PDRN quality evaluation classification system based on multi-branch convolutional neural network
Through the PDRN quality detection method based on multi-branch convolutional neural network, the problems of PDRN quantitative, qualitative inaccuracy and poor reproducibility in the prior art are solved, and the accurate measurement and identification of PDRNs within the range of 20bp-3000bp is achieved, which improves the Accuracy and stability of detection.
Patent Information
- Application Number
- CN202411104551.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-13
- Publication Date
- 2024-11-26
AI Technical Summary
The prior art cannot accurately measure the content and molecular weight of PDRN in the range of 20 bp-3000 bp, and the traditional methods have problems such as quantitative and qualitative inaccurate, large influence of environmental factors, and poor reproducibility.
Using the PDRN quality detection method based on multi-branch convolutional neural network, a multi-branch convolutional neural network system is constructed by collecting Raman spectral images, extracting spatial spectral features, and using adaptive weights and multi-scale and multi-context attention features to classify and recognition overcomes the limitations of fixed input windows of traditional neural networks.
Accurate quality evaluation of PDRN is achieved, quantitative and qualitative accuracy is improved, the influence of environmental factors is reduced, the reproducibility of detection is enhanced, and the PDRN measured each time has the same physical and chemical properties and pharmacological activities.
Smart Images

Figure CN119027387A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of image recognition, and in particular to a PDRN Raman spectrum image processing method and system based on a multi-branch convolutional neural network. Background Art
[0002] DNA is an important biological macromolecule, widely present in animals, plants and microorganisms. As a carrier of genetic information, it plays an irreplaceable role in maintaining the normal growth and development of the body. PDRN polydeoxyribonucleotide, also known as polydeoxyribonucleotide, mainly refers to a mixture of DNA with a length of 20bp-3000bp obtained after DNA is treated with acid, alkali and enzyme. PDRN from different sources has different pharmacological activities. Studies have confirmed that PDRN from salmon semen and testis has significant effects in anti-inflammatory, repair and anti-aging. Single-stranded PDRN from pig intestine has the effect of inhibiting fibrinolytic enzymes to prevent thrombosis. And the study revealed that PDRN from the same source, different molecular weight or different sources has different physiological activities. At present, the main methods for detecting DNA include ultraviolet spectrophotometry, diphenylamine method, and agarose gel electrophoresis. For single-component deoxyribonucleotides, the above methods can accurately determine the content and molecular weight, but for PDRN with a relatively large molecular weight span of 20bp-3000bp, the above methods cannot accurately measure the content and molecular weight of the product. How to accurately and quickly measure the content and molecular weight of PDRN to ensure that the PDRN obtained each time has the same physicochemical properties and pharmacological activity?
[0003] When light passes through biological macromolecules such as nucleic acids and proteins, Raman scattering occurs, forming a Raman spectrum. This spectrum, as a fingerprint of biological macromolecules, can accurately reflect the physical and chemical characteristics of the substance being tested, and can be used to identify and characterize the substance being tested. This Raman spectrum has the characteristics of high sensitivity, strong specificity, simple sample preparation, no interference from aqueous solutions, fast detection speed, and low cost. Although Raman spectroscopy is a powerful technology, the processing of raw spectral data is very complicated. Traditional experiments, calculation methods, and qualitative evaluation methods require a long processing time and can easily lead to erroneous analysis results. Convolutional neural networks (CNNs) consist of convolutional layers, fully connected layers, and pooling layers. They have the advantages of multi-level feature extraction capabilities, data-driven properties, and end-to-end input and output, which makes them a powerful tool for biospectral analysis and biospectral imaging. Summary of the invention
[0004] The object of the present invention is to provide a PDRN quality detection method and system based on a multi-branch convolutional neural network, which solves the shortcomings of the existing nucleic acid quality evaluation methods, such as inaccurate quantitative and qualitative analysis, large environmental factors and poor reproducibility. At the same time, by introducing a multi-branch convolutional neural network, the CNN adaptive position and size input window is improved, which effectively overcomes the limitation of the fixed input window of the traditional neural network.
