Image Recognition Method and System Based on Multi-Task Autoencoder

Abstract

This invention discloses an image recognition method and system based on a multi-task autoencoder, relating to the field of image recognition technology. The method includes the following steps: receiving an image dataset; preprocessing the image dataset to obtain an image processing dataset; inputting the image processing dataset into a pre-established autoencoder to extract effective latent vectors; inputting the effective latent vectors into a pre-established deep metric learning network for training to obtain trained latent vectors; classifying and recognizing the trained latent vectors to ultimately obtain normal and abnormal states, thereby achieving the effect of pathological image recognition. This invention effectively extracts meaningful features from pathological image data, and the proposed method demonstrates the possibility of processing discriminative information in latent representations.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, specifically to an image recognition method and system based on a multi-task autoencoder. Background Technology

[0002] Existing methods for identifying pathological images mostly employ traditional machine learning algorithms. For example, Cai Hanshu extracts two linear features (time domain and frequency domain) and five nonlinear features (c0 complexity, correlation dimension, Shannon entropy, Kolmogorov entropy, and power spectral entropy). Then, filters and wrapper classification are used to reduce the dimensionality of the features. Support vector machines, K-nearest neighbors, decision trees, logistic regression, and random forests are applied to classify the image features. The identification of the above-mentioned methods focuses on manual feature extraction and the use of general feature representations. The feature extraction process is cumbersome and labor-intensive, and has certain limitations. Summary of the Invention

[0003] To address the shortcomings mentioned in the background section, the present invention aims to provide an image recognition method and system based on a multi-task autoencoder.

[0004] The objective of this invention can be achieved through the following technical solution: an image recognition method based on a multi-task autoencoder, the method comprising the following steps:

[0005] Receive the image dataset, preprocess the image dataset to obtain the image processing dataset;

[0006] The image processing dataset is input into a pre-built autoencoder, and the autoencoder extracts effective latent vectors.

[0007] The effective latent vectors are input into a pre-built deep metric learning network for training, and the trained latent vectors are obtained.

[0008] The trained latent vectors are classified and identified to obtain normal and abnormal states.

[0009] Preferably, the preprocessing of the image dataset includes data filtering, data decomposition, and artifact removal.

[0010] Preferably, the data filtering process uses a Hamming windowed sincFIR filter to filter the image dataset and remove 50Hz power interference. The data decomposition process converts the continuous image dataset into epoch data, extracts the data epochs with lock times [-100, 500] into specific event types, removes the channel baseline mean [-100, 0] from the epoched image dataset, and then performs independent component analysis decomposition on the image dataset to remove specified components. The artifact removal uses an automatic EEG artifact detector algorithm based on the combined use of spatial and temporal features. It identifies independent artifact components by combining stereotyped artifacts with specific spatial and temporal features, and then removes them.

[0011] Preferably, the format of the image processing dataset is [number of trials, number of sample points, number of channels].

[0012] Preferably, the autoencoder includes an encoder and a decoder. The encoder is a process of transforming a high-dimensional EEG signal into a latent vector, and the decoder decodes the latent vector encoded by the encoder to obtain a high-dimensional EEG signal.

[0013] Preferably, the encoder has two identical CNN blocks; each CNN block contains a 2D convolutional layer, a batch normalization layer, an average pooling layer, and a dropout layer. First, a 2D convolutional filter is fitted, and data dimensionality reduction and feature extraction are performed along the time dimension based on all channel directions of the EEG data. The output signal is constructed into a new time series signal, and spatial information features are extracted from the feature channels. Then, a batch normalization layer is added to accelerate the training of the deep network and improve the model's generalization ability. After that, a 2D average pooling layer is used to extract important features of the input signal, while reducing parameters and sharing weights among all channels of the input signal. A dropout layer is added to help with regularization and prevent overfitting. Then, the flattened vector is mapped to a vector representation. The flattened vector is input into a fully connected layer with a hidden size of z, and effective latent vectors are embedded and generated in the latent space.

