CN114112992A - Detection method and device for blue pigment of cream and storage medium - Google Patents
Detection method and device for blue pigment of cream and storage medium Download PDFInfo
- Publication number
- CN114112992A CN114112992A CN202210064928.0A CN202210064928A CN114112992A CN 114112992 A CN114112992 A CN 114112992A CN 202210064928 A CN202210064928 A CN 202210064928A CN 114112992 A CN114112992 A CN 114112992A
- Authority
- CN
- China
- Prior art keywords
- model
- training
- initial
- data
- detection model
- 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
- 238000001514 detection method Methods 0.000 title claims abstract description 91
- 239000006071 cream Substances 0.000 title claims abstract description 50
- 239000001055 blue pigment Substances 0.000 title claims abstract description 35
- 238000012549 training Methods 0.000 claims abstract description 106
- 230000003595 spectral effect Effects 0.000 claims abstract description 60
- 238000000034 method Methods 0.000 claims abstract description 54
- 238000001228 spectrum Methods 0.000 claims abstract description 52
- 238000012360 testing method Methods 0.000 claims abstract description 38
- 230000009467 reduction Effects 0.000 claims abstract description 31
- 238000004458 analytical method Methods 0.000 claims abstract description 29
- 238000012795 verification Methods 0.000 claims abstract description 24
- 238000007781 pre-processing Methods 0.000 claims abstract description 17
- 239000000049 pigment Substances 0.000 claims abstract description 14
- 238000004422 calculation algorithm Methods 0.000 claims description 26
- 238000003062 neural network model Methods 0.000 claims description 24
- 230000006870 function Effects 0.000 claims description 23
- 230000008569 process Effects 0.000 claims description 20
- SGHZXLIDFTYFHQ-UHFFFAOYSA-L Brilliant Blue Chemical compound [Na+].[Na+].C=1C=C(C(=C2C=CC(C=C2)=[N+](CC)CC=2C=C(C=CC=2)S([O-])(=O)=O)C=2C(=CC=CC=2)S([O-])(=O)=O)C=CC=1N(CC)CC1=CC=CC(S([O-])(=O)=O)=C1 SGHZXLIDFTYFHQ-UHFFFAOYSA-L 0.000 claims description 19
- 239000000126 substance Substances 0.000 claims description 8
- 238000004590 computer program Methods 0.000 claims description 7
- 238000013135 deep learning Methods 0.000 claims description 4
- 239000008267 milk Substances 0.000 claims description 2
- 210000004080 milk Anatomy 0.000 claims description 2
- 235000013336 milk Nutrition 0.000 claims description 2
- 239000011022 opal Substances 0.000 claims 1
- 238000004364 calculation method Methods 0.000 abstract description 7
- 230000035945 sensitivity Effects 0.000 abstract description 6
- 238000002329 infrared spectrum Methods 0.000 description 9
- 230000009471 action Effects 0.000 description 4
- 230000004913 activation Effects 0.000 description 3
- 238000013528 artificial neural network Methods 0.000 description 3
- 238000013480 data collection Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 238000012545 processing Methods 0.000 description 3
- 210000004027 cell Anatomy 0.000 description 2
- 230000006835 compression Effects 0.000 description 2
- 238000007906 compression Methods 0.000 description 2
- 230000001419 dependent effect Effects 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 230000003287 optical effect Effects 0.000 description 2
- 102220102979 rs766288960 Human genes 0.000 description 2
- 238000004088 simulation Methods 0.000 description 2
- 238000012800 visualization Methods 0.000 description 2
- 238000004477 FT-NIR spectroscopy Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 239000003995 emulsifying agent Substances 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 238000007477 logistic regression Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000002156 mixing Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 230000003094 perturbing effect Effects 0.000 description 1
- 238000000053 physical method Methods 0.000 description 1
- 239000003755 preservative agent Substances 0.000 description 1
- 238000003908 quality control method Methods 0.000 description 1
- 239000011347 resin Substances 0.000 description 1
- 229920005989 resin Polymers 0.000 description 1
- 230000000717 retained effect Effects 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
- 235000013311 vegetables Nutrition 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
- G01N21/17—Systems in which incident light is modified in accordance with the properties of the material investigated
- G01N21/25—Colour; Spectral properties, i.e. comparison of effect of material on the light at two or more different wavelengths or wavelength bands
- G01N21/31—Investigating relative effect of material at wavelengths characteristic of specific elements or molecules, e.g. atomic absorption spectrometry
- G01N21/35—Investigating relative effect of material at wavelengths characteristic of specific elements or molecules, e.g. atomic absorption spectrometry using infrared light
- G01N21/359—Investigating relative effect of material at wavelengths characteristic of specific elements or molecules, e.g. atomic absorption spectrometry using infrared light using near infrared light
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
- G01N21/17—Systems in which incident light is modified in accordance with the properties of the material investigated
- G01N21/25—Colour; Spectral properties, i.e. comparison of effect of material on the light at two or more different wavelengths or wavelength bands
- G01N21/31—Investigating relative effect of material at wavelengths characteristic of specific elements or molecules, e.g. atomic absorption spectrometry
- G01N21/35—Investigating relative effect of material at wavelengths characteristic of specific elements or molecules, e.g. atomic absorption spectrometry using infrared light
- G01N21/3577—Investigating relative effect of material at wavelengths characteristic of specific elements or molecules, e.g. atomic absorption spectrometry using infrared light for analysing liquids, e.g. polluted water
-
- 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
Abstract
The invention provides a method, a device and a storage medium for detecting a blue pigment of cream, belonging to the technical field of data detection, wherein the method comprises the following steps: s1: acquiring data of cream to be detected by a near-infrared spectrometer to obtain original spectrum data; s2: preprocessing the original spectrum data to obtain preprocessed spectrum data; s3: performing dimensionality reduction analysis on the preprocessed spectral data to obtain dimensionality reduced spectral data; s4: randomly dividing all the spectral data subjected to dimensionality reduction into a training set, a test set or a verification set; s5: constructing a training model, and obtaining a target detection model through training analysis of the training set and the test set by the training model; s6: and detecting the verification set through a target detection model to obtain a detection result of the blue pigment of the cream. The invention improves the resolution of the spectrum, reduces the complexity of the model, improves the robustness of the model, improves the sensitivity of the spectrum to the concentration of the artificial pigment, reduces the calculation cost and consumption and improves the precision.