[0005] The technical solution adopted by the present invention is:
[0006] A PDRN quality evaluation classification system based on a multi-branch convolutional neural network comprises the following steps:
[0007] The following steps are involved:
[0008] S1. Collect Raman spectra of PDRN aqueous solutions with different concentrations of 1mg / ml-20mg / ml under standard conditions
[0009] S2. Further, the obtained images are analyzed to construct a PDRN molecular property prediction system based on a multi-branch convolutional neural network. Different branches are used to extract the spatial spectral features of the Raman spectrum, emphasizing the more discriminative branches and suppressing the branches with low discriminative rates. Branches are conducive to classification. Different branches use voting strategies to provide the prediction probability of the most likely category.
[0010] S3. Further, different weights are assigned to different branches according to their discriminative capabilities. The adaptive weights generated by each branch are multiplied by the extracted features; these weighted features are connected to generate multi-scale and multi-context attention features for each branch. For the i-th sample, the weighted feature Mi is formulated as follows:
[0011]
[0012] represents the weight of the qth branch, Represents the multi-layer spatial-spectral features extracted by the q-th branch.
[0013] S4. Further, provide the complementary information of shallow, middle and deep semantic features for the i-th sample to learn. Formulated as follows:
[0014]
[0015] represents the shallow, medium and deep spectral features extracted by the convolutional layer in the qth branch. Ls(·), Lm(·), Ld(·) represent the spectral features extracted from the shallow, medium and deep layers, P(·) represents the global average pooling operation, concat(·) represents the cascade operation, Represents the multi-layer spatial-spectral features of the qth branch of the i-th sample.
[0016] S5. Further, the weight of the qth branch of the i-th sample is provided, and the calculation formula is as follows:
[0017]
[0018] is the predicted probability of class li in the qth branch.
[0019] S6. Further, in R d×1 In the feature space, there are M Raman spectra X = {x1, x2, ..., x M}, the corresponding label is Y = {y1, y2…y M Given a sample xi (1≤i≤M), Qi represents the data set of the neighborhood spectrum of the non-local candidate region centered on xi. For all labeled samples, a new data set Q = {Q1, Q2…Q M} can be divided into and a+b=M.
[0020] S7. Further, in the non-local candidate region Qi, m spatial windows need to be searched. represents the Qth spatial window centered at xi. For the qth spatial window, its size and position are encoded as a vector
[0021] ,in and, are the horizontal and vertical coordinates of the center sample in the spatial window, is the size of the spatial window. In the multi-branch convolutional neural network, the convolutional layer and the corresponding pooling layer are stacked to extract the hierarchical features of the non-local candidate region Qi. Then, these features are mapped to a one-dimensional vector, which is used as the input of the next fully connected layer. Finally, after the fully connected layer, we obtain Formulated as follows:
[0022]
[0023] Where f is a composite function consisting of multiple convolution and pooling operations, and W and b are the corresponding weights and biases. FC and b FC are the weights and biases of the fully connected layer, and g[·] represents its activation function.
[0024] S8. Further, a softmax layer is constructed as the main classifier to perform comprehensive classification using the data extracted from the multi-scale and multi-context attention network layers and the multi-layer spatial spectrum features of the convolutional neural network. During the training process, the loss function of the multi-branch convolutional neural network is defined as:
[0025]
[0026] in Represents the main classifier for the i-th training sample The first term in the above formula is the loss of the main classifier, which is calculated by the cross entropy of the real class probability and the output class probability; the second term is the loss of the m branch classifiers calculated in the same way.
[0027] S9. Further, in the model testing phase, the i-th test sample is predicted by calculating the maximum probability of different classes. The label is calculated as:
[0028]
[0029] Where c = 1, 2, 3, ..., C, represents the predicted probability of the cth class from the main classifier.
[0030] A PDRN quality evaluation and classification system based on a multi-branch convolutional neural network, the PDRN quality evaluation and classification system based on a multi-branch convolutional neural network comprising: a data acquisition module, used to collect the Raman spectrum of the PDRN; a feature extraction module, used to input the Raman spectrum of the PDRN to be identified into a pre-trained multi-branch convolutional neural network, based on a multi-branch attention network layer and a non-local adaptive spatial segmentation strategy, to obtain the physicochemical characteristic information of PDRN with different molecular weights and different concentrations, input the PDRN Raman spectrum into a basic, convolutional block to obtain basic feature information, and then input three branches of the TBP block, namely, a bottleneck structure branch, an inverse bottleneck structure branch, and a non-convolutional branch, to obtain all feature information extracted by the three branches, and merge all feature information into multi-scale feature information;
[0031] The sample identification module is used to determine the molecular weight and content of the PDRN sample in a pre-set multi-branch convolutional neural network database according to the PDRN Raman spectrum characteristic information, so as to finally determine the physical and chemical characteristics of the measured PDRN.
