Torreya grandis species identification method, electronic device and readable storage medium

Abstract

The present application discloses a method for identifying the species of Torreya grandis, an electronic device and a readable storage medium; the method includes presetting a Torreya grandis identification model, identifying features in a Torreya grandis image through the Torreya grandis identification model, and identifying the species of Torreya grandis according to the features in the Torreya grandis image; a feature extraction module, which optimizes the neural network parameters of feature extraction based on a simulated quantum optimization method, and extracts key features in Torreya grandis images; a feature dimension reduction module, which performs feature dimension reduction on key features based on a sparse coding autoencoder; a classifier module, which classifies the key features after dimension reduction based on a smooth approximation extreme learning machine algorithm, and outputs the species of Torreya grandis. The present application effectively extracts the key features of Torreya grandis images, while reducing computational complexity; improves the generalization ability and training efficiency of the model; and reduces the influence of noise on recognition accuracy.

Description

Technical Field

[0001] The present application relates to the technical field of Torreya grandis species identification, and in particular to a Torreya grandis species identification method, electronic device and readable storage medium. Background Art

[0002] The Chinese invention patent with application number CN202211306819.1 proposes a multi-grained parametric modeling method for real-life trees, including: combining the watershed algorithm and the clustering method to segment and cluster the plant canopy height model and the plant area point cloud model, and realize the single tree segmentation of the plant area three-dimensional point cloud; through the single tree multi-view image recognition and feature extraction, determine the tree species information and individual tree characteristic parameters in the plant area; establish a tree growth rule equation, and describe the growth of the plant through the variable relationship expressed by the equation; construct a tree parameter information database, classify and store the tree category characteristic parameters, trunk and leaf material mapping; read the tree parameter information database according to the tree species information to obtain matching information, and combine the extracted tree individual characteristic parameters to realize the construction steps, assembly steps and material mapping steps of multi-grained tree three-dimensional models; generate and display tree scenes. This application can realize the rapid construction of multi-grained tree three-dimensional models and the efficient and smooth display of large-scale tree scenes.

[0003] The Chinese invention patent application number CN201910481805.5 proposes a method for extracting trees in outdoor point cloud scenes based on shape classification and combination. It uses the optimal feature set and a classification algorithm to improve the accuracy of outdoor scene classification, obtains classified data, and completes the problem of extracting single trees in outdoor scene point cloud data through positioning, filtering, matching and other steps. If there is a problem of overlapping tree crowns in the data, further optimization processing is performed to optimize and improve the final result to complete the extraction of single trees. This application solves the problem of inseparable trees with overlapping tree crowns in outdoor point cloud scenes.

[0004] The Chinese invention patent with application number CN202310511450.6 proposes a plant species identification method, device, storage medium and electronic device, wherein the method includes: determining plant feature data of the plant to be identified; inputting the plant feature data into the plant species identification model to obtain the plant species of the plant to be identified output by the plant species identification model; wherein the plant species identification model is constructed based on a neural network model and an Adaboost model. The method and device provided by this application improve the efficiency and accuracy of identifying plant species.

[0005] However, these technical solutions generally have problems in processing and feature extraction, generalization ability and training efficiency. They are also inefficient and inaccurate in identifying plant species. Summary of the invention

[0006] The main technical problem solved by the present application is to provide a method for identifying Torreya grandis species, an electronic device and a readable storage medium, so as to solve the problem that Torreya grandis species identification processing and feature extraction, generalization ability and training efficiency are insufficient, resulting in low efficiency and poor accuracy.

[0007] In order to solve the above technical problems, a technical solution adopted in the present application is to provide a method for identifying Torreya grandis species, comprising the steps of:

[0008] A Torreya grandis recognition model is preset, features in a Torreya grandis image are recognized by the Torreya grandis recognition model, and the type of Torreya grandis is recognized according to the features in the Torreya grandis image;

[0009] Torreya grandis identification model includes:

[0010] Feature extraction module: The feature extraction module optimizes the neural network parameters of the feature extraction module based on the simulated quantum optimization method to extract the key features in the Torreya grandis image;

[0011] Feature dimensionality reduction module, which is based on sparse coding autoencoder to reduce the dimensionality of key features;

[0012] Classifier module,The classifier module classifies the key features after dimensionality reduction based on the smooth approximation extreme learning machine algorithm, and outputs the type of Torreya grandis.

[0013] In some embodiments, a Torreya grandis recognition model is trained using a training data set, and the type of Torreya grandis is identified using the trained Torreya grandis recognition model; the training data set is a preprocessed Torreya grandis image, and the Torreya grandis image includes Torreya grandis in different geographical locations, different climatic conditions, and different growth stages; the Torreya grandis image contains different parts of Torreya grandis.

[0014] In some embodiments, the preprocessing steps of the Torreya grandis image include: data annotation and data standardization;

[0015] In the step of labeling the data of Torreya grandis images, the Torreya grandis images in the training data set are labeled to distinguish different Torreya grandis species; the labeling category is the specific species of Torreya grandis, and each labeling corresponds to the Torreya grandis species of the corresponding Torreya grandis image;

[0016] In the data standardization step of Torreya grandis image, it includes adjusting the pixel mean and standard deviation; image standardization is used for standardization, and the standardization method is expressed as follows:

[0017]

[0018] in, is the original Torreya grandis image, is the pixel mean of Torreya grandis image, is the pixel standard deviation of Torreya grandis image; This is the standardized image of Torreya grandis.