Description
Technical Field
The invention mainly relates to the technical field of data detection, in particular to a method and a device for detecting a blue pigment of cream and a storage medium.
Background
At present, a large amount of vegetable cream is used in cream products on the market, and some pigments, preservatives, emulsifiers and the like are added to achieve a taste close to that of animal cream. The artificial pigment in the plant cream in the cream product is mainly artificially added, and the pigment and the cream are mixed to be uniformly stirred when the cream is whipped. The inspection and quality control of artificial pigments are also problems that need to be solved for the safety of consumers.
The traditional method has the defects of long period, low precision, strong destructiveness and the like, so that a quick and high-precision alternative scheme is urgently needed. The near infrared spectrum analysis technology is a physical measurement technology for rapidly measuring the content and the characteristics of one or more chemical components in a sample by utilizing the optical characteristics of chemical substances in a near infrared spectrum region. The near infrared spectrum technology has the characteristics of stability, high analysis speed, high analysis efficiency and low analysis cost, and is convenient for realizing online analysis and typical nondestructive analysis.
The method has the problems of difficult acquisition of data of the near infrared spectrum, small data volume, unbalanced positive and negative samples, high data dimensionality and the like. In the near infrared spectrum, in addition to processing useful chemical information, a large amount of interference information such as background noise and other irrelevant miscellaneous items is also included, and how to extract useful chemical information is also a problem that needs to be focused.
Disclosure of Invention
The invention aims to solve the technical problem of the prior art and provides a method and a device for detecting a blue pigment of cream and a storage medium.
The technical scheme for solving the technical problems is as follows: a detection method of the brilliant blue pigment of the cream comprises the following steps:
s1: performing data acquisition on cream to be detected through a near-infrared spectrometer to obtain a plurality of original spectrum data;
s2: respectively preprocessing each original spectrum data to obtain preprocessed spectrum data corresponding to each original spectrum data;
s3: performing dimensionality reduction analysis on each preprocessed spectral data to obtain dimensionality reduced spectral data corresponding to each preprocessed spectral data;
s4: randomly dividing all the spectral data subjected to dimensionality reduction into a training set, a testing set or a verification set, wherein the quantity proportion of the training set, the testing set and the verification set after division meets a preset proportion;
s5: constructing a training model, and performing training analysis on the training set and the test set through the training model to obtain a target detection model;
s6: and detecting the verification set through the target detection model to obtain a detection result of the blue pigment of the cream.
Another technical solution of the present invention for solving the above technical problems is as follows: a cream brilliant blue element detection device comprises:
the data acquisition module is used for acquiring data of cream to be detected through a near-infrared spectrometer to obtain a plurality of original spectrum data;
the data preprocessing module is used for respectively preprocessing each original spectrum data to obtain preprocessed spectrum data corresponding to each original spectrum data;
the dimension reduction analysis module is used for respectively carrying out dimension reduction analysis on each preprocessed spectral data to obtain dimension reduced spectral data corresponding to each preprocessed spectral data;
the random division module is used for randomly dividing all the dimensionality-reduced spectral data into a training set, a test set or a verification set, and the quantity proportion of the training set, the test set and the verification set after division meets a preset proportion;
the training analysis module is used for constructing a training model, and performing training analysis on the training set and the test set through the training model to obtain a target detection model;
and the detection result obtaining module is used for detecting the verification set through the target detection model to obtain the detection result of the blue pigment of the cream.
Another technical solution of the present invention for solving the above technical problems is as follows: a detection apparatus for creme brilliant blue pigment, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the detection method for creme brilliant blue pigment as described above.
Another technical solution of the present invention for solving the above technical problems is as follows: a computer-readable storage medium, storing a computer program which, when executed by a processor, implements the method of detecting creme brilliant blue pigment as described above.
The invention has the beneficial effects that: the cream to be measured is collected by the near infrared spectrometer to obtain a plurality of original spectrum data, and the spectrum data after pretreatment is obtained by respectively pretreating each original spectrum data, thereby improving the spectrum resolution, reducing the complexity of the model, and the robustness of the model is improved, the dimension-reduced spectral data is obtained by respectively carrying out dimension reduction analysis on each preprocessed spectral data, the influence of various non-target factors on a target spectrum is weakened, effective information is reserved, the sensitivity of the spectrum to the concentration of the artificial pigment is improved, all the dimension-reduced spectral data are randomly divided into a training set, a test set or a verification set, the target detection model is obtained through training and analyzing the training set and the testing set by the training model, and the detection result of the creamy brilliant blue pigment is obtained through detecting the verification set by the target detection model, so that the calculation cost and consumption are reduced, and the precision is improved to a certain extent.
Drawings
FIG. 1 is a schematic flow chart of a method for detecting brilliant blue pigment in cream according to an embodiment of the present invention;
fig. 2 is a block diagram of a detection apparatus for detecting a blue pigment in cream according to an embodiment of the present invention.
Detailed Description
The principles and features of this invention are described below in conjunction with the following drawings, which are set forth by way of illustration only and are not intended to limit the scope of the invention.
Fig. 1 is a schematic flow chart of a method for detecting a blue pigment in cream according to an embodiment of the present invention.
As shown in fig. 1, a method for detecting the brilliant blue pigment of milk oil comprises the following steps:
s1: performing data acquisition on cream to be detected through a near-infrared spectrometer to obtain a plurality of original spectrum data;
s2: respectively preprocessing each original spectrum data to obtain preprocessed spectrum data corresponding to each original spectrum data;
s3: performing dimensionality reduction analysis on each preprocessed spectral data to obtain dimensionality reduced spectral data corresponding to each preprocessed spectral data;
s4: randomly dividing all the spectral data subjected to dimensionality reduction into a training set, a testing set or a verification set, wherein the quantity proportion of the training set, the testing set and the verification set after division meets a preset proportion;
s5: constructing a training model, and performing training analysis on the training set and the test set through the training model to obtain a target detection model;
s6: and detecting the verification set through the target detection model to obtain a detection result of the blue pigment of the cream.
It should be understood that the preset ratio may be 6:3: 1.
It should be understood that the artificial pigment brilliant blue in the cream (i.e. the cream to be tested) is rapidly scanned near-infrared by the fourier transform near-infrared spectroscopy technique.
