CN113469027A - Pulse map quality detection method based on deep learning - Google Patents
Pulse map quality detection method based on deep learning Download PDFInfo
- Publication number
- CN113469027A CN113469027A CN202110735237.4A CN202110735237A CN113469027A CN 113469027 A CN113469027 A CN 113469027A CN 202110735237 A CN202110735237 A CN 202110735237A CN 113469027 A CN113469027 A CN 113469027A
- Authority
- CN
- China
- Prior art keywords
- pulse
- pulse map
- map
- layer
- convolution
- 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 52
- 238000013135 deep learning Methods 0.000 title claims abstract description 34
- 238000011176 pooling Methods 0.000 claims abstract description 35
- 238000000034 method Methods 0.000 claims abstract description 20
- 238000007781 pre-processing Methods 0.000 claims abstract description 6
- 238000012545 processing Methods 0.000 claims abstract description 5
- 239000013598 vector Substances 0.000 claims description 38
- 230000009466 transformation Effects 0.000 claims description 32
- 230000004913 activation Effects 0.000 claims description 26
- 230000006870 function Effects 0.000 claims description 20
- 238000005070 sampling Methods 0.000 claims description 6
- 230000002087 whitening effect Effects 0.000 claims description 3
- 230000007423 decrease Effects 0.000 claims description 2
- 238000010586 diagram Methods 0.000 abstract description 27
- 239000000523 sample Substances 0.000 description 16
- 239000003814 drug Substances 0.000 description 9
- 238000012549 training Methods 0.000 description 8
- 230000008569 process Effects 0.000 description 7
- 230000009286 beneficial effect Effects 0.000 description 4
- 230000008901 benefit Effects 0.000 description 4
- 238000011156 evaluation Methods 0.000 description 3
- 238000005259 measurement Methods 0.000 description 2
- 238000013441 quality evaluation Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 230000003213 activating effect Effects 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 238000003745 diagnosis Methods 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 230000001902 propagating effect Effects 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 238000012163 sequencing technique Methods 0.000 description 1
Images
Classifications
-
- 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
- G06N3/048—Activation functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformation in the plane of the image
- G06T3/40—Scaling the whole image or part thereof
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformation in the plane of the image
- G06T3/60—Rotation of a whole image or part thereof
-
- G06T5/70—
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
- G06F2218/10—Feature extraction by analysing the shape of a waveform, e.g. extracting parameters relating to peaks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
- G06F2218/14—Classification; Matching by matching peak patterns
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30004—Biomedical image processing
- G06T2207/30101—Blood vessel; Artery; Vein; Vascular
-
- 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
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
- Y02P90/30—Computing systems specially adapted for manufacturing
Abstract
The invention discloses a pulse map quality detection method based on deep learning, and belongs to the field of image processing. The pulse diagram quality detection method comprises the following steps: determining the median values of all pulse waves of the pulse map sample through the input pulse map sample to determine the wave crest of the pulse map sample and generate a pulse map image; preprocessing the pulse map image by an ImageDataGenerator method to obtain an enhanced pulse map image; inputting the enhanced pulse map image into a convolution network model for identification and classification to obtain a standard quality pulse map; the convolutional network model comprises two convolutional layers, two pooling layers, two active layers and at least one full-connection layer. The pulse map quality detection method based on deep learning provided by the invention can eliminate subjective interference caused by manual detection and objectively evaluate the pulse map quality, so that a high-quality pulse map is identified, the consistency of the pulse map quality can be ensured, and the time for detecting and judging the pulse map quality can be greatly shortened.
Description
Technical Field
The invention relates to the field of image processing, in particular to a pulse map quality detection method based on deep learning.
Background
Due to the interference of factors such as environment and equipment, the quality of the pulse map is difficult to control in the pulse map acquisition process, so that the quality difference of the pulse map acquisition in clinic is large. The quality of the pulse map plays a very important role in the subsequent processes of extracting the characteristics of the pulse map, analyzing the pulse condition and the like, and the analysis of the pulse condition and the result of the pulse condition are seriously influenced when the quality of the pulse map is low.
Moreover, because the present pulse diagram quality detection mainly depends on manual work, and the manual work is interfered by various factors in the process of the judgment, the detection of the pulse diagram quality is time-consuming and labor-consuming, and the pulse diagram quality after the manual detection is also uneven.
Therefore, it is necessary to provide a pulse diagram quality detection method capable of rapidly and objectively extracting a high-quality pulse diagram.
