Picture feature extraction and recognition method and system based on transfer learning method

Abstract

The invention belongs to the field of image recognition, and provides an image feature extraction and recognition method and system based on a transfer learning method. The method comprises the following steps: selecting a picture recognition pre-training model through a pre-training model selection module, extracting features, inputting the extracted features into a self-encoding deep neural network through a self-encoding deep neural network feature learning module, designing the self-encoding deep neural network, training by using a sample picture data set, and extracting the features again. Then migrating the extracted features and a sample picture data set to a target task through a migration learning module, designing a new deep neural network, training the new deep neural network by using a training data set, optimizing a model through a model optimization module, and finally verifying an optimization effect through a verification module. According to the method, pre-training model optimization, feature extraction and transfer learning strategies are fused, and the accuracy and efficiency of image recognition can be effectively improved.

Description

Technical Field

[0001] The present invention belongs to the field of image recognition, and specifically relates to a method and system for extracting and recognizing image features based on a transfer learning method. Background Art

[0002] With the rapid development of information technology, image data has shown explosive growth. From social media to security monitoring, from medical imaging to autonomous driving, image recognition and feature extraction technology have become the core of many fields. However, traditional image processing methods often face problems such as high computational complexity and low recognition accuracy when facing large-scale, high-dimensional image data.

[0003] The main shortcomings of existing image recognition technology can be analyzed from the following perspectives:

[0004] (1) Technical bottlenecks and algorithmic uniformity: Although image recognition technology has made significant progress, my country's computer image recognition intelligent processing technology is still in its infancy. The uniformity of technical algorithms has constrained the development of image recognition technology to a certain extent. In addition, deep learning methods are extremely dependent on large-scale labeled data. This defect greatly limits the application of deep learning methods in actual image recognition tasks.

[0005] (2) Inadequate recognition of low-quality images: Image recognition technology exhibits deficiencies when processing low-quality images, such as blur, noise, and contrast changes, which reduce the recognition rate. For example, low-resolution or low-quality facial images are very common in practical applications, such as video surveillance, but the processing and recognition of these images remains a challenge.

[0006] (3) Sensitivity to specific conditions: Image recognition technology is very sensitive to factors such as lighting conditions, facial expressions, and age changes, which can affect the accuracy of recognition. For example, the recognition rate of a face recognition system will drop significantly when lighting conditions change, facial expressions change, or the face is blocked.

[0007] (4) Small sample learning problem: The small sample learning problem refers to the challenge of training a recognition model with only a small number of labeled samples. Deep learning methods have limitations in dealing with small sample learning problems, which limits their application in certain specific fields.

[0008] (5) Semantic gap problem: A major problem in image retrieval research is that the subjectivity and low-level features of keywords cannot fully describe the high-level concepts in the minds of users, resulting in an inconsistency between the interpretation of image descriptors and high-level semantics, the so-called semantic gap.

[0009] (6) Vulnerability to adversarial attacks: Image recognition systems, especially face recognition systems, are vulnerable to adversarial attacks such as morphological attacks, which can generate almost realistic facial images, posing challenges to the systems.

[0010] As an emerging machine learning method, transfer learning can effectively utilize existing knowledge, accelerate the learning process of new tasks, and improve the performance of the model. Therefore, studying the intelligent extraction and recognition methods of image features using transfer learning methods has important theoretical value and practical application significance.

[0011] In recent years, transfer learning has made significant progress in the field of image recognition. By using deep neural network models pre-trained on large-scale data sets, such as VGG and ResNet, transfer learning can quickly adapt to new image recognition tasks and achieve high accuracy. How to combine transfer learning with deep learning to further improve the performance of image recognition is an urgent problem to be solved. Summary of the invention

[0012] The present invention aims to address the defects of the prior art and provide a method and system for image feature extraction and recognition based on transfer learning methods, which integrates pre-training model optimization, feature extraction and transfer learning strategies, and performs targeted optimization by evaluating model performance, thereby effectively improving the accuracy and efficiency of image recognition.

