Image recognition method and system based on deep neural network optimization algorithm

Abstract

The invention discloses an image recognition method based on a deep neural network optimization algorithm, and relates to the technical field of image recognition, and the method comprises the steps: carrying out the preprocessing of an image data set, such as adaptive normalization and Hash coding label; coding a network structure according to a graph embedding technology, and mixing and initializing a weight threshold value; algorithm optimization features are fused, particle swarm initial determination is carried out, and then channel characterization fine tuning is carried out; using the training set to improve back propagation and adaptive moment estimation algorithm training, fusing a loss function and dynamically adjusting a learning rate; calculating evaluation and optimization of multiple indexes in the verification set, such as a genetic algorithm search structure and a Bayesian modulation learning rate; and inputting a to-be-recognized image into the optimized network, and obtaining a result by using an integrated recognition method. According to the invention, by optimizing the links of image preprocessing, neural network and recognition process, the image quality and data processing efficiency are improved, the features are accurately determined, the network training is accelerated, the network is comprehensively evaluated and optimized, the reliability of the recognition result is improved, and the image recognition precision and efficiency are improved.

Description

Technical Field

[0001] The present invention relates to the field of image recognition technology, and in particular to an image recognition method and system based on a deep neural network optimization algorithm. Background Art

[0002] As one of the core technologies in the field of computer vision, image recognition technology has been widely used in many fields, such as security monitoring, autonomous driving, medical imaging diagnosis, industrial detection, etc. With the increasing complexity of application scenarios and the continuous improvement of recognition accuracy requirements, traditional image recognition methods face many challenges.

[0003] Traditional image recognition methods are mainly based on manual feature extraction and machine learning classifiers. For example, by extracting features such as color, texture, and shape of an image, and then using classifiers such as support vector machines and decision trees for classification. However, these manual features often require manual design and selection, and their expressive power is limited for complex image data, making it difficult to capture the deep semantic information of the image.

[0004] The emergence of deep neural networks has brought a revolutionary breakthrough in image recognition. Deep neural networks have powerful automatic feature learning capabilities and can automatically learn effective feature representations from a large amount of image data. For example, convolutional neural networks (CNNs) automatically extract local features of images through convolutional layers and pooling layers, and fully connected layers integrate and classify these features. However, as the depth of the network increases, deep neural networks also face some problems.

[0005] On the one hand, the training process of deep neural networks usually requires a lot of computing resources and time. Large-scale image datasets and complex network structures make the training process very time-consuming, especially when hardware resources are limited. On the other hand, deep neural networks are prone to overfitting, that is, they perform well on the training set but degrade on the test set. The overfitting problem is mainly due to too many network parameters, which causes the model to overfit the training data and fail to generalize to new data.

[0006] In order to solve these problems, researchers have proposed various deep neural network optimization algorithms. For example, by adjusting the network structure, such as using lightweight design to reduce network parameters; using regularization technology to prevent overfitting; using advanced optimizers to speed up training, etc. However, existing optimization algorithms often only improve one or several problems, lacking a comprehensive and efficient optimization method that can simultaneously solve multiple problems of deep neural networks in image recognition and improve the accuracy and efficiency of image recognition.

[0007] Therefore, the present invention aims to provide an image recognition system and method based on a deep neural network optimization algorithm, which comprehensively considers multiple links such as image preprocessing, neural network initialization, feature optimization, network training, model evaluation and optimization, and improves the performance of deep neural networks in image recognition through innovative algorithms and technologies to meet the increasingly complex practical application needs. Summary of the invention

[0008] The present invention proposes an image recognition method and system based on a deep neural network optimization algorithm to solve the problems mentioned in the above-mentioned prior art.

[0009] In order to achieve the above object, the present invention adopts the following technical solution: an image recognition method based on a deep neural network optimization algorithm, comprising the following steps:

[0010] S1. Image preprocessing steps:

[0011] Divide the input image dataset into training set, validation set and test set according to the distribution characteristics and category ratio of the data to ensure that each set is representative;

[0012] The images in the training set and the validation set are normalized. The improved adaptive normalization algorithm is used to dynamically adjust the normalization parameters according to the local features of the image. The formula is: Where μ(x, r) and σ(x, r) are the mean and standard deviation of image x in local region r, respectively;

[0013] One-hot encoding is performed on the label data. In order to improve the encoding efficiency, hash encoding technology is used to compress and store the one-hot encoding results.

[0014] According to the noise type and distribution of the image, a hybrid denoising method is used to denoise the image by combining wavelet transform and non-local mean filtering to improve the image quality.

[0015] For high-dimensional image data, a dimensionality reduction method based on manifold learning is used to map the image data to a low-dimensional manifold space, thereby reducing the data dimension while retaining key features;

[0016] S2, neural network initialization steps:

[0017] Encode the deep neural network structure according to an innovative encoding method. This encoding method is based on graph embedding technology in graph theory, which maps the layer structure and node connection relationship of the neural network to a low-dimensional vector space to improve the representation ability of the network structure.