Disclosure of Invention
In order to solve at least one aspect of the above problems and disadvantages in the prior art, the present invention provides a deep learning based pulse map quality detection method. The technical scheme is as follows:
the invention aims to provide a pulse map quality detection method based on deep learning.
According to one aspect of the invention, a pulse map quality detection method based on deep learning is provided, and the pulse map quality detection method comprises the following steps:
step S1: determining the median values of all pulse waves of the pulse map sample through the input pulse map sample to determine the wave crest of the pulse map sample and generate a pulse map image;
step S2: preprocessing the pulse map image by an imagedata generator method to obtain an enhanced pulse map image;
step S3: inputting the enhanced pulse map image into a convolution network model for recognition and classification so as to obtain a standard quality pulse map;
the convolution network model sequentially comprises two convolution layers, two pooling layers, two activation layers and at least one full-connection layer, and convolution kernels arranged in the two convolution layers are different.
Specifically, in step S1, determining the median of all the pulse waves of the pulse map sample from the input pulse map sample to determine the peak of the pulse map sample and generate the pulse map image includes the following steps:
step S11: sampling a pulse map sample using a scipy, find _ peaks _ cwt () function in Python to obtain peaks of all pulses in the pulse map sample;
step S12: sorting wave crests of all pulse waves and obtaining median values of all pulse waves;
step S13: taking wave crests of a plurality of pulse waves forwards and backwards respectively through the median value, wherein the wave crests of the plurality of pulse waves form the pulse map image;
and the peak in the pulse map sample corresponding to the median is the optimal peak.
Preferably, in step S2, preprocessing the pulse map image by the imagedata generator method comprises the steps of:
step S21: and sequentially carrying out rotation transformation, symmetrical transformation, image whitening processing, cross-cut transformation and scaling transformation on the pulse map image.
Specifically, the rotation angle range of the rotation transformation is 20-40 degrees, the symmetry transformation comprises horizontal transformation and vertical transformation, the offset amplitude range of the horizontal transformation is 0.05-0.2, the vertical offset amplitude range of the vertical transformation is 0.05-0.2, the shear strength range of the shear transformation is 0.1-0.3, and the scaling amplitude range of the scaling transformation is 0.1-0.3.
Preferably, in step S3, the step of inputting the enhanced pulse map image into a convolutional network model for identification and classification includes the following steps:
step S31: inputting the enhanced pulse map image into a convolution network model to carry out convolution, pooling, activation, convolution, pooling and activation in sequence so as to obtain a high-dimensional feature vector;
step S32: flattening the high-dimensional feature vector to obtain a one-dimensional feature vector;
step S33: and respectively outputting the predicted maximum probability values of the standard quality pulse map and the low quality pulse map by the one-dimensional characteristic vector through a sigmoid function of a full connection layer.
Preferably, in step S31, the sizes of the convolution kernels set in the first convolution layer and the second convolution layer are both set to 5 × 5 and the step size is set to 2, the number of convolution kernels of the first convolution layer is set to 20, the output vector dimension is 32, the number of second convolution layers is set to 50, and the output vector dimension is 48.
Preferably, in step S31, all pooling layers down-sample the pulse map feature map by the maximum pooling method, the step size of the first pooling layer and the second pooling layer is set to 2, the convolution kernel size of the first pooling layer is set to 2 × 2, the number of convolution kernels is set to 32, the convolution kernel size of the second pooling layer is set to 2 × 2, and the number of convolution kernels is set to 48.
Further, in step S31, the activation functions used in all activation layers are relu functions;
in step S33, the full-connection layer is set to be a four-layer network, the vector dimension of the output of the four-layer network decreases layer by layer, and the vector dimension of the output after passing through the four-layer network is 2.
Specifically, in step S31, the activation functions x of all the activation layersn (t)Expressed as:
wherein t represents the number of convolutional layers, qmnRepresents a convolution kernel connecting the feature n of layer 1 with the feature map M of layer t-1, Mt-1Input feature maps representing the t-1 th layer selection indicate convolution, b represents bias, and f (-) represents a function of nonlinear activation.
Further, in step S13, the pulse map image is a time-series image.