[0014] Preferably, the decoder consists of two CNN blocks. First, the latent vector is passed through a fully connected layer and then input into a reshaping layer to reshape the input data to the appropriate dimensions. Next, the input data reshaped to the appropriate dimensions is passed through a Dropout layer and input into a transposed convolutional layer and an ELU layer. After that, it passes through a Dropout layer again and is input into the transposed convolutional layer and ELU layer to reconstruct the input EEG signal.

[0015] Preferably, the training process for the effective latent vector is as follows:

[0016] Semi-hard Tripletloss is used to reflect the relative distance between effective latent vectors of different classes. During training, a set of triplet vectors is extracted from the effective latent vectors and divided into anchor samples, positive samples and negative samples. The distance between anchor samples and positive samples is reduced, and the distance between anchor samples and negative samples is increased to enhance the discriminability of different classes. Finally, the trained latent vectors are obtained.

[0017] Preferably, the process of classifying and recognizing the trained latent vectors is as follows:

[0018] The extracted latent vectors are activated by Softmax and fed into a fully connected layer to obtain the weights important for each class. Then, classification is performed. Specifically, for a given latent vector, the Softmax classifier first calculates the score for each class, then applies an exponential function transformation to these scores, and finally normalizes them to probability values ​​so that their sum is 1. The Softmax classifier classifies the latent vector into the class with the highest probability. The Softmax classifier distinguishes different classes using the weights obtained during training. Different classes correspond to different weight vectors, and the values ​​of these weight vectors enable samples of the same class to be better distinguished, while samples of different classes receive different scores. The specific representation process is as follows:

[0019]

[0020] Where z represents the feature vector extracted by the multi-task autoencoder, and w and b represent the weight matrix and bias vector, respectively. This represents the sample predicted by the Softmax classifier;

[0021] This process incorporates a cross-entropy loss function for optimization. Cross-entropy is used to measure the difference between the probability distribution of the model's output and the probability distribution of the true labels. Therefore, minimizing the cross-entropy loss function during training makes the probability distribution of the model's output closer to the probability distribution of the true labels. The formula for the cross-entropy loss function is as follows:

[0022]

[0023] Among them, y n Represents a real sample. This represents the predicted sample, and class represents the number of categories.

[0024] Secondly, in order to achieve the above objectives, the present invention discloses an image recognition system based on a multi-task autoencoder, comprising:

[0025] Data preprocessing module: Used to receive image datasets, preprocess the image datasets, and obtain image processing datasets;

[0026] Autoencoder module: Used to input image processing datasets into a pre-built autoencoder, which extracts effective latent vectors.

[0027] Deep Metric Learning Module: Used to input effective latent vectors into a pre-built deep metric learning network for training, and obtain trained latent vectors;

[0028] Classification and recognition module: used to classify and recognize the latent vectors after training, and finally obtain the normal state and abnormal state.

[0029] The beneficial effects of this invention are:

[0030] This invention proposes a novel end-to-end architecture that can effectively extract meaningful features from pathological image data without requiring highly complex EEG preprocessing, and achieves good performance in subject-independent experiments. The proposed method applies deep metric learning to multi-task autoencoders to improve EEG classification performance, demonstrating the possibility of processing discriminative information in latent representations. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0033] Figure 2 This is a block diagram of the identification system of the present invention;

[0034] Figure 3 This is a block diagram of the sparse automatic encoder module structure of the present invention;

[0035] Figure 4 This is a schematic diagram of the deep metric learning module of the present invention;

[0036] Figure 5 This is an example of the mixing matrix diagram of the present invention;

[0037] Figure 6 This is a diagram showing the original feature distribution of an example of the present invention;

[0038] Figure 7 This is the feature distribution map after training the example model of this invention;

[0039] Figure 8 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] like Figure 1 As shown, an image recognition method based on a multi-task autoencoder includes the following steps:

[0042] Receive the image dataset, preprocess the image dataset to obtain the image processing dataset;

