An intelligent identification method for Radix Isatidis based on a fusion computing model of ultraviolet spectroscopy, fluorescence spectroscopy and image recognition

Abstract

This invention discloses a Radix Isatidis identification method based on a fusion computational model for ultraviolet and fluorescence spectral image recognition. This method fuses the collected Radix Isatidis UV and fluorescence spectra with images using a transformer model to generate identification results, such as Radix Isatidis grade and origin authenticity. This method simplifies the model structure, improving training speed and identification accuracy.

Description

Technical Field

[0001] The present invention belongs to the technical field of identification methods for the traditional Chinese medicine Radix Isatidis, and specifically relates to an intelligent identification method for Radix Isatidis based on a fusion operation model of ultraviolet spectrum, fluorescence spectrum and image recognition. Background Art

[0002] Current methods for evaluating the quality of traditional Chinese medicine (TCM) face numerous challenges: There is a lack of precise standards for evaluating properties, and evaluation experience is not readily shared or passed down. Qualitative and quantitative analyses based on indicator components lack specificity, making it difficult to truly reflect the intrinsic quality of the medicinal material holistically. Methods that monitor multiple indicators are mostly qualitative and isolated, lacking the quantitative, integrated indicators needed for comprehensive, multi-evaluation approaches to scientifically and comprehensively characterize TCM quality. Property evaluation and chemical evaluation reveal the quality of TCM from different perspectives. Clarifying the objective evaluation indicators of these two models, scientifically correlating these indicators, and identifying the correlation between "property information, digital information, and intrinsic material foundations" to establish an accurate, rapid, and simple comprehensive evaluation system for TCM are key to achieving holistic quality control of TCM and a pressing need for its modernization.

[0003] Isatidis Radix (Bai Lan Radix) is bitter and cold in nature, entering the Heart, Liver, and Stomach meridians. It possesses antibacterial and anti-inflammatory properties, including clearing heat and detoxifying, cooling blood and relieving sore throats. Modern pharmacological research has shown that Isatidis Radix possesses antibacterial and antiviral properties, with anti-influenza virus activity being particularly pronounced. As a heat-clearing and detoxifying herb, Isatidis Radix is effective in preventing respiratory infections such as influenza. Among the 100 most commonly used Chinese medicinal herbs for the prevention and treatment of viral pneumonia, Isatidis Radix ranks 9th, and 10% to 15% of patented Chinese medicine formulas use Isatidis Radix. Isatidis Radix also appears frequently in Traditional Chinese Medicine formulas for the treatment of pneumonia caused by the novel coronavirus. This demonstrates the increasing recognition and widespread clinical application of Isatidis Radix. However, cultivation variations and a variety of adulterated products flooding the market have made it difficult to distinguish the quality of Isatidis Radix, significantly impacting the stability of the finished product and jeopardizing public safety.

[0004] Convolutional neural networks in computer vision are translationally invariant and locally sensitive, lacking holistic perception and macroscopic understanding of images. Expanding the receptive field requires using larger convolution kernels and deeper convolutions. However, this significantly reduces computational efficiency, dramatically increases model complexity, and can even lead to the curse of dimensionality, preventing convergence in training.

[0005] The Transformer model is a simple yet flexible approach. If abstracted as a series of embeddings, it can be applied to any type of data. The self-attention-based Transformer model utilizes only the most primitive Transformer encoder-decoder structure and is capable of learning a global understanding of images. When the model parameters are large and sufficient data is available, the model's performance in tasks such as image classification and object detection can rival or even surpass that of fine-tuning parameters. Furthermore, the model's structure is simpler, training is faster, and it consumes less computer resources.

[0006] Currently, there is no research on the identification method of Isatis indigotica using the transformer model. Summary of the Invention

[0007] Purpose of the invention: In order to solve the above technical problems, the present invention provides a method for identifying Isatis root based on a fusion operation model of ultraviolet spectrum and fluorescence spectrum image recognition. This method can simplify the model structure, improve the training speed and identification accuracy.

[0008] Technical solution: The above invention objectives have not been achieved. The present invention adopts the following technical solution:

[0009] A method for identifying the Chinese medicinal material Radix Isatidis includes fusing a collected ultraviolet spectrum, a fluorescence spectrum, and an image of the Radix Isatidis using a transformer model to obtain an identification result of the Radix Isatidis.