The pulse map quality detection method based on deep learning has at least one of the following advantages:
(1) the pulse map quality detection method based on deep learning provided by the invention can eliminate subjective interference caused by manual detection to objectively evaluate the pulse map quality, so as to identify a high-quality pulse map, which is beneficial to improving the acquisition quality of the traditional Chinese medicine pulse map;
(2) the pulse map quality detection method based on deep learning can ensure the consistency of the pulse map quality;
(3) the pulse map quality detection method based on deep learning can greatly shorten the time for detecting and judging the pulse map quality;
(4) the pulse map quality detection method based on deep learning can convert time-series pulse map data into enhanced image data, then uses a convolution network model to fit the pulse map data, obtains model weight data and finally realizes pulse map quality evaluation, and the convolution network model can directly endow the experience of expert detection and pulse map judgment to the convolution network model;
(5) the pulse map quality detection method based on deep learning provided by the invention utilizes the deep learning technology to objectively detect and evaluate the pulse map quality, has very important significance for detecting and evaluating the quality of the traditional Chinese medicine pulse map, and is beneficial to improving the acquisition quality of the traditional Chinese medicine pulse map.
Drawings
These and/or other aspects and advantages of the present invention will become apparent and readily appreciated from the following description of the preferred embodiments, taken in conjunction with the accompanying drawings of which:
FIG. 1 is a flow diagram of a deep learning based pulse map quality detection method according to one embodiment of the invention;
FIG. 2 shows the pulse quality measurement accuracy and AUC (i.e., model evaluation index).
FIG. 3 is a schematic diagram of the pulse quality detection results (wherein the left pulse diagram is the high-quality pulse diagram obtained by the detection method of the present invention and the detection probability of the model using the present invention, and the right pulse diagram is the low-quality pulse diagram and the corresponding detection probability)
Detailed Description
The technical scheme of the invention is further specifically described by the following embodiments and the accompanying drawings. In the specification, the same or similar reference numerals denote the same or similar components. The following description of the embodiments of the present invention with reference to the accompanying drawings is intended to explain the general inventive concept of the present invention and should not be construed as limiting the invention.
Referring to fig. 1, a flow of a deep learning based pulse map quality detection method according to an embodiment of the present invention is shown. The pulse diagram quality detection method mainly comprises the following steps:
step S1: determining the median values of all pulse waves of the pulse map sample through the input pulse map sample to determine the wave crest of the pulse map sample and generate a pulse map image;
step S2: preprocessing the pulse map image by an imagedata generator method to obtain an enhanced pulse map image;
step S3: and inputting the enhanced pulse map image into a convolution network model for identification and classification so as to obtain a standard quality pulse map.
In one example, the convolutional neural network model includes a convolutional layer, a pooling layer, an activation layer, and four fully-connected layers, which are arranged in sequence. In one example, the fully connected layer may also be configured as 1 layer, 2 layers or 3 layers, and this example is only an illustrative example, and those skilled in the art should not be construed as a limitation to the present invention.
In one example, the size of the convolution kernels in both convolutional layers is set to 5 × 5 and the step size is set to 2, the number of convolution kernels for the first convolutional layer is set to 20, the output vector dimension is 32, the number of second convolutional layers is set to 50, and the output vector dimension is 48.
In one example, the pulse map feature map is downsampled by maximum pooling for both pooling layers, the step size for the first pooling layer and the second pooling layer are set to 2, the convolution kernel size for the first pooling layer is set to 2 × 2, the number of convolution kernels is set to 32, the convolution kernel size for the second pooling layer is set to 2 × 2, and the number of convolution kernels is set to 48.
In one example, determining the pulse wave peaks of all pulse map samples from the input pulse map samples and generating the pulse map image comprises the following steps:
step S11: sampling the pulse wave by adopting a pulse diagnosis probe, wherein the sampling frequency is 200Hz, the error is 0.02s and the allowable error is within 4 sampling points, and then obtaining the wave crests of all the highest points of a single sample by using a scipy, find _ peaks _ cwt () function in Python, thereby obtaining the pulse wave crests in each pulse map sample;
step S12: sequencing the pulse wave crests of all samples from large to small or from small to large, and taking the median of all the wave crests, thereby obtaining the median of all the pulse wave crests, wherein the wave crest corresponding to the median is the optimal wave crest;
step S13: the pulse wave peaks are divided into complete pulse map bands by taking 99 pulse wave peaks forward and 150 pulse wave peaks backward respectively through the median, so that the taken pulse wave peaks form main wave position data of all pulse map samples, the main wave position data are time sequence data, and then the main wave position data are converted into a time sequence pulse map image in a picture format, for example, the time sequence pulse map image is converted into a time-peak pulse map image.
