Card and certificate splicing identification method, electronic equipment and storage medium

Through pre-trained card certificate identification model and multi-scale feature processing technology, the problem of inaccurate identification of card certificate splicing identification in complex scenarios is solved, and the identification accuracy and robustness are improved. It is suitable for card certificate verification in financial services. .

CN119091215APending Publication Date: 2024-12-06PING AN TECH (SHENZHEN) CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411203393.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-29
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

The existing card certificate splicing identification technology is inaccurate in the case of complex scenes, lighting changes and poor image quality, and has insufficient ability to fight attacks, high resource consumption, and limited applicability.

Method used

The pre-trained target card authentication identification model is adopted to downsample the card authentication image through the target backbone network, generate a multi-scale mask feature map, and use a classification network to authenticate it, and combine it with the contour extraction network to extract the splicing contour information to improve the identification accuracy sex.

Benefits of technology

It improves the accuracy and robustness of card certificate splicing identification, can more accurately verify the effectiveness of financial card certificates, reduces the demand for resources, and enhances the defense ability to fight attacks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119091215A_ABST
    Figure CN119091215A_ABST
Patent Text Reader

Abstract

The invention relates to the field of finance and the technical field of artificial intelligence, in particular to a card and certificate splicing identification method, electronic equipment and a storage medium. According to the card splicing identification method, a target card image needs to be acquired firstly, and a pre-trained target card identification model is further acquired; wherein the target card identification model comprises a target backbone network and a target classification network; performing down-sampling processing on the target card image through the target backbone network to obtain at least two mask feature maps; and performing card splicing identification on each mask feature map through the target classification network to obtain a card identification result, the card identification result being used for representing that the target card image is a spliced image or a non-spliced image. Therefore, the card identification result can reflect the splicing condition of the target card image more accurately, and the validity of the financial card can be verified quickly and accurately in the financial service.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the fields of finance and artificial intelligence technology, and in particular to a card splicing and identification method, electronic equipment, and storage medium. Background Art

[0002] Card splicing prevention is a security technology designed to prevent abnormal editing or splicing of images or digital copies of cards to ensure the authenticity and integrity of cards. In the industry, there are some technologies used to protect images or videos from abnormal splicing or tampering. These technologies can be implemented using methods such as image processing, digital watermarking, and encryption algorithms. In financial business, whether online or offline, it is often necessary to verify the identity of customers. Traditional manual review methods are not only inefficient but also prone to errors. By applying automated card splicing identification technology, financial institutions can quickly and accurately verify whether the identity documents submitted by customers are spliced.

[0003] The card and certificate anti-splicing technology in related technologies may cause inaccurate recognition when dealing with complex scenes, lighting changes, and poor image quality. Therefore, how to improve the accuracy of card and certificate splicing identification has become an issue that needs to be further resolved in the industry. Summary of the invention

[0004] The present application aims to solve at least one of the technical problems existing in the prior art. To this end, the present application proposes a card splicing identification method, electronic device, and storage medium, which can improve the accuracy of card splicing identification. Therefore, in financial business, it is also more helpful to quickly and accurately verify the validity of financial cards.

[0005] According to the first aspect of the present application, a card splicing identification method includes:

[0006] Get the target card image;

[0007] Obtain a pre-trained target card identification model; wherein the target card identification model includes a target backbone network and a target classification network;

[0008] Downsampling the target card image through the target backbone network to obtain at least two mask feature maps; wherein each of the mask feature maps has a corresponding feature resolution, and the feature resolutions of any two mask feature maps are different;

[0009] The target classification network performs card splicing identification on each of the mask feature maps to obtain a card identification result, and the card identification result is used to indicate whether the target card image is a spliced ​​image or a non-spliced ​​image.

[0010] According to some embodiments of the present application, obtaining a pre-trained target card identification model includes:

[0011] Acquire a training sample set; wherein the training sample set includes a plurality of card sample images and a sample image annotation corresponding to each of the card sample images;

[0012] Input the card sample image into a preset original card identification model for current round identification processing to obtain the current round identification result;

[0013] Compare the sample image annotation corresponding to the card sample image with the current round identification result to obtain comparison deviation data;

[0014] The model parameters of the original card identification model are updated based on the comparison deviation data, and the process returns to inputting the card sample image into the original card identification model for this round of identification processing until the original card identification model meets the preset training termination condition, and the original card identification model is determined as the pre-trained target card identification model.

[0015] According to some embodiments of the present application, the original card identification model includes an original backbone network, an original classification network, and a contour extraction network, and the current round identification result includes the current round splicing contour information and the current round card category information;

[0016] The step of inputting the card sample image into a preset original card identification model for current round identification processing to obtain the current round identification result includes:

[0017] Downsampling each of the card sample images through the original backbone network to obtain at least two sample feature maps; wherein each of the sample feature maps corresponds to one of the feature resolutions;

[0018] Extracting the current round of stitching contour information from at least two of the sample feature maps through the contour extraction network;

[0019] The card category information of this round is extracted from at least two of the sample feature maps through the original classification network.