[0013] In order to achieve the above technical objectives, the present invention adopts the following technical solutions:

[0014] A method for extracting and identifying image features based on a transfer learning method comprises the following steps:

[0015] S1. Select a pre-trained image recognition model according to the current task requirements, remove the classification layer of the pre-trained model, input the data set into the modified pre-trained model, and extract features;

[0016] S2, input the features extracted in step S1 into the autoencoder deep neural network, design the autoencoder deep neural network, use the sample image data set to train the autoencoder deep neural network, and use the encoder part of the trained autoencoder deep neural network to further extract features of the sample image data set;

[0017] S3, design a new deep neural network, and migrate the features and sample image dataset extracted in step S2 to the target task through transfer learning, as input of the new deep neural network, and train the new deep neural network using the training dataset;

[0018] S4. Use the validation data set to verify the accuracy of the new deep neural network, evaluate the performance of the new deep neural network, adjust the model parameters or training process as needed, and optimize the model;

[0019] S5. Train the optimized model on the training data set, evaluate the optimized model on the validation data set, compare the model performance before and after optimization, and verify the optimization effect.

[0020] Preferably, in step S1, the function expression of the image recognition pre-training model is selected according to the current task requirements:

[0021] argmax m∈M [sim(D(m),T)+sim(A(m),T)+G(m)]

[0022] Where M is the set of pre-trained models, m∈M, m is the preset model, T is the current task requirement, D(m) is the pre-trained dataset of the model, A(m) is the architecture of the model, G(m) is the generalization ability of the model, and sim(·,·) represents the similarity function, which is used to evaluate the matching degree between the pre-trained model and the current task requirement.

[0023] Preferably, in step S1, f(x; θ) is set as the selected pre-trained model, wherein x is the input data, θ is the model parameter, and after removing the classification layer of the pre-trained model, a modified pre-trained model is obtained;

[0024] Set the modified pre-trained model as the feature extractor f feat (x; θ), the data set X = {x1, x2, ..., x n} Input to feature extractor f feat (x; θ), extract features, and the feature extraction process is expressed as:

[0025] Features = {f feat (x1;θ),f feat (x2;θ),...,f feat (x n ;θ)},

[0026] Among them, Features is the extracted features, X = {x1, x2, ..., x n} is the data set and θ is the model parameter.

[0027] Preferably, in step S2, the extracted features Features are input into an autoencoder deep neural network to design an autoencoder deep neural network;

[0028] Let g(r; φ) be the designed autoencoder neural network, and use the sample image dataset to train the autoencoder neural network g(r; φ). The objective function of training the autoencoder neural network is expressed as:

[0029] minφΣ r∈SmallDataset / / rg dec (g enc (r; φ enc );φ dec ) / / 2

[0030] Where SmallDataset represents the sample image dataset from which features need to be extracted, r is the input data, and the sample image dataset SmallDataset = {r1, r2, ..., r n}, φ is the network parameter, g enc (r; φ enc ) is the encoder part, g dec (z;φ dec ) is the decoder part, where z = g enc (r; φ enc ) is the encoded feature;

[0031] After training is completed, the encoder part is used to extract features. The feature extraction process is expressed as:

[0032] Features'={g enc (r1;φ enc ),g enc (r2;φ enc ),...,g enc (r n ; φ enc )}

[0033] Among them, Features' is the extracted features, SmallDataset = {r1, r2, ..., r n} is a sample image dataset, and φ is a network parameter.

[0034] Preferably, in step S3, the extracted features Features' and the sample image dataset are migrated to the target task as inputs of a new deep neural network, and a new deep neural network is designed;

[0035] Let g(z';φ c ) is a new deep neural network, and the objective function of training the new model is expressed as:

[0036] minφ c Σ (u,v)∈TrainDataset L(g(z';φ c ),v),

[0037] Where TrainDataset is the training set, u is the training input data, v is the training target value, (u,v)∈TrainDataset, z'=g enc (u;φ enc ), φ c are the parameters of the new model, and L(·,·) is the loss function.

[0038] Preferably, in step S4, the calculation formula for verifying the accuracy of the new model using the verification set is:

[0039] Accuracy=(1 / |ValDataset|)·(∑ (p,q)∈ValDataset I(g(g enc (p; φ enc );φ c )=q)),

[0040] Where Accuracy represents the accuracy, ValDataset is the validation set, p is the validation input data, q is the validation target value, (p,q)∈ValDataset, I(·,·) is the indicator function, which returns 1 when the prediction is correct and 0 otherwise.