[0018] The weights and thresholds of the neural network are initialized using a hybrid initialization method that combines normal distribution initialization and uniform distribution initialization. The initialization method is dynamically selected according to the different functions and positions of the network layer. For the network layer close to the input layer, uniform distribution initialization is used. The formula is W [l] ~U(-a, a), where a is a constant determined according to the range of input data; for the network layer close to the output layer, the normal distribution is used for initialization, and the formula is in is the standard deviation determined empirically or experimentally;

[0019] S3, feature optimization steps:

[0020] Determine the abstract dimension and required number of features of the deep neural network for image data through an innovative optimization algorithm that combines the advantages of the particle swarm algorithm and the algorithm based on channel characterization capability;

[0021] Firstly, the global search capability of the particle swarm algorithm is used to preliminarily determine the approximate range and number of features. The abstract dimension of the image dataset features is represented as the dimension of the particle position in the particle swarm algorithm. The number of features in each dimension of the image dataset is represented as the size of each dimension of the particle position in the particle swarm algorithm. The position and speed of each particle in the particle swarm are randomly initialized within the initial range. In the optimization process of the particle swarm algorithm, dynamically adjusted acceleration constants c1 and c2 are used. The adjustment formula is: and in and is the initial acceleration constant, t is the current iteration number, and T is the total iteration number; the inertia weight w adopts the adaptive adjustment formula w(t)=w 0 *e -λt , where w 0 is the initial inertia weight, λ is the attenuation coefficient;

[0022] Then, the algorithm based on the channel representation ability makes fine adjustments to the preliminarily determined features. Insert batch normalization layers, linear rectification layers, and average pooling layers into the deep neural network model to construct an auxiliary classifier, and insert a cross entropy loss function to form a target loss function with the reconstruction loss function of the deep neural network model; use the target loss function combined with the channel representation ability to train the auxiliary classifier to obtain the target model; during the training process, obtain the channel selection vector that represents the importance of each channel in the deep neural network model, use the target loss function to optimize the channel selection vector and model parameters, determine the channels with vector elements of the optimized channel selection vector being 0 as redundant channels, and obtain the optimized model structure by pruning the channels. In the algorithm based on channel representation ability, an adaptive adjustment strategy is used for the weight of the cross entropy loss function, and the value is dynamically adjusted according to the change of the loss function during the training process;

[0023] S4. Network training steps:

[0024] The optimized neural network is trained using the training set. During the training process, the improved back propagation algorithm is used to optimize the weights and bias values ​​of the neural network. In the back propagation algorithm, the formulas for calculating the weight gradient, bias gradient, and input gradient of the layer are respectively anddA [l-1] =W [l]T ·dZ [l] , where m is the batch size, np is the sum function in the numpy library, and dZ [l] is the output gradient of the lth layer. At the same time, in order to improve the training efficiency, a variant algorithm of stochastic gradient descent is used, specifically: the adaptive moment estimation (Adam) algorithm, combined with the momentum term and the adaptive learning rate adjustment mechanism, to accelerate the network convergence;

[0025] The loss function is calculated based on the output of the neural network and the actual label. The loss function uses a fusion loss function, which is a weighted combination of the cross entropy loss function and the mean square error loss function. The formula is L = α*L CE +(1-α)*L MSE , where L CE is the cross entropy loss function, L MSE is the mean square error loss function, α is the weight coefficient, the value of α is dynamically adjusted according to the error distribution during training, the parameters of the neural network are adjusted according to the value of the loss function, and a dynamic learning rate adjustment strategy is adopted. The learning rate adjustment formula is: where η 0 is the initial learning rate, t is the current iteration number, T is the total iteration number, and β is the adjustment coefficient.

[0026] S5. Model evaluation and optimization steps:

[0027] The accuracy, recall, F1 value and other evaluation indicators of the neural network are calculated on the validation set to comprehensively evaluate the performance of the neural network. The accuracy calculation formula is: The recall calculation formula is: The formula for calculating the F1 value is: in

[0028] The neural network is optimized according to the evaluation results. The optimization methods include adjusting the structure of the neural network, adjusting the learning rate, and increasing the number of training times. When adjusting the structure of the neural network, a structure search method based on a genetic algorithm is used to encode the structural parameters of the neural network into chromosomes, and the optimal network structure is found through the selection, crossover and mutation operations of the genetic algorithm. When adjusting the learning rate, a learning rate adjustment method based on Bayesian optimization is used to dynamically adjust the value of the learning rate according to the changes in the evaluation indicators. When increasing the number of training times, the number of additional training times is dynamically determined according to the convergence of the evaluation indicators to avoid overtraining.

[0029] S6, image recognition step:

[0030] The image to be recognized is input into the optimized neural network to obtain the recognition result of the image. During the recognition process, in order to improve the reliability of the recognition result, an integrated recognition method is adopted to integrate multiple optimized neural networks, perform multiple recognitions on the image, and then determine the final recognition result based on a voting mechanism or a weighted average mechanism.

