG, ET.AL: "Deep Domain Confusion: Maximizing for Domain Invariance", ARXIV 2014, 10 December 2014 (2014-12-10), pages 1 - 9 * |
| FARAHANI A, ET.AL: "A brief review of doamin adaptation", ADVANCES IN DATA SCIENCE AND INFORMATION ENGINEERING:PROCEEDINGS FROM ICDATA 2020 AND IKE 2020, 2021, 31 October 2020 (2020-10-31), pages 877 - 894 * |
| KONSTANTINOS BOUSMALIS, ET.AL: "Domain Separation Networks", 29TH CONFERENCE ON NEURAL INFORMATION PROCESSING SYSTEMS, 22 August 2016 (2016-08-22), pages 1 - 9 * |
| SONGLIN JIN, ET.AL: "Spatial-spectral feature extraction of hyperspectral images for wheat seed identification", COMPUTERS AND ELECTRICAL ENGINEERING, vol. 101 * |
| XIN ZHAO, HAOTIAN QUE, ET.AL: "Hybrid convolutional network based on hyperspectral imaging for wheat seed varieties classification", INFRARED PHYSICS AND TECHNOLOGY, 20 June 2022 (2022-06-20), pages 1 - 8 * |
| 王广来，等: "基于高光谱图像技术与迁移学习的水晶梨早期损伤检测", 光谱学与光谱分析, vol. 42, no. 11, 7 November 2022 (2022-11-07), pages 3626 - 3630 * |
| 阙昊天等: "基于特征波段卷积神经网络的小麦种子品种分类研究", 中国粮油学报, vol. 38, no. 5, pages 2 - 3 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN116883735B (en) | 2024-03-08 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Arivazhagan et al. | Detection of unhealthy region of plant leaves and classification of plant leaf diseases using texture features | |
| Srinidhi et al. | Plant pathology disease detection in apple leaves using deep convolutional neural networks: Apple leaves disease detection using efficientnet and densenet | |
| Yang et al. | High-throughput soybean seeds phenotyping with convolutional neural networks and transfer learning | |
| CN110647932B (en) | Planting crop structure remote sensing image classification method and device | |
| Padilla et al. | Portable yellow spot disease identifier on sugarcane leaf via image processing using support vector machine | |
| Zingaretti et al. | Automatic fruit morphology phenome and genetic analysis: An application in the octoploid strawberry | |
| CN116559111A (en) | Sorghum variety identification method based on hyperspectral imaging technology | |
| CN113420614A (en) | Method for identifying mildewed peanuts by using near-infrared hyperspectral images based on deep learning algorithm | |
| Jenifa et al. | Classification of cotton leaf disease using multi-support vector machine | |
| CN113920376A (en) | Method for identifying wheat seed varieties based on light-weight convolutional neural network | |
| Loddo et al. | An effective and friendly tool for seed image analysis | |
| Wang et al. | Intelligent detection of hard seeds of snap bean based on hyperspectral imaging | |
| CN116883735B (en) | Domain self-adaptive wheat seed classification method based on public features and private features | |
| CN116433528A (en) | Image detail enhancement display method and system for target area detection | |
| Filipović et al. | Significance of morphological features in rice variety classification using hyperspectral imaging | |
| CN110751144B (en) | Canopy plant hyperspectral image classification method based on sparse representation | |
| Hu et al. | Application of Extreme Learning Machine to Visual Diagnosis of Rapeseed Nutrient Deficiency | |
| Indukuri et al. | Paddy Disease Classifier using Deep learning Techniques | |
| Meena et al. | Plant Diseases Detection Using Deep Learning | |
| Leiva et al. | ScabyNet, a user-friendly application for detecting common scab in potato tubers using deep learning and morphological traits | |
| Hamzah et al. | Investigation of RGB to HSI Conversion Methods for Early Plant Disease Detection Using Hierarchical Synthesis Convolutional Neural Networks | |
| CN116310810B (en) | Cross-domain hyperspectral image classification method based on spatial attention-guided variable convolution | |
| Sambrani et al. | Chili Disease Detection and Classification using Various Machine Learning Techniques | |
| CN112560644B (en) | Crop disease and insect pest automatic identification method suitable for field | |
| Rampal et al. | Enhancement of Corn image quality using Very-Deep Super-Resolution (VDSR) neural network |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |