CN116310568A - Image anomaly identification method, device, computer readable storage medium and equipment - Google Patents

Abstract

The application provides an image anomaly identification method, an image anomaly identification device, a computer-readable storage medium and electronic equipment, and relates to the technical field of machine learning.

Description

Image abnormality recognition method, device, computer-readable storage medium and equipment

technical field

The present application relates to the technical field of machine learning, and in particular, to a method, device, computer-readable storage medium and equipment for identifying abnormal images.

Background technique

With the continuous development of machine learning technology, neural networks can achieve more and more technical purposes, such as image detection, semantic recognition, etc. In the segmented field of image detection, the neural network can identify abnormal areas in the image (such as tearing abnormal areas), so that relevant personnel can adjust equipment parameters to avoid abnormal areas in the image.

In related technologies, an image with an abnormal area is usually input to a neural network, so that the neural network extracts image features, and identifies the abnormal area and the abnormal type to which the abnormal area belongs based on the image features. However, this method has the problem of low recognition accuracy.

It should be noted that the information disclosed in the above background technology section is only used to enhance the understanding of the background of the application, and therefore may include information that does not constitute the relevant technology known to those of ordinary skill in the art.

Contents of the invention

The purpose of this application is to provide a method, device, computer-readable storage medium, and electronic device for identifying abnormal images, which can first identify the abnormal areas in the image to be processed, and then extract the spectral features, edge features, and statistical features of the abnormal areas. , that is to extract multi-dimensional features that can more comprehensively characterize abnormal areas, and then based on spectral features, edge features, and statistical features, the abnormal type of abnormal areas can be more accurately determined. Compared with related technologies, this application can extract abnormal areas Multi-dimensional features and abnormal type identification based on multi-dimensional features, instead of abnormal type identification based on a single image feature, so the identification accuracy of abnormal types can be improved.

Other features and advantages of the present application will become apparent from the following detailed description, or in part, be learned by practice of the present application.

According to one aspect of the present application, a method for identifying abnormal images is provided, the method comprising:

Identify abnormal regions in the image to be processed;

Extract spectral features, edge features, and statistical features of abnormal regions;

Determine the abnormal type of the abnormal region based on the spectral characteristics, edge characteristics, and statistical characteristics.

According to one aspect of the present application, a device for identifying an abnormal image is provided, the device comprising:

An abnormal area identification unit, configured to identify an abnormal area in the image to be processed;

A multi-dimensional feature extraction unit is used to extract spectral features, edge features, and statistical features of abnormal regions;

An anomaly type determining unit, configured to determine an anomaly type of an abnormal area based on frequency spectrum features, edge features, and statistical features.

According to an aspect of the present application, there is provided a computer program product or computer program comprising computer instructions stored in a computer readable storage medium. The processor of the computer device reads the computer instruction from the computer-readable storage medium, and the processor executes the computer instruction, so that the computer device executes the methods provided in the various optional implementation manners above.

According to one aspect of the present application, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, any one of the above-mentioned methods is implemented.

According to an aspect of the present application, there is provided an electronic device, including: a processor; and a memory for storing executable instructions of the processor; wherein, the processor is configured to perform any one of the above-mentioned methods by executing the executable instructions .

Exemplary embodiments of the present application may have some or all of the following beneficial effects:

In the image abnormality recognition method provided in an exemplary embodiment of the present application, the abnormal region in the image to be processed can be identified first, and then the spectral features, edge features, and statistical features of the abnormal region can be extracted, that is, the extraction can be more comprehensive. Characterize the multi-dimensional features of abnormal areas, and then based on spectral features, edge features, and statistical features, the abnormal type of abnormal areas can be more accurately determined. Compared with related technologies, this application can extract multi-dimensional features for abnormal areas and perform Abnormal type identification, instead of abnormal type identification based on a single image feature, can improve the identification accuracy of abnormal types. In addition, since this application subdivides the identification process of abnormal areas and the identification process of abnormal types, and based on this, it only performs feature extraction for abnormal areas. Identification and abnormal type identification, the present application can further improve the identification accuracy of abnormal types.

It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the application.

Description of drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the application and together with the description serve to explain the principles of the application. Apparently, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative efforts.

FIG. 1 schematically shows a flow chart of a method for identifying abnormal images according to an embodiment of the present application;

Fig. 2 schematically shows a schematic diagram of a feature pyramid structure according to an embodiment of the present application;

Fig. 3 schematically shows a schematic diagram of image slices according to an embodiment of the present application;

FIG. 4 schematically shows a flow chart of a method for identifying an image abnormality according to another embodiment of the present application;

Fig. 5 schematically shows a schematic structural view of an image abnormality identification device according to an embodiment of the present application;

Fig. 6 schematically shows a structural diagram of a computer system suitable for implementing the electronic device of the embodiment of the present application.

Detailed ways

Example embodiments will now be described more fully with reference to the accompanying drawings. Example embodiments may, however, be embodied in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that this application will be thorough and complete, and will fully convey the concepts of example embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided in order to give a thorough understanding of embodiments of the present application. However, those skilled in the art will appreciate that the technical solutions of the present application can be practiced without one or more of the specific details, or other methods, components, devices, steps, etc. can be used. In other instances, well-known technical solutions have not been shown or described in detail to avoid obscuring aspects of the application.

Please refer to FIG. 1 , which schematically shows a flow chart of a method for identifying an image abnormality according to an embodiment of the present application. As shown in Figure 1, the method includes the following steps.

Step S110: Identify abnormal regions in the image to be processed.

Step S120: Extracting spectral features, edge features, and statistical features of the abnormal region.

Step S130: Determine the abnormal type of the abnormal region based on the spectral features, edge features, and statistical features.

By implementing the method shown in Figure 1, the abnormal area in the image to be processed can be identified first, and then the spectral features, edge features, and statistical features of the abnormal area can be extracted, that is, multi-dimensional features that can more comprehensively characterize the abnormal area can be extracted, and then based on the spectrum Features, edge features, and statistical features can more accurately determine the abnormal type of abnormal areas. Compared with related technologies, this application can extract multi-dimensional features for abnormal areas and identify abnormal types based on multi-dimensional features instead of based on a single image The features are used to identify abnormal types, so the recognition accuracy for abnormal types can be improved. In addition, since this application subdivides the identification process of abnormal areas and the identification process of abnormal types, and based on this, it only performs feature extraction for abnormal areas. Identification and abnormal type identification, the present application can further improve the identification accuracy of abnormal types.

Next, the above-mentioned steps of this exemplary embodiment will be described in more detail.

In step S110, abnormal regions in the image to be processed are identified.