Specifically, the method comprises the following steps of carrying out data acquisition on cream to be detected through a near-infrared spectrometer:
step 1.1: obtaining resin cream used for determination (namely the cream to be determined); step 1.2: mixing cream with standard concentration difference brilliant blue pigment prepared in advance to prepare a sample to be detected; step 1.3: placing a sample to be detected into a near infrared spectrometer for near infrared scanning; step 1.4: the spectrometer records the near infrared spectrum image of the sample to obtain the measured data (namely the original spectrum data), and the integral data set can be objectively evaluated.
It should be understood that the pigment used in the cream to be tested is brilliant blue, and the purity is above 99%.
It should be understood that each of the raw spectral data is preprocessed by using a fourier transform near infrared spectral algorithm, so as to obtain preprocessed spectral data corresponding to each of the raw spectral data.
It should be appreciated that step S6 can result in a more accurate prediction of the light bluish pixels.
In the above embodiment, the cream to be detected is collected by the near-infrared spectrometer to obtain a plurality of original spectrum data, each original spectrum data is preprocessed to obtain preprocessed spectrum data, the resolution of the spectrum is improved, the complexity of the model is reduced, the robustness of the model is improved, each preprocessed spectrum data is subjected to dimensionality reduction analysis to obtain dimensionality reduction spectrum data, the influence of various non-target factors on a target spectrum is weakened, effective information is retained, the sensitivity of the spectrum to the artificial pigment concentration is improved, all the dimensionality reduction spectrum data are randomly divided into a training set or a test set or a verification set, a target detection model is obtained through the training analysis of the training model on the training set and the test set, the detection result of the cream bright blue pigment is obtained through the detection of the target detection model on the verification set, and the calculation cost and consumption are reduced, meanwhile, the precision is improved to a certain extent.
Optionally, as an embodiment of the present invention, the process of step S3 includes:
constructing an initial automatic encoder, and updating parameters of the initial automatic encoder by using an unsupervised learning algorithm to obtain an updated automatic encoder;
and respectively carrying out dimensionality reduction on each preprocessed spectral data through the updated automatic encoder to obtain dimensionality reduced spectral data corresponding to each preprocessed spectral data.
It should be appreciated that, for the problem of high dimensionality of near infrared spectral data, the target data (i.e., the preprocessed spectral data) is subjected to non-linearly dependent dimensionality reduction preprocessing by an automatic encoder.
It should be appreciated that pre-training of the auto-encoder (i.e., the initial auto-encoder) without supervised learning results in initial weights that are closer to the optimal solution.
It will be appreciated that the original data is compressed by the auto-encoder (i.e. the updated auto-encoder) into a short code ignoring noise, which is then decompressed by the algorithm to generate an image as close as possible to the original input. The high-dimensional data is encoded into a low-dimensional data through a multi-layer neural network, thereby reconstructing the high-dimensional data.
Specifically, before a training model is established, an automatic encoder is required to perform spectrum preprocessing work, the influence of various non-target factors on a target spectrum is weakened, effective information is reserved, and the sensitivity of the spectrum to the concentration of the artificial pigment is improved.
It should be appreciated that the auto-encoder can minimize the loss of information and remove noise from the original data.
It should be understood that there are two networks from the encoder (i.e. the updated automatic encoder), and firstly the encoding network is a deep automatic encoder with three layers, the first layer is an input layer of a feature dimension 746, the feature extraction is performed to enter a second layer dimension 378, the last layer dimension is reduced to 189, then we perform inverse reconstruction through the feature, and sequentially activate scaling under the action of RBM unit, and restore the data to the original spectral information through a 2-layer decoding network. The decoding network carries out forward propagation through digital coding, and merging is carried out after RBM unit probability in neurons of each layer is not activated, and finally data after dimensionality reduction (namely the spectral data after dimensionality reduction) is obtained.
It should be understood that the spectrometer data is scanned in a post-fixed dimension 746, followed by a halving of each layer.
Specifically, the process of performing the dimension reduction processing on each preprocessed spectral data through the updated automatic encoder includes the following steps:
firstly, after the pre-training process of the network is completed, the decoding and encoding parts are taken back again to be unfolded to form the whole network, then the real data is used as a sample label to finely adjust the parameters of the network, and the error between a reconstruction item and the original data is minimized to train the weight;
second, for continuous spectral data, the hidden layer of the first RBM is still binary, but its visible layer elements are linear elements with white gaussian noise. If the noise is unit variance, the updating rule of the hidden unit is the same, the updating rule of the ith visualization layer unit is sampled from Gaussian noise, the variance of the noise is unit variance, and the mean is
Average value of (d);
finally, each RBM's visible layer element has a true [0, 1] internal activation value, and for a higher layer RBM, its visible layer element is the activation probability of the previous RBM's hidden layer element, but except for the top RBM, the other RBM's hidden layer elements are random binary values. The hidden cell of the top RBM is a random real-valued state, which is sampled from unit variance noise, the mean value of the unit variance noise is determined by the visible cell of the RBM, and the cross entropy error formula is as follows:
where pi is the reconstructed value of the input data.
It will be appreciated that the goal of the auto-encoder is to map high dimensional data to a low dimensional space through projection, and it is desirable to obtain the maximum data variance in the projection dimension in order to preserve more features of the original data points with less data dimension used.
In the embodiment, the unsupervised learning algorithm is used for updating the parameters of the initial automatic encoder to obtain the updated automatic encoder, and the updated automatic encoder is used for respectively performing dimensionality reduction on each preprocessed spectral data to obtain the dimensionality reduced spectral data, so that the information loss and the noise of the original data can be reduced to the maximum extent, the effective information is reserved, and the sensitivity of the spectrum to the artificial pigment concentration is improved.
Optionally, as an embodiment of the present invention, the constructing an initial automatic encoder, and updating parameters of the initial automatic encoder by using an unsupervised learning algorithm to obtain an updated automatic encoder includes:
importing pre-constructed model parameters, and performing random initialization on the model parameters to obtain a plurality of initial node parameters and a plurality of initial weight parameters;
calculating an energy function for the plurality of initial node parameters and the plurality of initial weight parameters through a first equation to obtain the energy function, wherein the first equation is as follows:
wherein the content of the first and second substances,
is the initial node parameter of the current layer,
is the initial node parameter of the next layer,
and
are all initial weight parameters, and are all the initial weight parameters,
as a weight value, the weight value,
is a function of energy;
constructing an initial automatic encoder through a plurality of initial node parameters, a plurality of initial weight parameters and the energy function;
and importing a plurality of updating weights, and sequentially inputting each updating weight into the initial automatic encoder according to a preset first iteration number to update the parameters of the automatic encoder, so as to obtain an updated automatic encoder.