[0019] In some embodiments, the preprocessing step of the Torreya grandis image further includes data expansion; data expansion is achieved based on a generative adversarial network with cyclic adjustment of the learning rate, and a learning rate cyclic adjustment mechanism is added during the training process of the generative adversarial network to dynamically adjust the learning rate according to different stages of training;

[0020] The process of the generative adversarial network algorithm based on learning rate cyclic adjustment is as follows:

[0021] The first step is initialization. Set the initial network parameters of the generator and discriminator. The generator parameters are ,in represents the parameter set of the generator; the discriminator parameters are ,in Represents the parameter set of the discriminator;

[0022] The second step is to set the learning rate cycle adjustment strategy to dynamically change the learning rate according to the training progress. The learning rate cycle adjustment strategy adopts a periodic adjustment method, which is expressed as:

[0023]

[0024] in, is the current learning rate, and are the minimum and maximum learning rates, respectively, is the current training iteration number, is the maximum number of iterations;

[0025] The third step is generator training. Random noise is used as input to generate new Torreya grandis image samples. The training process of the generator is expressed as:

[0026]

[0027]

[0028]

[0029] in, is the random noise of the input, New samples generated by the generator, is the loss function of the generator, is the batch size;

[0030] The fourth step is discriminator training. The generated samples and real samples are evaluated and the discriminator is optimized to more accurately distinguish between generated samples and real samples. The training process of the discriminator is as follows:

[0031]

[0032]

[0033]

[0034]

[0035] in, is the real sample data, and are the outputs of the discriminator for real and generated samples, respectively. is the loss function of the discriminator;

[0036] The fifth step is adversarial training, which repeats the training of the generator and the discriminator until a balance is reached;

[0037] Step 6: Adjust the learning rate cyclically according to the strategy Adjust the learning rate during training to improve the diversity and quality of generated samples;

[0038] The seventh step is to output the data expansion results, add the optimized generated samples to the training data set to form an expanded training data set; input the expanded training data set into the Torreya grandis recognition model to train the Torreya grandis recognition model.

[0039] In some embodiments, the feature extraction module includes an input layer, a first hidden layer, a second hidden layer, and a third hidden layer, and has the following structure:

[0040] Input layer: contains Nin neurons, where Nin is used to stretch the data input to the Torreya grandis recognition model into the dimension of a one-dimensional vector;

[0041] First hidden layer: Each layer of the first hidden layer has 128 neurons and uses ReLU as the activation function;

[0042] Second hidden layer: Each layer of the first hidden layer has 256 neurons and uses ReLU as the activation function;

[0043] Third hidden layer: Each layer of the first hidden layer has 100 neurons and uses ReLU as the activation function.

[0044] In some embodiments, the specific steps of optimizing the neural network parameters for optimizing feature extraction are as follows:

[0045] The first step is parameter initialization. A quantum state is initialized for each parameter. Each quantum state represents a possible parameter value and an initial oscillation frequency is preset.

[0046] The second step is to explore multiple quantum states. In each round of iteration, a new random offset is introduced for each quantum state to simulate the oscillation in the quantum system. Given the parameters of a neural network In its quantum state In the equation, the offset is expressed as:

[0047]

[0048] in, is the oscillation frequency; is a random number drawn from a uniform random distribution;

[0049] Calculate the new parameter value and the corresponding fitness value. The way to update the parameter is expressed as:

[0050]

[0051] in, It is the new parameter value generated after exploring multiple quantum states; is the parameter value before multiple quantum state exploration; is the update amount of the parameter;

[0052] Each parameter has its corresponding quantum state, and the fitness function is defined as follows:

[0053]

[0054] in is the fitness function; is the mean square error loss function; is the quantum condensation term, which measures the synergistic effect of different quantum states; is a set that contains all quantum states; is a hyperparameter used to adjust the weight of the quantum condensation term;

[0055] The quantum condensation term is defined as follows:

[0056]

[0057] in, is the number of quantum states; is the quantum condensation function, which is used to measure the two quantum states and The degree of cohesion between

[0058] definition for:

[0059]

[0060] in, It is a quantum state and The Euclidean distance between is a constant;

[0061] The third step is dynamic frequency adjustment. According to the changes in the fitness function, the oscillation frequency of each parameter is dynamically adjusted. According to the changes in the fitness function, the method of dynamically adjusting the frequency is expressed as:

[0062]

[0063] in, The frequency before dynamic frequency adjustment; The frequency after dynamic frequency adjustment; It is a manually preset learning rate; is the rate of change of the fitness function, which is calculated as follows:

[0064]

[0065] in, and are the losses of the current and previous iterations respectively;

[0066] The fourth step, quantum collision, select two or more quantum states to collide; select two quantum states and The collision is simulated as follows:

[0067]

[0068] in, is the quantum state after the quantum collision, is a coefficient drawn from a uniform random distribution;

[0069] Evaluate the newly generated quantum state and calculate its corresponding fitness value; the way to evaluate the newly generated quantum state is expressed as:

[0070]

[0071] in, is the new parameter value generated after the quantum collision;

[0072] Calculate the new fitness value if but Alternative ; If this new state leads to a smaller fitness value, the collision is considered successful and the new quantum state is used for the next iteration;

[0073] The fifth step is to simulate the interference effect. The quantum states of all parameters are merged to simulate quantum interference. Consider two quantum states and , and its interference result is expressed as:

[0074]

[0075] in, The interference coefficient is set by humans, which determines and weight in intervention; Updated parameter values ​​for interference effects;

[0076] Select those quantum states that produce lower fitness values ​​and eliminate the rest;

[0077] According to the selected optimal quantum state, update the parameters of the neural network, considering all and , select the quantum state that leads to the minimum fitness value for parameter update, which is expressed as:

[0078]

[0079] in, Neural network parameters determined for this iteration; and is the set sum of the parameters of the neural network and its quantum state in this iteration;

[0080] Step 6, termination condition, repeat steps 2 to 5. If the preset number of iterations is reached or the fitness function reaches the preset threshold, the algorithm is terminated, which means that the training of the feature extraction module is completed.