[0020] According to some embodiments of the present application, extracting the current round of stitching contour information from at least two of the sample feature maps by using the contour extraction network includes:

[0021] Determine a resolution ranking based on a high-low relationship between the feature resolutions of the sample feature graphs and the feature resolutions;

[0022] Based on the resolution sorting, the sample feature maps corresponding to the feature resolutions are fused to obtain the current round of stitching contour information.

[0023] According to some embodiments of the present application, the resolution-based sorting, fusing the sample feature maps corresponding to the feature resolutions to obtain the current round of stitching contour information, includes:

[0024] For the sample feature map with the lowest feature resolution in the resolution sorting, input the sample feature map with the lowest feature resolution into the contour extraction network to extract contour features, so as to obtain contour sub-feature information corresponding to the sample feature map;

[0025] For the sample feature map of the non-lowest feature resolution in the resolution sorting, obtain the sample feature map of the feature resolution lower than the lowest feature resolution Figure 1 The contour sub-feature information of the first level is obtained, and the contour sub-feature information is merged with the sample feature map of non-lowest feature resolution to obtain the contour sub-feature information corresponding to the sample feature map;

[0026] According to the sample feature map with the highest feature resolution in the resolution sorting, the contour sub-feature information corresponding to the sample feature map with the highest feature resolution is determined as the current round of stitching contour information.

[0027] According to some embodiments of the present application, the sample image annotation includes a splicing contour annotation and a card category annotation of the card sample image, and the comparison deviation data includes contour deviation data and category deviation data;

[0028] The step of comparing the sample image annotation corresponding to the card sample image with the current round identification result to obtain comparison deviation data includes:

[0029] Performing a contour deviation comparison based on the splicing contour annotation and the current round of splicing contour information to obtain the contour deviation data;

[0030] A category deviation comparison is performed based on the card category label and the card category information of this round to obtain the category deviation data.

[0031] According to some embodiments of the present application, the comparing contour deviation based on the splicing contour annotation and the current round of splicing contour information to obtain the contour deviation data includes:

[0032] Performing scale matching processing on the splicing contour annotation to obtain an annotation feature map; wherein the feature resolution of the annotation feature map matches the current round of splicing contour information;

[0033] A contour deviation comparison is performed based on the marked feature map and the current round of splicing contour information to obtain the contour deviation data.

[0034] According to some embodiments of the present application, until the original card identification model satisfies a preset training termination condition, determining the original card identification model as the pre-trained target card identification model includes:

[0035] In response to the original card authentication model satisfying the training termination condition, removing the contour extraction network from the original card authentication model;

[0036] The original card authentication model with the contour extraction network removed is determined as the pre-trained target card authentication model.

[0037] In a second aspect, an embodiment of the present application provides an electronic device, including: a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the card splicing identification method as described in any one of the embodiments of the first aspect of the present application.

[0038] In a third aspect, an embodiment of the present application provides a computer-readable storage medium, wherein the storage medium stores a program, and the program is executed by a processor to implement the card splicing identification method as described in any one of the embodiments of the first aspect of the present application.

[0039] The card splicing identification method, electronic device, and storage medium according to the embodiments of the present application have at least the following beneficial effects:

[0040] According to the card splicing identification method of the embodiment of the present application, it is necessary to first obtain the target card image, and then downsample the target card image through the target backbone network to obtain at least two mask feature maps; wherein each of the mask feature maps has a corresponding feature resolution, and the feature resolutions of any two mask feature maps are different. Further, the card splicing identification is performed on the mask feature map corresponding to each feature resolution through the classification network of the target card identification model to obtain the card identification result. The card identification result is used to indicate whether the target card image is a spliced ​​image or a non-spliced ​​image. In this way, the pre-trained target card identification model is used to perform card splicing identification on the mask feature maps of different feature resolutions, and then obtain the card identification result. wherein, since the card identification result integrates the local details and global overview of the mask feature map, the card identification result can more accurately reflect the splicing of the target card image, and in financial services, it is also more helpful to quickly and accurately verify the validity of financial cards.

[0041] Additional aspects and advantages of the present application will be given in part in the description below, and in part will become apparent from the description below, or will be learned through the practice of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:

[0043] Figure 1 A schematic diagram of the process of the card splicing identification method provided in the embodiment of the present application;

[0044] Figure 2 Another schematic diagram of the process of the card splicing identification method provided in the embodiment of the present application;

[0045] Figure 3 yes Figure 2 Schematic diagram of the process of step S202;

[0046] Figure 4 An example diagram of a backbone network provided in an embodiment of the present application;

[0047] Figure 5 yes Figure 3 Schematic diagram of the process of step S302;

[0048] Figure 6 yes Figure 5 Schematic diagram of the process of step S502;

[0049] Figure 7 An example diagram of extracting the card category information and the splicing profile information of this round in the embodiment of the present application;

[0050] Figure 8 yes Figure 2 Schematic diagram of the process of step S203;

[0051] Fig. 9 yes Figure 8 Schematic diagram of the process of step S801;

[0052] Fig.10 An example diagram of an original card authentication model provided in an embodiment of the present application;

[0053] Fig.11 yes Figure 2 Schematic diagram of the process of step S204;

[0054] Fig.12 It is a schematic diagram of the hardware structure of the electronic device provided in the embodiment of the present application. DETAILED DESCRIPTION

[0055] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present application, and cannot be understood as limiting the present application.