[0010] Preferably, the identification results of Radix Isatidis include identification results such as grade and origin authenticity.

[0011] Preferably, the identification method of the Chinese medicinal material Radix Isatidis comprises the following steps:

[0012] (1) Construct an Isatis Radix image dataset, classify the collected image data, and combine them into an image database that can be used by the algorithm;

[0013] (2) Determine the UV-visible absorption spectrum and fluorescence spectrum of Radix Isatidis, classify the two types of spectral image data collected, and combine them into a UV spectrum database and a fluorescence spectrum database that can be used by the algorithm;

[0014] (3) Using the above image database, ultraviolet spectrum database and fluorescence spectrum database, a transformer model for the classification of Radix Isatidis was constructed to identify Radix Isatidis.

[0015] Further preferably, in step (1), the collected spectral image data is classified and paired non-TCM data is generated simultaneously, such as corresponding categories, labels, picture numbers, etc., and finally combined into an image database that can be used by the algorithm.

[0016] Further preferably, in step (2), the collected spectral image data are classified separately, and paired non-traditional Chinese medicinal material data are generated simultaneously, such as corresponding categories, labels, picture numbers, etc., and finally combined into an ultraviolet spectrum database and a fluorescence spectrum database that can be used by the algorithm.

[0017] Further preferably, in step (2), the preparation method of the Radix Isatidis test solution is as follows: the medicinal material powder is dissolved in methanol to a fixed volume, ultrasonically treated, supplemented to a fixed weight, filtered, and diluted to prepare a test solution.

[0018] Further preferably, in step (2), the measurement conditions of the UV-visible absorption spectrum are: a scanning range of 200 to 350 nm, a scanning interval of 1.0 nm, and a reference solution of 80% methanol solution.

[0019] More preferably, in step (2), the fluorescence spectrum is measured at a scanning speed of 1200 nm·min -1 , the scanning range is 370-600 nm, the excitation wavelength is 340 nm, and the slit width is 15 × 15 nm.

[0020] Further preferably, in step (3), the parameters of the converter model are:

[0021] The block size of the image block is 16×16, the number of heads of the multi-head self-attention layer is 8, the size of the multi-layer perceptron (MLP) is 32, the hidden layer size is 75; the number of layers of the transformer encoder is 2.

[0022] Further preferably, in step (3), the transformer model construction method for Radix Isatidis classification is as follows:

[0023] First, the input image is divided into fixed-size blocks. The size of each block is p×p×3, and a total of N blocks can be divided, where N=HW / P 2 Each p×p×3 block is flattened into a one-dimensional vector. A linear transformation is performed on each vector and the position information is added to form a fully connected layer and a multi-head self-attention layer, forming a transform encoder. The output of the encoder is layer-normalized and then fed to the MLP layer and hidden layer for layer-normalization. The result is then fed to the feedforward layer for layer-normalization. This is repeated a custom number of times to obtain the trained model.

[0024] The present invention comprehensively considers the role of trait characteristics and content characteristics in the quality evaluation of Isatis Root, and creatively proposes an Isatis Root identification method based on a fusion computing model of ultraviolet spectroscopy, fluorescence spectroscopy, and image recognition. It uses a transformer model to grasp the global understanding of the image, simplify the model structure, improve computational efficiency and scalability, establish a multi-dimensional Isatis Root quality evaluation method, and improve the accuracy of the identification results.

[0025] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0026] 1. Ultraviolet-visible absorption spectroscopy, also known as electronic (absorption) spectroscopy, is an absorption spectrum based on intramolecular electron transitions. It is widely used in the identification of traditional Chinese medicines and the quality evaluation of similar medicinal materials. It boasts rapid analysis and simple operation, as well as inexpensive instrumentation and minimal sample requirements. Fluorescence fingerprints can be presented as three-dimensional maps or contour maps, providing information such as excitation spectra, emission spectra, and luminescence intensity. They offer high sensitivity, selectivity, rapidity, good reproducibility, and ease of sampling, making them applicable to numerous testing applications.