[0081] In some embodiments, the training process of the optimized sparse coding autoencoder dimensionality reduction algorithm is as follows:

[0082] The first step is to initialize the autoencoder and the encoder parameters of the autoencoder network. and decoder parameters ;

[0083] The second step is sparsity adjustment, which is based on the sparsity adjustment mechanism of molecular dynamics principles and sets the sparsity parameters in the encoding process. ; Sparse energy function Defined as:

[0084]

[0085] in, Indicates the encoded Features, is the sparsity adjustment parameter;

[0086] The third step is forward propagation. In the forward propagation process of the sparse autoencoder, filters of different scales are used. Extracting input data The characteristics are expressed as:

[0087]

[0088] in, Indicates different scales;

[0089] The feature fusion weight calculation is performed, and the adaptive weight calculation method for each scale feature is expressed as:

[0090]

[0091] in, It is The variance of the scale feature, is the adjustment parameter;

[0092] The fusion feature representation is performed, and the fusion feature is obtained by combining the features of different scales and the corresponding weights, which is expressed as:

[0093]

[0094] Input data and fusion features Input to the encoding of the autoencoder to get the encoding , the encoding process is expressed as:

[0095]

[0096] in, is the activation function, is the encoder bias term;

[0097] After the decoder reconstructs , the decoding process is expressed as:

[0098]

[0099] in, is the bias term of the decoder;

[0100] The fourth step is to calculate the loss function and calculate the reconstruction loss. And sparsity loss, the loss function contains reconstruction loss and sparsity loss, expressed as:

[0101]

[0102] in, is the number of input features, is a hyperparameter that controls the importance of sparsity loss;

[0103] The fifth step is policy iteration optimization. The policy iteration method is applied to dynamically adjust the codec parameters according to the current loss. The policy iteration process is used to dynamically adjust the codec parameters as follows:

[0104]

[0105]

[0106] in, is the learning rate, which is used to control the amplitude of parameter updates;

[0107] Step 6: Output the reduced dimension data. The trained and optimized autoencoder outputs the reduced dimension features. .

[0108] In some embodiments, the training process of the extreme learning machine algorithm based on smooth approximation is as follows:

[0109] The first step is initialization. Initialize the hidden layer weights of the extreme learning machine and bias , the initialization method is random initialization, expressed as:

[0110]

[0111]

[0112] Among them, Rand() represents the random initialization function, and Represents the hidden layer weights after random initialization and bias ;

[0113] The second step is hidden layer conversion, which is used to transform the input data Apply hidden layer transformation to get hidden layer representation , expressed as:

[0114]

[0115] in, is the Sigmoid activation function;

[0116] The third step is to calculate the output weight using a smooth approximation method. , the calculation method is expressed as:

[0117]

[0118] in, is the regularization parameter;

[0119] Solving the inverse matrix In the process, , using Cholesky decomposition, where is decomposed into the product of two triangular matrices , then solve and The reverse, when When approaching singularity or instability, add regularization terms To improve numerical stability; this process is expressed as:

[0120]

[0121]

[0122] beg and , to obtain ;

[0123] The fourth step is adaptive neural regulation. The adaptive neural regulation mechanism of the evolutionary algorithm is used to adjust the hidden layer parameters. The adaptive neural regulation mechanism of the evolutionary algorithm is expressed as:

[0124]

[0125] in, It is a parameter adjustment function based on evolutionary algorithm;

[0126] The fifth step is classifier training. The classifier is trained using the optimized extreme learning machine, which is expressed as:

[0127]

[0128] Among them, TrainELM() is the function for training the extreme learning machine;

[0129] Step 6: Output the classification results and use the trained classifier to classify the new input data. Classification is performed, expressed as:

[0130]

[0131] Among them, Classify() is a function that uses the trained classifier for classification.

[0132] In order to solve the above technical problems, another technical solution adopted in the present application is: an electronic device is provided, comprising a memory and a processor coupled to each other, the processor being used to execute program instructions stored in the memory to implement the above-mentioned method for identifying Torreya grandis species.

[0133] In order to solve the above technical problems, another technical solution adopted by the present application is: providing a non-volatile computer-readable storage medium on which program instructions are stored, and when the program instructions are executed by a processor, the above-mentioned method for identifying Torreya grandis species is implemented.

[0134] The beneficial effects of the present application are as follows: the present application adopts an optimized neural network model and feature dimensionality reduction technology to effectively extract the key features of Torreya grandis images while reducing computational complexity. The neural network parameter optimization method based on simulated quantum optimization and the extreme learning machine algorithm based on smooth approximation are intended to improve the generalization ability and training efficiency of the model. The technical solution of the present application reduces the impact of noise on recognition accuracy through an optimized algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0135] Figure 1 is a structural block diagram of a feature extraction module according to an embodiment of the present application;

[0136] Figure 2 It is a flow chart of a parameter optimization method of a feature extraction module according to an embodiment of the present application. DETAILED DESCRIPTION

[0137] In order to facilitate the understanding of the present application, the present application is described in more detail below in conjunction with the accompanying drawings and specific embodiments. The preferred embodiments of the present application are provided in the accompanying drawings. However, the present application can be implemented in many different forms and is not limited to the embodiments described in this specification. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present application more thorough and comprehensive.

[0138] It should be noted that, unless otherwise defined, all technical and scientific terms used in this specification have the same meaning as those commonly understood by those skilled in the art of the present application. The terms used in the specification of this application are only for the purpose of describing specific embodiments and are not intended to limit this application. The term "and / or" used in this specification includes any and all combinations of one or more related listed items.

[0139] Figure 1 and Figure 2 An embodiment of the method for identifying the species of Torreya grandis of the present application is shown, comprising: presetting a Torreya grandis identification model, identifying features in a Torreya grandis image through the Torreya grandis identification model, and identifying the species of Torreya grandis according to the features in the Torreya grandis image.

[0140] The Torreya grandis recognition model includes: a feature extraction module, which is based on the simulated quantum optimization method to optimize the neural network parameters of the feature extraction module and extract the key features in the Torreya grandis image; a feature dimension reduction module, which is based on the sparse coding autoencoder to perform feature dimension reduction on the key features; a classifier module, which is based on the smooth approximation extreme learning machine algorithm to classify the key features after dimension reduction and output the type of Torreya grandis.

[0141] This application uses an optimized neural network model and feature dimensionality reduction technology to effectively extract the key features of Torreya grandis images while reducing computational complexity. A neural network parameter optimization method based on simulated quantum optimization and an extreme learning machine algorithm based on smooth approximation are designed to improve the generalization ability and training efficiency of the model. The technical solution of this application reduces the impact of noise on recognition accuracy through an optimized algorithm.

[0142] Before using the Torreya grandis recognition model, the Torreya grandis recognition model can be trained to improve the recognition accuracy.