[0032] On the other hand, the present invention also provides a storage medium, including one or more programs stored in the storage medium, which, when executed, implements the PDRN quality detection method based on a multi-branch convolutional neural network as described above.
[0033] Another aspect of the present invention provides a computer device, the computer device comprising a memory and a processor.
[0034] Wherein: the memory is used to store computer programs;
[0035] When the processor is used to execute the computer program stored in the memory, it implements the PDRN quality detection method based on the multi-branch convolutional neural network as described above.
[0036] The beneficial effects of adopting the above technical solution are:
[0037] The purpose of the present invention is to provide a PDRN quality evaluation classification system based on a multi-branch convolutional neural network, which integrates non-local adaptive spatial window search, multi-scale and multi-context attention feature extraction and classification into a unified framework. The higher the discriminative ability of the branch, the greater the weight of the branch, which ensures that the branch with stronger discriminative ability contributes more to the classification.
[0038] The shortcomings of existing nucleic acid quality assessment methods, such as inaccurate quantitative and qualitative analysis, large environmental influences, and poor reproducibility, are solved. At the same time, by introducing a multi-branch convolutional neural network and improving the CNN adaptive position and size input window, the limitations of the traditional neural network fixed input window are effectively overcome. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] Figure 1 A schematic diagram of the structure of a PDRN quality detection method and system based on a multi-branch convolutional neural network proposed in an embodiment of the present invention.
[0040] Figure 2 This is a confusion matrix diagram of the prediction results of the multi-branch convolutional neural network model for the test set data according to an embodiment of the present invention. DETAILED DESCRIPTION
[0041] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.
[0042] In order to facilitate the understanding of the present invention, the present invention is described more thoroughly and comprehensively below. Figure 1 A schematic diagram of a specific embodiment of the present invention is given in FIG. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein.
[0043] A PDRN quality evaluation classification system based on a multi-branch convolutional neural network comprises the following steps:
[0044] The following steps are involved:
[0045] S1 collects Raman spectra of PDRN aqueous solutions with different concentrations of 1mg / ml-20mg / ml under standard conditions
[0046] S2 further analyzes the obtained images and constructs a PDRN molecular property prediction system based on a multi-branch convolutional neural network. Different branches are used to extract the spatial spectral features of the Raman spectrum, emphasizing the more discriminative branches and suppressing the branches with low discriminative rates. Branches are conducive to classification. Different branches use voting strategies to provide the prediction probability of the most likely category.
[0047] S3 further assigns different weights to different branches according to their discriminative capabilities. The adaptive weights generated by each branch are multiplied by the extracted features; these weighted features are connected to generate multi-scale and multi-context attention features for each branch. For the i-th sample, the weighted feature Mi is formulated as follows:
[0048]
[0049] represents the weight of the qth branch, Represents the multi-layer spatial-spectral features extracted by the q-th branch.
[0050] S4 further provides complementary information for learning shallow, middle and deep semantic features of the i-th sample. It is formulated as follows:
[0051]
[0052] represents the shallow, medium and deep spectral features extracted by the convolutional layer in the qth branch. Ls(·), Lm(·), Ld(·) represent the spectral features extracted from the shallow, medium and deep layers, P(·) represents the global average pooling operation, concat(·) represents the cascade operation, Represents the multi-layer spatial-spectral features of the qth branch of the i-th sample.
[0053] S5 further provides the weight of the qth branch of the i-th sample, and the calculation formula is as follows:
[0054]
[0055] is the predicted probability of class li in the qth branch.
[0056] S6 further, in R d×1 In the feature space, there are M Raman spectra X = {x1, x2, ..., x M}, the corresponding label is Y = {y1, y2…yM Given a sample xi (1≤i≤M), Qi represents the data set of the neighborhood spectrum of the non-local candidate region centered on xi. For all labeled samples, a new data set Q = {Q1, Q2…Q M} can be divided into and
[0057] S7: Further, in the non-local candidate region Qi, m spatial windows need to be searched. represents the Qth spatial window centered at xi. For the qth spatial window, its size and position are encoded as a vector
[0058] ,in and, are the horizontal and vertical coordinates of the center sample in the spatial window, is the size of the spatial window. In the multi-branch convolutional neural network, the convolutional layer and the corresponding pooling layer are stacked to extract the hierarchical features of the non-local candidate region Qi. Then, these features are mapped to a one-dimensional vector, which is used as the input of the next fully connected layer. Finally, after the fully connected layer, we obtain Formulated as follows:
[0059]
[0060] Where f is a composite function consisting of multiple convolution and pooling operations, and W and b are the corresponding weights and biases. FC and b FC are the weights and biases of the fully connected layer, and g[·] represents its activation function.