[0031] Furthermore, in the image preprocessing step:

[0032] The local area in the adaptive normalization algorithm is dynamically determined according to the resolution and feature distribution of the image. Generally, a small local area of ​​the image is taken, with a size of K*K pixels, where the value range of K is [3, 15];

[0033] Hash coding technology uses the MurmurHash algorithm to compress and store one-hot coding results, with a compression ratio of more than 80%.

[0034] In the hybrid denoising method, the weights of wavelet transform and non-local mean filter are dynamically adjusted according to the noise level of the image. When the noise level is high, the weight of wavelet transform is increased; when the noise level is low, the weight of non-local mean filter is increased.

[0035] The dimensionality reduction method based on manifold learning uses the local linear embedding (LLE) algorithm or the isometric mapping (Isomap) algorithm to map the image data into a low-dimensional manifold space. The dimension of the low-dimensional space is determined according to the intrinsic dimension of the data and application requirements, and is generally set to [50, 200].

[0036] Furthermore, in the neural network initialization step:

[0037] In the encoding method based on graph embedding technology in graph theory, the layer structure and node connection relationship of the neural network are mapped to a low-dimensional vector space with a dimension of [5, 20], and the embedding algorithm used is the node2vec algorithm or the deepwalk algorithm.

[0038] In the hybrid initialization method, for the network layer close to the input layer, the constant of uniform distribution initialization is determined according to the range of input data, and the value range is [0.1, 1]. For the network layer close to the output layer, the standard deviation of normal distribution initialization is determined according to the output range of the network layer and the expected output variance, and the value range is [0.01, 0.1].

[0039] Furthermore, in the feature optimization step:

[0040] In the particle swarm algorithm, the initial acceleration constant and The value range is [1.2, 1.8], and the initial inertia weight w 0 The value range is [0.6, 0.9].

[0041] In the algorithm based on channel representation capability, the weight of the cross entropy loss function ranges from [0.2, 0.8], and the value is dynamically adjusted according to the change of the loss function during training, with an adjustment step size of [0.01, 0.1].

[0042] Furthermore, in the network training step:

[0043] In the adaptive moment estimation Adam algorithm, the coefficient of the momentum term ranges from [0.9, 0.99], and the coefficient of the adaptive learning rate adjustment mechanism ranges from [0.999, 0.1].

[0044] Furthermore, in the model evaluation and optimization step:

[0045] In the structure search method based on genetic algorithm, the coding length of the chromosome is determined according to the structural complexity of the neural network, and the general value is [100, 500]. The population size of the genetic algorithm is [50, 200], the crossover probability is [0.6, 0.9], and the mutation probability is [0.01, 0.1].

[0046] In the learning rate adjustment method based on Bayesian optimization, the prior distribution adopts Gaussian distribution with a mean of [0.01, 0.1] and a variance of [0.0001, 0.001], and the acquisition function adopts the expected improved EI function or the probability improved PI function.

[0047] Furthermore, a system for implementing the image recognition method based on the deep neural network optimization algorithm includes:

[0048] Image preprocessing module: used to preprocess the input image data set, including dividing it into training set, validation set and test set, and using adaptive normalization algorithm, hash coding technology, hybrid denoising method and dimensionality reduction method based on manifold learning to process the image;

[0049] Neural network initialization module: used to encode the deep neural network structure according to the encoding method based on the graph embedding technology in graph theory, and use the hybrid initialization method to initialize the weights and thresholds of the neural network;

[0050] Feature optimization module: used to determine the abstract dimension and required number of features of the deep neural network for image data through an optimization algorithm that combines the advantages of the particle swarm algorithm and the algorithm based on channel representation ability;

[0051] Network training module: used to train the optimized neural network using the training set, and optimize the weights and bias values ​​of the neural network using the adaptive moment estimation Adam algorithm and the fusion loss function;

[0052] Model evaluation and optimization module: used to calculate the accuracy, recall rate, and F1 value evaluation indicators of the neural network on the validation set, and optimize the neural network based on the evaluation results using a structure search method based on a genetic algorithm, a learning rate adjustment method based on Bayesian optimization, and a method of dynamically increasing the number of training times;

[0053] Image recognition module: used to input the image to be recognized into the optimized neural network and obtain the recognition result of the image using the integrated recognition method.

[0054] Furthermore, in the image preprocessing module:

[0055] The adaptive normalization algorithm unit dynamically determines the local area according to the resolution and feature distribution of the image, and uses the corresponding formula to perform normalization processing;

[0056] The hash coding technology unit uses the MurmurHash algorithm to compress and store the one-hot coding results;

[0057] The hybrid denoising method unit dynamically adjusts the weights of wavelet transform and non-local mean filter according to the noise level of the image to perform denoising;

[0058] The dimensionality reduction method unit based on manifold learning uses the Local Linear Embedding (LLE) algorithm or the Isomap algorithm to map the image data into a low-dimensional manifold space, and determines the dimension of the low-dimensional space according to the intrinsic dimension of the data and application requirements.