[0041] Preferably, in step S4, the indicators for evaluating the performance of the new deep neural network include accuracy, recall, and F1 score;

[0042] The calculation formula of the accuracy is:

[0043] Accuracy'=(TP+TN) / (FP+FN+TP+TN),

[0044] The calculation formula of the recall rate is:

[0045] Recall=TP / (TP+FN),

[0046] The calculation formula of the F1 score is:

[0047] Precision = TP / (TP+FP),

[0048] F1Score=2×[Precision×Recall / (Precision+Recall)],

[0049] Among them, Accuracy' is the accuracy, Recall is the recall rate, Precision is the precision, F1Score is the F1 score, TP represents true positive examples, that is, the number of instances correctly predicted by the model as positive, TN represents true negative examples, that is, the number of instances correctly predicted by the model as negative, FP represents false positive examples, that is, the number of negative examples incorrectly predicted by the model as positive, and FN represents false negative examples, that is, the number of positive examples incorrectly predicted by the model as negative.

[0050] Preferably, in step S4, the optimization model is optimized by regularization, and the optimization formula is:

[0051] L(φ c )=L original (φ c )+λΩ(φ c ),

[0052] Among them, L original (φ c ) is the original loss function, λ is the regularization coefficient, Ω(φ c ) is the regularization term.

[0053] The present invention also provides a picture feature extraction and recognition system based on a transfer learning method, comprising a pre-trained model selection module, an autoencoder deep neural network feature learning module, a transfer learning module, a model optimization module and a verification module;

[0054] The pre-trained model selection module is used to select a pre-trained image recognition model suitable for the current task requirements and extract features;

[0055] The autoencoder deep neural network feature learning module is used to input the features extracted by the pre-trained model selection module into the autoencoder deep neural network, design the autoencoder deep neural network, train the autoencoder deep neural network using the sample image data set, and use the encoder part of the trained autoencoder deep neural network to extract features;

[0056] The transfer learning module is used to use the features extracted by the autoencoder deep neural network feature learning module as input of a new deep neural network, design a new deep neural network, and train the new deep neural network using a training data set;

[0057] The model optimization module is used to optimize the model by adjusting model parameters or training process;

[0058] The verification module is used to compare the model performance before and after optimization and verify the optimization effect.

[0059] Compared with the prior art, the beneficial effects of the present invention are:

[0060] (1) The present invention is a method and system for image feature extraction and recognition based on a transfer learning method, which removes the classification layer and inputs data to obtain a modified pre-trained model, integrates pre-trained model optimization, feature extraction and transfer learning strategies, selects a suitable pre-trained model for feature extraction, constructs a data set and trains and optimizes the transfer learning model, which can accelerate the learning process and effectively improve the accuracy and efficiency of image recognition, especially when the amount of data is limited or the data is unevenly distributed, image recognition can be performed quickly and accurately;

[0061] (2) The image feature extraction and recognition method and system based on the transfer learning method of the present invention can further improve the performance of the model by evaluating the model performance and performing targeted optimization, thereby further improving the accuracy and efficiency of image recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 This is a flow chart of a method for image feature extraction and recognition based on a transfer learning method according to an embodiment of the present invention;

[0063] Figure 2 Schematic diagram of a system for extracting and identifying image features based on a transfer learning method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0064] The following will be combined with the attached Figure 1-2 , the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. 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.

[0065] Example 1

[0066] Combination Figure 1 As shown, according to a specific implementation scheme of the present invention, a method for extracting and identifying image features based on a transfer learning method of the present invention is described in detail below.

[0067] The present invention provides a method for image feature extraction and recognition based on a transfer learning method, comprising the following steps:

[0068] S1. Select a pre-trained model for image recognition based on the current task requirements. Specifically, when selecting a pre-trained model based on the current task requirements, you need to consider the model's pre-trained dataset, model architecture, and model generalization ability. Pre-trained models are usually trained on large-scale datasets, such as the ImageNet dataset, so as to learn rich image representation knowledge. For example, ResNet50 and MobileNetV2 are models pre-trained on ImageNet.

[0069] In this embodiment, the function expression of the image recognition pre-training model is selected according to the current task requirements:

[0070] argmax m∈M [sim(D(m),T)+sim(A(m),T)+G(m)]

[0071] Where M is the set of pre-trained models, m∈M, m is the preset model, T is the current task requirement, D(m) is the pre-trained dataset of the model, A(m) is the architecture of the model, G(m) is the generalization ability of the model, and sim(·,·) represents the similarity function, which is used to evaluate the matching degree between the pre-trained model and the current task requirement;

[0072] Remove the classification layer of the pre-trained model, input the data set into the modified pre-trained model, and extract features. Specifically, in terms of feature extraction, deep convolutional neural network (CNN) is widely used in the field of image classification due to its excellent feature extraction ability;

[0073] In this embodiment, let f(x; θ) be the selected pre-trained model, where x is the input data, θ is the model parameter, and after removing the classification layer of the pre-trained model, a modified pre-trained model is obtained;

[0074] Set the modified pre-trained model as the feature extractor f feat (x; θ), the data set X = {x1, x2, ..., x n} Input to feature extractor f feat (x; θ), extract features, and the feature extraction process is expressed as:

[0075] Features = {f feat (x1;θ),f feat (x2;θ),...,f feat (x n ;θ)},

[0076] Among them, Features is the extracted features, X = {x1, x2, ..., x n} is the data set and θ is the model parameter.