After obtaining the pulse map image, the pulse map image is preprocessed by the imagedata generator method, which comprises the following steps:
step S21: the pulse image is sequentially subjected to rotation transformation, symmetrical transformation, image whitening processing, cross-cut transformation and scaling transformation, so that the number of pictures can be increased, the interference of partial noise on the pulse image can be reduced, and the quality detection and identification capability of the pulse image can be further improved. In one example, the rotation is transformed into a horizontal rotation which is randomly performed, the angle of the picture randomly rotates by 30 ° when the data is promoted, although those skilled in the art can set the random rotation angle to be 20 °, 40 °, this example is only an illustrative example, and those skilled in the art should not be construed as a limitation to the present invention. In one example, during the symmetric transformation data lifting process, the offset amplitude during the picture horizontal transformation is, for example, 0.1, but those skilled in the art can adjust the offset amplitude to, for example, 0.05, 0.15, 0.2 as needed. The amplitude of the offset is, for example, 0.1 when the picture is vertically transformed, but those skilled in the art can adjust the offset to, for example, 0.05, 0.15, 0.2 as needed. The shear strength (the angle of the shear transformation in the counterclockwise direction) at the time of the shear transformation is, for example, 0.2, but those skilled in the art can adjust the shear strength to, for example, 0.1 or 0.3 as needed. The magnitude of the random scaling in the scaling transform is, for example, 0.2, although the skilled person can adjust it to, for example, 0.1, 0.3, if necessary. It should be understood by those skilled in the art that the horizontal and vertical offset magnitudes in the present example are only examples, those skilled in the art can set the horizontal offset magnitude and the vertical offset magnitude to be the same or different according to the requirement, and the miscut strength and the scaling magnitude are only examples, and those skilled in the art should not understand the illustration in the present example as a limitation to the present example.
After being preprocessed, an enhanced pulse map image is obtained, and then the pulse map image is input into a convolution network model to be identified and classified, so that a pulse map with standard quality (namely a high-quality pulse map) is obtained. The high-quality and low-quality pulse diagrams are the pulse diagrams classified by the experts of traditional Chinese medicine according to the following standards, wherein the high-quality pulse diagram refers to the pulse diagram which is a single peak, a double peak or a triple peak; the baseline of the low-quality pulse map is not stable (as shown in fig. 2).
In one example, the convolutional network model is built based on a Let deep neural network model, and the distinguishing experience and standard of a traditional Chinese medicine expert on a pulse diagram are obtained through training and learning. In one example, the model is trained by dividing the generated pulse map image into two data sets, the first data set is a training data set, the second data set is a testing data set, and the number of pulse map images in the two data sets is empirically divided by a ratio of 8: 2. Because the pulse map has high individuation characteristics, the traditional measurement indexes are complicated, and the establishment of a standard image library through a reference value range is very inefficient. Therefore, according to the common characteristics and the standard, the traditional Chinese medicine experts interpret and label the pulse map images in the data set with data, namely, the pulse map images are labeled as a high-quality pulse map and a low-quality pulse map. Expert expertise is directly given to the convolutional network model through repeated iteration in training.
After obtaining the enhanced pulse map image, inputting the enhanced pulse map image into a convolution network model, and performing the following operations:
step S31: inputting the enhanced pulse map image into a convolution network model to be sequentially subjected to convolution, pooling, activation, convolution, pooling and activation to obtain a high-dimensional feature vector; for example, after the first convolution, a feature vector with dimension (128, 128, 32) is obtained, then the feature vector is input into the pooling layer for pooling, the feature vector with dimension (64, 64, 32) is output, then the feature vector with dimension (63, 63, 48) is activated by the activation function and input into the second layer convolution layer, then the feature vector with dimension (31,31,48) is output, and then the feature vector with dimension (31,31,48) is input into the pooling layer.
Activation function x of two activation layers in step 31n (t)Are all expressed as:
in the formula (1), t represents the number of the convolutional layers, qmnRepresents a convolution kernel connecting the feature n of layer 1 with the feature map M of layer t-1, Mt-1Input feature maps representing the t-1 th layer selection indicate convolution, b represents bias, and f (-) represents a function of nonlinear activation.