Specifically, images captured by terminal devices (such as mobile phones, tablet computers, etc.) sometimes have abnormal conditions, such as abnormal color blocks, abnormal colors, dislocation abnormalities, blurred screen abnormalities, stripe abnormalities, tearing abnormalities, etc. These abnormalities conditions can affect the quality of the image. The reason for these abnormalities may be that the built-in camera module parameters of the terminal device are abnormal, the camera module hardware is abnormal, or the screen quality of the terminal device is not good, or the shooting algorithm of the terminal device is not good. In order to solve this problem, it is necessary to identify the abnormality of the image taken by the device that needs to be debugged, and then the above-mentioned reasons can be debugged based on the result of the abnormality identification to correct the abnormality.

However, in related technologies, the neural network is usually trained through images with abnormal regions, so that the neural network can learn these abnormal regions. In the training process of the neural network, the image features are usually extracted first, and then the abnormal area and the abnormal type of the abnormal area are identified at the same time based on the image features. , the abnormal area was not identified, the abnormal type was identified incorrectly, etc.

The present application believes that there is room for optimization in the above process, that is, on the one hand, the abnormal area identification process and the abnormal type identification process can be divided to avoid identifying the abnormal type based on the feature extraction results for the whole image. Considering this, it is only to divide the steps of related technologies. In order to further improve the recognition accuracy, on the other hand, multi-dimensional feature extraction can be performed based on the identified abnormal area. The abnormal type of the abnormal area can be more accurately identified through multi-dimensional features. . Based on this, the present application takes the above-mentioned color block abnormalities, heterochromatic abnormalities, dislocation abnormalities, blurred screen abnormalities, stripe abnormalities, tearing abnormalities, etc. as abnormal types to train the neural network.

First, the application can identify abnormal regions in the image to be processed based on a detection network (eg, YoloV7 network); wherein, the detection network can set a corresponding feature extraction layer based on actual needs, which is not limited in the embodiment of the application. In addition, the image to be processed can be in the format of JPEG, GIF, PNG, BMP, etc., and the number of abnormal regions in the image to be processed can be one or more, and the sizes of the abnormal regions in the same image to be processed can be the same or different.

In step S120, spectral features, edge features, and statistical features of the abnormal region are extracted.

Specifically, spectral features can be used to characterize the intensity of grayscale changes to illustrate image textures; edge features can be used to represent edge information that is clearly different from normal images; statistical features can be used to characterize statistical feature summaries.

As an optional embodiment, extracting the spectral features of the abnormal region includes: performing region segmentation on the abnormal region to obtain a set of sub-regions; performing Fourier transform on the abnormal region and the set of sub-regions to obtain a set of spectral images; The collection determines the spectral signature of the anomalous region. In this way, the spectral features of the abnormal area can be obtained based on Fourier transform. The spectral features can describe the abnormal area from the frequency domain. Using the spectral feature as the identification condition of the abnormal type can improve the identification accuracy of the abnormal type.

Specifically, for each identified abnormal region in the image to be processed, the corresponding spectral features can be obtained through the above steps.

Wherein, performing region segmentation on the abnormal region to obtain a set of subregions includes: performing region segmentation processing on the abnormal region according to a preset division rule (for example, generating 4 parts/16 parts) to obtain a set of subregions, and in the set of subregions Each sub-region is derived from the anomalous region.

Furthermore, the Fourier transform can be performed on the abnormal region and each subregion in the subregion set, so as to obtain a spectrum image set, which not only includes the spectrum images of each subregion but also includes the spectrum image of the abnormal region; wherein, The spectral features recorded by each spectral image may be different.

Furthermore, the image filter can be triggered to extract the spectral features of each spectral image in the spectral image set; furthermore, the spectral features of each spectral image can be spliced into a spectral image of an abnormal region, or the spectral features of each spectral image can be fused into an abnormal region. The spectral image of the region, the spectral feature is a high-dimensional vector; wherein, the image filter can be regarded as a square ring from the middle to the surrounding (for example, the filter length and width are both a, and the image filter is composed of a/2 rings), and each ring shares weights to reduce filter parameters.

As an optional embodiment, extracting edge features of the abnormal region includes: acquiring depth features of the abnormal region; performing edge detection on the abnormal region to obtain reference edge features; and determining edge features of the abnormal region based on the depth features and the reference edge features. In this way, the edge features of the abnormal area can be obtained based on edge detection. The edge features can highlight and describe the edge of the abnormal area. Using the spectral feature as the identification condition of the abnormal type can improve the identification accuracy of the abnormal type.

Specifically, since the edge of the image with the abnormal area has information that is obviously different from the normal image, the depth features (R, G, B) of the abnormal area can be obtained. The way to obtain the depth feature of the abnormal area is to perform a A series of processing, so that the processing results contain deep features. Among them, a series of processing methods are as follows: filter the abnormal area through Gaussian filter to smooth the non-edge area with weak texture, which is beneficial to obtain more accurate edge features; After the horizontal gradient G _x and vertical gradient G _y of pixels, the magnitude G and direction θ of each pixel can be calculated based on the horizontal gradient G _x and vertical gradient G _y ; furthermore, each pixel in the abnormal area can be traversed, In order to suppress the pixels of the local maximum in the positive/negative gradient direction within each predetermined range (eg, 3*3 pixel range); furthermore, based on the specified gradient magnitude thresholds G ₁ and G ₂ , the values larger than G ₁ can be The edge with amplitude G is set as a strong edge, the edge with amplitude G that is smaller than G ₁ and greater than G ₂ is set as a virtual edge, and the edge with amplitude G that is smaller than G ₂ is suppressed; where G ₁ is greater than G ₂ . Furthermore, virtual edges can be traversed to suppress virtual edges that are not connected with strong edges.

As an optional embodiment, performing edge detection on the abnormal region to obtain reference edge features includes: determining the maximum and minimum edge magnitude values of the abnormal region; based on the maximum and minimum edge magnitude values, dividing the abnormal region Each edge amplitude value in is updated as the reference edge feature. In this way, each edge amplitude value can be updated based on the maximum edge amplitude value and the minimum edge amplitude value, so as to obtain a reference edge feature, which can improve the characterization accuracy of the finally determined edge feature of the abnormal region.

Specifically, the amplitude G of each edge of the abnormal area can be counted to determine the maximum edge amplitude value G _min and the minimum edge amplitude value G _max . Based on this, based on the maximum edge amplitude value and the minimum edge amplitude value, each edge amplitude value in the abnormal area is updated to the reference edge feature, including: through the linear difference method, all the amplitudes G are based on the maximum edge amplitude value G _min and the minimum edge amplitude value The edge magnitude value G _max is transformed into an integer value on the interval [0,255], ie the reference edge feature G _new .

Based on this, the edge features of the abnormal area are determined based on the depth feature and the reference edge feature, including: based on the depth feature (R, G, B) and the reference edge feature G _new , each pixel in the abnormal area can be expressed as (R, G, B, G _new ); furthermore, based on the feature pyramid structure (FPN, Feature Pyramid Network), the edge features of the abnormal region expressed as (R, G, B, G _new ) are extracted.