[0027] 2. Sensory-based identification methods are easily influenced by subjective factors such as the identifier's environment, preferences, experience, mental state, and health status, resulting in low reliability and difficulty in establishing a unified standard. Radix Isatidis images contain multiple features such as color, shape, and texture, reflecting the overall appearance of the herb. Image recognition can overcome the subjectivity of sensory identification and is highly accurate, effectively replacing the human eye. This facilitates the formation of systematic quality evaluation standards and promotes the standardization of the Radix Isatidis industry.

[0028] 3. Active ingredient content control and property identification are the main methods of quality evaluation. The present invention uses ultraviolet spectroscopy combined with image recognition to evaluate the quality of Isatis root from different dimensions, which is conducive to improving the accuracy of identification and forming a systematic quality evaluation standard.

[0029] 4. The Transformer model is a simple yet flexible approach that can be applied to any type of data. The self-attention-based Transformer model utilizes only the most primitive Transformer encoder-decoder structure and is capable of learning a global understanding of images. When the model parameters are large and sufficient data is available, the model's performance in tasks such as image classification and object detection can rival or even surpass that of fine-tuning parameters. Furthermore, the model's structure is simpler, training is faster, and it consumes less computer resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 Radix Isatidis image dataset.

[0031] Figure 2 Ultraviolet spectrum database, including (a) Anhui, (b) Gansu, (c) Heilongjiang, (d) Shandong, and (e) southern isatis root.

[0032] Figure 3 Fluorescence spectrum database, including (a) Anhui, (b) Gansu, (c) Heilongjiang, (d) Shandong, and (e) southern isatis root. DETAILED DESCRIPTION

[0033] The following is a comprehensive description of the present invention. The implementation cases described are the most preferred implementation methods of the present invention, but the present invention is not limited to the following examples.

[0034] The isatis root slices in the present invention come from Anhui, Gansu, Shandong and Heilongjiang provinces in China, and the southern isatis root comes from Guangxi, China. The isatis root is preferably the dried root of Isatis indigotica Fort., a plant of the Cruciferae family, commonly known as "Northern Isatis"; the southern isatis root is preferably the dried rhizome and root of Baphicacanthus cusia (Nees) Bremek., a plant of the Acanthaceae family.

[0035] Example

[0036] (1) Constructing a Radix Isatidis image dataset

[0037] Isatis root slices were sourced from Anhui, Gansu, Shandong, and Heilongjiang provinces in China, while southern isatis root was sourced from Guangxi, China. The original image data was captured using a macro lens in a light box containing an LED ring light source, covering different shooting angles and light intensities of isatis root. The data for each image was manually labeled, i.e., the collected image data was classified and paired non-TCM data was generated simultaneously, such as corresponding categories, labels, and image numbers. This was ultimately combined into an image database that could be used by the algorithm, such as Figure 1 As shown in the figure, after screening, a total of 1920 images of Radix Isatidis and Radix Adenophorae were obtained, with 384 images of Radix Isatidis from each origin, representing a uniformly distributed dataset. 60% of the samples from each category were randomly selected as the training set for the transformer model, 20% as the test set, and 20% as the validation set.

[0038] (2) Constructing UV spectrum library and spectrum database

[0039] 2.1 Preparation of Isatis Radix Solution

[0040] Radix Isatidis and Radix Adenophorae with acquired images were used as the spectrum acquisition objects. Grind and accurately weigh 1 g of the medicinal material powder (passed through a No. 4 sieve), dissolve it in 80% methanol and make up the volume to a 25 ml volumetric flask. Ultrasonic treatment was performed for 30 minutes, and the weight was made up. The solution was filtered and diluted to 0.5 mg ml -1 of the test sample solution.

[0041] 2.2 Preparation of reference solution

[0042] Take (R, S)-cosidine reference substance and add 80% methanol to prepare a 0.5 mg / ml reference substance solution.

[0043] 2.3 Preparation of control medicinal material solution

[0044] Take 1 g of control medicinal material (identified indigo root medicinal material) powder, add 25 ml of 80% methanol, ultrasonically treat for 30 minutes, make up the weight, filter, and obtain a control medicinal material solution.