[0077] S2, input the features extracted in step S1 into the autoencoder deep neural network, design the autoencoder deep neural network, use the sample image data set to train the autoencoder deep neural network, and use the encoder part of the trained autoencoder deep neural network to further extract features of the sample image data set, specifically,

[0078] Input the extracted features into the autoencoder deep neural network and design the autoencoder deep neural network;

[0079] Let g(r; φ) be the designed autoencoder neural network, and use the sample image dataset to train the autoencoder neural network g(r; φ). The objective function of training the autoencoder neural network is expressed as:

[0080] minφ∑ r∈SmallDataset / / rg dec (g enc (r; φ enc );φ dec ) / / 2

[0081] Where SmallDataset represents the sample image dataset from which features need to be extracted, r is the input data, and the sample image dataset SmallDataset = {r1, r2, ..., r n}, φ is the network parameter, g enc (r; φ enc ) is the encoder part, g dec (z;φ dec ) is the decoder part, where z = g enc (r; φ enc ) is the encoded feature;

[0082] After training is completed, the encoder part is used to extract features. The feature extraction process is expressed as:

[0083] Features'={g enc (r1;φ enc ),g enc (r2;φ enc ),...,g enc (r n ; φ enc )}

[0084] Among them, Features' is the extracted features, SmallDataset = {r1, r2, ..., r n} is a sample image dataset, and φ is a network parameter.

[0085] S3, design a new deep neural network, and migrate the features and sample image dataset extracted in step S2 to the target task through transfer learning, as the input of the new deep neural network, and use the training dataset to train the new deep neural network. You can choose to freeze some or all layers of the new deep neural network to prevent overfitting. Specifically,

[0086] Transfer the extracted features and sample image dataset Features' to the target task as input of the new deep neural network, and design a new deep neural network;

[0087] Let g(z';φ c ) is a new deep neural network, and the objective function of training the new model is expressed as:

[0088] minφ c ∑ (u,v)∈TrainDataset L(g(z';φ c ),v),

[0089] Where TrainDataset is the training set, u is the training input data, v is the training target value, (u,v)∈TrainDataset, z'=g enc (u;φ enc ), φ c are the parameters of the new model, and L(·,·) is the loss function.

[0090] S4. Use the validation data set to verify the accuracy of the new deep neural network and evaluate the performance of the new deep neural network. Specifically,

[0091] The calculation formula for the accuracy of the new model using the validation set is:

[0092] Accuracy=(1 / |ValDataset|)·(∑ (p,q)∈ValDataset I(g(g enc (p; φ enc );φ c )=q)),

[0093] Where Accuracy represents the accuracy, ValDataset is the validation set, p is the validation input data, q is the validation target value, (p,q)∈ValDataset, I(·,·) is the indicator function, which returns 1 when the prediction is correct, otherwise it returns 0;

[0094] The metrics used to evaluate the performance of new deep neural networks include accuracy, recall, and F1 score;

[0095] The calculation formula of the accuracy is:

[0096] Accuracy'=(TP+TN) / (FP+FN+TP+TN),

[0097] The calculation formula of the recall rate is:

[0098] Recall=TP / (TP+FN),

[0099] The calculation formula of the F1 score is:

[0100] Precision = TP / (TP+FP),

[0101] F1Score=2×[Precision×Recall / (Precision+Recall)],

[0102] Among them, Accuracy' is the accuracy, Recall is the recall rate, Precision is the precision, F1Score is the F1 score, TP represents true positive examples, that is, the number of instances correctly predicted by the model as positive, TN represents true negative examples, that is, the number of instances correctly predicted by the model as negative, FP represents false positive examples, that is, the number of negative examples incorrectly predicted by the model as positive, and FN represents false negative examples, that is, the number of positive examples incorrectly predicted by the model as negative.