In one example, to avoid the gradient attenuation or even disappearance with the increase of the number of layers in the model, and thus the convergence speed of the training becomes slower and slower, the activation function used in the activation layer in step S31 is a relu function, and the mathematical expression thereof is:
f(x)＝max(0，x) (2)
equation (2) indicates that if the input data is greater than 0, the input data is output as the output data, otherwise 0 is output.
In one example, the pooling layer performs down-sampling by maximum pooling, that is, samples of the previous convolution are sampled, the layer performs non-overlapping segmentation on the convolution image according to a certain proportion, and maximizes the rectangular segmented according to a certain proportion, so that the maximum advantage of the pooling operation is that the number of feature maps is not changed, but the dimensionality of the features is reduced.
Step S32: after obtaining the high latitude characteristic vector, the high-dimensional characteristic vector is flattened to obtain a one-dimensional characteristic vector; for example, the feature vector with dimension (31,31,48) is flattened to obtain a one-dimensional feature vector, and the dimension size of the one-dimensional feature vector becomes 46128.
Step S33: and (3) sequentially passing the one-dimensional feature vectors through four fully-connected layers, outputting feature vectors with the dimension of 2 at the last layer, then performing secondary classification by activating a sigmoid function, and outputting the predicted maximum probability values of the standard quality pulse diagram and the low quality pulse diagram. For example, outputting a feature vector with dimension size of 256 after passing through a feature vector with dimension size of 46128 after flattening through a first layer full connection layer, then outputting a feature vector with dimension size of 84 through a second layer full connection layer, then outputting a feature vector with dimension size of 10 through a third layer full connection layer, and finally outputting a feature vector with dimension size of 2 through a fourth layer full connection layer, and performing secondary classification through sigmoid function activation. And establishing connection relations among the nodes in the convolution layer and the pooling layer and all the adjacent nodes through the multi-layer full connection layer.
After the full connection of the above processes, the accuracy (accuracy rate) evaluation layer and the AUC (model evaluation index) layer are used to display the accuracy rate of the training process and the stability of the AUC model (as shown in fig. 2). Where the AUC (i.e., model assessment index) value is 1, which performs best, and 0.5, which is a random chance choice, ranked as follows: 0.5-0.6 means extremely poor precision, 0.6-0.7 means poor precision, 0.7-0.8 means general precision, 0.8-0.9 means good precision), and 0.9-1.0 means excellent precision. Where AUC was 0.967, accuracy was 0.918.
In order to verify the performance of the pulse diagram quality detection model, the pulse diagram picture is put into a set network for training, the input pulse diagram picture is firstly transmitted to an output layer in the forward direction, then output is obtained, and the error between an output value and a true value is calculated. And (4) propagating the calculated errors backwards layer by layer in a back propagation mode according to the principle of minimum error, and automatically updating the weight among the networks in the back propagation process. And repeating iteration to finally obtain an identifier capable of rapidly classifying the pulse condition quality pictures, determining the optimal parameters of model training by calculating the AUC and the accuracy of the used pulse condition pictures, and finally obtaining a picture quality detection model with higher accuracy and without overfitting. Where the AUC using the methods and models of the invention was 0.967, the accuracy was 0.918. The AUC during training and the accuracy of the test results with increasing number of samples are shown in fig. 2.
As shown in fig. 3, the left pulse diagram is a high-quality pulse diagram obtained by the method of the present invention, and the detection probability thereof is 95% or more, and the right pulse diagram is a low-quality pulse diagram obtained without the method of the present invention, and the detection probability thereof is 70% to 90%. It can be seen that the pulse pattern obtained by the method of the present invention is of good and stable quality, whereas the pulse pattern obtained by the conventional method is of very unstable and uneven quality.
The pulse map quality detection method based on deep learning has at least one of the following advantages:
(1) the pulse map quality detection method based on deep learning provided by the invention can eliminate subjective interference caused by manual detection to objectively evaluate the pulse map quality, so as to identify a high-quality pulse map, which is beneficial to improving the acquisition quality of the traditional Chinese medicine pulse map;
(2) the pulse map quality detection method based on deep learning can ensure the consistency of the pulse map quality;
(3) the pulse map quality detection method based on deep learning can greatly shorten the time for detecting and judging the pulse map quality;
(4) the pulse map quality detection method based on deep learning can convert time-series pulse data into polynomial-added image data, then a convolution network model is used for fitting the pulse map data to obtain model weight data, and finally pulse map quality evaluation is realized;
(5) the pulse map quality detection method based on deep learning provided by the invention utilizes the deep learning technology to objectively detect and evaluate the pulse map quality, has very important significance for detecting and evaluating the quality of the traditional Chinese medicine pulse map, and is beneficial to improving the acquisition quality of the traditional Chinese medicine pulse map.