[0061] S8 further constructs a softmax layer as the main classifier to perform comprehensive classification using the data extracted from the multi-scale and multi-context attention network layers and the multi-layer spatial spectrum features of the convolutional neural network. During the training process, the loss function of the multi-branch convolutional neural network is defined as:
[0062]
[0063] in Represents the main classifier for the i-th training sample The first term in the above formula is the loss of the main classifier, which is calculated by the cross entropy of the real class probability and the output class probability; the second term is the loss of the m branch classifiers calculated in the same way.
[0064] S9 Further, in the model testing phase, the i-th test sample is predicted by calculating the maximum probability of different classes. The label is calculated as:
[0065]
[0066] Where c = 1, 2, 3, ..., C, represents the predicted probability of the cth class from the main classifier.
[0067] A PDRN quality evaluation and classification system based on a multi-branch convolutional neural network, the PDRN quality evaluation and classification system based on a multi-branch convolutional neural network comprising: a data acquisition module, used to collect the Raman spectrum of the PDRN; a feature extraction module, used to input the Raman spectrum of the PDRN to be identified into a pre-trained multi-branch convolutional neural network, based on a multi-branch attention network layer and a non-local adaptive spatial segmentation strategy, to obtain the physicochemical characteristic information of PDRN with different molecular weights and different concentrations, input the PDRN Raman spectrum into a basic, convolutional block to obtain basic feature information, and then input three branches of the TBP block, namely, a bottleneck structure branch, an inverse bottleneck structure branch, and a non-convolutional branch, to obtain all feature information extracted by the three branches, and merge all feature information into multi-scale feature information;
[0068] The sample identification module is used to determine the molecular weight and content of the PDRN sample in a pre-set multi-branch convolutional neural network database according to the PDRN Raman spectrum characteristic information, so as to finally determine the physical and chemical characteristics of the measured PDRN.
[0069] On the other hand, the present invention also provides a storage medium, including one or more programs stored in the storage medium, which, when executed, implements the PDRN quality detection method based on a multi-branch convolutional neural network as described above.
[0070] Another aspect of the present invention provides a computer device, the computer device comprising a memory and a processor.
[0071] Wherein: the memory is used to store computer programs;
[0072] When the processor is used to execute the computer program stored in the memory, it implements the PDRN quality detection method based on the multi-branch convolutional neural network as described above.
[0073] The specific implementation of the embodiment is as follows.
[0074] 1. Build a multi-branch convolutional neural network model. Raman spectroscopy is a powerful tool for studying hydrogen bonding, main chain conformation and base stacking of nucleic acids. However, factors such as content, molecular weight, and pH value will affect the fingerprint feature information of Raman spectra. In addition, Raman spectra themselves have shortcomings such as unstable signals. Traditional chemometric methods cannot accurately extract subtle changes in complex physical and chemical features hidden in the spectrum. On the one hand, the deep learning model based on convolutional neural networks can independently learn features by optimizing the weights of each network layer to avoid the omission of feature information. On the other hand, it does not require understanding the relationship between input and output or related parameter knowledge, making it easier to obtain fingerprint feature information.
[0075] 1.1 Embedding the convolutional layers and fully connected layers in a multi-branch convolutional neural network, searching for the size and position of the spatial window according to its spatial distribution. These spatial windows are used as inputs to different branches of the multi-branch convolutional neural network. The convolutional layers contain 512 and 256 neurons, supplemented by 1×3 convolutional filters in each layer, and undergo a linear rectification function (ReLU) activation before transitioning to the maximum pooling layer. The spectrum uses the VGG residual operation in the convolutional layer before entering the pooling layer. For the qth spatial window, its size and position are encoded as a vector
[0076] 1.2 Convolutional long short-term memories (ConvLSTMs) are added to the shallow, middle, and deep layers of each branch in a multi-branch convolutional neural network to extract joint spatial-spectral features and make full use of the complementary information between different layers.