It should be understood that the pre-constructed model parameters are a 128 x 128 number matrix, which is initially 1, and which is randomized after the randomization function is called.
It should be understood that the update weight is a randomized parameter with a weight of 1.
It will be appreciated that to model large data sets, a multi-layer network is used, with the output of the first layer network being the input to the second layer network. And when a network layer is added, the lower bound probability value of the network for input data reconstruction is improved, and the upper layer network can extract the higher-order characteristics of the lower layer network.
Specifically, firstly, reconstructing a binary vector through a 2-layer network, and associating a random binary pixel point with a random binary feature detector through symmetric weighted connection; secondly, the pixel points are equivalent to visualization units of the RBM, and the feature detector is equivalent to a hidden unit. The energy (energy function) between the joint system (v, h) of visual and hidden units is expressed as:
vi and hj are states of the ith visible layer unit (i.e., the initial node parameter of the current layer) and the jth hidden layer unit (i.e., the initial node parameter of the next layer), bi and bj are bias terms (i.e., the initial weight parameters), wij is a weight, and the network obtains the probability of each possible image through the energy function; and finally, the weight and the bias value of the network are adjusted to ensure that the energy of the network to the input image is the lowest.
It should be appreciated that the operation of the self-encoder to reconstruct the input during the pre-training phase, i.e., the portion of unsupervised training, amounts to forcing the data hiding features of the net learning input before the output layer. The output layer would then reverse model the original input based on the input's eigen-expression representation. In actual operation, the input information is first encoded, the information is extracted through the hidden layer, and finally the output information is decoded. The determined distributed representation has a lower dimension than the original input, thereby achieving dimensionality reduction.
In the embodiment, the unsupervised learning algorithm is used for updating the parameters of the initial automatic encoder to obtain the updated automatic encoder, so that the accuracy of the model is improved, the information loss of the original data can be reduced to the maximum extent, the noise is removed, and the effective information is reserved.
Optionally, as an embodiment of the present invention, the process of step S5 includes:
constructing a full-connection neural network model based on a deep learning algorithm, and training the training set and the test set for the first time through the full-connection neural network model to obtain an initial detection model and a residual prediction deviation value;
judging whether the residual prediction deviation value is larger than a preset deviation threshold value, if so, compressing the initial detection model and the training set through a knowledge evolution algorithm to obtain a target detection model; if not, the process returns to step S4.
Preferably, the preset deviation threshold may be 3.0.
The method has the advantages that the spectral data are subjected to regression prediction by deep learning, the sigmoid function is adopted in the logistic regression as an activation function, a fully-connected neural network model is established, the problem that the deep learning is large in required data volume and high in difficulty in manually collecting near-infrared samples to achieve the required data volume is solved, the model is compressed by using a knowledge evolution method when the deep learning model is established, the common neural network and a residual error network can be seamlessly integrated by the knowledge evolution, and the burden of overfitting and data collection is reduced.
It should be appreciated that knowledge evolution is used to compress the model (i.e., the initial detection model), the training of which is based on weight initialization and allows for more compact models to be trained without affecting accuracy.
In the above embodiment, the initial detection model and the residual prediction deviation value are obtained by training the training set and the test set for the first time through the fully connected neural network model, whether the residual prediction deviation value is greater than the preset deviation threshold value is judged, and the target detection model is obtained by compressing the initial detection model and the training set through the knowledge evolution algorithm, so that the networks can be seamlessly connected, and the burden of overfitting and data collection is reduced.
Optionally, as an embodiment of the present invention, the process of training the training set and the test set for the first time through the fully-connected neural network model to obtain an initial detection model and a residual prediction deviation value includes:
s511: importing a plurality of real values corresponding to the dimensionality reduced spectral data in the training set, and performing model training on the training set and all the real values through the fully-connected neural network model to obtain an initial detection model;
s512: and analyzing the residual prediction deviation value of the test set and all the real values through the initial detection model to obtain a residual prediction deviation value.
It should be understood that the real value is the concentration of the pigment in the sample, and can be understood as the concentration of the brilliant blue pigment in the cream to be measured.
In the embodiment, the initial detection model is obtained by training the training set and the models of all real values through the fully-connected neural network model, and the residual prediction deviation value is obtained by analyzing the residual prediction deviation values of the initial detection model on the testing set and all real values.
Optionally, as an embodiment of the present invention, in step S511, performing model training on the training set and all real values through the fully-connected neural network model, and obtaining an initial detection model includes:
s5111: predicting each post-dimensionality reduction spectral data in the training set according to a forward propagation algorithm and the full-connection neural network model to obtain a first predicted value corresponding to each post-dimensionality reduction spectral data;
s5112: calculating the difference between each first predicted value and the true value corresponding to the dimension-reduced spectral data respectively to obtain an error value corresponding to each dimension-reduced spectral data;
s5113: introducing a loss function, and performing back propagation on all error values according to the loss function to obtain a gradient value;
s5114: and updating parameters of the fully-connected neural network model according to the gradient descent algorithm and the gradient value to obtain an updated fully-connected neural network model, returning to the step S5111 until a preset second iteration number is reached, and taking the updated fully-connected neural network model as an initial detection model.
It should be understood that the preset second iteration number may be the same as or different from the preset first iteration number.
It should be understood that the predicted value (i.e. the first predicted value) is calculated by a forward propagation algorithm, and the predicted value (i.e. the first predicted value) and the true value are compared to obtain the difference therebetween.
Specifically, the gradient of the loss function for each parameter is calculated through a back propagation algorithm (backpropagation), and then each parameter is updated by using a gradient descent algorithm (gradient count) according to the gradient and the learning rate. The popular understanding is that the gradient descent algorithm is mainly used for optimizing the value of a single parameter, and the back propagation algorithm provides an efficient way to use the gradient descent algorithm on all parameters, so that the loss function of the neural network model on the training data is as small as possible.
In the embodiment, the initial detection model is obtained by training the training set and all real value models through the fully connected neural network model, so that the loss function of the neural network model on the training data is as small as possible, and the accuracy of model detection is improved.