[0059] Furthermore, in the neural network initialization module:

[0060] The coding unit based on graph embedding technology in graph theory maps the layer structure and node connection relationship of the neural network to a low-dimensional vector space and embeds it using the node2vec algorithm or the deepwalk algorithm;

[0061] The hybrid initialization method unit uses corresponding initialization methods to initialize weights and bias values ​​according to the position and function of the network layer.

[0062] Compared with the prior art, the present invention has the following beneficial effects:

[0063] Image preprocessing: The improved adaptive normalization algorithm, hybrid denoising method and dimensionality reduction method based on manifold learning are used to improve image quality, reduce data dimension, and retain key features, providing a better data foundation for subsequent neural network training.

[0064] Neural network initialization: The graph theory-based encoding method and hybrid initialization method improve the representation ability of the network structure, make the network initialization more reasonable, and help improve the training effect and generalization ability of the network.

[0065] Feature optimization: Combining the advantages of the particle swarm algorithm and the channel characterization capability-based algorithm, it can more accurately determine the abstract dimension of network features and the required number of features, reduce redundant features, and improve the feature expression ability of the network.

[0066] Network training: The improved back-propagation algorithm, adaptive moment estimation algorithm and fusion loss function are used to accelerate the network convergence speed and improve the training efficiency. At the same time, the dynamic learning rate adjustment strategy and loss function weight adjustment are used to further optimize the network training effect.

[0067] Model evaluation and optimization: Taking into account multiple evaluation indicators such as accuracy, recall rate, F1 value, etc., the use of structural search methods based on genetic algorithms, learning rate adjustment methods based on Bayesian optimization, and methods of dynamically increasing the number of training times can more comprehensively evaluate network performance and effectively optimize the network based on the evaluation results, thereby improving the network's generalization ability and recognition accuracy.

[0068] Image recognition: The integrated recognition method is used to improve the reliability and accuracy of the recognition results, which can better meet the image recognition needs in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 This is a schematic block diagram of an image recognition method based on a deep neural network optimization algorithm proposed in the present invention. DETAILED DESCRIPTION

[0070] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0071] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as limiting the present invention.

[0072] In addition, the terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined. In addition, the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, and it can be the internal connection of two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances. The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0073] Reference Figure 1 :An image recognition method based on a deep neural network optimization algorithm comprises the following steps:

[0074] S1. Image preprocessing steps:

[0075] Divide the input image dataset into training set, validation set and test set according to the distribution characteristics and category ratio of the data to ensure that each set is representative;

[0076] The images in the training set and the validation set are normalized. The improved adaptive normalization algorithm is used to dynamically adjust the normalization parameters according to the local features of the image. The formula is: Where μ(x, r) and σ(x, r) are the mean and standard deviation of image x in local region r, respectively;

[0077] One-hot encoding is performed on the label data. In order to improve the encoding efficiency, hash encoding technology is used to compress and store the one-hot encoding results.

[0078] According to the noise type and distribution of the image, a hybrid denoising method is used to denoise the image by combining wavelet transform and non-local mean filtering to improve the image quality.

[0079] For high-dimensional image data, a dimensionality reduction method based on manifold learning is used to map the image data to a low-dimensional manifold space, thereby reducing the data dimension while retaining key features;

[0080] S2, neural network initialization steps:

[0081] Encode the deep neural network structure according to an innovative encoding method. This encoding method is based on graph embedding technology in graph theory, which maps the layer structure and node connection relationship of the neural network to a low-dimensional vector space to improve the representation ability of the network structure.

[0082] The weights and thresholds of the neural network are initialized using a hybrid initialization method that combines normal distribution initialization and uniform distribution initialization. The initialization method is dynamically selected according to the different functions and positions of the network layer. For the network layer close to the input layer, uniform distribution initialization is used. The formula is W [l] ~U(-a, a), where a is a constant determined according to the range of input data; for the network layer close to the output layer, the normal distribution is used for initialization, and the formula is in is the standard deviation determined empirically or experimentally;

[0083] S3, feature optimization steps:

[0084] Determine the abstract dimension and required number of features of the deep neural network for image data through an innovative optimization algorithm that combines the advantages of the particle swarm algorithm and the algorithm based on channel characterization capability;

[0085] Firstly, the global search capability of the particle swarm algorithm is used to preliminarily determine the approximate range and number of features. The abstract dimension of the image dataset features is represented as the dimension of the particle position in the particle swarm algorithm. The number of features in each dimension of the image dataset is represented as the size of each dimension of the particle position in the particle swarm algorithm. The position and speed of each particle in the particle swarm are randomly initialized within the initial range. In the optimization process of the particle swarm algorithm, dynamically adjusted acceleration constants c1 and c2 are used. The adjustment formula is: and in and is the initial acceleration constant, t is the current iteration number, and T is the total iteration number; the inertia weight w adopts the adaptive adjustment formula w(t)=w 0 *e -λt , where w 0 is the initial inertia weight, λ is the attenuation coefficient;