[0077] 1.3 The attention branch mechanism allows each branch to adaptively adjust its weight. It dynamically fuses multi-branch features for classification. The multi-branch convolutional neural network is optimized using the losses of the branches and the main classifier.
[0078]
[0079]
[0080] 1.4 The main classifier classifies the test samples.
[0081] 2. Training and learning
[0082] 2.1 Prepare PDRN-1 (20bp-300bp) solutions and PDRN-2 (300bp-2000bp) samples with concentrations of 0.1mg / ml, 1mg / ml, 5mg / ml, 10mg / ml, and 20mg / ml, respectively. After the samples are completely dissolved, use the LABRAM HR800 laser confocal micro-Raman spectrometer of Jobin Yvon, France, at a temperature of 25±3℃, an excitation line wavelength of 514.5nm, an output power of 10mW, and a slit width of 200μm to obtain the Raman spectra of each sample and output the spectrum, with a peak position error of <±3cm -1
[0083] 2.2 Obtain the PDRN-1 and PDRN-2 at 400-3400cm -1 Raman spectra in the range of 600-1000, 1500-2000 and 2900-3400cm -1 Raman peaks in the range have unique PDRN "fingerprints" that can show fine-scale specific features of different concentrations and molecular weights, making them easier to distinguish.
[0084] 2.3 Spectral input multi-branch convolutional neural network system, activate the adaptive search mechanism, locate the most representative feature area of spectral data through iterative refinement, divide it into three independent data blocks, and perform multi-branch self-attention operation on it, using its multi-branch structure to capture subtle spectral features at different scales. After a series of operations such as convolution, pooling, and dropout, a comprehensive and multi-level spatial-spectral feature is finally extracted.
[0085] 2.4 Collect 1000 spectra for each PDRN and allocate them in the ratio of 6:2:2 among training set, validation set and test set. After feature extraction and model debugging and optimization of 8000 spectra in the training set and validation set, the multi-branch convolutional neural network model is input into the test set of 2000 spectra to verify the final performance of the model.
[0086] 2.5 Use the multi-branch convolutional neural network model generated after training to test the test data set and carry out the identification of 10 types of PDRN. Figure 2 The confusion matrix in Figure 2 shows the performance breakdown of the multi-branch convolutional neural network for each PDRN recognition analysis, with the entries on the diagonal representing the recognition accuracy for each category: the average recognition accuracy of the model is 98.2%.
[0087] Figure 2Note: PDRN-1A 0.1mg / ml; PDRN-1B 1mg / ml; PDRN-1C 5mg / ml; PDRN-1D10mg / ml; PDRN-1E 20mg / ml; PDRN-2A 0.1mg / ml; PDRN-2B 1mg / ml; PDRN-2C 5mg / ml; PDRN-2D10mg / ml; PDRN-2E 20mg / ml.
[0088] The above examples of the present invention are described in detail, but the content is only a preferred embodiment of the present invention and cannot be considered to limit the scope of implementation of the present invention. All equivalent changes and improvements made according to the scope of application of the present invention should still fall within the scope of the patent coverage of the present invention.
Claims
1. A PDRN quality evaluation classification system based on a multi-branch convolutional neural network, characterized by: It includes a data acquisition module for collecting Raman spectra of PDRN; A feature extraction module is used to input the Raman spectrum of the PDRN to be identified into a pre-trained multi-branch convolutional neural network, based on a multi-branch attention network layer and a non-local adaptive spatial segmentation strategy to obtain the physical and chemical characteristic information of PDRN of different molecular weights and concentrations, and input the PDRN Raman spectrum into the basic, convolution block to obtain basic feature information, and then input the three branches of the TBP block, namely the bottleneck structure branch, the reverse bottleneck structure branch, and the non-convolution branch, to obtain all the feature information extracted by the three branches, and merge all the feature information into multi-scale feature information; The sample identification module is used to determine the molecular weight and content of the PDRN sample in a pre-set multi-branch convolutional neural network database according to the PDRN Raman spectrum characteristic information, so as to finally determine the physical and chemical characteristics of the measured PDRN.
