Background similarity recognition method, system, device and medium

Abstract

The present invention provides a method, system, device and medium for background similarity recognition, including: inputting an image to be recognized into a trained real-time instance segmentation model to mask a human body in the image to be recognized, obtaining a masked image to be recognized, extracting features from the masked image to be recognized, and obtaining a first feature; inputting the image to be recognized into a trained autoencoder model to extract features from the image to be recognized, and obtaining a second feature; obtaining a third feature based on the first feature and the second feature; performing a similarity comparison between the third feature and the feature of each background image in a background image data set, and searching for a number of background images that are most similar to the third feature. The first feature and the second feature are obtained respectively through the instance segmentation model and the autoencoder model, and the third feature is obtained based on the two. The vector dimension of the third feature is high and the information is richer. When it is input for similarity comparison, similarity recognition with higher accuracy can be achieved.

Description

Technical Field

[0001] The present invention relates to the field of visual detection, and more specifically, to a method, system, device and medium for background similarity recognition. Background Art

[0002] Background recognition, especially background recognition in video scenes, is one of the common application scenarios in visual detection. In the process of background recognition, instance segmentation technology is often used to pre-segment the background and objects / people in the image and mask things other than the background. Currently, the most widely used instance segmentation algorithm is the YOLACT (You Only Look At CoefficienTs) algorithm, also known as the real-time instance segmentation algorithm. However, instance segmentation in video scenes places higher requirements on algorithm speed and accuracy. Although the YOLACT algorithm can meet the speed requirements, its performance is slightly low, and there will be problems with inaccurate edge segmentation.

[0003] Secondly, another commonly used technology is feature extraction technology, which is used to extract key features of the background. Currently, the most widely used one is CNN (Convolutional Neural Network), which mainly extracts the fully connected layer features of CNN derivative algorithms. However, this series of algorithms has drawbacks. The models of this series of algorithms are all pre-trained based on ImageNet (a large visualization database used for visual object recognition software research). Therefore, there will be problems such as the extracted features not fitting the production data themselves and supervised learning requiring a large amount of annotation.

[0004] Finally, an indispensable technology in image recognition is vector search, which is used to perform similarity search after extracting background features. When the amount of data is large and the vector dimension is high, higher requirements are placed on the vector search framework. However, the prior art lacks a background recognition tool that can solve all of the above problems at the same time. Therefore, the background recognition tools in the prior art have low recognition accuracy and slow running speed. Summary of the invention

[0005] The present invention aims to overcome at least one defect of the above-mentioned prior art and provide a method, system, device and medium for background similarity recognition, which is used to solve the problem of low recognition accuracy when background recognition is performed using instance segmentation technology combined with feature extraction technology in the prior art.

[0006] The technical solution adopted by the present invention includes:

[0007] In a first aspect, the present invention provides a method for background similarity recognition, comprising:

[0008] Inputting the image to be identified into a trained real-time instance segmentation model to mask a human body in the image to be identified to obtain a masked image to be identified, and performing feature extraction on the masked image to be identified to obtain a first feature;

[0009] Inputting the image to be identified into a trained autoencoder model to extract features of the image to be identified to obtain a second feature;

[0010] Obtaining a third feature according to the first feature and the second feature;

[0011] The third feature is compared with the feature of each background image in the background image data set for similarity, and a plurality of background images that are most similar to the third feature are searched for.

[0012] In a second aspect, the present invention provides a system for background similarity recognition, comprising:

[0013] A first feature acquisition module is used to input the image to be identified into the trained real-time instance segmentation model to mask the human body in the image to be identified to obtain the masked image to be identified, and perform feature extraction on the masked image to be identified to obtain a first feature;

[0014] A second feature acquisition module is used to input the image to be identified into a trained autoencoder model to extract features of the image to be identified to obtain a second feature;

[0015] A third feature acquisition module, used to obtain a third feature according to the first feature and the second feature;

[0016] The similar background determination module is used to perform similarity comparison between the third feature and the feature of each background image in the background image data set, and search for a number of background images that are most similar to the third feature.

[0017] In a third aspect, the present invention provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and wherein the processor implements the above-mentioned background recognition method when executing the computer program.

[0018] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the above-mentioned background recognition method when executed by a processor.