Optionally, as an embodiment of the present invention, the process of step S512 includes:
inputting the test set into the initial detection model according to a preset third iteration number to perform model test to obtain a plurality of second predicted values;
calculating residual prediction deviation values of all the second predicted values and all the real values through a second formula to obtain residual prediction deviation values, wherein the second formula is as follows:
wherein the content of the first and second substances,
in order to be the true value of the value,
in order to be the second predicted value,
is the average value of the values,
the deviation values are predicted for the residuals.
It should be understood that the formula for the residual prediction bias (RPD) is as follows:
the formula for the standard deviation (mean square error) is as follows:
the root mean square error is given by:
where X represents the true value, X represents the predicted value (i.e., the second predicted value), and E represents the expectation, or average, of the sample.
In the above embodiment, the test set is input into the initial detection model according to the preset third iteration number to perform model test to obtain a plurality of second predicted values, and the residual prediction deviation values of all the second predicted values and all the real values are calculated by the second formula to obtain the residual prediction deviation values.
Optionally, as an embodiment of the present invention, the process of compressing the initial detection model and the training set by using a knowledge evolution algorithm to obtain a target detection model includes:
according to the parameter setting proportion of the network space volume of the initial detection model, segmenting the initial detection model to obtain a first sub-network and a second sub-network;
performing parameter randomization on the first sub-network to obtain a randomized first sub-network;
constructing a detection model to be trained through the randomized first sub-network and the second sub-network;
and training the training set again through the detection model to be trained to obtain a target detection model.
It should be understood that the parameter randomization process is a randomization of the invocation of an existing randomization function.
It should be understood that the Knowledge Evolution (KE) method separates the deep network into two hypotheses, a fitting hypothesis (i.e., the first subnetwork) and a resetting hypothesis (i.e., the second subnetwork). We iteratively develop knowledge within the fitting hypothesis by perturbing the multi-generation reset hypothesis. The method not only improves the performance, but also learns a thin network with low calculation cost. The KE can be seamlessly integrated with the normal neural network and the residual network and the KE reduces the burden of overfitting and data collection.
Specifically, to evolve knowledge in deep networks, the network is split into two hypotheses, fitting hypothesis H4 (i.e., the first subnetwork) and resetting hypothesis H0 (i.e., the second subnetwork); step 3.2.2: we develop knowledge inside H4 (i.e. the first sub-network) by retraining the multi-generation network. For each generation of training, we perturb the weights inside H0 (i.e., the second sub-network) to encourage H4 to learn independent knowledge information. The network is split into two blocks for H4 (i.e. the first sub-network) and H0 (i.e. the second sub-network), i.e. H4= mn and H0= (1-m) n. By both H4 (the second sub-network) and H0 (i.e. the second sub-network) being randomly initialized, we constantly initialize H0 (i.e. the second sub-network) in training, encouraging H4 (i.e. the first sub-network) to be able to gain the impact of greater weight in the network and to gain higher simulation and alternatives to the entire network; and the training N is an e epoch, the training networks are called the first generation H14= MN1 (i.e. the randomized first subnetwork) and H1O = (1-M) N1 (i.e. the second subnetwork); thereafter, epochs are trained by successive iterations and using an initialization function F1= M1F1+ (1-M1) F1r, where F1 is the convolution filter at layer l, M1 is the corresponding binary mask, and F1r is a randomly initialized decimal. The three tensors (F1, F1r and M1) have the same size. F1r is initialized with a default initialization profile. And continuing training iteration on the initialized model so as to transfer the knowledge learned by H4 to the next generation of H4. H4 can be regarded as a complete lightweight model alone, reduce the dependence on the data volume, and because its data volume is greatly reduced, the computational cost is also reduced.
Specifically, the specific steps of the knowledge evolution method are as follows:
the training path symbol includes a deep network n of layer 1 is assumed. The network n has a convolution filter f, a batch norm z, and a fully connected layer with weight w, offset b terms;
starting with the conceptual decomposition of the deep network n into two exclusive hypotheses (subnets), the fitting hypothesis H4 (i.e., the first subnetwork) and the resetting hypothesis H0 (i.e., the second subnetwork). These assumptions are summarized by a binary mask m, 1 for H4 (i.e. the first sub-network) and 0 for H0 (i.e. the second sub-network), i.e. H4= mn and H0= (1-m) n. After the summary assumption, the random initialization network n, i.e. H4 (i.e. the first sub-network) and H0 (i.e. the second sub-network) are both randomly initialized. Taking training N as an e epoch, and referring training networks as a first generation H14= MN1 and H1O = (1-M) N1;
to learn a better network (next generation), the network N is reinitialized using H4 (i.e., the first subnetwork) and then retrained N-learning N2. The network N is first reinitialized with N1 fitting the convolution filter f and weights w in the hypothesis H4 (i.e., the first subnetwork), and then the rest of the network is randomly initialized. Formally, we reinitialize each layer 1 as shown below
F1= M1F1 + (1-M1) F1r；
Where F1 is the convolution filter at layer 1, M1 is the corresponding binary mask, and F1r is the randomly initialized decimal. The three tensors (F1, F1r and M1) have the same size. F1r is initialized with a default initialization profile;
the weights W1 and bias B1 are reinitialized by the corresponding binary mask. The structure has a bias term only in the single last fully connected layer. Thus, for these architectures, all bias terms belong to the fitting hypothesis, i.e., H4, learning the bulk norm across generations without randomization;
after reinitialization, n is retrained as the e epoch to learn the second generation G2. To learn a better network, the n generations of G are repeatedly initialized and retrained. Basically, knowledge (convolution filters and weights) is passed from one generation to the next by fitting the hypothesis H4 (i.e. the first subnetwork). Therefore, the purpose of model compression is achieved, and the trained H4 can be used as a complete lightweight model, so that the dependence on data volume can be reduced, and the problems of difficulty in data acquisition and small data volume during the passing of the near infrared spectrum are exactly met due to the fact that the data volume is greatly reduced and the calculation cost is reduced.
In the embodiment, the initial detection model is divided into the first sub-network and the second sub-network according to the parameter setting proportion of the network space volume of the initial detection model, the randomized first sub-network is obtained through the parameter randomization of the first sub-network, the detection model to be trained is obtained through the construction of the randomized first sub-network and the second sub-network, and the target detection model is obtained through the retraining of the detection model to be trained on the training set, so that the purpose of model compression is achieved, a complete lightweight model is formed, the dependence on data volume can be reduced, and the calculation cost is reduced due to the fact that the data volume is greatly reduced, and the problems that data obtaining is difficult and the data volume is small when the near infrared spectrum passes through are solved.
