CN118505600A - Unified anomaly detection method based on multi-source uncertainty mining - Google Patents

Abstract

The invention belongs to the technical field of visual anomaly detection, and discloses a unified anomaly detection method based on multi-source uncertainty mining. According to the invention, in the abnormal segmentation network training process, the multi-source uncertainty mining network and the abnormal segmentation network are introduced, and the multi-source uncertainty mining network and the abnormal segmentation network are interactively trained layer by layer based on a cross attention mechanism, so that the abnormal segmentation network can be assisted to better pay attention to global features and abnormal local features of an image, and thus more accurate positioning is realized.

Description

Unified anomaly detection method based on multi-source uncertainty mining

Technical Field

The present invention belongs to the technical field of visual anomaly detection and relates to a unified anomaly detection method based on multi-source uncertainty mining.

Background Art

Anomaly detection plays a key role in computer vision and industrial applications. The main goal of visual anomaly detection is to accurately identify abnormal images and precisely locate abnormal areas. Existing visual anomaly detection methods can be divided into three paradigms: unsupervised, semi-supervised, and fully supervised. Unsupervised methods can effectively model normal samples, do not rely on prior anomaly information, and can detect unknown abnormal forms in products. However, their detection performance is not satisfactory. Semi-supervised and fully supervised methods focus on modeling abnormal samples, which requires specifying the abnormal form and labeling the abnormal instances. These methods have good detection performance for known anomalies. However, since they are limited to predefined abnormal forms, their detection performance is greatly weakened when faced with unknown abnormal forms.

In the industrial field, unsupervised methods are usually used to quickly build detection models. Subsequently, by introducing some abnormal samples, a transition to semi-supervised or fully supervised methods is made to further improve the anomaly detection performance. However, there is a challenge with this strategy: the two stages are separated, that is, the model obtained in the unsupervised stage cannot be carried over to the subsequent supervised learning. In addition, a unified model for anomaly detection of multiple categories of products can complete the detection of multiple products with only a single model, which is more suitable for the actual production environment. Existing unified models usually adopt the feature reconstruction method, but only the unification is completed in the feature reconstruction part, while a unified decision boundary is lacking in the detection process. Specifically, when all categories of products are detected using a unified decision boundary, the performance is lower than that of detecting each category separately.

Summary of the invention

The purpose of the present invention is to address the above-mentioned problems existing in the prior art and to provide a unified anomaly detection method based on multi-source uncertainty mining, which can identify various anomaly forms without pre-definition and unify decision boundaries.

In order to achieve the above object, the present invention adopts the following technical solutions to achieve it.

The present invention provides a unified anomaly detection method based on multi-source uncertainty mining, which comprises the following steps:

S1 trains the anomaly segmentation network based on a semi-supervised learning method based on multi-source uncertainty mining, including the following steps:

S11 uses data containing normal samples and abnormal samples to construct a training set, and obtains multi-source pseudo labels through several pre-trained basic models;

S12 uses any basic model to obtain a reconstructed image of a sample in a training set, and uses the Euclidean distance between the sample and its reconstructed image as a biased difference to input an anomaly segmentation network to obtain an anomaly segmented image; the anomaly segmentation network includes a plurality of feature extraction stages arranged in sequence, and the output of the previous stage is used as the input of the next stage; at least part of the image features extracted in the feature extraction stage are output to a multi-source uncertainty mining network;

S13 multi-source pseudo labels are input into the multi-source uncertainty mining network, combined with the image features output from the abnormal segmentation network, the global attention distribution is obtained based on the cross-attention mechanism, and an uncertainty weight map is generated;

S14 constructs a loss function and obtains a loss value based on multi-source pseudo labels, abnormal segmentation images, and uncertainty weight maps;

S15 uses the loss value to update the parameters of the anomaly segmentation network and the multi-source uncertainty mining network;

Repeat the above steps S12-S15 until the loss function converges to obtain the abnormal segmentation network of the training number;

S2 performs unified anomaly detection on the image to be detected, including the following steps:

S21 obtains a reconstructed image of the image to be detected by using the basic model, and uses the Euclidean distance between the image to be detected and its reconstructed image as a biased difference;

S22 inputs the biased difference into the anomaly segmentation network to obtain an anomaly segmented image.