[0045] 2.4 Preparation of standard curve for reference medicinal materials

[0046] Accurately pipette 1.0ml, 2.0ml, 3.0ml, 4.0ml, and 5.0ml of the reference solution into 25ml volumetric flasks, then add 80% methanol to the mark. Plot a regression curve of concentration X against absorbance Y, and calculate the regression equation. The regression equation is: y = 0.107x - 0.0869, r = 0.9993. The results show that (R, S)-gaitinib at 1.79μg·ml -1 ~8.96μg·ml -1 The linear relationship is good within the range.

[0047] 2.5 UV and fluorescence determination of Isatis solution

[0048] The reference substance, reference herb, and test sample solutions were placed in 1 cm quartz cuvettes and scanned at the full UV wavelength. All three exhibited maximum absorption at 204 nm. Measurement conditions: a scan range of 200-350 nm, a scan interval of 1.0 nm, and an 80% methanol solution as the reference solution. UV-Vis absorption spectra of Radix Isatidis and Radix Isatidis Australis were measured.

[0049] Take the reference substance, reference medicinal material, and test solution for fluorescence scanning. The measurement conditions are: scanning speed 1200nm·min -1 , the scanning range was 370-600 nm, the excitation wavelength was 340 nm, and the slit width was 15×15 nm, and the fluorescence spectra of Radix Isatidis and Radix Adenophorae were measured.

[0050] The collected UV and fluorescence spectral images are manually labeled, and each sample image is matched with the spectrum one by one. The collected spectral image data are classified, and the matched non-TCM data are generated simultaneously, such as the corresponding category, label, picture number, etc., and finally combined into a spectral database that can be used by the algorithm, such as Figure 2 、 Figure 3 .

[0051] (3) Constructing a transformer model for Isatis indigotica classification

[0052] The images for transformer model training are stored in the data training set folder; the images are divided into 5 categories; the number of images in any category is no less than 100; the images are obtained through the training set file list, test set file list and validation set file list respectively; the training set, test set and validation set file lists each include several rows of records; the content of any row of records is the file name + space character + category number; the file name includes the relative path of the image; the category name is the source of the sample.

[0053] First, the image is divided into 14 blocks, the sliding window moves in steps of 16, and the size of each block is 16×16 pixels. Each block is a color image with three RGB channels. At this time, each block is a tensor. The tensor needs to be converted into a vector, so the 14 small blocks become 14 vectors x1-x 14 , integrated into a one-dimensional sequence. Use the fully connected layer to perform a linear transformation on the vector x, calculate the matrix Wx1+b to get the vector z1, and the fully connected layer shares parameters (W and b are the same). Use the same fully connected layer to perform a linear transformation on the 14 x vectors to get the vector z1-z 14 Then the position of the image block is encoded. The position is an integer between 1 and 14. Each position is encoded as a vector. The vector size is the same as the z vector. The vector obtained by position encoding is added to the z vector to obtain vectors z1~z containing content information and position information. 14 .

[0054] The UV spectrum and fluorescence spectrum use the same feature extraction module, including 5 convolutional layers, 3 pooling layers and 1 fully connected layer. After multiple convolution and pooling operations, the two parallel feature extraction modules map the original data into their respective hidden layer feature spaces.

[0055] At the same time, in order to enhance the effectiveness of Radix Isatidis feature information and reduce the deviation in image understanding, the feature superposition module adds the feature maps output by the first pooling layer of each of the UV spectrum and fluorescence spectrum feature sub-networks to obtain superimposed features. The superimposed features are then input into the convolution-pooling combination operation for composite feature extraction. Similarly, in the second and third pooling layers, the three parallel output feature maps are superimposed again and passed to their fully connected layers in the superimposed feature extraction module. Finally, a multimodal superimposed feature description vector is obtained. In the feature fusion module, vector splicing is used to connect the image and spectral superimposed features, and the spectra are fused into a sequence.