Optionally, as another embodiment of the present invention, the present invention preprocesses the near infrared spectrum (i.e., the original spectral data), so as to weaken the influence of various non-target factors on the target spectrum, retain effective information, improve the sensitivity of the spectrum to the artificial pigment concentration, and simultaneously, the preprocessing can also improve the resolution of the spectrum, reduce the complexity of the model, and improve the robustness of the model, and the optimal preprocessing method obtained through repeated experiments is selected as the preprocessing method selected in the data preprocessing stage for preprocessing by the nonlinear-dependent automatic encoder, and the automatic encoder can reduce the information loss and remove noise of the original data to the maximum extent, wherein the automatic encoder reconstructs input operations in the pre-training stage, that is, the unsupervised training portion is equivalent to the hidden features of the data input by network learning before the forcing output layer; then, the output layer can represent reverse simulation original input based on the input feature expression, in actual operation, input information is encoded, information is extracted through a hidden layer, and finally output information is decoded; the dimensionality of the determined distributed representation is lower than that of the original input, so that dimensionality reduction is realized; secondly, because the AutoEncoder trains end-to-end, the accuracy rate is continuously improved, and the input and output are closer and closer by designing the encode and decode process, the method is an unsupervised learning process. It is similar to the PCA (programmable automation controller) algorithm in machine learning, the main function of AutoEncoder is to compress the data to obtain the feature value of reduced dimension, and this intermediate result is just similar to the result of PCA, which is the most essential feature of the original data. Knowledge evolution learns a more portable network, reduces calculation cost and consumption, and improves precision to a certain extent.
Fig. 2 is a block diagram of a detection apparatus for detecting a blue pigment in cream according to an embodiment of the present invention.
Alternatively, as another embodiment of the present invention, as shown in fig. 2, a device for detecting cream brilliant blue pigment includes:
the data acquisition module is used for acquiring data of cream to be detected through a near-infrared spectrometer to obtain a plurality of original spectrum data;
the data preprocessing module is used for respectively preprocessing each original spectrum data to obtain preprocessed spectrum data corresponding to each original spectrum data;
the dimension reduction analysis module is used for respectively carrying out dimension reduction analysis on each preprocessed spectral data to obtain dimension reduced spectral data corresponding to each preprocessed spectral data;
the random division module is used for randomly dividing all the dimensionality-reduced spectral data into a training set, a test set or a verification set, and the quantity proportion of the training set, the test set and the verification set after division meets a preset proportion;
the training analysis module is used for constructing a training model, and performing training analysis on the training set and the test set through the training model to obtain a target detection model;
and the detection result obtaining module is used for detecting the verification set through the target detection model to obtain the detection result of the blue pigment of the cream.
Optionally, another embodiment of the present invention provides a detection apparatus for creulenin, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and when the processor executes the computer program, the detection method for creulenin as described above is implemented. The device may be a computer or the like.
Alternatively, another embodiment of the present invention provides a computer-readable storage medium storing a computer program which, when executed by a processor, implements the method for detecting a blue pigment in cream as described above.
It is noted that, herein, relational terms such as first and second, and the like may be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Also, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus.
It is clear to those skilled in the art that, for convenience and brevity of description, the specific working processes of the above-described apparatuses and units may refer to the corresponding processes in the foregoing method embodiments, and are not described herein again.
In the several embodiments provided in the present application, it should be understood that the disclosed apparatus and method may be implemented in other ways. For example, the above-described apparatus embodiments are merely illustrative, and for example, a division of a unit is merely a logical division, and an actual implementation may have another division, for example, a plurality of units or components may be combined or integrated into another system, or some features may be omitted, or not executed.
Units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of the embodiment of the present invention.
In addition, functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may exist alone physically, or two or more units are integrated into one unit. The integrated unit can be realized in a form of hardware, and can also be realized in a form of a software functional unit.
The integrated unit, if implemented in the form of a software functional unit and sold or used as a stand-alone product, may be stored in a computer readable storage medium. Based on such understanding, the technical solution of the present invention essentially or partially contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product stored in a storage medium and including instructions for causing a computer device (which may be a personal computer, a server, or a network device) to execute all or part of the steps of the method according to the embodiments of the present invention. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a magnetic disk or an optical disk, and other various media capable of storing program codes.
The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents, improvements and the like that fall within the spirit and principle of the present invention are intended to be included therein.
Claims (10)
1. A detection method of the brilliant blue pigment of the cream is characterized by comprising the following steps:
s1: performing data acquisition on cream to be detected through a near-infrared spectrometer to obtain a plurality of original spectrum data;
s2: respectively preprocessing each original spectrum data to obtain preprocessed spectrum data corresponding to each original spectrum data;
s3: performing dimensionality reduction analysis on each preprocessed spectral data to obtain dimensionality reduced spectral data corresponding to each preprocessed spectral data;
s4: randomly dividing all the spectral data subjected to dimensionality reduction into a training set, a testing set or a verification set, wherein the quantity proportion of the training set, the testing set and the verification set after division meets a preset proportion;
s5: constructing a training model, and performing training analysis on the training set and the test set through the training model to obtain a target detection model;
s6: and detecting the verification set through the target detection model to obtain a detection result of the blue pigment of the cream.
2. The method for detecting cream brilliant blue pigment according to claim 1, wherein the process of step S3 includes:
constructing an initial automatic encoder, and updating parameters of the initial automatic encoder by using an unsupervised learning algorithm to obtain an updated automatic encoder;
and respectively carrying out dimensionality reduction on each preprocessed spectral data through the updated automatic encoder to obtain dimensionality reduced spectral data corresponding to each preprocessed spectral data.