In the above step S12, the basic model is selected from EdgRec, DRAEM, FastFlow or MSTAD, etc. Each feature extraction stage in the anomaly segmentation network includes several convolution modules; each convolution module consists of a convolution layer, a batch normalization layer and a ReLU activation function. After the other feature extraction stages except the last feature extraction stage, the next feature extraction stage is entered through the downsampling layer. The last feature extraction stage obtains the anomaly segmentation image through the upsampling layer and the convolution module. Except for the first feature extraction stage, the image features extracted in the remaining feature extraction stages are output to the multi-source uncertainty mining network.

In the above step S13, the multi-source uncertainty mining network includes an encoder, several cross-attention modules and a decoder arranged in sequence; the encoder is used to encode the input image features; the cross-attention module is used to obtain the global attention distribution of the image features output by the corresponding feature extraction stage of the abnormal segmentation network for the multi-source pseudo-labels based on the cross-attention mechanism, and superimpose it with the encoder output or the output of the previous cross-attention module, and then perform patch fusion; the decoder is used to decode the input image features to obtain an uncertainty weight map.

In a preferred implementation, the encoder includes three convolutional layers and two downsampling layers, the two downsampling layers are respectively located between two adjacent convolutional layers, and their main function is to extract features and change the size of the input feature image. The decoder includes a convolutional layer and an upsampling layer, which is mainly used to output an uncertainty weight map that can accurately represent the pixel-level uncertainty mining content.

In a preferred implementation, the number of the cross attention modules is the same as the number of feature extraction stages accessing the multi-source uncertainty mining network. The cross attention module includes a cross attention layer, a feedforward neural network and a patch fusion layer.

In the above step S14, the loss function constructed is:

Where, θ _S and θ _∑ are the trainable parameters of the anomaly segmentation network and the multi-source uncertainty mining network, respectively; Indicates that the i-th pixel obtained based on the m-th pseudo label uses the pseudo label and the cross entropy loss calculated with the unnormalized score ^Si ; represents the predicted log variance, Indicates the use of pseudo labels The uncertainty weight of the i-th pixel is obtained; M represents the number of pseudo labels, H represents the height of the image, and W represents the width of the image.

In the above step S2, in order to improve the abnormality detection of the image to be detected, this step also includes:

S23 performs global average pooling processing on the abnormal segmented image, and takes the maximum value as the abnormal score of the image to be detected;

S24 performs weighted fusion on the abnormal segmentation image and the image to be detected to obtain an abnormal heat map.

Compared with the prior art, the unified anomaly detection method based on multi-source uncertainty mining provided by the present invention has the following beneficial effects:

1) In the training process of the anomaly segmentation network, the present invention introduces a multi-source uncertainty mining network and anomaly segmentation network. Based on the cross-attention mechanism, the multi-source uncertainty mining network and the anomaly segmentation network are interactively trained layer by layer, which can help the anomaly segmentation network to better focus on the global features of the image and the local features of the anomaly, thereby achieving more accurate positioning;

2) The present invention uses the basic model obtained by unsupervised learning to obtain the biased difference of anomalies; by using unlabeled abnormal samples, the performance of the basic model in anomaly detection is further improved;

3) The present invention utilizes an uncertainty-weighted loss function and regards the pseudo-labels of the model as Gibbs distribution in a Bayesian framework; through this loss function, a multi-source uncertainty mining network and anomaly segmentation network can be trained simultaneously, and the anomaly segmentation network can perform anomaly detection under unified decision boundary conditions.