Among them, the structure of FPN can refer to Figure 2, as shown in Figure 2, you can input pictures (that is, abnormal regions) to FPN, so as to perform feature extraction operations based on each feature extraction layer in the residual neural network (ResNet), in ResNet , features C1, C2, C3, C4, and C5 corresponding to different depths can be obtained through multiple feature extractions. After performing 1×1 convolution on C5 to obtain P5, the upsampling result of P5 can be generated; After the P5 obtained after the convolution of ×1, the P4 upsampling result can be generated based on the P5 upsampling result; after the P4 obtained after the 1×1 convolution of C3, the P3 upsampling result can be generated based on the P4 upsampling result; After performing 1×1 convolution on C2 to obtain P3, the P2 upsampling result can be generated based on the P3 upsampling result. Furthermore, a 3×3 convolution can be performed on each upsampling result to eliminate the overlap effect caused by the upsampling process, and then the multi-scale edge features of the abnormal region can be generated based on each 3×3 convolution result.

As an optional embodiment, extracting the statistical features of the abnormal area includes: performing feature extraction on each feature point in the abnormal area to obtain a feature point vector set; performing mixed Gaussian distribution calculation on the feature point vector set to obtain a distribution function; The likelihood function is expressed based on the distribution function, and the weight partial derivative, mean partial derivative, and variance partial derivative of the likelihood function are calculated; the weight partial derivative, mean partial derivative, and variance partial derivative are fused into the statistical characteristics of the abnormal area. In this way, the statistical characteristics of the abnormal area can be determined based on Gaussian distribution, likelihood function, partial derivative, etc. The statistical characteristics can describe the abnormal area from the statistical dimension, and the statistical characteristics can be used as the identification condition of the abnormal type, which can improve the identification accuracy of the abnormal type .

以得到分布函数p(x)；其中，N(x|μ_k,∑_k)表示高斯分布的概率密度函数，K表示高斯模型的个数，π_k、μ_k、∑_k分别表示第k个高斯模型的权重、均值和方差。进而，基于分布函数p(x)可以表示似然函数/>

其中，λ＝{π_k,μ_k,∑_k,k＝1,…,K}。进而，可以对似然函数求权重偏导、均值偏导、方差偏导，并将权重偏导、均值偏导、方差偏导归一化后拼接为定长的特征向量，作为异常区域的统计特征。Specifically, feature extraction can be performed on each feature point in the abnormal area to obtain a feature point vector set X={x _t ,t=1,...,T}, where the feature point in the abnormal area refers to the gray value of the image A point that changes drastically or a point with a large curvature on the edge of the image (ie, the intersection of two edges), T represents the feature point, and x _t represents the feature point vector (SIFT) of the t-th feature point. Furthermore, the parameters in the mixed Gaussian distribution can be calculated based on X={x _t ,t=1,...,T}

To get the distribution function p(x); among them, N(x|μ _k ,∑ _k ) represents the probability density function of Gaussian distribution, K represents the number of Gaussian models, π _k , μ _k , ∑ _k represent the kth Gaussian model weights, mean and variance. Furthermore, the likelihood function can be expressed based on the distribution function p(x) />

Wherein, λ={π _k , μ _k , ∑ _k , k=1, . . . , K}. Furthermore, weight partial derivatives, mean partial derivatives, and variance partial derivatives can be calculated for the likelihood function, and the weight partial derivatives, mean partial derivatives, and variance partial derivatives can be normalized and spliced into fixed-length feature vectors as the statistics of abnormal regions feature.

As an optional embodiment, identifying the abnormal region in the image to be processed includes: obtaining a sample data set and an enhanced data set; training a detection network based on the sample data set and the enhanced data set; identifying the image to be processed based on the trained detection network Anomalies in the image. In this way, the classification network can be trained with a large sample through the sample data set and the enhanced data set, which can improve the detection accuracy of the detection network for abnormal areas. The number of samples in the related technology is insufficient. Compared with the related technology, this application can increase the sample size. Increase the training intensity of the detection network.

Specifically, the sample data in the sample data set is the image containing the abnormal area. This type of image can be captured by the terminal device, or it can be obtained from an external shooting terminal device when the terminal displays content, or in other ways. This application Examples are not limited. In addition, the enhanced data in the enhanced data set may be generated based on sample data, or generated based on other images containing abnormal regions, or in other ways, which are not limited in this embodiment of the present application. Training the detection network based on the sample data set and the enhanced data set can enable the detection network to obtain detection results with higher accuracy when detecting abnormal regions.

As an optional embodiment, obtaining an enhanced data set includes: generating one type of enhanced data through an adversarial network; generating two types of enhanced data through region transformation; obtaining three types of enhanced data through image slices; At least one of the second type of enhanced data and the third type of enhanced data is used to determine the enhanced data set. In this way, a variety of enhanced data can be obtained in various ways to enrich the sample types, and based on the diversified enhanced data, it can help improve the accuracy of the classification network.

Specifically, the confrontation network may include a generator and a discriminator. The generator is used to generate abnormal images, and the discriminator is used to distinguish whether the received image is a real image or a fake image generated by the generator. The region transformation method is used to simulate an image containing an abnormal region through a variety of image processing methods. The image slicing method is used to cut out multiple sub-images from the image containing the abnormal region. Optionally, the image slicing method can be implemented as a slice-assisted hyper-inference algorithm (SlicingAided Hyper Inference, SAHI) or other algorithms. limited.

As an optional embodiment, generating a type of enhanced data through an adversarial network includes: training the adversarial network through a type of specified image set; generating a type of enhanced data through a generator in the trained adversarial network. In this way, an adversarial network for generating images containing abnormal regions can be trained, and the adversarial network can be used to generate enhanced data to train the classification network, which can increase the training intensity for the classification network.

Specifically, a class of designated image sets may contain multiple images with abnormal regions.

The generator in the confrontation network includes multiple feature extraction layers, and the parameters such as the convolution kernel size and step size of each feature extraction layer can be set arbitrarily according to actual needs. For example, in the first feature extraction layer, the convolution kernel size is 4, the step size is 1, and no padding is performed; in the second feature extraction layer to the fourth feature extraction layer, the convolution kernel size is 4, the step size is 2, and the filling is 1; after the fourth feature extraction layer, it includes a normalization layer, a linear rectification layer (ReLU) and a hyperbolic tangent function layer (Tanh). For the generator, a high-dimensional random vector can be obtained, and the feature map representation that can be processed by the convolution operation can be obtained through the deconvolution operation, that is, a three-channel color image.