[0086] Then, the algorithm based on the channel representation ability makes fine adjustments to the preliminarily determined features. Insert batch normalization layers, linear rectification layers, and average pooling layers into the deep neural network model to construct an auxiliary classifier, and insert a cross entropy loss function to form a target loss function with the reconstruction loss function of the deep neural network model; use the target loss function combined with the channel representation ability to train the auxiliary classifier to obtain the target model; during the training process, obtain the channel selection vector that represents the importance of each channel in the deep neural network model, use the target loss function to optimize the channel selection vector and model parameters, determine the channels with vector elements of the optimized channel selection vector being 0 as redundant channels, and obtain the optimized model structure by pruning the channels. In the algorithm based on channel representation ability, an adaptive adjustment strategy is used for the weight of the cross entropy loss function, and the value is dynamically adjusted according to the change of the loss function during the training process;

[0087] S4. Network training steps:

[0088] The optimized neural network is trained using the training set. During the training process, the improved back propagation algorithm is used to optimize the weights and bias values ​​of the neural network. In the back propagation algorithm, the formulas for calculating the weight gradient, bias gradient, and input gradient of the layer are respectively anddA [l-1] =W [l]T ·dZ [l] , where m is the batch size, np is the sum function in the numpy library, and dZ [l] is the output gradient of the lth layer. At the same time, in order to improve the training efficiency, a variant algorithm of stochastic gradient descent is used, specifically: the adaptive moment estimation (Adam) algorithm, combined with the momentum term and the adaptive learning rate adjustment mechanism, to accelerate the network convergence;

[0089] The loss function is calculated based on the output of the neural network and the actual label. The loss function uses a fusion loss function, which is a weighted combination of the cross entropy loss function and the mean square error loss function. The formula is L = α*L CE +(1-α)*L MSE , where L CE is the cross entropy loss function, L MSEis the mean square error loss function, α is the weight coefficient, the value of α is dynamically adjusted according to the error distribution during training, the parameters of the neural network are adjusted according to the value of the loss function, and a dynamic learning rate adjustment strategy is adopted. The learning rate adjustment formula is: where η 0 is the initial learning rate, t is the current iteration number, T is the total iteration number, and β is the adjustment coefficient.

[0090] S5. Model evaluation and optimization steps:

[0091] The accuracy, recall, F1 value and other evaluation indicators of the neural network are calculated on the validation set to comprehensively evaluate the performance of the neural network. The accuracy calculation formula is: The recall calculation formula is: The formula for calculating the F1 value is: in

[0092] The neural network is optimized according to the evaluation results. The optimization methods include adjusting the structure of the neural network, adjusting the learning rate, and increasing the number of training times. When adjusting the structure of the neural network, a structure search method based on a genetic algorithm is used to encode the structural parameters of the neural network into chromosomes, and the optimal network structure is found through the selection, crossover and mutation operations of the genetic algorithm. When adjusting the learning rate, a learning rate adjustment method based on Bayesian optimization is used to dynamically adjust the value of the learning rate according to the changes in the evaluation indicators. When increasing the number of training times, the number of additional training times is dynamically determined according to the convergence of the evaluation indicators to avoid overtraining.

[0093] S6, image recognition step:

[0094] The image to be recognized is input into the optimized neural network to obtain the recognition result of the image. During the recognition process, in order to improve the reliability of the recognition result, an integrated recognition method is adopted to integrate multiple optimized neural networks, perform multiple recognitions on the image, and then determine the final recognition result based on a voting mechanism or a weighted average mechanism.

[0095] In the present invention, in the image preprocessing step:

[0096] The local area in the adaptive normalization algorithm is dynamically determined according to the resolution and feature distribution of the image. Generally, a small local area of ​​the image is taken, with a size of K*K pixels, where the value range of K is [3, 15];

[0097] Hash coding technology uses the MurmurHash algorithm to compress and store one-hot coding results, with a compression ratio of more than 80%.

[0098] In the hybrid denoising method, the weights of wavelet transform and non-local mean filter are dynamically adjusted according to the noise level of the image. When the noise level is high, the weight of wavelet transform is increased; when the noise level is low, the weight of non-local mean filter is increased.

[0099] The dimensionality reduction method based on manifold learning uses the local linear embedding (LLE) algorithm or the isometric mapping (Isomap) algorithm to map the image data into a low-dimensional manifold space. The dimension of the low-dimensional space is determined according to the intrinsic dimension of the data and application requirements, and is generally set to [50, 200].

[0100] In the present invention, in the neural network initialization step:

[0101] In the encoding method based on graph embedding technology in graph theory, the layer structure and node connection relationship of the neural network are mapped to a low-dimensional vector space with a dimension of [5, 20], and the embedding algorithm used is the node2vec algorithm or the deepwalk algorithm.

[0102] In the hybrid initialization method, for the network layer close to the input layer, the constant of uniform distribution initialization is determined according to the range of input data, and the value range is [0.1, 1]. For the network layer close to the output layer, the standard deviation of normal distribution initialization is determined according to the output range of the network layer and the expected output variance, and the value range is [0.01, 0.1].

[0103] In the present invention, in the feature optimization step:

[0104] In the particle swarm algorithm, the initial acceleration constant and The value range is [1.2, 1.8], and the initial inertia weight w 0 The value range is [0.6, 0.9].