4) Extensive experiments are conducted on the MVTec AD dataset and the effectiveness of the proposed method is demonstrated; in addition, when a self-motivated approach is used, i.e., pseudo-labels that do not rely on other models, the anomaly detection performance can still be improved.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a unified anomaly detection method based on multi-source uncertainty mining;

Fig. 2 shows the original image and the corresponding labels; (a) is the original image, (b) is the real label, (c) is the positioning label obtained by the method of the present invention, (d) is the pseudo label obtained by the EdgRec model, (e) is the pseudo label obtained by the DRAEM model, (f) is the pseudo label obtained by the FastFlow model, and (g) is the pseudo label obtained by the MSTAD model;

Figure 3 shows the original image and the abnormality localization maps obtained by different methods.

DETAILED DESCRIPTION

The following will clearly and completely describe the technical solutions of various embodiments of the present invention in conjunction with the accompanying drawings. 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 making creative work are within the scope of protection of the present invention.

Example 1

This paper proposes a novel semi-supervised anomaly detection strategy that uses a pixel reconstruction model or a feature reconstruction model as a basic model. The multi-source uncertainty mining network (MUMNet) and the anomaly segmentation network (ASNet) are introduced. During the training process, based on the cross-attention mechanism, MUMNet and ASNet interact layer by layer, helping ASNet to better focus on the global features of the image and the local features of the anomaly, thereby achieving more accurate positioning.

The data used in this embodiment is the MVTec AD dataset, which contains 3629 normal images, covering 5 different texture categories and 10 unique object categories. At the same time, the test set includes 467 normal images and 1258 abnormal images. For each abnormal sample in the test set, the dataset provides the image label and segmentation information as the true value.

1. Basic model

The basic model used in this embodiment is selected from EdgRec, DRAEM, FastFlow or MSTAD.

These basic models are all unsupervised reconstruction models that use the Euclidean distance between the input and output of the above basic models to represent the biased differences of various forms of anomalies. This method eliminates the reliance on manually labeled anomalies. The unified anomaly detection method (MUM-UAD) based on multi-source uncertainty mining in this embodiment is built on this basis and adopts a progressive learning strategy. However, since the unsupervised model is only learned through normal samples and cannot utilize abnormal samples, it cannot be used in the subsequent semi-supervised stage. To solve this problem, MUM-UAD adopts an uncertainty mining learning strategy, using the basic model as a connection between the unsupervised learning and semi-supervised learning stages, thereby improving detection performance and better adapting to actual production environments.

2. Anomaly Segmentation Network (ASNet)

The main goal of the anomaly segmentation network (ASNet) is to generate a high-precision anomaly segmentation map in the final decision-making process. To achieve this goal, this embodiment makes appropriate modifications to the network architecture of ResNet34.

The anomaly segmentation network (ASNet) provided in this embodiment, as shown in Figure 1(a), includes four feature extraction stages (Stage0-Stage3) arranged in sequence, and the output of the previous stage is used as the input of the next stage. Each feature extraction stage includes a number of convolution modules, and the number is expressed as _N0 = 3, _N1 = 4, _N2 = 6 and _N3 = 3 respectively; the network structure can be adjusted accordingly according to the type of basic model adopted. Each convolution module consists of a convolution layer, a batch normalization layer and a ReLU activation function. After the feature extraction stages except the last feature extraction stage, the other feature extraction stages enter the next feature extraction stage through the downsampling layer, and the last feature extraction stage obtains the anomaly segmentation image through the upsampling layer and the convolution module. Except for the first feature extraction stage (Stage0), the image features extracted by the remaining feature extraction stages (Stage1-Stage3) are output to the multi-source uncertainty mining network.

For feature reconstruction, ASNet adopts the inverted structure of ResNet34. Specifically, all downsampling operations are replaced by upsampling, and the order of all stages is reversed. For pixel reconstruction models or cases without base models, ASNet follows the standard ResNet34 structure, but replaces the last linear output layer with an upsampling layer to adjust the output size. Here, ASNet is used to extract multi-scale features from the biased differences obtained from the base model, resulting in an unbiased anomaly segmentation map.

In this embodiment, ASNet is initialized using the pre-trained weights of ResNet34, and the size of the deepest feature map is limited to 14×14.