[0143] The Torreya grandis recognition model is trained using a training data set, which is a preprocessed Torreya grandis image. The Torreya grandis images include Torreya grandis in different geographical locations, different climatic conditions, and different growth stages. The images contain different parts of Torreya grandis, such as leaves, fruits, and trunks, to ensure that the Torreya grandis recognition model can fully recognize the various features of Torreya grandis.

[0144] The preprocessing steps for Torreya grandis images include: data labeling, data standardization and data expansion.

[0145] In the step of data annotation of Torreya grandis images: the Torreya grandis images in the training data set are manually annotated to distinguish different Torreya grandis species. The annotation categories are specific types of Torreya grandis, such as fine-leaf Torreya grandis, broad-leaf Torreya grandis, rice Torreya grandis, sesame Torreya grandis, ivory Torreya grandis, small round Torreya grandis, round Torreya grandis, large round Torreya grandis, etc. In one embodiment, for Torreya grandis images, let aX={image 1, image 2, ..., image N}, where each image is a high-resolution image of Torreya grandis; aY={annotation 1, annotation 2, ..., annotation N}, where each annotation corresponds to the Torreya grandis species of the corresponding image.

[0146] In the data standardization step of Torreya grandis images: the collected Torreya grandis images are standardized, including adjusting the pixel mean and standard deviation, in preparation for subsequent processing.

[0147] The image is standardized and the standardization can be expressed as follows:

[0148]

[0149] in, is the original Torreya grandis image, is the pixel mean of Torreya grandis image, is the pixel standard deviation of Torreya grandis image. This is the standardized image of Torreya grandis.

[0150] In this application task, the collection, acquisition, labeling and preprocessing of training data are time-consuming and labor-intensive, and insufficient training samples can easily lead to poor generalization ability of the model and affect the accuracy of the model.

[0151] In order to solve the above problems, in the data expansion step of Torreya grandis images, this application proposes a generative adversarial network based on learning rate cyclic adjustment for sample generation, thereby realizing data expansion. In the training process of the generative adversarial network, a learning rate cyclic adjustment mechanism is introduced to dynamically adjust the learning rate according to different stages of training to improve the quality and diversity of generated samples. This solves the problems of insufficient data volume and insufficient sample diversity in traditional methods.

[0152] Specifically, the training process of the generative adversarial network algorithm based on learning rate cyclic adjustment is as follows:

[0153] 1. Initialization

[0154] Set the initial network parameters of the generator and discriminator. The generator parameters are ,in represents the parameter set of the generator; the discriminator parameters are ,in Represents the set of parameters of the discriminator.

[0155] 2. Learning rate adjustment strategy setting

[0156] Set the learning rate cycle adjustment strategy to dynamically change the learning rate according to the training progress. The learning rate cycle adjustment strategy adopts a periodic adjustment method, which can be expressed as:

[0157]

[0158] in, is the current learning rate, and are the minimum and maximum learning rates, respectively, is the current training iteration number, is the maximum number of iterations. In one embodiment, Can be set to , Can be set to If the iteration plan is 10,000 times, then , is the current number of iterations, for example, at the 2000th iteration, .

[0159] 3. Generator training

[0160] Using random noise as input, new Torreya grandis image samples are generated. The training process of the generator can be expressed as:

[0161]

[0162]

[0163]

[0164] in, is the random noise of the input, New samples generated by the generator, is the loss function of the generator, is the batch size. Random noise Usually generated by a specific distribution (such as Gaussian distribution), you can set a mean and standard deviation to generate this noise, for example, use a standard normal distribution (mean is 0, standard deviation is 1) to generate .

[0165] 4. Discriminator training

[0166] The generated samples and real samples are evaluated, and the discriminator is optimized to more accurately distinguish between the generated samples and the real samples. The training process of the discriminator is as follows:

[0167]

[0168]

[0169]

[0170]

[0171] in, is the real sample data, and are the outputs of the discriminator for real and generated samples, respectively. is the loss function of the discriminator.

[0172] 5. Adversarial Training

[0173] The training of the generator and the discriminator is repeated until a balance is reached, that is, the discriminator cannot significantly distinguish between generated samples and real samples.

[0174] 6. Learning rate cycle adjustment

[0175] According to the strategy The learning rate is adjusted during training to improve the diversity and quality of generated samples.

[0176] 7. Output data expansion results

[0177] The optimized generated samples are added to the training dataset to form an expanded training dataset.

[0178] The expanded training data set is input into the Torreya grandis recognition model to train the Torreya grandis recognition model, so as to improve the training efficiency of the Torreya grandis recognition model and the accuracy of Torreya grandis recognition.

[0179] The Torreya grandis recognition model includes a feature extraction module, a feature dimension reduction module and a classifier. The training of the Torreya grandis recognition model is the training of the feature extraction module, the feature dimension reduction module and the classifier. The feature extraction module is used to extract the key features in the Torreya grandis image. The feature dimension reduction module is used to reduce the dimension of the key features extracted by the feature extraction module to improve the recognition speed. The classifier is used to classify the key features after dimension reduction and output the type of Torreya grandis according to the classification results.

[0180] The feature extraction module uses an optimized neural network model to extract key features from Torreya grandis images. The simulated quantum optimization method is used to optimize the parameters of the neural network to improve the efficiency and accuracy of feature extraction.

[0181] The expanded Torreya grandis image is input into the feature extraction module to extract the features of the Torreya grandis image, so as to train the Torreya grandis recognition model. The feature extraction module of the present application adopts an optimized neural network model for feature extraction. The traditional neural network model adopts a gradient descent-based method to update parameters during training, but it is easy to produce gradient vanishing and gradient explosion phenomena. At the same time, it is also easy to fall into a local optimal solution during the parameter optimization process.

[0182] like Figure 1 As shown, the feature extraction module of the present application includes an input layer, a first hidden layer, a second hidden layer and a third hidden layer, and the structure is as follows:

[0183] Input layer: contains Nin neurons, where Nin stretches the data input to the Torreya grandis recognition model into the dimension of a one-dimensional vector.

[0184] First hidden layer: The first hidden layer has 128 neurons per layer. ReLU is used as the activation function.