[0043] It should be further explained that, in this embodiment,

[0044] First, the image dataset is converted to an average value for reference. Second, a Hamming windowed sincFIR filter with a frequency range of 0.3–100 Hz is used to filter the data and remove 50 Hz power interference. Then, the continuous image dataset is converted to epoch data, and the data epochs with a lock time of [-100, 500] are extracted as specific event types. Next, the channel baseline mean of [-100, 0] is removed from the epoched EEG dataset. Then, independent component analysis is performed on the image dataset to remove specified components. Finally, a fast, efficient, and automatic method is used to remove artifacts, namely the Automatic EEG Artifact Detector (ADJUST) algorithm based on the combined use of spatial and temporal features. It identifies independent artifact components by combining the specific spatial and temporal features of stereotyped artifacts, thus obtaining the image processing dataset. The format of the image processing dataset is [number of trials, number of samples, number of channels], where the number of trials is 480, the number of samples is 125, and the number of channels is 128.

[0045] The image processing dataset is input into a pre-built autoencoder, and the autoencoder extracts effective latent vectors.

[0046] In this embodiment, the autoencoder includes an encoder and a decoder. The encoder transforms a high-dimensional EEG signal into a low-dimensional latent vector, that is, it maps a latent representation in a specific latent space. The decoder component processes the signal in the reverse manner, that is, it decodes the latent vector encoded by the encoder to obtain the high-dimensional EEG signal.

[0047] Further explanation is needed regarding the encoder block, which contains two identical CNN blocks. Each CNN block includes 2D convolutional layers, batch normalization layers, average pooling layers, and dropout layers. First, a 2D convolutional filter (Conv2D) of size (1, 64) is fitted, performing dimensionality reduction and feature extraction along the time dimension based on all channel directions of the image data. This constructs a new time-series signal from the output signal, while simultaneously extracting spatial features from the feature channels. Then, a batch normalization (BN) layer is added to accelerate the training of the deep network and improve the model's generalization ability. Next, a 2D average pooling layer (Average Pooling2D) of size (1, 5) is used to extract important features from the input signal, while reducing parameters and sharing weights across all channels of the input signal. Finally, a dropout layer is added to aid regularization and prevent overfitting. The result is then mapped to a vector representation through a flattening operation. Finally, the flattened vector is input to a fully connected layer with a hidden size of z, embedding and generating a latent vector z in the latent space. Here, the latent vector z can become a meaningful feature representation of the association between the EEG signal and depression. The decoder also consists of two CNN blocks, whose structure matches the encoder components in a symmetrical manner. First, to match the input dimension of the CNN, the latent vector is passed through a fully-connected (FC) layer and then fed into a reshape layer to reshape the input data to the appropriate dimension. Next, the reshaped input data is passed through a dropout layer and fed into a transposed convolutional layer (Conv2DTranspose) with a stride of 5 and a kernel size of (1, 64) and an ELU layer. After that, it passes through another dropout layer and is fed into a transposed convolutional layer with a stride of 5 and a kernel size of (1, 32) and an ELU layer to reconstruct the input EEG signal. The transposed convolutional layer helps to extract useful features and reduce useless features, which is beneficial for reconstructing the hidden vector. The training objective of the entire sparse autoencoder module is to minimize the error between the input data and the reconstructed data. In this process, the mean squared error (MSE) is chosen as the loss function. The specific structure of a sparse auto encoder is as follows: Figure 3 As shown.

[0048] The effective latent vectors are input into a pre-built deep metric learning network for training, and the trained latent vectors are obtained.

[0049] In this embodiment, the training process for the effective latent vector is as follows:

[0050] Semi-hard Tripletloss is used to reflect the relative distance between effective latent vectors of different classes. During training, a set of triplet vectors within the effective latent vectors is extracted and divided into anchor samples, positive samples, and negative samples. The distance between anchor samples and positive samples is reduced, and the distance between anchor samples and negative samples is increased, thereby enhancing the discriminative power between different classes. The final trained latent vectors are obtained. The principle is as follows: Figure 4 As shown.