[0105] In the algorithm based on channel representation capability, the weight of the cross entropy loss function ranges from [0.2, 0.8], and the value is dynamically adjusted according to the change of the loss function during training, with an adjustment step size of [0.01, 0.1].

[0106] In the present invention, in the network training step:

[0107] In the adaptive moment estimation Adam algorithm, the coefficient of the momentum term ranges from [0.9, 0.99], and the coefficient of the adaptive learning rate adjustment mechanism ranges from [0.999, 0.1].

[0108] In the present invention, in the model evaluation and optimization step:

[0109] In the structure search method based on genetic algorithm, the coding length of the chromosome is determined according to the structural complexity of the neural network, and the general value is [100, 500]. The population size of the genetic algorithm is [50, 200], the crossover probability is [0.6, 0.9], and the mutation probability is [0.01, 0.1].

[0110] In the learning rate adjustment method based on Bayesian optimization, the prior distribution adopts Gaussian distribution with a mean of [0.01, 0.1] and a variance of [0.0001, 0.001], and the acquisition function adopts the expected improved EI function or the probability improved PI function.

[0111] In the present invention, a system for implementing the image recognition method based on the deep neural network optimization algorithm comprises:

[0112] Image preprocessing module: used to preprocess the input image data set, including dividing it into training set, validation set and test set, and using adaptive normalization algorithm, hash coding technology, hybrid denoising method and dimensionality reduction method based on manifold learning to process the image;

[0113] Neural network initialization module: used to encode the deep neural network structure according to the encoding method based on the graph embedding technology in graph theory, and use the hybrid initialization method to initialize the weights and thresholds of the neural network;

[0114] Feature optimization module: used to determine the abstract dimension and required number of features of the deep neural network for image data through an optimization algorithm that combines the advantages of the particle swarm algorithm and the algorithm based on channel representation ability;

[0115] Network training module: used to train the optimized neural network using the training set, and optimize the weights and bias values ​​of the neural network using the adaptive moment estimation Adam algorithm and the fusion loss function;

[0116] Model evaluation and optimization module: used to calculate the accuracy, recall rate, and F1 value evaluation indicators of the neural network on the validation set, and optimize the neural network based on the evaluation results using a structure search method based on a genetic algorithm, a learning rate adjustment method based on Bayesian optimization, and a method of dynamically increasing the number of training times;

[0117] Image recognition module: used to input the image to be recognized into the optimized neural network and obtain the recognition result of the image using the integrated recognition method.

[0118] In the present invention, in the image preprocessing module:

[0119] The adaptive normalization algorithm unit dynamically determines the local area according to the resolution and feature distribution of the image, and uses the corresponding formula to perform normalization processing;

[0120] The hash coding technology unit uses the MurmurHash algorithm to compress and store the one-hot coding results;

[0121] The hybrid denoising method unit dynamically adjusts the weights of wavelet transform and non-local mean filter according to the noise level of the image to perform denoising;

[0122] The dimensionality reduction method unit based on manifold learning uses the Local Linear Embedding (LLE) algorithm or the Isomap algorithm to map the image data into a low-dimensional manifold space, and determines the dimension of the low-dimensional space according to the intrinsic dimension of the data and application requirements.

[0123] In the present invention, in the neural network initialization module:

[0124] The coding unit based on graph embedding technology in graph theory maps the layer structure and node connection relationship of the neural network to a low-dimensional vector space and embeds it using the node2vec algorithm or the deepwalk algorithm;

[0125] The hybrid initialization method unit uses corresponding initialization methods to initialize weights and bias values ​​according to the position and function of the network layer.