[0185] Second hidden layer: Each layer of the first hidden layer has 256 neurons. ReLU is used as the activation function.

[0186] Third hidden layer: Each layer of the first hidden layer has 100 neurons. ReLU is used as the activation function.

[0187] The weight and bias parameters are initialized using Xavier initialization.

[0188] The present invention uses an optimized neural network model as a feature extraction module for feature extraction. This model combines the method of simulated quantum optimization to improve the efficiency and accuracy of feature extraction, while alleviating the problems of gradient vanishing and gradient exploding.

[0189] Furthermore, during the training process of the feature extraction module, this application proposes a neural network parameter optimization method based on simulated quantum optimization. Inspired by quantum physics, the quantum state represents the state of a system, and the oscillation of the quantum state describes the transition of the quantum state between different states. Drawing on the characteristics of quantum oscillation, each neural network parameter is placed in a quantum state, and these parameters are optimized by simulating quantum oscillations.

[0190] like Figure 2 As shown, the specific steps of the optimized neural network parameter optimization method are as follows:

[0191] 1. Parameter initialization

[0192] A quantum state is initialized for each parameter (including the weight parameters and bias parameters of the neural network). Each quantum state represents a possible parameter value, and an initial oscillation frequency is preset artificially.

[0193] 2. Exploration of multiple quantum states

[0194] In each iteration, a new random shift is introduced to each quantum state to simulate the oscillations in the quantum system. In its quantum state The offset can be expressed as:

[0195]

[0196] in, is the oscillation frequency; is a random number drawn from a uniform random distribution.

[0197] Furthermore, the new parameter values ​​are calculated and the corresponding fitness values ​​are calculated. The method of updating the parameters can be expressed as:

[0198]

[0199] in, It is the new parameter value generated after exploring multiple quantum states; is the parameter value before multiple quantum state exploration; is the parameter update amount.

[0200] Furthermore, the fitness value is calculated by the fitness function. The traditional fitness function usually uses the inverse of the loss function, which easily leads to slow convergence of the neural network. The present application proposes a quantum condensation fitness function, which draws on the phenomenon of quantum condensation to enhance the synergy of different quantum states in the algorithm.

[0201] Specifically, the parameter optimization of the neural network in this application is a multi-parameter optimization problem, each parameter has its corresponding quantum state, and the fitness function is defined as follows:

[0202]

[0203] in is the fitness function; is the mean square error loss function; is the quantum condensation term, which measures the synergistic effect of different quantum states; is a set that contains all quantum states; is a hyperparameter used to adjust the weight of the quantum condensation term.

[0204] Furthermore, the quantum condensation term is defined as follows:

[0205]

[0206] in, is the number of quantum states; is the quantum condensation function, which is used to measure the two quantum states and The degree of cohesion between them.

[0207] Furthermore, define for:

[0208]

[0209] in, It is a quantum state and The Euclidean distance between is a very small constant that ensures that the denominator is not zero.

[0210] This fitness function attempts to enhance the cohesion between different quantum states while minimizing the basic loss. By adding the quantum cohesion term, the algorithm will tend to select those quantum states that are closer together and therefore more likely to work together, resulting in better solutions, helping the algorithm to explore more effectively in high-dimensional parameter space, and possibly avoid falling into local minima, speeding up the convergence of the training process.

[0211] 3. Dynamic frequency adjustment

[0212] According to the changes in the fitness function, the oscillation frequency of each parameter is dynamically adjusted. According to the changes in the fitness function, the way to dynamically adjust the frequency can be expressed as:

[0213]

[0214] in, The frequency before dynamic frequency adjustment; The frequency after dynamic frequency adjustment; It is a manually preset learning rate; is the rate of change of the fitness function, which is calculated as follows:

[0215]

[0216] in, and are the losses of the current and previous iterations respectively.

[0217] If the fitness value decreases, the frequency is increased to speed up the exploration; if the fitness value increases, the frequency is decreased to slow down the exploration.

[0218] 4. Quantum collision

[0219] To simulate a quantum collision, two or more quantum states are selected for collision. During the collision simulation, information is exchanged between these selected quantum states. This exchange is nonlinear and may produce a new quantum state. and To perform an impact, the way to simulate the impact can be expressed as:

[0220]

[0221] in, is the quantum state after the quantum collision, is a coefficient drawn from a uniform random distribution.

[0222] Furthermore, the newly generated quantum state is evaluated and its corresponding fitness value is calculated. The method of evaluating the newly generated quantum state can be expressed as:

[0223]

[0224] in, is the new parameter value generated after the quantum collision.

[0225] Further, calculate the new fitness value, if but Alternative If this new state leads to a smaller fitness value, the collision is considered successful and the new quantum state is used for the next iteration.

[0226] 5. Interference effect simulation

[0227] The quantum states of all parameters are merged to simulate quantum interference. Consider two quantum states and , and its interference result can be expressed as:

[0228]

[0229] in, The interference coefficient is set by humans, which determines and weight in intervention; Updated parameter values ​​for interference effects.

[0230] Furthermore, those quantum states that produce lower fitness values ​​are selected and the rest are eliminated.

[0231] Further, according to the selected optimal quantum state, the parameters of the neural network are updated. Consider all and , select the quantum state that leads to the minimum fitness value for parameter update, which can be expressed as:

[0232]

[0233] in, Neural network parameters determined for this iteration; and is the set of parameters of the neural network and its quantum state for this iteration.

[0234] 6. Termination conditions

[0235] Repeat steps 2 to 5. If the preset number of iterations is reached or the fitness function reaches the preset threshold, the algorithm is terminated, indicating that the training of the feature extraction module is completed.

[0236] After the feature extraction module extracts the key features, the feature dimension reduction module reduces the dimension of the key features. The feature dimension reduction module uses the optimized sparse coding autoencoder algorithm to reduce the dimension of the high-dimensional vector after feature extraction. Using the optimized sparse coding autoencoder for feature dimension reduction, combined with the molecular dynamics principle and adaptive multi-scale feature fusion technology, can effectively reduce the computational complexity while retaining the key feature information.