[0019] Compared with the prior art, the present invention has the following beneficial effects:

[0020] The present invention provides a method, system, device and medium for background similarity recognition. On the one hand, a human body in an image to be recognized is segmented and masked through a real-time instance segmentation model, thereby obtaining a mask image that basically only contains background information, and obtaining the feature of the mask image as a first feature; on the other hand, the feature of the image to be recognized is obtained as a second feature through an autoencoder model, and a third feature is obtained based on the first feature and the second feature, so that the information between the two features complements each other, and the vector dimension of the third feature is high and the information is richer. When the similarity is compared with the data in the background image data set, a higher accuracy similarity recognition can be achieved. And the method can be applied to the video audit system of a bank, and is used to reduce the risk of intermediary gangs guiding users to purchase goods by performing similarity recognition on the background image of the user. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a flow chart of steps S110-S120, T110, and D110-D120 of the method provided in Example 1 of the present invention.

[0022] Figure 2 It is a schematic diagram of the framework of the real-time instance model in Embodiment 1 of the present invention.

[0023] Figure 3 Schematic diagram of the architecture of the autoencoder model in Example 1 of the present invention.

[0024] Figure 4 This is a flow chart of steps S210-S220, T210, and D210-D220 of the method provided in Example 2 of the present invention.

[0025] Figure 5 Another schematic flow chart of the method provided in Example 2 of the present invention.

[0026] Figure 6 This is a schematic diagram of the module composition of the system provided in Example 3 of the present invention. DETAILED DESCRIPTION

[0027] The drawings of the present invention are only for illustrative purposes and should not be construed as limiting the present invention. In order to better illustrate the following embodiments, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of the actual product; it is understandable to those skilled in the art that some well-known structures and their descriptions in the drawings may be omitted.

[0028] Example 1

[0029] The present embodiment provides a method for background similarity recognition, which is used to enrich the image features after the instance segmentation algorithm mask by using the autoencoder model to extract the features of the image to be recognized, so as to improve the accuracy of background recognition.

[0030] like Figure 1 As shown, the method includes:

[0031] S110, inputting the image to be identified into the trained real-time instance segmentation model to mask the human body in the image to be identified, thereby obtaining the masked image to be identified;

[0032] In this step, the part to be identified in the image to be identified is the background part in the image, which is input into the real-time instance segmentation model. The model can segment the part of the image to be identified except the background, especially the human body part, so as to generate a corresponding mask. The mask is combined with the original image to obtain the masked image to be identified. The masked image basically only contains the background information in the original image.

[0033] Specifically, the real-time instance segmentation model is pre-trained, and the pre-training process is: inputting a first data set into the real-time instance segmentation model to train it. The first data set is a data set of images containing human bodies, that is, the data set is composed of images, and each image contains a human body, which is used to enable the real-time instance segmentation model to learn to recognize the contour edges of the human body. After the training is completed, the human body instance can be accurately segmented in the image. The first data set can be person class data in the MS COCO (Microsoft Common Objects in Context) data set.

[0034] Specifically, Figure 2 As shown in the figure, the real-time instance segmentation model processes the image as follows:

[0035] Step 1: The feature extraction network in the real-time instance segmentation model has 5 convolution modules. After the image is input into the model, each convolution module will output a result. The results output by the 5 convolution modules are C1 to C5 respectively.

[0036] In this step, the feature extraction network is also called a backbone network, specifically a neural network used to extract features from an image. Preferably, in order to improve the real-time performance of the processing, the feature extraction network can use a lightweight neural network architecture, such as a MobilenetV3 network.

[0037] Step 2: The result output by the convolution module is input into the FPN (Feature Pyramid Networks) network in the real-time instance model, where the output result C5 is passed through a convolution layer of the FPN network to obtain the result P5, amplified by a bilinear interpolation, and added to the result C4 to obtain the result P4, and the result P4 is processed in the same way to obtain the result P3. In addition, the result P5 is convolved to obtain the result P6, and the result P6 is convolved to obtain the result P7;

[0038] In this step, the FPN network is a feature map pyramid network, which is mainly used to solve the multi-scale problem in object detection.

[0039] Step 3: Input the result P3 into the Protonet (Prototypical Network) network to generate K 138*138 proto prototypes.