[0126] The above are only preferred specific implementation modes of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical solutions and inventive concepts of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. An image recognition method based on a deep neural network optimization algorithm, characterized in that: The following steps are involved: S1. Image preprocessing steps: Divide the input image data set into training set, validation set and test set according to the distribution characteristics and category ratio of the data; The images in the training set and the validation set are normalized. The improved adaptive normalization algorithm is used to dynamically adjust the normalization parameters according to the local features of the image. The formula is: Where μ(x, r) and σ(x, r) are the mean and standard deviation of image x in local region r, respectively; One-hot encodes the label data and uses hash coding technology to compress and store the one-hot encoding results; According to the noise type and distribution of the image, a hybrid denoising method is used to denoise the image by combining wavelet transform and non-local mean filtering. For high-dimensional image data, a dimensionality reduction method based on manifold learning is used to map the image data to a low-dimensional manifold space, thereby reducing the data dimension while retaining key features; S2, neural network initialization steps: Encode the deep neural network structure. The encoding method is based on the graph embedding technology in graph theory, which maps the layer structure and node connection relationship of the neural network into a low-dimensional vector space. The weights and thresholds of the neural network are initialized using a hybrid initialization method that combines normal distribution initialization and uniform distribution initialization. The initialization method is dynamically selected according to the different functions and positions of the network layer. For the network layer close to the input layer, uniform distribution initialization is used. The formula is W [l] ~U(-a, a), where a is a constant determined according to the range of input data; for the network layer close to the output layer, the normal distribution is used for initialization, and the formula is in is the standard deviation determined empirically or experimentally; S3, feature optimization steps: The abstract dimension and required number of features of the deep neural network for the image data are determined by the optimization algorithm. The optimization algorithm combines the advantages of the particle swarm algorithm and the algorithm based on channel characterization capability, including: The global search capability of the particle swarm algorithm is used to preliminarily determine the range and number of features. The abstract dimension of the image data set features is represented as the dimension of the particle position in the particle swarm algorithm. The number of features in each dimension of the image data set is represented as the size of each dimension of the particle position in the particle swarm algorithm. The position and speed of each particle in the particle swarm are randomly initialized within the initial range. In the optimization process of the particle swarm algorithm, dynamically adjusted acceleration constants c1 and c2 are used. The adjustment formula is: and in and is the initial acceleration constant, t is the current iteration number, and T is the total iteration number; the inertia weight w adopts the adaptive adjustment formula w(t)=w 0 *e -λt , where w 0 is the initial inertia weight, λ is the attenuation coefficient; The algorithm based on channel representation ability adjusts the initially determined features, inserts batch normalization layer, linear rectification layer, and average pooling layer into the deep neural network model to construct an auxiliary classifier, and inserts the cross entropy loss function to form a target loss function with the reconstruction loss function of the deep neural network model; the auxiliary classifier is trained using the target loss function combined with the channel representation ability to obtain the target model; during the training process, a channel selection vector representing the importance of each channel in the deep neural network model is obtained, the channel selection vector and model parameters are optimized using the target loss function, and the channels whose vector elements of the optimized channel selection vector are 0 are determined as redundant channels, and the optimized model structure is obtained by pruning the channels. In the algorithm based on channel representation ability, an adaptive adjustment strategy is adopted for the weight of the cross entropy loss function, and the value is dynamically adjusted according to the change of the loss function during the training process; S4. Network training steps: The optimized neural network is trained using the training set. During the training process, the improved back propagation algorithm is used to optimize the weights and bias values ​​of the neural network. In the back propagation algorithm, the formulas for calculating the weight gradient, bias gradient, and input gradient of the layer are respectively anddA [l-1] =W [l] T·dZ [l] , where m is the batch size, np is the sum function in the numpy library, and dZ [l] is the output gradient of the lth layer. At the same time, in order to improve the training efficiency, a variant algorithm of stochastic gradient descent is used, specifically: the adaptive moment estimation Adam algorithm, combined with the momentum term and the adaptive learning rate adjustment mechanism, to accelerate the network convergence; The loss function is calculated based on the output of the neural network and the actual label. The loss function uses a fusion loss function, which is a weighted combination of the cross entropy loss function and the mean square error loss function. The formula is L = α*L CE +(1-α)*L MSE , where L CE is the cross entropy loss function, L MSE is the mean square error loss function, α is the weight coefficient, the value of α is dynamically adjusted according to the error distribution during training, the parameters of the neural network are adjusted according to the value of the loss function, and a dynamic learning rate adjustment strategy is adopted. The learning rate adjustment formula is: where η 0 is the initial learning rate, t is the current number of iterations, T is the total number of iterations, and β is the adjustment coefficient; S5. Model evaluation and optimization steps: The accuracy, recall, F1 value and other evaluation indicators of the neural network are calculated on the validation set to comprehensively evaluate the performance of the neural network. The accuracy calculation formula is: The recall calculation formula is: The formula for calculating the F1 value is: in The neural network is optimized according to the evaluation results. The optimization methods include adjusting the structure of the neural network, adjusting the learning rate, and increasing the number of training times. When adjusting the structure of the neural network, a structure search method based on a genetic algorithm is used to encode the structural parameters of the neural network into chromosomes, and the network structure is found through the selection, crossover and mutation operations of the genetic algorithm. When adjusting the learning rate, a learning rate adjustment method based on Bayesian optimization is used to dynamically adjust the value of the learning rate according to the changes in the evaluation indicators. When increasing the number of training times, the number of training times to be increased is dynamically determined according to the convergence of the evaluation indicators. S6, image recognition step: The image to be recognized is input into the optimized neural network to obtain the recognition result of the image. During the recognition process, in order to improve the reliability of the recognition result, an integrated recognition method is adopted to integrate multiple optimized neural networks, perform multiple recognitions on the image, and then determine the final recognition result based on a voting mechanism or a weighted average mechanism.

2. The image recognition method based on a deep neural network optimization algorithm according to claim 1, characterized in that: In the image preprocessing step: The local area in the adaptive normalization algorithm is dynamically determined according to the resolution and feature distribution of the image. The local area of ​​the image is taken with a size of K*K pixels, where the value range of K is [3, 15]; Hash coding technology uses the MurmurHash algorithm to compress and store one-hot coding results, with a compression ratio of more than 80%; In the hybrid denoising method, the weights of wavelet transform and non-local mean filtering are dynamically adjusted according to the noise level of the image. When the noise level is high, the weight of wavelet transform is increased; When the noise level is low, increase the weight of the non-local means filter; The dimensionality reduction method based on manifold learning uses the local linear embedding (LLE) algorithm or the isometric mapping (Isomap) algorithm to map the image data into a low-dimensional manifold space. The dimension of the low-dimensional space is determined according to the intrinsic dimension of the data and application requirements, and is generally set to [50, 200].