Furthermore, the discriminator in the adversarial network can judge whether the three-channel color image is a fake image generated by the generator. If the discriminator cannot correctly distinguish the real image from the fake image, then adjust the parameters of the discriminator until the discriminator can correctly distinguish real and fake images. Wherein, the discriminator may include multiple feature extraction layers, and parameters such as convolution kernel size and step size of each feature extraction layer may be set arbitrarily according to actual needs. For example, in the first feature extraction layer, the convolution kernel size is 4, the stride is 2, and the padding is 1; in the second to fourth feature extraction layers, the convolution kernel size is 4, and the stride is 1. The length is 1, no padding is performed; after the fourth feature extraction layer, it includes a normalization layer and a linear rectification layer (ReLU), a leaky linear rectification function layer (Leaky ReLU), and a S-type function layer (Sigmoid).

As an optional embodiment, training the adversarial network through a class of specified image sets includes: processing a class of specified image sets as adversarial training samples; triggering the generator in the adversarial network to generate reference training samples; The discriminator in the confrontation network is trained against the training samples; the generator is trained based on the discriminant result of the discriminator. In this way, the adversarial network training based on a specified image set can be realized, so that the generator in the adversarial network can generate realistic abnormal images for the training of the detection network.

Specifically, the generator can generate reference training samples based on high-dimensional random vectors, and then, the reference training samples can be input into the discriminator, and the discriminator should judge that the reference training samples are fake images, otherwise adjust parameters. In addition, it is also necessary to input the images in a specified image set into the discriminator, and the discriminator should determine that the images in a specified image set are real images, otherwise adjust the parameters.

As an optional embodiment, processing a class of specified image sets as adversarial training samples includes: cropping abnormal regions for each image in a class of specified image sets; if the size of the abnormal region is larger than the preset size, then The size cuts the abnormal region into multiple subgraphs. If the size of the abnormal region is smaller than or equal to the preset size, the abnormal region is scaled to obtain the abnormal subgraph; at least one of the multiple subgraphs and the abnormal subgraph is determined as an adversarial Training samples. In this way, each image in a specified image set can be adjusted so that each image meets the sample requirements, and then, training the adversarial network based on each adjusted image can improve training efficiency.

Specifically, when the size of the abnormal area is larger than the preset size (eg, 256*256), the abnormal area can be cut into multiple sub-images satisfying the preset size according to the preset size, and after obtaining the multiple sub-images, further The abnormal region can be scale-transformed based on a preset size, and then the scale-transformed abnormal region is also determined as an adversarial training sample. When the size of the abnormal region is smaller than or equal to the preset size, in addition to performing scale transformation on the abnormal region, the abnormal region can also be cut into multiple reference subimages according to the preset number of divisions (for example, 16). Since the original abnormal region does not meet the predetermined The size is set, so multiple reference subgraphs must not meet the preset size. At this time, multiple reference subgraphs can be scaled based on the preset size, so that multiple reference subgraphs after scale transformation are also determined as confrontation training. samples, which can enrich the sample size for adversarial training.

As an optional embodiment, the region transformation method includes a mosaic processing method and/or a random color change processing method, and generating the second type of enhanced data through the region transformation method includes: performing region selection on the images in the second type of designated image set to obtain The area to be processed; through the mosaic processing method and/or the random color change processing method, perform area transformation on the area to be processed to obtain the area transformation result; replace each area to be processed with the corresponding area transformation result to obtain the second type of enhanced data. In this way, the types of enhanced data can be enriched by region transformation, and the problem of large data volume differences between abnormal types can be avoided. Furthermore, if the detection network is trained based on diversified enhanced data, the reliability and robustness of the detection network can be improved.

Specifically, the region transformation manner may also include other processing manners, such as background blurring, which is not limited in this embodiment of the present application. In addition, the images in the second-type specified image set may be normal regions (ie, images not containing abnormal regions) or images with abnormal regions, which is not limited in this embodiment of the present application.

Wherein, area selection is carried out to the images in the second class designated image set to obtain the area to be processed, including: randomly selecting the area based on the preset size (eg, 128*128) to the images in the second class designated image set to obtain the area to be processed; Wherein, the region selection times for the same image may be one or more times, and if it is multiple times, there may or may not be content overlap among the multiple regions selected for the same image.

In addition, performing area transformation on the area to be processed by mosaic processing and/or random color change processing to obtain the area transformation result, including: performing area transformation on the area to be processed by mosaic processing to obtain the area transformation result; or, using random color The processing method is changed, and the region to be processed is transformed to obtain a region transformation result; or, the region to be processed is transformed to obtain a region transformation result through the mosaic processing method and the random color change processing method.

Among them, optionally, the area to be processed is transformed by mosaic processing to obtain the result of the area transformation, including: reducing the scale of the area to be processed to a preset pixel size by linear interpolation, and converting the area to be processed to a preset pixel size by nearest neighbor interpolation. The processing area is enlarged to the size of the original area by transforming the size, and the result of the area transformation is obtained.

Wherein, optionally, through the random color change processing method, the region to be processed is transformed to obtain the region transformation result, including: determining the height (height) of the region to be processed and the random number x; The pixel value at the pixel position of height is exchanged with the pixel value at the pixel position of 0～x·height to obtain the area transformation result; wherein, the random number x belongs to [0,1].

Furthermore, the result of the region transformation can be overlaid on the original region to be processed to obtain a new image as the second type of enhanced data.

As an optional embodiment, the three types of enhanced data are obtained by means of image slices, including: performing abnormal area scale transformation on each image in the three types of specified image sets to obtain the area to be processed; sliding the area to be processed through a preset sliding window Slice to obtain image slices; determine each image slice as three types of enhanced data. In this way, based on the image slicing method, the problem of lack of pixel details of the training samples in the abnormal area is too small, and the sliding slice of the area to be processed by the preset sliding window can enrich and enhance the sample size of the small abnormal area in the data, thereby effectively It is beneficial to improve the detection accuracy of the detection network for small-sized abnormal regions.

Specifically, the abnormal region of each image in the three types of designated image sets may be scale-transformed to a preset size (eg, 480*480), so as to obtain the region to be processed. Furthermore, the area to be processed can be slided and sliced through a preset sliding window to obtain an image slice. Please refer to FIG. 3 for details. FIG. 3 schematically shows an image slice according to an embodiment of the present application. As shown in FIG. 3 , the image slice 310 and the image slice 320 can be obtained after the area to be processed 300 is slidingly sliced based on the preset sliding window. The image slice 310 and the image slice 320 here are only for illustration. In the actual application process, Any number of image slices can be fetched. The size of the preset sliding window (for example, 256*256) can be set arbitrarily according to actual needs, and is not limited here.

In addition, there may or may not be an image intersection among the designated image sets of the first class, the designated image sets of the second class, and the designated image sets of the third class.

In step S130, the abnormal type of the abnormal region is determined based on the spectral features, edge features, and statistical features.