[0237] This application proposes an optimized sparse coding autoencoder dimensionality reduction algorithm. The autoencoder is a neural network that is used to learn efficient representation (encoding) of input data and then reconstruct (decode) the data. On this basis, the sparse coding autoencoder adds sparsity constraints to force the network to learn more compact and meaningful data representations. This application improves the structure of the autoencoder and introduces the principle of molecular dynamics to adjust the sparsity in the encoding process, thereby achieving more efficient feature representation. In addition, this application applies the strategy iteration method to the training process of the autoencoder to dynamically adjust the encoding and decoding process to optimize the performance of feature dimensionality reduction.

[0238] Specifically, the training process of the optimized sparse coding autoencoder dimensionality reduction algorithm is as follows:

[0239] 1. Initialize the autoencoder

[0240] Initialize the parameters of the autoencoder network and (representing the parameters of the encoder and decoder respectively).

[0241] 2. Sparsity Adjustment

[0242] Apply the sparsity adjustment mechanism based on the principle of molecular dynamics to set the sparsity parameters in the encoding process Specifically, the sparse energy function Defined as:

[0243]

[0244] in, Indicates the encoded Features, is a sparsity adjustment parameter. In one embodiment, the parameter Can be set to .

[0245] 3. Forward Propagation

[0246] In the forward propagation process of the sparse autoencoder, this application introduces an adaptive multi-scale feature fusion technology that can capture and fuse data features from different scales to enrich the final feature representation. Specifically, filters of different scales are used Extracting input data The characteristics can be expressed as:

[0247]

[0248] in, Indicates different scales.

[0249] Furthermore, feature fusion weight calculation is performed, and the method of calculating the adaptive weight of each scale feature can be expressed as:

[0250]

[0251] in, It is The variance of the scale feature, is a tuning parameter. In one embodiment, the parameter Can be set to .

[0252] Furthermore, fusion feature representation is performed, and the fusion feature is obtained by combining features of different scales and corresponding weights, which can be expressed as:

[0253]

[0254] Furthermore, input data and fusion features Input to the encoding of the autoencoder to get the encoding , the encoding process can be expressed as:

[0255]

[0256] in, is the activation function, is the encoder bias term.

[0257] Further, After the decoder reconstructs , the decoding process can be expressed as:

[0258]

[0259] in, is the decoder bias term.

[0260] 4. Loss function calculation

[0261] Calculate the reconstruction loss And sparsity loss, the loss function includes reconstruction loss and sparsity loss, which can be expressed as:

[0262]

[0263] in, is the number of input features, is a hyperparameter that controls the importance of sparsity loss. In one embodiment, the parameter Can be set to .

[0264] 5. Strategy Iteration Optimization

[0265] Apply the policy iteration method to dynamically adjust the codec parameters according to the current loss. The way the policy iteration process is used to dynamically adjust the codec parameters can be expressed as:

[0266]

[0267]

[0268] in, is the learning rate, which is used to control the magnitude of parameter updates. In one embodiment, the learning rate Can be set to .

[0269] 6. Output dimensionality reduction data

[0270] The output of the trained and optimized autoencoder is a reduced dimension feature .

[0271] The feature dimension reduction module outputs the key features after dimension reduction, which are input into the classifier for classification to determine the category of the Torreya grandis image. The key features after dimension reduction are input into the classifier trained by the extreme learning machine algorithm based on smooth approximation. Through the optimized extreme learning machine algorithm based on smooth approximation, the classification accuracy is improved, the noise sensitivity is reduced, and the generalization ability of the model is improved.

[0272] The present application proposes an extreme learning machine algorithm based on smooth approximation. The extreme learning machine is an efficient single-layer feedforward neural network classifier that randomly initializes the hidden layer weights and analytically calculates the output weights. The present application improves the training process of the extreme learning machine by combining the smooth approximation technology of high-order derivatives, thereby improving the classification accuracy and reducing the noise sensitivity.

[0273] Specifically, the training process of the extreme learning machine algorithm based on smooth approximation is as follows:

[0274] 1. Initialization

[0275] Initialize the hidden layer weights of the extreme learning machine and bias , the initialization method is random initialization, which can be expressed as:

[0276]

[0277]

[0278] Among them, Rand() represents the random initialization function, and Represents the hidden layer weights after random initialization and bias .

[0279] 2. Hidden layer conversion

[0280] For input data Apply hidden layer transformation to get hidden layer representation , which can be expressed as:

[0281]

[0282] in, is the Sigmoid activation function.

[0283] 3. Output weight calculation

[0284] Use smooth approximation method to calculate output weights , the calculation method can be expressed as:

[0285]

[0286] in, is a regularization parameter that controls model complexity to prevent overfitting. In one embodiment, the parameter Can be set to .

[0287] Furthermore, in solving the inverse matrix In the process, , this application uses Cholesky decomposition, where is decomposed into the product of two triangular matrices , then solve and The reverse, when When approaching singularity or instability, regularization terms are added To improve numerical stability. This process can be expressed as:

[0288]

[0289]

[0290] Further, ask and , to obtain .

[0291] 4. Adaptive neuromodulation

[0292] The adaptive neural regulation mechanism of the evolutionary algorithm is applied to adjust the hidden layer parameters. The adaptive neural regulation mechanism of the evolutionary algorithm used in this application can be expressed as:

[0293]

[0294] in, It is a parameter adjustment function based on evolutionary algorithm.

[0295] 5. Classifier training

[0296] The classifier is trained using the optimized ELM algorithm, which can be expressed as:

[0297]

[0298] Among them, TrainELM() is the function for training the extreme learning machine.

[0299] 6. Output classification results

[0300] Use the trained classifier to classify new input data For classification, it can be expressed as:

[0301]

[0302] Among them, Classify() is a function that uses the trained classifier for classification.

[0303] After the training of the feature extraction module, feature dimensionality reduction module and classifier is completed, the training of the Torreya grandis recognition model is completed, and the Torreya grandis species in the Torreya grandis image can be recognized by the trained Torreya grandis recognition model. The Torreya grandis image to be recognized is input into the trained Torreya grandis recognition model, and the Torreya grandis species is obtained by using the trained feature extraction module, feature dimensionality reduction module and classifier in the Torreya grandis recognition model.