Although a few embodiments of the present general inventive concept have been shown and described, it will be appreciated by those skilled in the art that changes may be made in these embodiments without departing from the principles and spirit of the general inventive concept, the scope of which is defined in the appended claims and their equivalents.
Claims (10)
1. A pulse map quality detection method based on deep learning comprises the following steps:
step S1: determining the median values of all pulse waves of the pulse map sample through the input pulse map sample to determine the wave crest of the pulse map sample and generate a pulse map image;
step S2: preprocessing the pulse map time-frequency image through an ImageDataGenerator method to obtain an enhanced pulse map image;
step S3: inputting the enhanced pulse map image into a convolution network model for recognition and classification so as to obtain a standard quality pulse map;
the convolution network model sequentially comprises two convolution layers, two pooling layers, two activation layers and at least one full-connection layer, and convolution kernels arranged in the two convolution layers are different.
2. The deep learning based pulse map quality detection method according to claim 1,
in step S1, determining the median of all the pulse waves of the pulse map sample from the input pulse map sample to determine the peak of the pulse map sample and generate the pulse map image includes the following steps:
step S11: sampling a pulse map sample using a scipy, find _ peaks _ cwt () function in Python to obtain peaks of all pulses in the pulse map sample;
step S12: sorting wave crests of all pulse waves and obtaining median values of all pulse waves;
step S13: taking wave crests of a plurality of pulse waves forwards and backwards respectively through the median value, wherein the wave crests of the plurality of pulse waves form the pulse map image;
and the peak in the pulse map sample corresponding to the median is the optimal peak.
3. The deep learning based pulse map quality detection method according to claim 2,
in step S2, preprocessing the pulse map image by the imagedata generator method includes the steps of:
step S21: and sequentially carrying out rotation transformation, symmetrical transformation, image whitening processing, cross-cut transformation and scaling transformation on the pulse map image.
4. The deep learning based pulse map quality detection method according to claim 3,
the rotation angle range of the rotation transformation is 20-40 degrees, the symmetry transformation comprises horizontal transformation and vertical transformation, the offset range of the horizontal transformation is 0.05-0.2, the vertical offset range of the vertical transformation is 0.05-0.2, the shear strength range of the shear transformation is 0.1-0.3, and the scaling range of the scaling transformation is 0.1-0.3.
5. The deep learning based pulse map quality detection method according to any one of claims 1-4,
in step S3, the step of inputting the enhanced pulse map image into a convolutional network model for identification and classification includes the following steps:
step S31: inputting the enhanced pulse map image into a convolution network model to carry out convolution, pooling, activation, convolution, pooling and activation in sequence so as to obtain a high-dimensional feature vector;
step S32: flattening the high-dimensional feature vector to obtain a one-dimensional feature vector;
step S33: and respectively outputting the predicted maximum probability values of the standard quality pulse map and the low quality pulse map by the one-dimensional characteristic vector through a sigmoid function of a full connection layer.
6. The deep learning based pulse map quality detection method according to claim 5,
in step S31, the sizes of the convolution kernels set in the first convolution layer and the second convolution layer are each set to 5 × 5 with the step size set to 2, the number of convolution kernels of the first convolution layer is set to 20, the output vector dimension is 32, the number of second convolution layers is set to 50, and the output vector dimension is 48.
7. The deep learning based pulse map quality detection method according to claim 6,
in step S31, the pulse map feature map is downsampled by the maximum pooling method for all pooling layers, the step size of each of the first pooling layer and the second pooling layer is set to 2, the convolution kernel size of the first pooling layer is set to 2 × 2, the number of convolution kernels is set to 32, the convolution kernel size of the second pooling layer is set to 2 × 2, and the number of convolution kernels is set to 48.
8. The deep learning based pulse map quality detection method according to claim 5,
in step S31, all the activation functions used in the activation layer are relu functions;
in step S33, the full-connection layer is set to be a four-layer network, the vector dimension of the output of the four-layer network decreases layer by layer, and the vector dimension of the output after passing through the four-layer network is 2.