Specifically, there is usually one abnormal type in an abnormal area, and optionally, there may be multiple abnormal types in an abnormal area, that is, there may be multiple abnormal types in the same abnormal area.

As an optional embodiment, determining the abnormal type of the abnormal region based on spectral features, edge features, and statistical features includes: fusing spectral features, edge features, and statistical features into target fusion features; determining the abnormal area based on target fusion features exception type. In this way, abnormal types can be identified based on the result of fusing multi-dimensional features, and the accuracy of identifying abnormal types can be improved.

Specifically, the fusion of spectral features, edge features, and statistical features into target fusion features includes: splicing spectral features (v ₁ ), edge features (v ₂ ), and statistical features (v ₃ ) to obtain a vector of length d t=[v ₁ ,v ₂ ,v ₃ ] ^T , determine the vector t as the target fusion feature; or, weight the spectral feature (v ₁ ), edge feature (v ₂ ), statistical feature (v ₃ ) and processing to obtain the target fusion features.

As an optional embodiment, determining the abnormality type of the abnormal region based on the target fusion features includes: acquiring target features corresponding to each image to be processed; wherein each target feature includes the target fusion feature; determining the abnormality based on each target feature The exception type for the zone. In this way, the abnormal type of the abnormal area of the current image to be processed can be determined based on the target features of the multiple images to be processed, and the recognition accuracy for the abnormal type can be improved.

Specifically, it may correspond to the target features [t ₁ , ..., t _N ] of each image to be processed [I ₁ , ..., I _N ]; wherein, N here may represent the number of images to be processed. Furthermore, the abnormal type of the abnormal region of the current image to be processed can be determined based on [t ₁ , . . . , t _N ].

As an optional embodiment, determining the abnormal type of the abnormal region based on each target feature includes: constructing a feature data matrix according to the feature mean value of each target feature; calculating the covariance matrix based on the feature data matrix, and extracting the A plurality of eigenvectors; a target matrix is generated based on a projection matrix and a feature data matrix constructed by the plurality of eigenvectors; and an abnormal type of the abnormal region is determined according to the target matrix. In this way, each target feature can be fused into a feature data matrix, and based on the feature data matrix and its corresponding covariance matrix, a target matrix for identifying abnormal types can be generated, which is conducive to improving the target features of other images for the abnormal type judgment of the current image. function to improve the recognition accuracy for abnormal types.

Specifically, the feature data matrix D=[t ₁ -δ, ...,t _{N -δ} ] can be constructed according to the feature mean δ of each target feature [t ₁ ,...,t _N ]. Furthermore, the covariance moment C=1/dDD ^T can be calculated based on the characteristic data matrix D, and the eigenvectors in the covariance matrix C can be sorted from large to small/small to large, and the top k (k>d) can be extracted eigenvectors, and based on the projection matrix P constructed by the first k eigenvectors, the PD is calculated by matrix multiplication to obtain the target matrix. Furthermore, it is also possible to perform dimension reduction processing on the target matrix, input the data matrix after dimension reduction processing to the classification network, and obtain the abnormal type of the abnormal area output by the classification network, and generate The loss function is used to realize the training of the classification network based on the loss function, so that the classification network has the ability to determine the abnormal type of the abnormal region.

Referring to FIG. 4 , FIG. 4 schematically shows a flowchart of a method for identifying an image abnormality according to another embodiment of the present application. As shown in FIG. 4 , the method for identifying abnormal images includes: step S400 to step S416 .

Step S400: Crop the abnormal region for each image in a specified image set. If the size of the abnormal region is larger than the preset size, then crop the abnormal region into multiple sub-images based on the preset size. If the size of the abnormal region is smaller than or equal to the preset size , then perform scale transformation on the abnormal region to obtain the abnormal submap. Furthermore, at least one of multiple subgraphs and abnormal subgraphs can be determined as an adversarial training sample, triggering the generator in the adversarial network to generate a reference training sample, and training the discriminator in the adversarial network through the reference training sample and the adversarial training sample , and the generator is trained based on the discriminative results of the discriminator. Furthermore, a class of augmented data is generated by the generator in the trained adversarial network.

Step S402: Select the region of the images in the second type of designated image set to obtain the region to be processed, and perform region transformation on the region to be processed by mosaic processing and/or random color change processing to obtain a region transformation result, and convert each The area to be processed is replaced by the corresponding area transformation result, and the second type of enhanced data is obtained.

Step S404: Scale-transform the abnormal region of each image in the three designated image sets to obtain the region to be processed, and perform sliding slices on the region to be processed through the preset sliding window to obtain image slices, and determine each image slice as the three types of enhanced data.

Step S406: Determine an enhanced data set based on at least one of the first type of enhanced data, the second type of enhanced data, and the third type of enhanced data, acquire a sample data set, and train a detection network based on the sample data set and the enhanced data set.

Step S408: Identify abnormal regions in the image to be processed based on the trained detection network.

Step S410: Segment the abnormal area to obtain a set of sub-areas, perform Fourier transform on the abnormal area and the set of sub-areas to obtain a set of spectral images, and then determine the spectral features of the abnormal area based on the set of spectral images.

Step S412: Obtain the depth feature of the abnormal area, determine the maximum and minimum edge amplitude values of the abnormal area, and then update each edge amplitude value in the abnormal area to the reference edge feature based on the maximum and minimum edge amplitude values, The edge features of abnormal regions are determined based on depth features and reference edge features.

Step S414: Perform feature extraction on each feature point in the abnormal area to obtain a feature point vector set, perform mixed Gaussian distribution calculation on the feature point vector set to obtain a distribution function, and then express the likelihood function based on the distribution function, and calculate the likelihood function Weight partial derivation, mean partial derivation, and variance partial derivation, fused weight partial derivation, mean partial derivation, and variance partial derivation into the statistical characteristics of the abnormal area.

Step S416: Fuse the spectral features, edge features, and statistical features into target fusion features; acquire target features corresponding to each image to be processed; wherein, each target feature includes target fusion features. Furthermore, a feature data matrix is constructed according to the feature mean value of each target feature, a covariance matrix is calculated based on the feature data matrix, and multiple feature vectors in the covariance matrix are extracted. And, a target matrix is generated based on the projection matrix constructed by multiple feature vectors and the feature data matrix, and the abnormal type of the abnormal region is determined according to the target matrix.

It should be noted that steps S400 to S416 correspond to the steps and their embodiments shown in FIG. 1 . For the specific implementation of steps S400 to S416, please refer to the steps and their embodiments shown in FIG. 1 . I won't repeat them here.