3. Multi-source Uncertainty Mining Network (MUMNet)

MUMNet aims to accurately identify the reliability of each pixel by analyzing pseudo labels and input images. Specifically, by simultaneously mining the commonalities and differences between multiple pseudo labels generated by multiple anomaly detection models, MUMNet can effectively capture reliable label samples.

The multi-source uncertainty mining network consists of an encoder, several cross-attention modules and a decoder arranged in sequence.

The encoder is used to encode the input image features. The encoder includes three convolutional layers and two downsampling layers. The two downsampling layers are located between two adjacent convolutional layers. Their main function is to extract features and change the size of the input feature image.

The decoder is used to decode the input image features to obtain an uncertainty weight map. The decoder includes a convolution layer and an upsampling layer, which is mainly used to output an uncertainty weight map that can accurately represent the pixel-level uncertainty mining content.

The number of cross-attention modules is used to obtain the global attention distribution of the image features output by the corresponding feature extraction stage of the abnormal segmentation network for multi-source pseudo-labels based on the cross-attention mechanism, and superimpose it with the encoder output or the output of the previous cross-attention module, and perform patch fusion. The number of cross-attention modules is the same as the number of feature extraction stages connected to the multi-source uncertainty mining network; specifically, the number of cross-attention modules in this embodiment is 3, namely the first cross-attention module to the third cross-attention module. As shown in Figure 1(a), the cross-attention module includes a cross-attention layer, a feedforward neural network and a patch fusion layer. The feature images extracted from Stage1 to Stage3 in the abnormal segmentation network are respectively input into the cross-attention layers of the first cross-attention module to the third cross-attention module.

The criss-cross attention layer uses linear projection to transform the image features extracted from ASNet into At the same time, the features derived from the pseudo-labels are recorded as The linear projection is and Here, i represents the stage number of the cross-attention module, n represents the number of elements in the feature map, and c represents the dimension of the feature. The detailed information of the i-th stage is as follows:

By utilizing the cross-attention mechanism, the global attention distribution of the image features for the multi-source pseudo-label features at this particular stage can be obtained. Subsequently, the information is integrated using a feedforward neural network (FFN) and downsampled through patch fusion to interact with the feature information in the next stage.

f ⁱ⁺¹ =PatchMerging(FFN(f′)) (2);

This module enhances the perception of abnormal information by features of different scales in ASNet and helps MUMNet generate more robust uncertainty maps.

In this embodiment, MUMNet is initialized with a uniform distribution, and the cross-attention module interacts with feature maps of three scales: 56×56, 28×28, and 14×14.

Based on the above explanation, the unified anomaly detection method based on multi-source uncertainty mining provided in this embodiment includes the following steps:

S1 trains the anomaly segmentation network based on a semi-supervised learning method based on multi-source uncertainty mining, including the following steps:

S11 uses data containing normal samples and abnormal samples to construct a training set, and obtains multi-source pseudo labels through several pre-trained basic models.

Since the MVTec AD dataset only contains abnormal samples in the test set, this embodiment redefines the dataset. The original MVTec AD training set contains only normal samples and is called the "basic training set", while the test set containing abnormal samples and a small number of normal samples is called the "basic test set". Then, several abnormal samples and normal samples are selected from each product category in the "basic test set" to form a "new training set". The remaining samples in the "basic test set" form the "new test set". At the same time, the resolution of all input images is adjusted to 224×224.

The above basic models are pre-trained using the "basic training set", and the parameter weights of each basic model are retained. In addition, the value covering 98% of the scores in the anomaly score map generated by the four models (EdgRec, DRAEM, FastFlow, MSTAD) trained on the "basic training set" is selected as the threshold.

The pre-trained basic model is used to detect the "new training set" and generate anomaly score maps. These anomaly score maps are then binarized according to the threshold and used as multi-source pseudo labels.

S12 uses any basic model to obtain the reconstructed image of the sample in the training set, and uses the Euclidean distance between the sample and its reconstructed image as the biased difference input into the anomaly segmentation network to obtain the anomaly segmentation image.