[0040] Step 4: Input the results P3 to P7 into the Prediction Head network to generate the target position, mask coefficient and classification confidence, and then input the above results into the NMS algorithm for processing;

[0041] In this step, the NMS algorithm is a non-maximum suppression algorithm. As the name implies, non-maximum suppression is to suppress elements that are not maximum values. Preferably, in this step, the NMS algorithm is replaced with the Cluster NMS algorithm, i.e., the clustered non-maximum suppression algorithm. After the replacement, the original real-time instance segmentation model can effectively solve the problem of inaccurate classification and inaccurate edge segmentation during target detection. According to experimental data, the speed of the model can reach 30FPS, the accuracy of human edge prediction has been improved, and the accuracy of human classification has been improved by 3%.

[0042] Step 5: The Cluster NMS algorithm combines the result obtained in step 4 with the k 138*138 proto prototypes generated in step 3 to obtain the result, which is input into the Crop module.

[0043] Step 6: The Crop module clears the mask outside the boundary of the image and inputs the result into the Threshold module. The Threshold module binarizes the generated mask to obtain the masked image to be identified.

[0044] S120, extracting features from the masked image to be identified to obtain a first feature;

[0045] In this step, feature extraction can be implemented using any applicable feature extraction network. Preferably, a neural network model with a VGG19 architecture is used to extract features from the masked image to be identified, thereby obtaining the first feature.

[0046] The method further includes the following step T110. Step T110 and steps S110 to S120 are two independent groups of steps. The two groups of steps may be executed in parallel or in sequence, which is not limited here.

[0047] T110, inputting the image to be identified into the trained autoencoder model to extract features of the image to be identified, and obtaining a second feature;

[0048] In this step, specifically, the autoencoder model is pre-trained, and the pre-training process is: inputting the second data set into the autoencoder model to train it. The second data set is a data set of images to which noise has been added and which simultaneously contains human bodies and production scenes, that is, the data set is composed of images, and each image contains human bodies and production scenes, and is used to enable the autoencoder model to learn to recognize human bodies and backgrounds in images simultaneously.

[0049] like Figure 3 As shown, the specific training process of the autoencoder model is:

[0050] Step 1: Add noise to the dataset containing both human body and production scene images to form the second dataset;

[0051] In this step, the noise added is Gaussian noise, which refers to a type of noise whose probability density function obeys Gaussian distribution.

[0052] Step 2: Input the second data set into the autoencoder model. The autoencoder model compresses the noisy data into the Encoder, extracts the key features of the compressed data, inputs the key features into the Decoder for decompression, and obtains the difference between the data before noisy and the decompressed data by comparison.

[0053] In this step, the purpose of the autoencoder model to compress the data first is to ensure that important features are not lost, reduce the dimension of the input information, and reduce the processing burden of the model.

[0054] After comparing the data before noise addition and the data after decompression, the autoencoder model determines the overall prediction error based on the difference, performs reverse propagation, and gradually improves the accuracy of the autoencoder.

[0055] The autoencoder model is trained based on the Gaussian autoencoder algorithm in the unsupervised model, and the trained autoencoder model performs excellently in feature extraction of the second data set.

[0056] The method further includes the following steps D110 to D120, wherein step D110 needs to be executed after step T110 and steps S110 to S120 are completed.

[0057] D110, obtaining a third feature according to the first feature and the second feature;

[0058] In this step, the third feature combines the information of the first feature and the second feature. Compared with using only the first feature or only the second feature for similarity identification, the third feature has richer information and a higher vector dimension, which can effectively improve the accuracy of similarity retrieval.

[0059] Specifically, in this step, the third feature is obtained by concatenating the first feature and the second feature, and the vector dimension of the third feature is the superposition of the vector dimensions of the first feature and the second feature. In a specific implementation, the vector dimension of the first feature obtained by steps S110 to S120 is 2048, and the vector dimension of the second feature obtained by step T110 is 1325, then the vector dimension of the third feature obtained is 1325+2048=3373.

[0060] D120, performing a similarity comparison between the third feature and the feature of each background image in the background image data set, and searching for a number of background images that are most similar to the third feature;

[0061] In this step, the background image data set is composed of background images, and the result of background similarity recognition selects the background image most similar to the third feature in the background image data set as the search result.