It can be seen that by implementing the method shown in Figure 4, the abnormal area in the image to be processed can be identified first, and then the spectral features, edge features, and statistical features of the abnormal area can be extracted, that is, multi-dimensional features that can more comprehensively characterize the abnormal area can be extracted, and then Based on spectral features, edge features, and statistical features, the abnormal type of abnormal areas can be more accurately determined. Compared with related technologies, this application can extract multi-dimensional features for abnormal areas and identify abnormal types based on multi-dimensional features, rather than based on a single The image features of the image are used to identify abnormal types, so the recognition accuracy of abnormal types can be improved. In addition, since this application subdivides the identification process of abnormal areas and the identification process of abnormal types, and based on this, it only performs feature extraction for abnormal areas. Identification and abnormal type identification, the present application can further improve the identification accuracy of abnormal types.

Please refer to FIG. 5 , which schematically shows a structural block diagram of an apparatus for identifying abnormal images according to an embodiment of the present application. As shown in FIG. 5 , the apparatus 500 for identifying abnormal images may include the following units.

An abnormal area identification unit 501, configured to identify an abnormal area in the image to be processed;

A multi-dimensional feature extraction unit 502, used to extract spectral features, edge features, and statistical features of abnormal regions;

An abnormality type determining unit 503, configured to determine the abnormality type of the abnormal area based on spectral features, edge features, and statistical features.

Among them, the abnormal types include: color block abnormality, heterochromatic abnormality, dislocation abnormality, blurred screen abnormality, stripe abnormality, and tearing abnormality.

It can be seen that the implementation of the device shown in Figure 5 can first identify the abnormal area in the image to be processed, and then extract the spectral features, edge features, and statistical features of the abnormal area, that is, extract the multi-dimensional features that can more comprehensively characterize the abnormal area, and then Based on spectral features, edge features, and statistical features, the abnormal type of abnormal areas can be more accurately determined. Compared with related technologies, this application can extract multi-dimensional features for abnormal areas and identify abnormal types based on multi-dimensional features, rather than based on a single The image features of the image are used to identify abnormal types, so the recognition accuracy of abnormal types can be improved. In addition, since this application subdivides the identification process of abnormal areas and the identification process of abnormal types, and based on this, it only performs feature extraction for abnormal areas. Identification and abnormal type identification, the present application can further improve the identification accuracy of abnormal types.

As an optional embodiment, the multi-dimensional feature extraction unit 502 extracts the spectral features of the abnormal region, including:

Segment the abnormal region to obtain a set of sub-regions;

Perform Fourier transform on the abnormal area and the sub-area set to obtain the spectrum image set;

Spectral features of abnormal regions are determined based on a set of spectral images.

It can be seen that the implementation of this optional embodiment can obtain the spectral features of the abnormal area based on Fourier transform, and the spectral features can describe the abnormal area from the frequency domain, and the spectral feature can be used as the identification condition of the abnormal type, which can improve the identification accuracy of the abnormal type .

As an optional embodiment, the multi-dimensional feature extraction unit 502 extracts edge features of the abnormal region, including:

Obtain the depth features of the abnormal region;

Perform edge detection on abnormal areas to obtain reference edge features;

The edge features of abnormal regions are determined based on depth features and reference edge features.

It can be seen that by implementing this optional embodiment, the edge features of the abnormal area can be obtained based on edge detection. The edge features can highlight and describe the edge of the abnormal area. Using the spectral feature as the identification condition of the abnormal type can improve the identification accuracy of the abnormal type.

As an optional embodiment, the multi-dimensional feature extraction unit 502 performs edge detection on the abnormal region to obtain reference edge features, including:

Determining the maximum and minimum edge amplitude values of the abnormal region;

Based on the maximum edge amplitude value and the minimum edge amplitude value, each edge amplitude value in the abnormal area is updated as a reference edge feature.

It can be seen that, implementing this optional embodiment, each edge amplitude value can be updated based on the maximum edge amplitude value and the minimum edge amplitude value, so as to obtain the reference edge feature, which can improve the characterization accuracy of the edge feature of the finally determined abnormal region .

As an optional embodiment, the multi-dimensional feature extraction unit 502 extracts statistical features of abnormal regions, including:

Perform feature extraction on each feature point in the abnormal area to obtain a set of feature point vectors;

Perform mixed Gaussian distribution calculation on the feature point vector set to obtain the distribution function;

Express the likelihood function based on the distribution function, and calculate the weight partial derivative, mean partial derivative, and variance partial derivative of the likelihood function;

The weight partial derivative, mean partial derivative, and variance partial derivative are fused into the statistical characteristics of abnormal regions.

It can be seen that, implementing this optional embodiment, the statistical characteristics of the abnormal region can be determined based on Gaussian distribution, likelihood function, partial derivative, etc. The statistical characteristics can describe the abnormal region from the statistical dimension, and the statistical characteristics can be used as the identification of the abnormal type Conditions can improve the recognition accuracy of abnormal types.

As an optional embodiment, the abnormal region identifying unit 501 identifies the abnormal region in the image to be processed, including:

Get sample dataset and augmented dataset;

Train the detection network based on the sample data set and the enhanced data set;

Identify abnormal regions in the image to be processed based on the trained detection network.

It can be seen that, implementing this optional embodiment, the classification network can be trained with a large sample through the sample data set and the enhanced data set, and the detection accuracy of the detection network for abnormal areas can be improved. The number of samples in the related art is insufficient. Because of related technologies, the sample size can be increased and the training intensity of the detection network can be improved.

As an optional embodiment, the multi-dimensional feature extraction unit 502 acquires an enhanced data set, including:

Generating a class of augmented data via an adversarial network;

Generate the second type of enhanced data by means of area transformation;

Obtain three types of enhanced data through image slicing;

The enhanced data set is determined based on at least one of the first type of enhanced data, the second type of enhanced data, and the third type of enhanced data.

It can be seen that, implementing this optional embodiment, various enhanced data can be obtained in various ways to enrich sample types, and based on the diversified enhanced data, it can be beneficial to improve the accuracy of the classification network.

As an optional embodiment, the multi-dimensional feature extraction unit 502 generates a type of enhanced data through an adversarial network, including:

Train the adversarial network through a class of specified image sets;

A class of augmented data is generated by a generator in a trained adversarial network.

It can be seen that by implementing this optional embodiment, an adversarial network for generating images containing abnormal regions can be trained, and the adversarial network can be used to generate enhanced data to train the classification network, which can increase the training intensity for the classification network.

As an optional embodiment, the multi-dimensional feature extraction unit 502 trains the confrontation network through a class of specified image sets, including:

Process a class of specified image sets as adversarial training samples;

Trigger the generator in the adversarial network to generate reference training samples;

Train the discriminator in the adversarial network by referring to training samples and adversarial training samples;

The generator is trained based on the discriminative results of the discriminator.

It can be seen that implementing this optional embodiment can realize adversarial network training based on a specified image set, so that the generator in the adversarial network can generate realistic abnormal images for training of the detection network.