3. The image recognition method based on a deep neural network optimization algorithm according to claim 1, characterized in that: In the neural network initialization step: In the encoding method based on graph embedding technology in graph theory, the layer structure and node connection relationship of the neural network are mapped to a low-dimensional vector space with a dimension of [5, 20], and the embedding algorithm used is the node2vec algorithm or the deepwalk algorithm; In the hybrid initialization method, for the network layer close to the input layer, the constant of uniform distribution initialization is determined according to the range of input data, and the value range is [0.1, 1]. For the network layer close to the output layer, the standard deviation of normal distribution initialization is determined according to the output range of the network layer and the expected output variance, and the value range is [0.01, 0.1].

4. The image recognition method based on a deep neural network optimization algorithm according to claim 1, characterized in that: In the feature optimization step: In the particle swarm algorithm, the initial acceleration constant and The value range is [1.2, 1.8], and the initial inertia weight w 0 The value range of is [0.6, 0.9]; In the algorithm based on channel representation capability, the weight of the cross entropy loss function ranges from [0.2, 0.8], and the value is dynamically adjusted according to the change of the loss function during training, with an adjustment step size of [0.01, 0.1].

5. The image recognition method based on deep neural network optimization algorithm according to claim 1, characterized in that: In the network training step: In the adaptive moment estimation Adam algorithm, the coefficient of the momentum term ranges from [0.9, 0.99], and the coefficient of the adaptive learning rate adjustment mechanism ranges from [0.999, 0.1].

6. The image recognition method based on deep neural network optimization algorithm according to claim 1, characterized in that: In the model evaluation and optimization step: In the structure search method based on genetic algorithm, the encoding length of chromosome is determined according to the structural complexity of neural network, and the value is [100, 500]. The population size of genetic algorithm is [50, 200], the crossover probability is [0.6, 0.9], and the mutation probability is [0.01, 0.1]. In the learning rate adjustment method based on Bayesian optimization, the prior distribution adopts Gaussian distribution with a mean of [0.01, 0.1] and a variance of [0.0001, 0.001], and the acquisition function adopts the expected improved EI function or the probability improved PI function.

7. A system for implementing the image recognition method based on a deep neural network optimization algorithm as described in any one of claims 1 to 6, characterized in that: include: Image preprocessing module: used to preprocess the input image data set, including dividing it into training set, validation set and test set, and using adaptive normalization algorithm, hash coding technology, hybrid denoising method and dimensionality reduction method based on manifold learning to process the image; Neural network initialization module: used to encode the deep neural network structure according to the encoding method based on the graph embedding technology in graph theory, and use the hybrid initialization method to initialize the weights and thresholds of the neural network; Feature optimization module: used to determine the abstract dimension and required number of features of the deep neural network for image data through an optimization algorithm that combines the advantages of the particle swarm algorithm and the algorithm based on channel representation ability; Network training module: used to train the optimized neural network using the training set, and optimize the weights and bias values ​​of the neural network using the adaptive moment estimation Adam algorithm and the fusion loss function; Model evaluation and optimization module: used to calculate the accuracy, recall rate, and F1 value evaluation indicators of the neural network on the validation set, and optimize the neural network based on the evaluation results using a structure search method based on a genetic algorithm, a learning rate adjustment method based on Bayesian optimization, and a method of dynamically increasing the number of training times; Image recognition module: used to input the image to be recognized into the optimized neural network and obtain the recognition result of the image using the integrated recognition method.

8. The image recognition system based on deep neural network optimization algorithm according to claim 7, characterized in that: In the image preprocessing module: The adaptive normalization algorithm unit dynamically determines the local area according to the resolution and feature distribution of the image, and uses the corresponding formula to perform normalization processing; The hash coding technology unit uses the MurmurHash algorithm to compress and store the one-hot coding results; The hybrid denoising method unit dynamically adjusts the weights of wavelet transform and non-local mean filter according to the noise level of the image to perform denoising; The dimensionality reduction method unit based on manifold learning uses the local linear embedding (LLE) algorithm or the isometric mapping (Isomap) algorithm to map the image data into a low-dimensional manifold space, and determines the dimension of the low-dimensional space according to the intrinsic dimension of the data and application requirements.

9. The image recognition system based on deep neural network optimization algorithm according to claim 7, characterized in that: In the neural network initialization module: The coding unit based on graph embedding technology in graph theory maps the layer structure and node connection relationship of the neural network to a low-dimensional vector space and embeds it using the node2vec algorithm or the deepwalk algorithm; The hybrid initialization method unit uses corresponding initialization methods to initialize weights and bias values ​​according to the position and function of the network layer.

Images