At the same time, the feature images extracted by the feature extraction stages Stage1-Stage3 in the anomaly segmentation network are output to the corresponding cross-attention module of the multi-source uncertainty mining network.

S13 multi-source pseudo labels are input into the multi-source uncertainty mining network, combined with the image features output from the anomaly segmentation network, and the global attention distribution is obtained based on the cross-attention mechanism to generate an uncertainty weight map.

This step uses the multi-source uncertainty mining network given above to generate an uncertainty weight graph.

The output of the last cross-attention module is decoded by the decoder to obtain an uncertainty weight map that can accurately represent the pixel-level uncertainty mining content.

S14 constructs a loss function and obtains the loss value based on multi-source pseudo labels, abnormal segmentation images and uncertainty weight maps.

Among the M anomaly detection methods (corresponding to M pseudo labels), the pseudo label of pixel i generated by the mth method is represented as Where c = 1 represents an abnormal area, c = 0 represents a normal area, The corresponding uncertainty weight map is given by Represents. The pseudo label is modeled as a random variable y that follows the Gibbs distribution under Bayesian theory. When the Softmax function is used to normalize the anomaly score, the probability distribution of y can be calculated as:

When given the observed pseudo-label The negative log-likelihood can be further derived as follows:

in, Indicates that the i-th pixel obtained based on the m-th pseudo label uses the pseudo label and the cross entropy loss calculated with the unnormalized scores ^Si (given by the anomaly segmentation images). However, in practice, one can use the predicted log variance Increase numerical stability during training, Indicates the use of pseudo labels The uncertainty weight of the i-th pixel is obtained.

Therefore, the loss can be restated as follows:

Where θ _S and θ _∑ are the trainable parameters of the anomaly segmentation network and the multi-source uncertainty mining network, respectively; Formula (5) is called the uncertainty weighted loss, which helps the joint learning of the entire network. Therefore, according to Formula (4) and Formula (5), and extended to all M pseudo labels, the final loss function is:

Under the supervision of formula (6), MUMNet and ASNet are trained together to enhance the ability of MUMNet to capture detailed abnormal feature information, thereby improving the performance of ASNet in anomaly detection.

S15 uses the loss value to update the parameters of the anomaly segmentation network and the multi-source uncertainty mining network.

Repeat the above steps S12-S15 until the loss function converges to obtain the abnormal segmentation network of the training number.

The convergence of the loss function can be judged by the change in the loss value or by setting an upper limit on the number of iterations. When the loss value tends to be stable or reaches the upper limit of the number of iterations, the loss function converges and the training ends.

In this embodiment, the learning rate of MUMNet and ASNet is set to 1e-4, and the Adam optimizer is used for end-to-end joint training. During the training process, the "new training set" is divided into 6 batches, and the network is trained in batches. Five different random seeds are used for evaluation. The entire training process takes about 200 epochs on a GeForce 4080 GPU.

S2 performs unified anomaly detection on the image to be detected.

In this embodiment, samples in the "new test set" are used as images to be detected, and unified anomaly detection is performed according to the following steps:

S21 obtains a reconstructed image of the image to be detected by using the basic model, and takes the Euclidean distance between the image to be detected and its reconstructed image as a biased difference.

S22 inputs the biased difference into the anomaly segmentation network to obtain an anomaly segmented image.

S23 performs global average pooling processing on the abnormal segmented image, and takes the maximum value as the abnormal score of the image to be detected; the obtained abnormal score can be used for abnormality detection.

In this embodiment, the size of the average pooling used to calculate the anomaly score is set to 80.

S24 performs weighted fusion on the abnormal segmentation image and the image to be detected to obtain an abnormal heat map; the obtained abnormal heat map can be used for abnormality positioning.

In this embodiment, the abnormal heat map is obtained according to α·I+(1-α)·S′, where I represents the input image, S′ represents the score map of the output abnormal segmentation image S after colormap conversion, and α represents the weight coefficient; specifically, it can refer to that shown in FIG. 1.