[0051] The trained latent vectors are classified and identified to obtain normal and abnormal states.

[0052] In the specific implementation process, the process of classifying and recognizing the trained latent vectors is as follows:

[0053] The extracted latent vectors are activated by Softmax and fed into a fully connected layer to obtain the weights important for each class. Then, classification is performed. Specifically, for a given latent vector, the Softmax classifier first calculates the score for each class, then applies an exponential function transformation to these scores, and finally normalizes them to probability values ​​so that their sum is 1. The Softmax classifier classifies the latent vector into the class with the highest probability. The Softmax classifier distinguishes different classes using the weights obtained during training. Different classes correspond to different weight vectors, and the values ​​of these weight vectors enable samples of the same class to be better distinguished, while samples of different classes receive different scores. The specific representation process is as follows:

[0054]

[0055] Where z represents the feature vector extracted by the multi-task autoencoder, and w and b represent the weight matrix and bias vector, respectively. This represents the sample predicted by the Softmax classifier;

[0056] This process incorporates a cross-entropy loss function for optimization. Cross-entropy is used to measure the difference between the probability distribution of the model's output and the probability distribution of the true labels. Therefore, minimizing the cross-entropy loss function during training makes the probability distribution of the model's output closer to the probability distribution of the true labels. The formula for the cross-entropy loss function is as follows:

[0057]

[0058] Among them, y n Represents a real sample. This represents the predicted sample, and class represents the number of categories.

[0059] This invention utilizes a Multi-modal Open Dataset for Mental-disorder Analysis (MODMA) provided by Lanzhou University. This dataset includes 128 channels of ERP records, comprising 24 MDD subjects and 29 healthy controls (HCs). The MDD subjects were aged 16-52 years, and the HCs were aged 19-52 years. Detailed information is shown in Table 1.

[0060] The hardware operating system for this experiment was Windows 10, and the network model was built using the Keras framework (with TensorFlow V2.2.0 as the backend). During the overall training process, the number of epochs was set to 100. The Leave Two-Subject-Out Cross Validation (LTSOCV) method was used as the experimental protocol. Twenty-four patients with depression were paired with healthy controls according to gender and age, resulting in 24 pairs. Each pair consisted of one patient with depression and one healthy individual, serving as the test set, representing a complete dataset. The remaining participants were used as the training set. Accuracy and F1 score were used as evaluation metrics, and the proposed model was compared with other models used for EEG signal classification (KNN, DT, RF, FBCSP-SVM, EEGNet, and DeepConvNet). In this experiment, the image data was re-divided into five frequency bands: delta (1-4Hz), theta (4-8Hz), alpha (8-13Hz), beta (13-30Hz), and gamma (30-50Hz), and then divided into training and testing datasets. The training and testing data came from different subjects and were used to conduct independent experiments on the subjects to analyze the performance of depression recognition in each frequency band.

[0061] Table 2 compares the recognition results of our method with other newer methods, showing the independent classification performance of EEG signals in five frequency bands and across the entire frequency band. It can be observed that the deep learning method outperforms the machine learning method in the five frequency bands. This is because the deep learning method extracts hidden features through neural networks and optimizes the model through an inverse mechanism. The results show that the end-to-end automatic feature extraction model is superior to the manual feature extraction model. Furthermore, the machine learning method achieves the best performance in the β band; KNN, DT, RF, and FBCSP-SVM achieve 4.34%, 2.6%, 1.65%, and 3.69% higher accuracy in the β band than in other frequency bands, respectively. This suggests that the machine learning method may have better adaptability to β band EEG signals. Moreover, the deep learning method achieves better results for high-frequency (especially gamma band and full band) EEG data, indicating a strong correlation between high-frequency EEG and depression, from which the deep learning structure can extract hidden features. The multi-task autoencoder fusion deep metric learning model used in this experiment can achieve good EEG feature extraction for depression and obtain good accuracy, indicating that the model can be used as an objective tool for depression detection.