[0103] The model parameters or training process are adjusted as needed to optimize the model. Specifically, when adjusting the model parameters, the model can be optimized by adjusting the learning rate, batch size, and regularization coefficient. The model can also be optimized by changing the model structure, such as modifying the number of layers, adding or deleting feature fusion modules, etc., and using more complex training strategies, such as introducing global context information and feature fusion modules, etc. to optimize the model. In this embodiment, the model is optimized by regularization, and the optimization formula is:

[0104] L(φ c )=L original (φ c )+λΩ(φ c ),

[0105] Among them, L original (φ c ) is the original loss function, λ is the regularization coefficient, Ω(φ c ) is the regularization term.

[0106] When using the introduction of global context information and feature fusion modules for model optimization, a more complex model structure is usually designed, which may include multiple feature extractors and fusion strategies. The optimization formula will vary depending on the specific implementation, but the general form can be:

[0107] F fused =Fuse(F1,F2,...F n),

[0108] Among them, F1, F2, ... F n is the feature map obtained from different layers or different feature extractors, and Fuse(·) is the fusion function.

[0109] S5. Train the optimized model on the training data set, evaluate the optimized model on the validation data set, compare the model performance before and after optimization, and verify the optimization effect.

[0110] Example 2

[0111] Combination Figure 2 As shown, according to a specific implementation scheme of the present invention, a picture feature extraction and recognition system based on a transfer learning method of the present invention is described in detail below.

[0112] The present invention provides a picture feature extraction and recognition system based on a transfer learning method, comprising a pre-training model selection module, an autoencoder deep neural network feature learning module, a transfer learning module, a model optimization module and a verification module;

[0113] The pre-trained model selection module is used to select a pre-trained image recognition model suitable for the current task requirements and extract features;

[0114] The autoencoder deep neural network feature learning module is used to input the features extracted by the pre-trained model selection module into the autoencoder deep neural network, design the autoencoder deep neural network, train the autoencoder deep neural network using the sample image data set, and use the encoder part of the trained autoencoder deep neural network to extract features;

[0115] The transfer learning module is used to use the features extracted by the autoencoder deep neural network feature learning module as input of a new deep neural network, design a new deep neural network, and train the new deep neural network using a training data set;

[0116] The model optimization module is used to optimize the model by adjusting model parameters or training process;

[0117] The verification module is used to compare the model performance before and after optimization and verify the optimization effect.

[0118] The above description is only an embodiment of the present invention and is not intended to limit the present invention. Any modification, equivalent replacement and improvement made within the application scope of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for image feature extraction and recognition based on transfer learning method, characterized in that: The following steps are involved: S1. Select a pre-trained image recognition model according to the current task requirements, remove the classification layer of the pre-trained model, input the image dataset into the modified pre-trained model, and extract image features; S2, input the features extracted in step S1 into the autoencoder deep neural network, design the autoencoder deep neural network, use the sample image data set to train the autoencoder deep neural network, and use the encoder part of the trained autoencoder deep neural network to further extract features of the sample image data set; S3, design a new deep neural network, and migrate the features and sample image dataset extracted in step S2 to the target task through transfer learning, as input of the new deep neural network, and train the new deep neural network using the training dataset; S4. Use the validation data set to verify the accuracy of the new deep neural network, evaluate the performance of the new deep neural network, adjust the model parameters or training process as needed, and optimize the model; S5. Train the optimized model on the training data set, evaluate the optimized model on the validation data set, compare the model performance before and after optimization, and verify the optimization effect.

2. The image feature extraction and recognition method based on the transfer learning method according to claim 1 is characterized in that: In step S1, the function expression of the image recognition pre-training model is selected according to the current task requirements: argmax m∈M [sim(D(m),T)+sim(A(m),T)+G(m)] Where M is the set of pre-trained models, m∈M, m is the preset model, T is the current task requirement, D(m) is the pre-trained dataset of the model, A(m) is the architecture of the model, G(m) is the generalization ability of the model, and sim(·,·) represents the similarity function, which is used to evaluate the matching degree between the pre-trained model and the current task requirement.

3. The image feature extraction and recognition method based on the transfer learning method according to claim 2 is characterized in that: In step S1, let f(x; θ) be the selected pre-trained model, where x is the input data, θ is the model parameter, and after removing the classification layer of the pre-trained model, a modified pre-trained model is obtained; Set the modified pre-trained model as the feature extractor f feat (x; θ), the data set X = {x1, x2, ..., x n } Input to feature extractor f feat (x; θ), extract features, and the feature extraction process is expressed as: Features={f feat (x1;θ),f feat (x2;θ),...,f feat (x n ;i)}, Among them, Features is the extracted features, X = {x1, x2, ..., x n } is the data set and θ is the model parameter.