[0062] In a preferred embodiment, in order to complete the search process more quickly and meet the requirements of real-time recognition, step D120 is implemented through the Milvus database. Milvus is a vector database that can be constructed based on a data set input by a user. It is mainly used for vector similarity search and accelerates unstructured data retrieval. It has the characteristics of being able to achieve millisecond-level search of trillion-level vectors and simple management of unstructured data. Based on this, before executing step D120, it is necessary to first convert each background image in the background image data set into a feature vector, use the feature vectors of all background images to construct a vector data set, and build a Milvus library based on the vector data set. After the Milvus library is constructed, the specific execution process of step D120 is:

[0063] The third feature is input into the constructed Milvus library, so that the Milvus library compares the third feature with the vector of each background image in the vector data set for similarity, and obtains background images corresponding to several vectors that are most similar to the third feature.

[0064] In this step, the Milvus library can be used to complete the search in a very short time with a very large sample size, and the accuracy is close to 90%. Simulation data shows that when the feature dimension is 3373 and the sample size in the background data set is 1 million, the vector query time is 0.008 seconds, and the average batch query time is 0.3-2 seconds, meeting the real-time requirements. The accuracy of the first result among the most similar results is 95%, and the accuracy of the result ranked first among the most similar results is 89.5%, which is very high.

[0065] This embodiment provides a method for background similarity recognition. On the one hand, a mask image after covering the human body is obtained by inputting the image to be recognized into a trained real-time instance segmentation model, and the mask image basically retains only the background information in the image, and the first feature is extracted. On the other hand, the image to be recognized is input into a trained autoencoder model to extract features from it to obtain a second feature. The first feature and the second feature are spliced ​​to obtain a third feature. Compared with the use of a Gaussian autoencoder model alone for feature extraction or a real-time instance segmentation model alone for instance segmentation and feature extraction and then similarity search, the search using the spliced ​​third feature can achieve a higher search accuracy, thereby determining several backgrounds that are most similar to the image to be recognized, which is convenient for subsequent in-depth image analysis based on the search results and the information of the image to be recognized itself. In order to improve the accuracy and efficiency of the overall method, the Cluster NMS algorithm is selected in the real-time instance segmentation model instead of the NMS algorithm, the first feature is obtained by using the network of the VGG19 architecture, the second feature is obtained by using the autoencoder model, and the Milvus library is used for vector similarity search, which can ensure the efficiency and accuracy of vector similarity search even when the sample size is very large.

[0066] Example 2

[0067] Based on the same concept as Example 1, this embodiment provides a background similarity identification method applied to an audit system. Specifically, the audit system refers to a video audit system for credit cards. A large number of video audits are conducted every day in the credit card audit process. In order to effectively prevent intermediaries from guiding customers to submit applications, it is particularly important to identify whether the customer video backgrounds are similar. Video background similarity means that the video backgrounds of multiple customers are similar in the activation process. When the similarity of the video backgrounds of several customers is high, it means that there is a high possibility that an intermediary gang has submitted applications. At present, this link is still mainly based on human eye recognition, so too much labor cost is consumed in the business process. The method provided in this embodiment is applied to the audit system, which can promptly identify the risks brought by intermediary submissions, and no manual review is required.

[0068] like Figure 4 , 5 As shown, the method includes:

[0069] S210, inputting the image to be identified into the trained real-time instance segmentation model to mask the human body in the image to be identified, and obtaining the masked image to be identified;

[0070] In this step, the image to be identified is one frame of the video image of the user acquired by the camera.

[0071] The real-time instance segmentation model is pre-trained, and the pre-training process is: input the first data set into the real-time instance segmentation model to train it. The first data set is the person class data in the MS COCO (Microsoft Common Objects in Context) data set. And the NMS algorithm used in the real-time instance segmentation model is replaced with the Cluster NMS algorithm, that is, the clustering non-maximum suppression algorithm. After the replacement, it can effectively solve the problems of inaccurate classification and inaccurate edge segmentation of the original real-time instance segmentation model during target detection. The speed of the model can reach 30FPS, the accuracy of human edge prediction is improved, and the accuracy of human classification is improved by 3%.

[0072] S220, input the masked image to be recognized into the neural network model of VGG19 architecture for feature extraction to obtain Feature1.