As an optional embodiment, the multi-dimensional feature extraction unit 502 processes a class of designated image sets as confrontation training samples, including:

Perform abnormal region cropping on each image in a specified image set;

If the size of the abnormal area is larger than the preset size, the abnormal area is cut into multiple sub-images based on the preset size, and if the size of the abnormal area is smaller than or equal to the preset size, the abnormal area is scale-transformed to obtain the abnormal sub-image;

Determining at least one of multiple subgraphs and abnormal subgraphs as adversarial training samples.

It can be seen that, implementing this optional embodiment, each image in a specified image set of a class can be adjusted so that each image meets the sample requirements, and then training an adversarial network based on each adjusted image can improve training efficiency.

As an optional embodiment, the region transformation method includes a mosaic processing method and/or a random color change processing method, and the multi-dimensional feature extraction unit 502 generates the second type of enhanced data through the region transformation method, including:

Perform region selection on the images in the second type of specified image set to obtain the region to be processed;

By means of mosaic processing and/or random color change processing, the area to be processed is transformed to obtain the result of the region transformation;

Each area to be processed is replaced by the corresponding area transformation result to obtain the second type of enhanced data.

It can be seen that the implementation of this optional embodiment can enrich the types of enhanced data through region transformation, avoiding the problem of large data volume differences between abnormal types, and furthermore, if the detection network is trained based on diversified enhanced data, the detection network can be improved. reliability and robustness.

As an optional embodiment, the multi-dimensional feature extraction unit 502 obtains three types of enhanced data through image slices, including:

Scale transformation of the abnormal region is performed on each image in the three specified image sets to obtain the region to be processed;

Slide and slice the area to be processed through the preset sliding window to obtain image slices;

Each image slice is identified as three types of augmented data.

It can be seen that the implementation of this optional embodiment can solve the problem of lack of pixel details in training samples with too small anomaly regions based on the image slicing method. Sliding and slicing the region to be processed through the preset sliding window can enrich and enhance data for small size The sample size of the abnormal region is beneficial to improve the detection accuracy of the detection network for small abnormal regions.

As an optional embodiment, the abnormal type determining unit 503 determines the abnormal type of the abnormal region based on spectral features, edge features, and statistical features, including:

Combine spectral features, edge features, and statistical features into target fusion features;

The abnormal type of the abnormal region is determined based on the target fusion features.

It can be seen that, implementing this optional embodiment, the abnormal type can be identified based on the result of fusing multi-dimensional features, and the identification accuracy of the abnormal type can be improved.

As an optional embodiment, the abnormal type determination unit 503 determines the abnormal type of the abnormal region based on the target fusion feature, including:

Acquiring target features corresponding to each image to be processed; wherein, each target feature includes a target fusion feature;

The abnormal type of the abnormal region is determined based on each target feature.

It can be seen that by implementing this optional embodiment, the abnormal type of the abnormal area of the current image to be processed can be determined based on the target features of multiple images to be processed, and the recognition accuracy of the abnormal type can be improved.

As an optional embodiment, the abnormal type determining unit 503 determines the abnormal type of the abnormal region based on each target feature, including:

Construct a feature data matrix according to the feature mean of each target feature;

Calculate the covariance matrix based on the characteristic data matrix, and extract multiple eigenvectors in the covariance matrix;

Generate a target matrix based on a projection matrix constructed from multiple eigenvectors and a feature data matrix;

The abnormal type of the abnormal region is determined according to the target matrix.

It can be seen that, implementing this optional embodiment, each target feature can be fused into a feature data matrix, and a target matrix for identifying abnormal types can be generated based on the feature data matrix and its corresponding covariance matrix, which is conducive to improving the target of other images. The role of features in determining the abnormal type of the current image to improve the recognition accuracy of abnormal types.

It should be noted that although several modules or units of the device for action execution are mentioned in the above detailed description, this division is not mandatory. Actually, according to the embodiment of the present application, the features and functions of two or more modules or units described above may be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided to be embodied by a plurality of modules or units.

Since each functional module of the image anomaly recognition device of the example embodiment of the present application corresponds to the steps of the above-mentioned example embodiment of the image anomaly recognition device, for details not disclosed in the device embodiment of the present application, please refer to the above-mentioned An embodiment of an image abnormality recognition device.

Please refer to FIG. 6 . FIG. 6 shows a schematic structural diagram of a computer system suitable for implementing an electronic device according to an embodiment of the present application.

It should be noted that the computer system 600 of the electronic device shown in FIG. 6 is only an example, and should not limit the functions and scope of use of this embodiment of the present application.

As shown in FIG. 6 , a computer system 600 includes a central processing unit (CPU) 601 that can execute programs according to programs stored in a read-only memory (ROM) 602 or loaded from a storage section 608 into a random-access memory (RAM) 603 Instead, various appropriate actions and processes are performed. In the RAM 603, various programs and data necessary for system operation are also stored. The CPU 601 , ROM 602 , and RAM 603 are connected to each other via a bus 604 . An input/output (I/O) interface 605 is also connected to the bus 604 .

The following components are connected to the I/O interface 605: an input section 606 including a keyboard, a mouse, etc.; an output section 607 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker; a storage section 608 including a hard disk, etc. and a communication section 609 including a network interface card such as a LAN card, a modem, or the like. The communication section 609 performs communication processing via a network such as the Internet. A drive 610 is also connected to the I/O interface 605 as needed. A removable medium 611, such as a magnetic disk, optical disk, magneto-optical disk, semiconductor memory, etc., is mounted on the drive 610 as necessary so that a computer program read therefrom is installed into the storage section 608 as necessary.

In particular, according to the embodiments of the present application, the processes described below with reference to the flowcharts can be implemented as computer software programs. For example, the embodiments of the present application include a computer program product, which includes a computer program carried on a computer-readable medium, where the computer program includes program codes for executing the methods shown in the flowcharts. In such an embodiment, the computer program may be downloaded and installed from a network via communication portion 609 and/or installed from removable media 611 . When the computer program is executed by a central processing unit (CPU) 601, various functions defined in the method and apparatus of the present application are performed.

As another aspect, the present application also provides a computer-readable medium. The computer-readable medium may be included in the electronic device described in the above embodiments; it may also exist independently without being assembled into the electronic device. middle. The above-mentioned computer-readable medium carries one or more programs, and when the above-mentioned one or more programs are executed by an electronic device, the electronic device is made to implement the methods in the above-mentioned embodiments.

It should be noted that the computer-readable medium shown in this application may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. A computer readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or device, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to, electrical connections with one or more wires, portable computer diskettes, hard disks, 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 of the above. In this application, a computer-readable storage medium may be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, however, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, in which computer-readable program codes are carried. Such propagated data signals may take many forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the foregoing. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, apparatus, or device. . Program code embodied on a computer readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wireline, optical fiber cable, RF, etc., or any suitable combination of the foregoing.