2. A PDRN quality evaluation classification system based on a multi-branch convolutional neural network according to claim 1, characterized in that: The steps include: S1, collect Raman spectra images of PDRN aqueous solutions with different concentrations; S2. Analyze the obtained images and construct a PDRN molecular property prediction system based on a multi-branch convolutional neural network. Different branches are used to extract the spatial spectral features of the Raman spectrum, highlight the more discriminative branches and suppress the branches with low discriminative rates. Different branches use voting strategies to provide prediction probabilities of their categories. S3, assigning different weights to different branches according to their discriminative capabilities, and multiplying the adaptive weights generated by each branch with the extracted features; These weighted features are concatenated to generate multi-scale and multi-context attention features of each branch; S4, provides complementary information for learning shallow, middle and deep semantic features of the i-th sample; S5, providing the weight of the qth branch of the i-th sample; S6. In R d×1 In the feature space, there are M Raman spectra X = {x1, x2, ..., x M }, the corresponding label is Y = {y1, y2…y M }, given a sample xi (1≤i≤M), Qi represents the data set of the neighborhood spectrum of the non-local candidate region centered on xi; for all labeled samples, a new data set Q = {Q1, Q2…Q M }Divided into and a+b=M; S7. In the non-local candidate region Qi, m spatial windows need to be searched. represents the Qth spatial window centered at xi. For the qth spatial window, its size and position are encoded as a vector in and, are the horizontal and vertical coordinates of the center sample in the spatial window, is the size of the spatial window; In the multi-branch convolutional neural network, the convolutional layers and the corresponding pooling layers are stacked to extract the hierarchical features of the non-local candidate regions Qi. Then, these features are mapped to a one-dimensional vector, which is used as the input of the next fully connected layer. Finally, after the fully connected layer, Formulated as follows: Where f is a composite function containing multiple convolution and pooling operations, W and b are the corresponding weights and biases, and W FC and b FC are the weights and biases of the fully connected layer, and g[·] represents its activation function; S8. Construct a softmax layer as the main classifier to perform comprehensive classification using the data extracted from the multi-scale and multi-context attention network layers and the multi-layer spatial spectrum features of the convolutional neural network; S9. In the model testing phase, the i-th test sample is predicted by calculating the maximum probability of different classes. .
3. A PDRN quality evaluation classification system based on a multi-branch convolutional neural network according to claim 1, characterized in that: In step S1, Raman spectral images of PDRN aqueous solutions with different concentrations of 1 mg / ml-20 mg / ml under standard conditions are collected.
4. A PDRN quality evaluation classification system based on a multi-branch convolutional neural network according to claim 1, characterized in that: In step S3, for the i-th sample, the weighted feature Mi is formulated as follows: represents the weight of the qth branch, Represents the multi-layer spatial-spectral features extracted by the q-th branch.
5. A PDRN quality evaluation classification system based on a multi-branch convolutional neural network according to claim 1, characterized in that: In step S4, the following is formulated: represents the shallow, medium and deep spectral features extracted by the convolutional layer in the qth branch, Ls(·), Lm(·), Ld(·) represent the spectral features extracted from the shallow, medium and deep layers, P(·) represents the global average pooling operation, concat(·) represents the cascade operation, Represents the multi-layer spatial-spectral features of the qth branch of the i-th sample.
6. A PDRN quality evaluation classification system based on a multi-branch convolutional neural network according to claim 1, characterized in that: In step S5, the calculation formula is as follows: is the predicted probability of class li in the qth branch.
7. A PDRN quality evaluation classification system based on a multi-branch convolutional neural network according to claim 1, characterized in that: Step S8, during the training process, the loss function of the multi-branch convolutional neural network is defined as: in Represents the main classifier for the i-th training sample The first term in the above formula is the loss of the main classifier, which is calculated by the cross entropy of the real class probability and the output class probability; the second term is the loss of the m branch classifiers calculated in the same way.
8. A PDRN quality evaluation classification system based on a multi-branch convolutional neural network according to claim 1, characterized in that: In step S9, the calculation formula is: Where c = 1, 2, 3, ..., C, represents the predicted probability of the cth class from the main classifier.
Citation Information
Cited By
Compound fenbendazole tablet quality evaluation method based on convolutional neural network
CN121306594A
Raman spectrum joint classification and quantification method and device based on prior fusion double-flow network, and medium
CN121723378A
Diesel oil Raman spectrum detection method, device and equipment based on convolutional neural network
CN121830628A