[0073] The method further includes the following step T210. Step T210 and steps S210 to S220 are two independent groups of steps, and the two groups of steps are executed in parallel.

[0074] T210, input the image to be identified into the trained autoencoder model to extract features of the image to be identified, and obtain Feature2;

[0075] In this step, specifically, the autoencoder model is pre-trained, and the pre-training process is: inputting the second data set into the autoencoder model to train it. The second data set is a data set to which Gaussian noise has been added and contains images of human bodies and production scenes at the same time, that is, the data set consists of images, and each image contains human bodies and production scenes.

[0076] The method further includes the following steps D210 to D220, wherein step D210 needs to be executed after step T210 and steps S210 to S220 are completed.

[0077] D210, concatenate Feature1 and Feature2 to obtain Feature3;

[0078] In this step, Feature3 combines the information of Feature1 and Feature2. Compared with using only Feature1 or only Feature2 for similarity identification, the information in Feature3 is richer and can effectively improve the accuracy of similarity retrieval.

[0079] Specifically, in this step, Feature3 is obtained by concatenating Feature1 and Feature2. The vector dimension of Feature1 obtained through steps S210 to S220 is 2048, and the vector dimension of Feature2 obtained through step T210 is 1325. Then the vector dimension of Feature3 obtained is 1325+2048=3373.

[0080] D220. Input Feature3 into the constructed Milvus library so that the Milvus library compares the similarity between Feature3 and the vector of each background image in the vector data set, and searches for background images corresponding to several vectors that are most similar to Feature3.

[0081] In this step, the background image dataset is composed of background images that have been reviewed by all review systems, so that the Milvus library can search for background images similar to the image to be identified in the background image dataset. If the similarity between the similar background image and the image to be identified is very high, then the user application corresponding to the image to be identified is very likely to be guided by an intermediary, which requires attention and can be used as basic information for early warning or alarm.

[0082] Before executing step D220, each background image in the background image data set needs to be converted into a feature vector, a vector data set is constructed using the feature vectors of all background images, and a Milvus library is constructed based on the vector data set.

[0083] This embodiment provides a background similarity recognition method applied to the audit system, which can effectively reduce the risk of intermediary gangs guiding customers to submit applications, provide banks with basic information that can be used as early warnings or alarms, reduce the cost of manual review, and improve the accuracy and efficiency of recognition.

[0084] The method provided in this embodiment is based on the same concept as the method provided in Embodiment 1. Except for the difference in application scenarios and the differences brought about by adapting to the scenarios, the steps or nouns in this embodiment that are the same as those in Embodiment 1, their definitions, functions, preferred or specific implementation methods, and the beneficial effects brought about are the same as those described in Embodiment 1 and will not be repeated here.

[0085] Example 3

[0086] Based on the same concept as in Example 1, this embodiment provides a system for background similarity recognition, such as Figure 6 As shown, the system includes:

[0087] The first feature acquisition module 310 is used to input the image to be identified into a trained real-time instance segmentation model to mask the human body in the image to be identified, obtain the masked image to be identified, and perform feature extraction on the masked image to be identified to obtain the first feature.

[0088] The trained real-time instance segmentation model is obtained by inputting a first data set into the real-time instance segmentation model to train it. The first data set is a data set containing images of human bodies. The masked image to be identified can be obtained by feature extraction using a neural network with a VGG19 architecture.

[0089] Preferably, the Cluster NMS algorithm is used to replace the NMS algorithm in the real-time instance segmentation model.

[0090] The second feature acquisition module 320 is used to input the image to be identified into the trained autoencoder model to extract features of the image to be identified and obtain a second feature.

[0091] The trained autoencoder model is obtained by inputting a second data set into the autoencoder model to train the model. The second data set is a data set to which Gaussian noise has been added and which contains images of human bodies and production scenes.

[0092] The third feature acquisition module 330 is used to obtain the third feature according to the first feature and the second feature.

[0093] The similar background determination module 340 is used to perform similarity comparison between the third feature and the feature of each background image in the background image data set, and search for a number of background images that are most similar to the third feature.

[0094] Specifically, the similar background determination module 340 is implemented by a pre-built Milvus library, which converts each background image in the background image data set into a feature vector, constructs a vector data set using the feature vector of the background image, constructs a Milvus library based on the vector data set, and inputs the third feature into the Milvus library so that the Milvus library compares the third feature with the vector of each background image in the vector data set for similarity, and searches for background images corresponding to several vectors that are most similar to the third feature.