9. The deep learning-based pulse map quality detection method according to claim 5, wherein in step S31, all activation functions x of the activation layersn (t)Expressed as:
wherein t represents the number of convolutional layers, qmnRepresents a convolution kernel connecting the feature n of the l-th layer with the feature map M of the t-1 th layer, Mt-1Input feature maps representing the t-1 th layer selection indicate convolution, b represents bias, and f (-) represents a function of nonlinear activation.
10. The deep learning based pulse map quality detection method according to claim 2,
in step S13, the pulse map image is a time-series image.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110735237.4A CN113469027B (en) | 2021-06-30 | 2021-06-30 | Pulse map quality detection method based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110735237.4A CN113469027B (en) | 2021-06-30 | 2021-06-30 | Pulse map quality detection method based on deep learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113469027A true CN113469027A (en) | 2021-10-01 |
| CN113469027B CN113469027B (en) | 2023-12-08 |
Family
ID=77874398
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110735237.4A Active CN113469027B (en) | 2021-06-30 | 2021-06-30 | Pulse map quality detection method based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113469027B (en) |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109034034A (en) * | 2018-07-12 | 2018-12-18 | 广州麦仑信息科技有限公司 | A kind of vein identification method based on nitrification enhancement optimization convolutional neural networks |
| CN111649779A (en) * | 2020-05-31 | 2020-09-11 | 天津大学 | Oil well oil content and total flow rate measuring method based on dense neural network and application |
| CN112560629A (en) * | 2020-12-07 | 2021-03-26 | 杭州电子科技大学 | Smart phone key detection method based on 1D-CNN |
-
2021
- 2021-06-30 CN CN202110735237.4A patent/CN113469027B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109034034A (en) * | 2018-07-12 | 2018-12-18 | 广州麦仑信息科技有限公司 | A kind of vein identification method based on nitrification enhancement optimization convolutional neural networks |
| CN111649779A (en) * | 2020-05-31 | 2020-09-11 | 天津大学 | Oil well oil content and total flow rate measuring method based on dense neural network and application |
| CN112560629A (en) * | 2020-12-07 | 2021-03-26 | 杭州电子科技大学 | Smart phone key detection method based on 1D-CNN |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113469027B (en) | 2023-12-08 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109493342B (en) | Skin disease picture lesion type classification method based on deep learning | |
| CN110040594B (en) | Convolutional neural network-based elevator operation detection system and method | |
| CN112304614A (en) | End-to-end rolling bearing intelligent fault diagnosis method adopting multi-attention machine system | |
| CN113052113A (en) | Depression identification method and system based on compact convolutional neural network | |
| CN108171119B (en) | SAR image change detection method based on residual error network | |
| CN112990082B (en) | Detection and identification method of underwater sound pulse signal | |
| CN110866490A (en) | Face detection method and device based on multitask learning | |
| CN111351860A (en) | Wood internal defect detection method based on Faster R-CNN | |
| CN112043260A (en) | Electrocardiogram classification method based on local mode transformation | |
| CN113283524A (en) | Anti-attack based deep neural network approximate model analysis method | |
| CN110458071B (en) | DWT-DFPA-GBDT-based optical fiber vibration signal feature extraction and classification method | |
| CN111310719A (en) | Unknown radiation source individual identification and detection method | |
| CN110288026A (en) | A kind of image partition method and device practised based on metric relation graphics | |
| CN113469027A (en) | Pulse map quality detection method based on deep learning | |
| CN105513058B (en) | A kind of brain active region detection method and device | |
| CN111289251A (en) | Rolling bearing fine-grained fault identification method | |
| CN116010798A (en) | YOLOv 5-based rapid identification and classification method for complex typical communication signals | |
| Kulevome et al. | Rolling bearing fault diagnostics based on improved data augmentation and ConvNet | |
| CN112861670B (en) | Transmission line hardware detection method and system | |
| CN110223295B (en) | Significance prediction method and device based on deep neural network color perception | |
| CN114970601A (en) | Power equipment partial discharge type identification method, equipment and storage medium | |
| Burlina et al. | Unsupervised deep novelty detection: Application to muscle ultrasound and myositis screening | |
| CN113326892A (en) | Relation network-based few-sample image classification method | |
| AlQaralleh et al. | Fusion Based Tongue Color Image Analysis Model for Biomedical Applications. | |
| Chen et al. | Radar Intra-pulse Modulation Signal Classification Using CNN Embedding and Relation Network under Small Sample Set |
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 |