3. The method for detecting the blue pigment in the milky oil according to claim 2, wherein the step of constructing the initial automatic encoder, updating parameters of the initial automatic encoder by using an unsupervised learning algorithm, and obtaining the updated automatic encoder comprises the following steps:
importing pre-constructed model parameters, and performing random initialization on the model parameters to obtain a plurality of initial node parameters and a plurality of initial weight parameters;
calculating an energy function for the plurality of initial node parameters and the plurality of initial weight parameters through a first equation to obtain the energy function, wherein the first equation is as follows:
wherein the content of the first and second substances,
is the initial node parameter of the current layer,
is the initial node parameter of the next layer,
and
are all initial weight parameters, and are all the initial weight parameters,
as a weight value, the weight value,
is a function of energy;
constructing an initial automatic encoder through a plurality of initial node parameters, a plurality of initial weight parameters and the energy function;
and importing a plurality of updating weights, and sequentially inputting each updating weight into the initial automatic encoder according to a preset first iteration number to update the parameters of the automatic encoder, so as to obtain an updated automatic encoder.
4. The method for detecting cream brilliant blue pigment according to claim 1, wherein the process of step S5 includes:
constructing a full-connection neural network model based on a deep learning algorithm, and training the training set and the test set for the first time through the full-connection neural network model to obtain an initial detection model and a residual prediction deviation value;
judging whether the residual prediction deviation value is larger than a preset deviation threshold value, if so, compressing the initial detection model and the training set through a knowledge evolution algorithm to obtain a target detection model; if not, the process returns to step S4.
5. The method for detecting the blue pigment in the milky oil according to claim 4, wherein the process of training the training set and the testing set for the first time by the fully connected neural network model to obtain an initial detection model and a residual prediction deviation value comprises:
s511: importing a plurality of real values corresponding to the dimensionality reduced spectral data in the training set, and performing model training on the training set and all the real values through the fully-connected neural network model to obtain an initial detection model;
s512: and analyzing the residual prediction deviation value of the test set and all the real values through the initial detection model to obtain a residual prediction deviation value.
6. The method for detecting the brilliant blue pigment in milk according to claim 5, wherein in step S511, the model training of the training set and all real values by the fully-connected neural network model to obtain an initial detection model comprises:
s5111: predicting each post-dimensionality reduction spectral data in the training set according to a forward propagation algorithm and the full-connection neural network model to obtain a first predicted value corresponding to each post-dimensionality reduction spectral data;
s5112: calculating the difference between each first predicted value and the true value corresponding to the dimension-reduced spectral data respectively to obtain an error value corresponding to each dimension-reduced spectral data;
s5113: introducing a loss function, and performing back propagation on all error values according to the loss function to obtain a gradient value;
s5114: and updating parameters of the fully-connected neural network model according to the gradient descent algorithm and the gradient value to obtain an updated fully-connected neural network model, returning to the step S5111 until a preset second iteration number is reached, and taking the updated fully-connected neural network model as an initial detection model.
7. The method for detecting cream brilliant blue pigment according to claim 5, wherein the process of step S512 comprises:
inputting the test set into the initial detection model according to a preset third iteration number to perform model test to obtain a plurality of second predicted values;
calculating residual prediction deviation values of all the second predicted values and all the real values through a second formula to obtain residual prediction deviation values, wherein the second formula is as follows:
wherein the content of the first and second substances,
in order to be the true value of the value,
in order to be the second predicted value,
is the average value of the values,
the deviation values are predicted for the residuals.
8. The method for detecting the blue pigment in the milky oil according to claim 4, wherein the step of compressing the initial detection model and the training set by a knowledge evolution algorithm to obtain a target detection model comprises:
according to the parameter setting proportion of the network space volume of the initial detection model, segmenting the initial detection model to obtain a first sub-network and a second sub-network;
performing parameter randomization on the first sub-network to obtain a randomized first sub-network;
constructing a detection model to be trained through the randomized first sub-network and the second sub-network;
and training the training set again through the detection model to be trained to obtain a target detection model.
9. A cream brilliant blue element detection device, characterized by comprising:
the data acquisition module is used for acquiring data of cream to be detected through a near-infrared spectrometer to obtain a plurality of original spectrum data;
the data preprocessing module is used for respectively preprocessing each original spectrum data to obtain preprocessed spectrum data corresponding to each original spectrum data;
the dimension reduction analysis module is used for respectively carrying out dimension reduction analysis on each preprocessed spectral data to obtain dimension reduced spectral data corresponding to each preprocessed spectral data;
the random division module is used for randomly dividing all the dimensionality-reduced spectral data into a training set, a test set or a verification set, and the quantity proportion of the training set, the test set and the verification set after division meets a preset proportion;
the training analysis module is used for constructing a training model, and performing training analysis on the training set and the test set through the training model to obtain a target detection model;
and the detection result obtaining module is used for detecting the verification set through the target detection model to obtain the detection result of the blue pigment of the cream.