[0056] Next, a one-dimensional convolutional layer is used to obtain the embedding of each sequence. Since the image is divided into 14 blocks, the model uses 14 projection layers. A distance method based on the center loss function is used to supervise the learning of the center sequence embedding layer. After the embedding position is added to the sequence, the sequence is input to the standard transformer encoder. The transformer encoder consists of 2 identical layers, each layer consists of two sublayers, namely a multi-head self-attention layer and a fully connected feedforward network. The number of heads of the multi-head self-attention layer is 8, the size of the multi-layer perceptron (MLP) is 32, and the hidden layer size is 75. Output c0-c 14 Input c0 into the softmax classifier, which outputs a vector p, where p is the classification result. Compare the cross-entropy between vector p and the true category (true label). Use the cross-entropy as the loss function, calculate the gradient of the loss function with respect to the neural network parameters, and use the gradient to update the neural network parameters. Applying parameter sharing and metric learning to this type of model can achieve better performance in terms of convergence rate and classification accuracy. The accuracy of distinguishing Isatis Root from Isatis Root (Southern Isatis) reached 91.31%, and the accuracy of distinguishing Isatis Root from different origins reached 89.58%.

[0057] The converter model parameters are shown in Table 1.

[0058] Table 1 Converter model parameters

[0059]

[0060] Table 2 Model classification accuracy

[0061]

[0062] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A method for identifying the Chinese medicinal material Radix Isatidis, characterized in that: The method includes fusing the collected Radix Isatidis ultraviolet spectrum, fluorescence spectrum and image using a transformer model to obtain an identification result of Radix Isatidis; the method includes the following steps: (1) Construct an Isatis Radix image dataset, classify the collected image data, and combine them into an image database that can be used by the algorithm; (2) Determine the UV-visible absorption spectrum and fluorescence spectrum of Radix Isatidis, classify the two types of spectral image data collected, and combine them into a UV spectrum database and a fluorescence spectrum database that can be used by the algorithm; (3) Using the above-mentioned image database, ultraviolet spectrum database, and fluorescence spectrum database, a transformer model for the classification of Radix Isatidis was constructed to identify Radix Isatidis. The parameters of the transformer model are: The block size of the image block is 16×16, the number of heads of the multi-head self-attention layer is 8, the size of the multi-layer perceptron MLP is 32, and the hidden layer size is 75; the number of layers of the transformer encoder is 2; The converter model is constructed as follows: First, the image is divided into 14 blocks, the sliding window moves in steps of 16, and the size of each block is 16×16 pixels; each block is a color image with three RGB channels. At this time, each block is a tensor. The tensor needs to be converted into a vector, so the 14 small blocks become 14 vectors X1-X 14 , integrated into a one-dimensional sequence; a linear transformation is performed on each vector and then the position information is added to form a fully connected layer and a multi-head self-attention layer to form a transform encoder. The output of the encoder is layer-normalized and then output to the multi-layer perceptron MLP and hidden layer for layer-normalization. The result is output to the feedforward layer for layer-normalization. This is repeated a custom number of times to obtain the trained model.

2. The identification method of the Chinese medicinal material Radix Isatidis according to claim 1, characterized in that: The identification results of Radix Isatidis include the authenticity identification results of grade and origin.

3. The identification method of the Chinese medicinal material Radix Isatidis according to claim 1, characterized in that: In step (1), the collected image data is classified and paired non-TCM data is generated simultaneously, including corresponding categories, labels, and picture numbers, and finally combined into an image database that can be used by the algorithm.

4. The identification method of the Chinese medicinal material Radix Isatidis according to claim 1, characterized in that: In step (2), the collected spectral image data are classified separately, and paired non-traditional Chinese medicinal material data are generated simultaneously, including corresponding categories, labels, and picture numbers, and finally combined into an ultraviolet spectrum database and a fluorescence spectrum database that can be used by the algorithm.

5. The identification method of the Chinese medicinal material Radix Isatidis according to claim 1, characterized in that: In step (2), the preparation method of the Radix Isatidis test solution is as follows: the powder of the medicinal material is dissolved in methanol to a fixed volume, ultrasonically treated, supplemented to a fixed weight, filtered, and diluted to prepare a test solution.

6. The identification method of the Chinese medicinal material Radix Isatidis according to claim 1, characterized in that: In step (2), the measurement conditions of the UV-visible absorption spectrum are: a scanning range of 200-350 nm, a scanning interval of 1.0 nm, and an 80% methanol solution as the reference solution.

7. The identification method of the Chinese medicinal material Radix Isatidis according to claim 1, characterized in that: In step (2), the fluorescence spectrum was measured at a scanning speed of 1200 nm min -1 , the scanning range is 370~600nm, the excitation wavelength is 340nm, and the slit width is 15×15nm.

Images