The flowchart and block diagrams in the Figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present application. In this regard, each block in a flowchart or block diagram may represent a module, program segment, or portion of code that includes one or more logical functions for implementing specified executable instructions. It should also be noted that, in some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or they may sometimes be executed in the reverse order, depending upon the functionality involved. It should also be noted that each block in the block diagrams or flowchart illustrations, and combinations of blocks in the block diagrams or flowchart illustrations, can be implemented by a dedicated hardware-based system that performs the specified function or operation, or can be implemented by a A combination of dedicated hardware and computer instructions.

The units described in the embodiments of the present application may be implemented by software or by hardware, and the described units may also be set in a processor. Wherein, the names of these units do not constitute a limitation of the unit itself under certain circumstances.

Other embodiments of the present application will be readily apparent to those skilled in the art from consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any modification, use or adaptation of the application, which follow the general principles of the application and include common knowledge or conventional technical means in the field not disclosed in the application. The specification and examples are to be considered exemplary only, with a true scope and spirit of the application indicated by the appended claims.

Claims (19)

1. A method for identifying abnormal images, comprising: Identify abnormal regions in the image to be processed; extracting spectral features, edge features, and statistical features of the abnormal region; An abnormal type of the abnormal area is determined based on the frequency spectrum feature, the edge feature, and the statistical feature. 2. The method according to claim 1, wherein extracting the spectral features of the abnormal region comprises: Performing region segmentation on the abnormal region to obtain a set of sub-regions; Performing Fourier transform on the abnormal area and the set of sub-areas to obtain a set of spectrum images; Spectral features of the abnormal region are determined based on the set of spectral images. 3. The method according to claim 1, wherein extracting the edge features of the abnormal region comprises: Obtaining the depth features of the abnormal region; performing edge detection on the abnormal region to obtain reference edge features; Edge features of the abnormal region are determined based on the depth features and the reference edge features. 4. The method according to claim 3, wherein performing edge detection on the abnormal region to obtain reference edge features comprises: determining a maximum edge amplitude value and a minimum edge amplitude value for the abnormal region; Based on the maximum edge magnitude value and the minimum edge magnitude value, each edge magnitude value in the abnormal area is updated as a reference edge feature. 5. The method according to claim 1, wherein extracting the statistical features of the abnormal region comprises: performing feature extraction on each feature point in the abnormal region to obtain a set of feature point vectors; Perform mixed Gaussian distribution calculation on the feature point vector set to obtain the distribution function; Expressing a likelihood function based on the distribution function, and calculating weight partial derivatives, mean partial derivatives, and variance partial derivatives of the likelihood function; The weight partial derivative, the mean partial derivative, and the variance partial derivative are fused into statistical features of the abnormal region. 6. The method according to claim 1, wherein determining the abnormal type of the abnormal region based on the spectral features, the edge features, and the statistical features comprises: Fusing the spectral features, the edge features, and the statistical features into target fusion features; The abnormal type of the abnormal region is determined based on the target fusion feature. 7. The method according to claim 6, wherein determining the abnormal type of the abnormal region based on the target fusion feature comprises: Acquiring target features corresponding to each image to be processed; wherein, each target feature includes the target fusion feature; The abnormal type of the abnormal area is determined based on the respective target features. 8. The method according to claim 7, wherein determining the abnormal type of the abnormal region based on the characteristics of each target comprises: Constructing a feature data matrix according to the feature mean value of each target feature; calculating a covariance matrix based on the characteristic data matrix, and extracting a plurality of eigenvectors in the covariance matrix; generating a target matrix based on the projection matrix constructed from the plurality of eigenvectors and the feature data matrix; An abnormality type of the abnormal area is determined according to the target matrix. 9. The method according to claim 1, wherein identifying the abnormal region in the image to be processed comprises: Get sample dataset and augmented dataset; training a detection network based on the sample data set and the enhanced data set; Identify abnormal regions in the image to be processed based on the trained detection network. 10. The method according to claim 9, wherein obtaining an enhanced data set comprises: Generating a class of augmented data via an adversarial network; Generate the second type of enhanced data by means of area transformation; Obtain three types of enhanced data through image slicing; An enhanced data set is determined based on at least one of the first type of enhanced data, the second type of enhanced data, and the third type of enhanced data. 11. The method according to claim 10, wherein generating a class of enhanced data through an adversarial network comprises: Train the adversarial network through a class of specified image sets; A class of augmented data is generated by a generator in a trained adversarial network. 12. The method according to claim 11, wherein training the confrontation network through a class of designated image sets includes: Process a class of specified image sets as adversarial training samples; Trigger the generator in the adversarial network to generate reference training samples; training a discriminator in an adversarial network through the reference training samples and the adversarial training samples; The generator is trained based on the discriminative results of the discriminator. 13. The method according to claim 12, wherein processing a class of specified image sets as an adversarial training sample includes: Perform abnormal region cropping on each image in a specified image set; If the size of the abnormal area is larger than the preset size, then crop the abnormal area into multiple sub-pictures based on the preset size, and if the size of the abnormal area is smaller than or equal to the preset size, then Scale transformation is performed on the region to obtain an abnormal submap; At least one of the plurality of subgraphs and the abnormal subgraph is determined as an adversarial training sample. 14. The method according to claim 10, characterized in that, the region transformation method includes a mosaic processing method and/or a random color change processing method, and the second type of enhanced data is generated through the region transformation method, including: Perform region selection on the images in the second type of specified image set to obtain the region to be processed; Using the mosaic processing method and/or the random color change processing method, performing region transformation on the region to be processed to obtain a region transformation result; Each area to be processed is replaced by the corresponding area transformation result to obtain the second type of enhanced data. 15. The method according to claim 10, wherein the three types of enhanced data are obtained through image slices, including: Scale transformation of the abnormal region is performed on each image in the three specified image sets to obtain the region to be processed; Sliding and slicing the area to be processed through a preset sliding window to obtain image slices; Each of the image slices is determined as three types of enhanced data. 16. The method according to any one of claims 1-15, wherein the abnormality types include: abnormal color block, abnormal color, abnormal dislocation, abnormal blurring, abnormal stripes, and abnormal tearing. 17. A device for identifying abnormal images, comprising: An abnormal area identification unit, configured to identify an abnormal area in the image to be processed; A multi-dimensional feature extraction unit for extracting spectral features, edge features, and statistical features of the abnormal region; An abnormality type determining unit, configured to determine the abnormality type of the abnormal region based on the frequency spectrum feature, the edge feature, and the statistical feature. 18. A computer-readable storage medium, on which a computer program is stored, wherein the computer program implements the method according to any one of claims 1-16 when executed by a processor. 19. An electronic device, comprising: processor; and a memory for storing executable instructions of the processor; Wherein, the processor is configured to execute the method according to any one of claims 1-16 by executing the executable instructions.

Images