10. A computer-readable storage medium, in which a computer program is stored, which, when being executed by a processor, carries out a method for detecting a blue opal pigment according to any one of claims 1 to 8.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210064928.0A CN114112992B (en) | 2022-01-20 | 2022-01-20 | Detection method and device for blue pigment of cream and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210064928.0A CN114112992B (en) | 2022-01-20 | 2022-01-20 | Detection method and device for blue pigment of cream and storage medium |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114112992A true CN114112992A (en) | 2022-03-01 |
| CN114112992B CN114112992B (en) | 2022-04-12 |
Family
ID=80360917
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210064928.0A Active CN114112992B (en) | 2022-01-20 | 2022-01-20 | Detection method and device for blue pigment of cream and storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114112992B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114371145A (en) * | 2022-03-21 | 2022-04-19 | 武汉工程大学 | Detection method and device for milk oil mixed pigment, electronic equipment and storage medium |
| CN115205716A (en) * | 2022-08-11 | 2022-10-18 | 北京林业大学 | Method, device and system for estimating oil content of olive fruits and storage medium |
| CN116735527A (en) * | 2023-06-09 | 2023-09-12 | 湖北经济学院 | Near infrared spectrum optimization method, device and system and storage medium |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104359882A (en) * | 2014-11-12 | 2015-02-18 | 江南大学 | Method for simultaneously measuring hybrid pigment by synchronous fluorescence spectroscopy with RBF (Radial Basis Function) neural network |
| CN106295199A (en) * | 2016-08-15 | 2017-01-04 | 中国地质大学（武汉） | Automatic history matching method and system based on autocoder and multiple-objection optimization |
| CN107833208A (en) * | 2017-10-27 | 2018-03-23 | 哈尔滨工业大学 | A kind of hyperspectral abnormity detection method based on changeable weight depth own coding |
| CN110782018A (en) * | 2019-10-28 | 2020-02-11 | 北京环境特性研究所 | Spectral dimension reduction method and device based on self-encoder |
| CN111044483A (en) * | 2019-12-27 | 2020-04-21 | 武汉工程大学 | Method, system and medium for determining pigment in cream based on near infrared spectrum |
| CN111144499A (en) * | 2019-12-27 | 2020-05-12 | 北京工业大学 | Fan blade early icing fault detection method based on deep neural network |
| AU2021101715A4 (en) * | 2021-04-03 | 2021-05-20 | Southeast University | A transfer learning based model for fatigue crack initiation sites detection |
| CN113008805A (en) * | 2021-02-07 | 2021-06-22 | 浙江工业大学 | Radix angelicae decoction piece quality prediction method based on hyperspectral imaging depth analysis |
| CN113408663A (en) * | 2021-07-20 | 2021-09-17 | 中国科学院地理科学与资源研究所 | Fusion model construction method, fusion model using device and electronic equipment |
-
2022
- 2022-01-20 CN CN202210064928.0A patent/CN114112992B/en active Active
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104359882A (en) * | 2014-11-12 | 2015-02-18 | 江南大学 | Method for simultaneously measuring hybrid pigment by synchronous fluorescence spectroscopy with RBF (Radial Basis Function) neural network |
| CN106295199A (en) * | 2016-08-15 | 2017-01-04 | 中国地质大学（武汉） | Automatic history matching method and system based on autocoder and multiple-objection optimization |
| CN107833208A (en) * | 2017-10-27 | 2018-03-23 | 哈尔滨工业大学 | A kind of hyperspectral abnormity detection method based on changeable weight depth own coding |
| CN110782018A (en) * | 2019-10-28 | 2020-02-11 | 北京环境特性研究所 | Spectral dimension reduction method and device based on self-encoder |
| CN111044483A (en) * | 2019-12-27 | 2020-04-21 | 武汉工程大学 | Method, system and medium for determining pigment in cream based on near infrared spectrum |
| CN111144499A (en) * | 2019-12-27 | 2020-05-12 | 北京工业大学 | Fan blade early icing fault detection method based on deep neural network |
| CN113008805A (en) * | 2021-02-07 | 2021-06-22 | 浙江工业大学 | Radix angelicae decoction piece quality prediction method based on hyperspectral imaging depth analysis |
| AU2021101715A4 (en) * | 2021-04-03 | 2021-05-20 | Southeast University | A transfer learning based model for fatigue crack initiation sites detection |
| CN113408663A (en) * | 2021-07-20 | 2021-09-17 | 中国科学院地理科学与资源研究所 | Fusion model construction method, fusion model using device and electronic equipment |
Non-Patent Citations (3)
| Title |
|---|
| ERRINGTON B, ET AL: "Micronised egyptian blue pigment: a novel near-infrared luminescent fingerprint dusting powder", 《DYES PIGM》 * |
| 刘军 等: "近红外光谱无损检测技术中数据的分析方法概述", 《武汉工程大学学报》 * |
| 姜囡: "《语音信号识别技术与实践》", 31 December 2019 * |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114371145A (en) * | 2022-03-21 | 2022-04-19 | 武汉工程大学 | Detection method and device for milk oil mixed pigment, electronic equipment and storage medium |
| CN115205716A (en) * | 2022-08-11 | 2022-10-18 | 北京林业大学 | Method, device and system for estimating oil content of olive fruits and storage medium |
| CN115205716B (en) * | 2022-08-11 | 2023-04-07 | 北京林业大学 | Method, device and system for estimating oil content of olive fruits and storage medium |
| CN116735527A (en) * | 2023-06-09 | 2023-09-12 | 湖北经济学院 | Near infrared spectrum optimization method, device and system and storage medium |
| CN116735527B (en) * | 2023-06-09 | 2024-01-05 | 湖北经济学院 | Near infrared spectrum optimization method, device and system and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114112992B (en) | 2022-04-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN114112992B (en) | Detection method and device for blue pigment of cream and storage medium | |
| Croitoru et al. | Diffusion models in vision: A survey | |
| CN113159051B (en) | Remote sensing image lightweight semantic segmentation method based on edge decoupling | |
| CN103278464B (en) | Flesh of fish detection method and device | |
| CN109145992A (en) | Cooperation generates confrontation network and sky composes united hyperspectral image classification method | |
| CN111127146B (en) | Information recommendation method and system based on convolutional neural network and noise reduction self-encoder | |
| CN106056647B (en) | A kind of magnetic resonance fast imaging method based on the sparse double-deck iterative learning of convolution | |
| CN108398268A (en) | A kind of bearing performance degradation assessment method based on stacking denoising self-encoding encoder and Self-organizing Maps | |
| CN112150568A (en) | Magnetic resonance fingerprint imaging reconstruction method based on Transformer model | |
| CN109389171B (en) | Medical image classification method based on multi-granularity convolution noise reduction automatic encoder technology | |
| CN111652049A (en) | Face image processing model training method and device, electronic equipment and storage medium | |
| Mdrafi et al. | Joint learning of measurement matrix and signal reconstruction via deep learning | |
| CN112529865A (en) | Mixed pixel bilinear deep layer de-mixing method, system, application and storage medium | |
| CN114565594A (en) | Image anomaly detection method based on soft mask contrast loss | |
| CN114049525A (en) | Fusion neural network system, device and method for identifying gas types and concentrations | |
| Zhou et al. | A hybrid denoising model using deep learning and sparse representation with application in bearing weak fault diagnosis | |
| CN108573512B (en) | Complex visual image reconstruction method based on depth coding and decoding dual model | |
| CN113887559A (en) | Brain-computer information fusion classification method and system for brain off-loop application | |
| CN115346091B (en) | Method and device for generating Mura defect image data set | |
| CN113887656B (en) | Hyperspectral image classification method combining deep learning and sparse representation | |
| CN113128459B (en) | Feature fusion method based on multi-level electroencephalogram signal expression | |
| EP4050518A1 (en) | Generation of realistic data for training of artificial neural networks | |
| Quan et al. | Unsupervised deep learning for phase retrieval via teacher-student distillation | |
| CN117095208B (en) | Lightweight scene classification method for photoelectric pod reconnaissance image | |
| Abderrazak et al. | Texture Synthesis Using Improved Transfer Learning |
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 |