[0304] The present invention adopts a generative adversarial network based on learning rate cyclic adjustment for data expansion, which can effectively generate high-quality and diverse training data, solving the problems of insufficient data volume and insufficient sample diversity in traditional methods. An optimized neural network model is used for feature extraction. This model combines the method of simulated quantum optimization to improve the efficiency and accuracy of feature extraction, while alleviating the problems of gradient vanishing and gradient explosion. The optimized sparse coding autoencoder is used for feature dimensionality reduction, combined with the principle of molecular dynamics and adaptive multi-scale feature fusion technology, which can effectively reduce the computational complexity while retaining key feature information. The classifier is trained using an extreme learning machine algorithm based on smooth approximation. This algorithm improves classification accuracy, reduces noise sensitivity, and enhances the generalization ability of the model. Finally, the accuracy and efficiency of Torreya grandis species identification are significantly improved, especially when processing Torreya grandis samples in different geographical locations, climatic conditions and growth stages, the recognition and classification performance can be effectively improved.

[0305] Based on the same inventive concept, the present invention also provides an electronic device including a memory and a processor coupled to each other, wherein the memory is used to store a computer program; the processor is used to read and execute the computer program stored in the memory, and when the computer program is executed, the processor executes the above-mentioned method for identifying Torreya grandis species.

[0306] If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods of each embodiment of the present application.

[0307] The above are merely embodiments of the present application, and are not intended to limit the patent scope of the present application. Any equivalent structural transformations made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for identifying Torreya grandis species, characterized in that: Includes steps: Presetting a Torreya grandis recognition model, identifying features in a Torreya grandis image by using the Torreya grandis recognition model, and identifying the type of the Torreya grandis according to the features in the Torreya grandis image; The Torreya grandis identification model comprises: A feature extraction module, wherein the feature extraction module optimizes the neural network parameters of the feature extraction module based on a simulated quantum optimization method to extract key features from the Torreya grandis image; A feature dimension reduction module, wherein the feature dimension reduction module performs feature dimension reduction on the key features based on a sparse coding autoencoder; A classifier module, wherein the classifier module classifies the key features after dimensionality reduction based on a smooth approximation extreme learning machine algorithm, and outputs the type of the Torreya grandis; The Torreya grandis recognition model is trained by a training data set, and the Torreya grandis type is identified by the trained Torreya grandis recognition model; the training data set is a preprocessed Torreya grandis image, and the Torreya grandis image includes Torreya grandis in different geographical locations, different climatic conditions and different growth stages; the Torreya grandis image includes different parts of the Torreya grandis; The preprocessing steps of the Torreya grandis image include: data labeling and data standardization; In the data annotation step, the Torreya grandis images in the training data set are annotated to distinguish different Torreya grandis species; the annotation category is the specific species of Torreya grandis, and each annotation corresponds to the Torreya grandis species of the corresponding Torreya grandis image; In the data standardization step: including adjusting the pixel mean and standard deviation, using image standardization to perform standardization processing, the standardization processing method is expressed as: in, is the original Torreya grandis image, is the pixel mean of the Torreya grandis image, is the pixel standard deviation of the Torreya grandis image; is the standardized image of Torreya grandis.

2. The method for identifying Torreya grandis species according to claim 1, characterized in that: The preprocessing step of the Torreya grandis image also includes data expansion; The data expansion is realized based on a generative adversarial network with cyclic adjustment of learning rate, and a learning rate cyclic adjustment mechanism is added to the training process of the generative adversarial network to dynamically adjust the learning rate according to different stages of training; The process of the generative adversarial network algorithm based on learning rate cyclic adjustment is as follows: The first step is initialization. Set the initial network parameters of the generator and discriminator. The generator parameters are ,in represents the parameter set of the generator; the discriminator parameters are ,in represents a set of parameters of the discriminator; The second step is to set a learning rate cyclic adjustment strategy to dynamically change the learning rate according to the training progress. The learning rate cyclic adjustment strategy adopts a periodic adjustment method, which is expressed as: in, is the current learning rate, and are the minimum and maximum learning rates, respectively, is the current training iteration number, is the maximum number of iterations; The third step is to train the generator, using random noise as input to generate new Torreya grandis image samples. The training process of the generator is expressed as: in, is the random noise of the input, is the new sample generated by the generator, is the loss function of the generator, is the batch size; The fourth step is to train the discriminator, evaluate the generated samples and the real samples, and optimize the discriminator to accurately distinguish the generated samples from the real samples. The training process of the discriminator is as follows: in, is the real sample data, and are the outputs of the discriminator for real and generated samples, respectively, is the loss function of the discriminator; The fifth step is adversarial training, repeating the training of the generator and the discriminator until a balance is reached; Step 6: Adjust the learning rate cyclically according to the strategy Adjust the learning rate during training; The seventh step is to output the data expansion result, add the optimized generated samples to the training data set to form an expanded training data set; input the expanded training data set into the Torreya grandis recognition model to train the Torreya grandis recognition model.

3. The method for identifying Torreya grandis species according to claim 2, characterized in that: The feature extraction module includes an input layer, a first hidden layer, a second hidden layer and a third hidden layer, and has the following structure: Input layer: contains Nin neurons, where Nin is used to stretch the data input to the Torreya grandis recognition model into the dimension of a one-dimensional vector; First hidden layer: Each layer of the first hidden layer has 128 neurons and uses ReLU as the activation function; Second hidden layer: Each layer of the first hidden layer has 256 neurons and uses ReLU as the activation function; Third hidden layer: Each layer of the first hidden layer has 100 neurons and uses ReLU as the activation function.