[0062] Figure 5 The diagram shows the confusion matrix of the identification rate of this method for MDD and HC subjects, where the identification rate for healthy individuals is 87.73% and the identification rate for patients with depression is 80.95%. The difference in the number of MDD and HC subjects in the database (depressed: 24; normal: 29) and the smaller number of categories in the model may be the reason for the poor learning and slightly different identification rates in some experiments.

[0063] To verify the ability of this method to extract highly discriminative features from EEG signals, the T-SNE method was used to visualize the latent features learned by the model in a two-dimensional embedding space. First, using the original EEG data as a baseline, the distribution of unlearned brain features was directly visualized using T-SNE, such as... Figure 6 As shown. Then, in the trained model, the T-SNE algorithm is used to visualize the high-dimensional embedding space at the input of the last fully connected layer. Figure 7 The diagram shows the feature distribution map learned after model training, which can effectively identify the good feature extraction and learning, and classify the MDD and HC features well.

[0064] Table 1. Detailed information of participants in the MODMA dataset.

[0065]

[0066] Table 2. Classification performance of different methods for EEG-induced depression in five frequency bands and across the entire frequency band.

[0067]

[0068] Secondly, in order to achieve the above objectives, such as Figure 8 As shown, this invention discloses an image recognition system based on a multi-task autoencoder, comprising:

[0069] Data preprocessing module: Used to receive image datasets, preprocess the image datasets, and obtain image processing datasets;

[0070] Autoencoder module: Used to input image processing datasets into a pre-built autoencoder, which extracts effective latent vectors.

[0071] Deep Metric Learning Module: Used to input effective latent vectors into a pre-built deep metric learning network for training, and obtain trained latent vectors;

[0072] Classification and recognition module: used to classify and recognize the latent vectors after training, and finally obtain the normal state and abnormal state.

[0073] Based on the same inventive concept, this invention also provides a computer device, comprising: one or more processors, and a memory for storing one or more computer programs; the programs include program instructions, and the processor executes the program instructions stored in the memory. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, used to implement one or more instructions, specifically for loading and executing one or more instructions stored in a computer storage medium to implement the above-described method.

[0074] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, performs the above-described method. This storage medium can be any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the present invention, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0075] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0076] The foregoing has shown and described the basic principles, main features, and advantages of this disclosure. Those skilled in the art should understand that this disclosure is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of this disclosure. Various changes and modifications can be made to this disclosure without departing from its spirit and scope, and all such changes and modifications fall within the scope of this disclosure as claimed.

Claims

1. An image recognition method based on a multi-task autoencoder, characterized in that, The method includes the following steps: Receive the image dataset, preprocess the image dataset to obtain the image processing dataset; The image processing dataset is input into a pre-built autoencoder, and the autoencoder extracts effective latent vectors. The autoencoder includes an encoder and a decoder. The encoder is the process of transforming a high-dimensional EEG signal into a latent vector. The decoder decodes the latent vector encoded by the encoder to obtain a high-dimensional EEG signal. The encoder has two identical CNN blocks; each CNN block contains a 2D convolutional layer, a batch normalization layer, an average pooling layer, and a dropout layer. First, a 2D convolutional filter is fitted, and data dimensionality reduction and feature extraction are performed along the time dimension based on all channel directions of the EEG data. The output signal is constructed into a new time series signal, and spatial information features are extracted from the feature channels. Then, a batch normalization layer is added to accelerate the training of the deep network and improve the model's generalization ability. After that, a 2D average pooling layer is used to extract important features of the input signal, while reducing parameters and sharing weights among all channels of the input signal. A dropout layer is added to help with regularization and prevent overfitting. Then, the flattening operation is used to map the vector representation. The flattened vector is input into a fully connected layer with a hidden size of z, and is embedded in the latent space to generate effective latent vectors. The effective latent vectors are input into a pre-built deep metric learning network for training, and the trained latent vectors are obtained. The trained latent vectors are classified and identified to obtain normal and abnormal states.