4. The image feature extraction and recognition method based on the transfer learning method according to claim 3 is characterized in that: In step S2, the extracted features Features are input into the autoencoder deep neural network to design the autoencoder deep neural network; Let g(r; φ) be the designed autoencoder neural network, and use the sample image dataset to train the autoencoder neural network g(r; φ). The objective function of training the autoencoder neural network is expressed as: minφΣ r∈SmallDataset / / rg dec (g enc (r;φ enc );f dec ) / / 2 Where SmallDataset represents the sample image dataset from which features need to be extracted, r is the input data, and the sample image dataset SmallDataset = {r1, r2, ..., r n }, φ is the network parameter, g enc (r; φ enc ) is the encoder part, g dec (z;φ dec ) is the decoder part, where z = g enc (r; φ enc ) is the encoded feature; After training is completed, the encoder part is used to extract features. The feature extraction process is expressed as: Features'={g enc (r1;φ enc ),g enc (r2;φ enc ),...,g enc (r n ;f enc )} Among them, Features' is the extracted features, SmallDataset = {r1, r2, ..., r n } is a sample image dataset, and φ is a network parameter.

5. The image feature extraction and recognition method based on the transfer learning method according to claim 4 is characterized in that: In step S3, the extracted features Features' and sample image dataset are transferred to the target task as inputs of a new deep neural network, and a new deep neural network is designed; Let g(z';φ c ) is a new deep neural network, and the objective function of training the new model is expressed as: minφ c S (u,v)∈TrainDataset L(g(z';φ c ),v), Where TrainDataset is the training set, u is the training input data, v is the training target value, (u,v)∈TrainDataset, z'=g enc (u;φ enc ), φ c are the parameters of the new model, and L(·,·) is the loss function.

6. The image feature extraction and recognition method based on transfer learning method according to claim 5, characterized in that: In step S4, the calculation formula for verifying the accuracy of the new model using the verification set is: Accuracy=(1 / |ValDataset)·(∑ (p,q)∈ValDataset I(g(g enc (p;φ enc );φ c )=q)), Where Accuracy represents the accuracy, ValDataset is the validation set, p is the validation input data, q is the validation target value, (p,q)∈ValDataset, I(·,·) is the indicator function, which returns 1 when the prediction is correct and 0 otherwise.

7. The image feature extraction and recognition method based on transfer learning method according to claim 6, characterized in that: In step S4, the performance indicators of the new deep neural network include accuracy, recall, and F1 score; The calculation formula of the accuracy is: Accuracy'=(TP+TN) / (FP+FN+TP+TN), The calculation formula of the recall rate is: Recall=TP / (TP+FN), The calculation formula of the F1 score is: Precision = TP / (TP+FP), F1Score=2×[Precision×Recall / (Precision+Recall)], Among them, Accuracy' is the accuracy, Recall is the recall rate, Precision is the precision, F1Score is the F1 score, TP represents true positive examples, that is, the number of instances correctly predicted by the model as positive, TN represents true negative examples, that is, the number of instances correctly predicted by the model as negative, FP represents false positive examples, that is, the number of negative examples incorrectly predicted by the model as positive, and FN represents false negative examples, that is, the number of positive examples incorrectly predicted by the model as negative.

8. The image feature extraction and recognition method based on transfer learning method according to claim 7, characterized in that: In step S4, the model is optimized by regularization, and the optimization formula is: L(φ c )=L original (f c )+λΩ(φ c ), Among them, L original (φ c ) is the original loss function, λ is the regularization coefficient, Ω(φ c ) is the regularization term.

9. A system for extracting and identifying image features based on a transfer learning method applicable to the method according to any one of claims 1 to 8, characterized in that: It includes pre-trained model selection module, auto-encoding deep neural network feature learning module, transfer learning module, model optimization module and verification module; The pre-trained model selection module is used to select a pre-trained image recognition model suitable for the current task requirements and extract features; The autoencoder deep neural network feature learning module is used to input the features extracted by the pre-trained model selection module into the autoencoder deep neural network, design the autoencoder deep neural network, train the autoencoder deep neural network using the sample image data set, and use the encoder part of the trained autoencoder deep neural network to extract features; The transfer learning module is used to use the features extracted by the autoencoder deep neural network feature learning module as input of a new deep neural network, design a new deep neural network, and train the new deep neural network using a training data set; The model optimization module is used to optimize the model by adjusting model parameters or training process; The verification module is used to compare the model performance before and after optimization and verify the optimization effect.

Images