4. The method for identifying Torreya grandis species according to claim 3, characterized in that: The specific steps of optimizing the neural network parameters of the feature extraction module are as follows: The first step is parameter initialization. A quantum state is initialized for each parameter. Each quantum state represents a possible parameter value and an initial oscillation frequency is preset. The second step is to explore multiple quantum states. In each round of iteration, a new random offset is introduced for each quantum state to simulate the oscillation in the quantum system. Given the parameters of a neural network In its quantum state In the equation, the offset is expressed as: in, is the oscillation frequency; is a random number drawn from a uniform random distribution; Calculate the new parameter value and the corresponding fitness value. The way to update the parameter is expressed as: in, It is the new parameter value generated after exploring multiple quantum states; is the parameter value before multiple quantum state exploration; is the update amount of the parameter; Each parameter has its corresponding quantum state, and the fitness function is defined as follows: in, is the fitness function; is the mean square error loss function; is the quantum condensation term, which measures the synergistic effect of different quantum states; is a set that contains all quantum states; is a hyperparameter used to adjust the weight of the quantum condensation term; The quantum condensation term is defined as follows: in, is the number of quantum states; is the quantum condensation function, which is used to measure the two quantum states and The degree of cohesion between definition for: in, It is a quantum state and The Euclidean distance between is a constant; The third step is dynamic frequency adjustment. According to the changes in the fitness function, the oscillation frequency of each parameter is dynamically adjusted. According to the changes in the fitness function, the method of dynamically adjusting the frequency is expressed as: in, The frequency before dynamic frequency adjustment; The frequency after dynamic frequency adjustment; It is a manually preset learning rate; is the rate of change of the fitness function, which is calculated as follows: in, and are the losses of the current and previous iterations respectively; The fourth step, quantum collision, select two or more quantum states to collide; select two quantum states and The collision is simulated as follows: in, is the quantum state after the quantum collision, is a coefficient drawn from a uniform random distribution; Evaluate the newly generated quantum state and calculate its corresponding fitness value; the way to evaluate the newly generated quantum state is expressed as: in, is the new parameter value generated after the quantum collision; Calculate the new fitness value if but Alternative ; If this new state leads to a smaller fitness value, the collision is considered successful and the new quantum state is used for the next iteration; The fifth step is to simulate the interference effect. The quantum states of all parameters are merged to simulate quantum interference. Consider two quantum states and , and its interference result is expressed as: in, The interference coefficient is set by humans, which determines and weight in intervention; Updated parameter values ​​for interference effects; Select those quantum states that produce lower fitness values ​​and eliminate the rest; According to the selected optimal quantum state, update the parameters of the neural network, considering all and , select the quantum state that leads to the minimum fitness value for parameter update, which is expressed as: in, Neural network parameters determined for this iteration; and is the set sum of the parameters of the neural network and its quantum state in this iteration; Step 6, termination condition, repeat steps 2 to 5. If the preset number of iterations is reached, or the fitness function reaches a preset threshold, the algorithm is terminated, indicating that the training of the feature extraction module is completed.

5. The method for identifying Torreya grandis species according to claim 4, characterized in that: The training process of the optimized sparse coding autoencoder dimensionality reduction algorithm is as follows: The first step is to initialize the autoencoder and the encoder parameters of the autoencoder network. and decoder parameters ; The second step is sparsity adjustment, which is based on the sparsity adjustment mechanism of molecular dynamics principles and sets the sparsity parameters in the encoding process. ; Sparse energy function Defined as: in, Indicates the encoded Features, is the sparsity adjustment parameter; The third step is forward propagation. In the forward propagation process of the sparse autoencoder, filters of different scales are used. Extracting input data The characteristics are expressed as: in, Indicates different scales; The feature fusion weight calculation is performed, and the adaptive weight calculation method for each scale feature is expressed as: in, It is The variance of the scale feature, is the adjustment parameter; The fusion feature representation is performed, and the fusion feature is obtained by combining the features of different scales and the corresponding weights, which is expressed as: Input data and fusion features Input into the encoding of the autoencoder to obtain the encoding , the encoding process is expressed as: in, is the activation function, is the encoder bias term; After the decoder reconstructs , the decoding process is expressed as: in, is the bias term of the decoder; The fourth step is to calculate the loss function and calculate the reconstruction loss. And sparsity loss, the loss function contains reconstruction loss and sparsity loss, expressed as: in, is the number of input features, is a hyperparameter that controls the importance of sparsity loss; The fifth step is policy iteration optimization. The policy iteration method is applied to dynamically adjust the codec parameters according to the current loss. The policy iteration process is used to dynamically adjust the codec parameters as follows: in, is the learning rate, which is used to control the amplitude of parameter updates; Step 6: Output the dimension reduction data. After training and optimization, the autoencoder outputs the dimension reduction features. .

6. The method for identifying Torreya grandis species according to claim 5, characterized in that: The training process of the extreme learning machine algorithm based on smooth approximation is as follows: The first step is initialization. Initialize the hidden layer weights of the extreme learning machine and bias , the initialization method is random initialization, expressed as: Among them, Rand() represents the random initialization function, and Represents the hidden layer weights after random initialization and bias ; The second step is hidden layer conversion, which is used to transform the input data Apply hidden layer transformation to get hidden layer representation , expressed as: in, is the Sigmoid activation function; The third step is to calculate the output weight using a smooth approximation method. , the calculation method is expressed as: in, is the regularization parameter; Solving the inverse matrix In the process, , using Cholesky decomposition, where is decomposed into the product of two triangular matrices , then solve and The reverse, when When approaching singularity or instability, add regularization terms To improve numerical stability; this process is expressed as: beg and , to obtain ; The fourth step is adaptive neural regulation. The adaptive neural regulation mechanism of the evolutionary algorithm is used to adjust the hidden layer parameters. The adaptive neural regulation mechanism of the evolutionary algorithm is expressed as: in, It is a parameter adjustment function based on evolutionary algorithm; The fifth step is classifier training, using the optimized extreme learning machine to train the classifier, which is expressed as: Among them, TrainELM() is the function for training the extreme learning machine; Step 6: Output the classification results and use the trained classifier to classify the new input data. Classification is performed, expressed as: Among them, Classify() is a function that uses the trained classifier for classification.

7. An electronic device, characterized in that: It comprises a memory and a processor coupled to each other, wherein the processor is used to execute program instructions stored in the memory to implement the method for identifying Torreya grandis species as described in any one of claims 1 to 6.

8. A computer-readable storage medium having program instructions stored thereon, characterized in that: When the program instructions are executed by a processor, the method for identifying Torreya grandis species described in any one of claims 1 to 6 is implemented.

Images