In this embodiment, experiments are conducted on four unsupervised anomaly detection reconstruction models, including models designed for single-class detection (DRAEM and EdgRerc), models designed for multi-class detection (UniAD and MSTAD), and two semi-supervised anomaly detection models (DRA and BGAD). Specifically, the unsupervised models were trained on the "basic training set" of MVTec AD, and their weights were retained. In the subsequent semi-supervised stage, one of the unsupervised models was selected as the base model, and MUMNet and ASNet were further trained using the "new training set" in combination with the MUM-UAD strategy. Before and after fusing MUM-UAD, the unsupervised models were tested with a unified decision boundary on the "new test set".

Table 1 compares the results of the unsupervised base model before and after the introduction of the MUMAD-UAD strategy under the unified decision boundary condition.

“Det.” denotes image-level AUROC, “Loc.” denotes pixel-level AUROC, the MUM-UAD strategy is denoted by “ours”, and the best result in each before-after comparison is highlighted in bold.

AUROC stands for the area under the receiver operating characteristic curve and is used to evaluate the performance of the proposed anomaly detection and localization methods.

Aupro stands for Area Under the Modulus Overlap Curve and is used to measure anomaly location assessment to ensure that anomalies of all sizes are treated equally during the assessment process.

Table 2 Comparison results of semi-supervised methods under unified decision boundary conditions

“B.M.” indicates the base model (e.g., MSTAD), while “ours” represents the MUM-UAD strategy.

Ours w/o B.M. represents the situation where the basic model is missing in this embodiment. At this time, based on the MUM-UAD strategy, MUMNet and ASNet are trained with the help of pseudo labels, and then the trained ASNet is tested.

The best results are highlighted in bold.

As shown in Table 1, when the unsupervised model is combined with the MUM-UAD strategy, it can be applied to semi-supervised tasks and produce different degrees of improvement in detection performance. Specifically, when the base model focuses on a single-class anomaly detection task, the combination of this strategy shows an improvement of more than 10% in image-level AUC or pixel-level AUC. This enhancement highlights the effectiveness of MUM-UAD in empowering a single-class anomaly detection model to solve multi-class anomaly problems. On the contrary, when the base model focuses on multi-class anomaly detection tasks, the method of the present invention shows more superior performance, highlighting its applicability to multi-class scenarios. We further studied the UniAD and MSTAD models. As shown in Figures 2 and 3, after combining the strategy of the present invention, these models established more accurate segmentation boundaries for anomalies, effectively reducing the focus on non-abnormal areas, thereby significantly improving detection performance. This further confirms that integrating MUM-UAD into ASNet can significantly enhance its local perception of anomalies and help improve its overall understanding of images.

For the semi-supervised models (DRA and BGAD), they are directly trained on the “new training set” and tested on the “new test set”. The proposed method performs well in the traditional semi-supervised anomaly detection methods, as shown in Table 2, especially in the absence of an unsupervised base model. In addition, when the knowledge provided by the base model is utilized, the proposed method achieves better results, further verifying its effectiveness.

Those skilled in the art will appreciate that the embodiments herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims (10)

1. The unified anomaly detection method based on multi-source uncertainty mining is characterized by comprising the following steps of:

s1, training an abnormal segmentation network based on a multi-source uncertainty mining semi-supervised learning method, wherein the method comprises the following steps of:

s11, constructing a training set by utilizing data comprising normal samples and abnormal samples, and acquiring multi-source pseudo tags through a plurality of pre-trained basic models;

S12, acquiring a reconstructed image of a sample in a training set by using any basic model, and inputting the Euclidean distance between the sample and the reconstructed image thereof as a biased difference into an abnormal segmentation network to obtain an abnormal segmentation image; the abnormal segmentation network comprises a plurality of characteristic extraction stages which are sequentially arranged, wherein the output of the former stage is used as the input of the latter stage; outputting the image features extracted in at least part of the feature extraction stage to a multi-source uncertainty mining network;

s13, inputting a multi-source pseudo tag into a multi-source uncertainty mining network, combining image features output from an anomaly segmentation network, acquiring global attention distribution based on a cross attention mechanism, and generating an uncertainty weight graph;