2. The image recognition method based on a multi-task autoencoder according to claim 1, characterized in that, The preprocessing of the image dataset includes data filtering, data decomposition, and artifact removal.

3. The image recognition method based on a multi-task autoencoder according to claim 2, characterized in that, The data filtering process uses a Hamming windowed sinc FIR filter to filter the image dataset and remove 50Hz power interference. The data decomposition process converts the continuous image dataset into epoch data, extracts the data epochs with a lock time of [-100-500] into specific event types, removes the channel baseline mean of [-1000] from the epoched image dataset, and then performs independent component analysis decomposition on the image dataset to remove specified components. The artifact removal uses an automatic EEG artifact detector algorithm based on the combined use of spatial and temporal features. It identifies independent artifact components by combining stereotyped artifacts with specific spatial and temporal features, and then removes them.

4. The image recognition method based on a multi-task autoencoder according to claim 1, characterized in that, The format of the image processing dataset is [number of trials, number of sample points, number of channels].

5. The image recognition method based on a multi-task autoencoder according to claim 1, characterized in that, The decoder consists of two CNN blocks. First, the latent vector is passed through a fully connected layer and then input into a reshaping layer to reshape the input data to the appropriate dimensions. Next, the input data reshaped to the appropriate dimensions is passed through a Dropout layer and input into a transposed convolutional layer and an ELU layer. After that, it passes through a Dropout layer again and is input into the transposed convolutional layer and ELU layer to reconstruct the input EEG signal.

6. The image recognition method based on a multi-task autoencoder according to claim 1, characterized in that, The training process for the effective latent vectors is as follows: Semi-hard Triplet loss is used to reflect the relative distance between effective latent vectors of different classes. During training, a set of triplet vectors is extracted from the effective latent vectors and divided into anchor samples, positive samples and negative samples. The distance between anchor samples and positive samples is reduced, and the distance between anchor samples and negative samples is increased to enhance the discriminability of different classes. Finally, the trained latent vectors are obtained.

7. The image recognition method based on a multi-task autoencoder according to claim 1, characterized in that, The process of classifying and recognizing the trained latent vectors is as follows: The extracted latent vectors are activated by Softmax and fed into a fully connected layer to obtain the weights important for each class. Then, classification is performed, as follows: Given a latent vector, the Softmax classifier first calculates a score for each class, then applies an exponential function transformation to these scores, and finally normalizes them to probability values ​​so that their sum equals 1. The Softmax classifier classifies the latent vector into the class with the highest probability. The Softmax classifier distinguishes different classes using weights obtained during training; different classes correspond to different weight vectors. The values ​​of these weight vectors enable better differentiation of samples within the same class, resulting in different scores for samples from different classes. The process is illustrated below: Where z represents the feature vector extracted by the multi-task autoencoder, and w and b represent the weight matrix and bias vector, respectively. This represents the sample predicted by the Softmax classifier; This process incorporates a cross-entropy loss function for optimization. Cross-entropy is used to measure the difference between the probability distribution of the model's output and the probability distribution of the true labels. Therefore, during training, minimizing the cross-entropy loss function ensures that the probability distribution of the model's output matches the probability distribution of the true labels. The formula for the cross-entropy loss function is as follows: Among them, y n Represents a real sample. This represents the predicted sample, and class represents the number of categories.

8. An image recognition system based on a multi-task autoencoder, employing the image recognition method based on a multi-task autoencoder as described in any one of claims 1 to 7, characterized in that, include: Data preprocessing module: Used to receive image datasets, preprocess the image datasets, and obtain image processing datasets; Autoencoder module: Used to input image processing datasets into a pre-built autoencoder, which extracts effective latent vectors. Deep Metric Learning Module: Used to input effective latent vectors into a pre-built deep metric learning network for training, and obtain trained latent vectors; Classification and recognition module: used to classify and recognize the latent vectors after training, and finally obtain the normal state and abnormal state.

Images