[0095] The system provided in this embodiment and the method provided in Example 1 are based on the same concept and are a system that applies the method provided in Example 1. Therefore, the steps or nouns in this embodiment that are the same as those in Example 1, their definitions, functions, preferred or specific implementation methods, and the beneficial effects brought about are the same as those described in Example 1 and will not be repeated here.

[0096] Example 4

[0097] Based on the same concept as embodiments 1 and 2, this embodiment provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the background recognition method provided in embodiment 1 or embodiment 2 when executing the computer program.

[0098] This embodiment further provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the background recognition method provided in Embodiment 1 or Embodiment 2 is implemented.

[0099] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the technical solution of the present invention, and are not intended to limit the specific implementation methods of the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the claims of the present invention shall be included in the protection scope of the claims of the present invention.

Claims

1. A method for background similarity recognition, characterized in that: include: Inputting the image to be identified into a trained real-time instance segmentation model to mask a human body in the image to be identified to obtain a masked image to be identified, and performing feature extraction on the masked image to be identified to obtain a first feature; Inputting the image to be identified into a trained autoencoder model to extract features of the image to be identified, thereby obtaining a second feature; Obtaining a third feature according to the first feature and the second feature; The third feature is compared with the feature of each background image in the background image data set for similarity, and several background images that are most similar to the third feature are searched for. Specifically, each background image in the background image data set is converted into a feature vector, a vector data set is constructed using the feature vector of the background image, a Milvus library is constructed based on the vector data set, and the third feature is input into the Milvus library so that the Milvus library compares the third feature with the vector of each background image in the vector data set for similarity, and background images corresponding to several vectors that are most similar to the third feature are obtained.

2. The method for background similarity recognition according to claim 1, characterized in that: The trained real-time instance segmentation model is obtained by inputting the first data set into the real-time instance segmentation model to train it; The trained autoencoder model is obtained by inputting the second data set into the autoencoder model to train it; The first data set is a data set containing images of human bodies, and the second data set is a data set containing images of human bodies and production scenes to which noise has been added.

3. The method for background similarity recognition according to claim 2, characterized in that: The trained autoencoder model is obtained by inputting the second data set into the autoencoder model to train it, specifically: The trained autoencoder model is obtained by adding noise to the second data set to obtain a noisy second data set, and inputting the noisy second data set into the autoencoder model to train the autoencoder model; The autoencoder model is used to compress the noisy data, extract key features of the compressed data, decompress the key features, and obtain the difference between the data before noisy and the decompressed data by comparison.

4. The method for background similarity recognition according to claim 2 or 3, characterized in that: The noise is Gaussian noise.

5. The method for background similarity recognition according to claim 1, characterized in that: The NMS algorithm used in the real-time instance segmentation model is the Cluster NMS algorithm.

6. The method for background similarity recognition according to claim 1, characterized in that: Performing feature extraction on the masked image to be identified to obtain a first feature, specifically: inputting the masked image to be identified into a neural network model of a VGG19 architecture to perform feature extraction to obtain the first feature.

7. A system for background similarity recognition, characterized in that: include: A first feature acquisition module is used to input the image to be identified into the trained real-time instance segmentation model to mask the human body in the image to be identified to obtain the masked image to be identified, and perform feature extraction on the masked image to be identified to obtain a first feature; A second feature acquisition module is used to input the image to be identified into a trained autoencoder model to extract features of the image to be identified to obtain a second feature; A third feature acquisition module, used to obtain a third feature according to the first feature and the second feature; A similar background determination module is used to perform similarity comparison between the third feature and the feature of each background image in the background image data set, and search for several background images that are most similar to the third feature. Specifically, each background image in the background image data set is converted into a feature vector, a vector data set is constructed using the feature vector of the background image, a Milvus library is constructed based on the vector data set, and the third feature is input into the Milvus library so that the Milvus library compares the third feature with the vector of each background image in the vector data set to obtain background images corresponding to several vectors that are most similar to the third feature.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the background recognition method according to any one of claims 1 to 6 is implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the background recognition method according to any one of claims 1 to 6 is implemented.

Images