S14, constructing a loss function and acquiring a loss value based on the multi-source pseudo tag, the abnormal segmentation image and the uncertainty weight graph;

s15, updating parameters of the abnormal segmentation network and the multi-source uncertainty mining network by using the loss value;

Repeating the steps S12-S15 until the loss function converges to obtain an abnormal segmentation network of the training number;

S2, carrying out unified anomaly detection on the image to be detected, wherein the method comprises the following sub-steps:

S21, acquiring a reconstructed image of an image to be detected by using a basic model, and taking the Euclidean distance between the image to be detected and the reconstructed image thereof as a biased difference;

s22, inputting the deviation into an anomaly segmentation network to obtain an anomaly segmentation image.

2. The unified anomaly detection method based on multi-source uncertainty mining of claim 1, wherein in step S12, the base model is selected from EdgRec, DRAEM, fastFlow or MSTAD.

3. The unified anomaly detection method based on multi-source uncertainty mining of claim 1, wherein each feature extraction stage in the anomaly segmentation network comprises a number of convolution modules; each convolution module consists of a convolution layer, a batch normalization layer, and a ReLU activation function.

4. A unified anomaly detection method based on multi-source uncertainty mining according to claim 3, wherein the next feature extraction stage is entered via a downsampling layer after the feature extraction stages other than the last feature extraction stage, and the last feature extraction stage obtains an anomaly segmented image via an upsampling layer and a convolution module.

5. The unified anomaly detection method based on multi-source uncertainty mining of any one of claims 1 to 4, wherein in step S13, the multi-source uncertainty mining network comprises an encoder, several cross-attention modules, and a decoder arranged in sequence; the encoder is used for encoding the input image characteristics; the cross attention module is used for acquiring the global attention distribution of the image features output by the corresponding feature extraction stage of the abnormal segmentation network on the multi-source pseudo tag based on a cross attention mechanism, and superposing the image features with the output of the encoder or the output of the previous cross attention module, and carrying out patch fusion; the decoder is used for decoding the input image characteristics to obtain an uncertainty weight graph.

6. The unified anomaly detection method based on multi-source uncertainty mining of claim 5, wherein the encoder comprises three convolutional layers and two downsampling layers, the two downsampling layers being located between two adjacent convolutional layers, respectively; the decoder includes a convolutional layer and an upsampling layer.

7. The unified anomaly detection method based on multi-source uncertainty mining of claim 5, wherein the cross-attention module comprises a cross-attention layer, a feed-forward neural network, and a patch fusion layer.

8. The unified anomaly detection method based on multi-source uncertainty mining of claim 5, wherein in step S14, the constructed loss function is:

Wherein, theta _S and theta _∑ are trainable parameters of the anomaly segmentation network and the multi-source uncertainty mining network respectively; Representing the use of pseudo tags by the ith pixel based on the mth pseudo tag And cross entropy loss calculated by the non-normalized score S ⁱ; Representing the predicted logarithmic variance of the prediction, Indicating use of pseudo tagsThe uncertainty weight of the ith pixel is obtained; m represents the number of pseudo tags, H represents the height of the image, and W represents the width of the image.

9. The unified anomaly detection method based on multi-source uncertainty mining of claim 1, wherein step S2 further comprises:

S23, carrying out global average pooling treatment on the abnormal segmentation image, and taking the maximum value as the abnormal score of the image to be detected;

s24, carrying out weighted fusion on the abnormal segmentation image and the image to be detected to obtain an abnormal heat map.

10. The unified anomaly detection method based on multi-source uncertainty mining of claim 8, wherein step S2 further comprises:

S23, carrying out global average pooling treatment on the abnormal segmentation image, and taking the maximum value as the abnormal score of the image to be detected;

s24, carrying out weighted fusion on the abnormal segmentation image and the image to be detected to obtain an abnormal heat map.