[0056] In the description of this application, "several" means one or more, "more" means more than two, "greater than", "less than", "exceed", etc. are understood to exclude the number itself, and "above", "below", "within", etc. are understood to include the number itself. If there is a description of "first" or "second", it is only used for the purpose of distinguishing technical features, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features or implicitly indicating the order of the indicated technical features.

[0057] In the description of the present application, it should be understood that descriptions involving orientation, such as up, down, left, right, front, back, etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present application.

[0058] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "illustrative embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials, or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0059] In the description of this application, it should be noted that, unless otherwise clearly defined, the terms such as setting, installing, connecting, etc. should be understood in a broad sense, and technicians in the relevant technical field can reasonably determine the specific meanings of the above terms in this application in combination with the specific content of the technical solution. In addition, the identification of specific steps below does not represent a limitation on the order of steps and execution logic. The execution order and execution logic between each step should be understood and inferred with reference to the contents described in the embodiment.

[0060] Card anti-splicing is a security technology designed to prevent abnormal editing or splicing of the image or digital copy of the card to ensure the authenticity and integrity of the card. In the industry, there are some technologies used to protect images or videos from being abnormally spliced ​​or tampered with. These technologies can be implemented using methods such as image processing, digital watermarking, and encryption algorithms.

[0061] Card splicing identification technology plays a vital role in the financial field, especially in identity verification and fraud detection. Financial transactions often involve a large amount of personal sensitive information. The security of this information is directly related to the user's property safety and the reputation of financial institutions. Card splicing identification technology can effectively identify card images that have been forged or tampered with by splicing, thereby improving the security of financial transactions.

[0062] In financial business, whether online or offline, it is often necessary to verify the identity of customers. The traditional manual review method is not only inefficient, but also prone to errors. By applying automated card splicing identification technology, financial institutions can quickly and accurately verify whether the identity documents submitted by customers are spliced. In addition, this technology can also be applied to the anti-counterfeiting of financial cards and certificates, ensuring the authenticity of the cards and certificates and preventing illegal copying and use. The role of card splicing identification technology in the financial field is indispensable. It not only improves the security and efficiency of financial services, but also helps to build a more secure and reliable financial ecological environment.

[0063] The card anti-splicing in the related technology may have inaccurate recognition problems when dealing with complex scenes, lighting changes, and poor image quality, which affects the robustness of the algorithm. Secondly, in the face of adversarial attacks from malicious users, the related technology may lack sufficient defense capabilities, and attackers may find ways to bypass the algorithm's protection measures. In addition, some anti-splicing algorithms have high requirements in terms of performance and computing resources, which not only requires strong processing power, but also may require a large amount of storage space, which limits the application of the algorithm in resource-constrained environments. Finally, the applicability of existing algorithms is limited, and the effects of different algorithms in different application scenarios may vary significantly.

[0064] Card anti-splicing is a security technology designed to prevent abnormal editing or splicing of the image or digital copy of the card to ensure the authenticity and integrity of the card. In the industry, there are some technologies used to protect images or videos from being abnormally spliced ​​or tampered with. These technologies can be implemented using methods such as image processing, digital watermarking, and encryption algorithms.

[0065] The card and certificate anti-splicing technology in related technologies may have inaccurate recognition problems when dealing with complex scenes, lighting changes, and poor image quality, which affects the robustness of the algorithm. Therefore, how to improve the accuracy of card and certificate splicing identification has become an issue that needs to be further resolved in the industry.

[0066] The present application aims to solve at least one of the technical problems existing in the prior art. To this end, the present application proposes a card splicing identification method, electronic device, and storage medium, which can improve the accuracy of card splicing identification. Therefore, in financial business, it is also more helpful to quickly and accurately verify the validity of financial cards.

[0067] Reference Figure 1 According to the card splicing identification method of the embodiment of the present application, it may include, but is not limited to, the following:

[0068] Step S101, obtaining a target card image;

[0069] Step S102, obtaining a pre-trained target card identification model; wherein the target card identification model includes a target backbone network and a target classification network;

[0070] Step S103, downsampling the target card image through the target backbone network to obtain at least two mask feature maps; wherein each mask feature map has a corresponding feature resolution, and the feature resolutions of any two mask feature maps are different;

[0071] Step S104, performing card splicing identification on each mask feature map through a target classification network to obtain a card identification result, which is used to indicate whether the target card image is a spliced ​​image or a non-spliced ​​image.

[0072] The card splicing identification method shown in step S101 to step S104 of the embodiment of the present application needs to first obtain the target card image, and then downsample the target card image through the target backbone network to obtain at least two mask feature maps; wherein each mask feature map has a corresponding feature resolution, and the feature resolutions of any two mask feature maps are different. Further, the card splicing identification is performed on the mask feature map corresponding to each feature resolution through the classification network of the target card identification model to obtain the card identification result. The card identification result is used to indicate whether the target card image is a spliced ​​image or a non-spliced ​​image. In this way, the pre-trained target card identification model is used to perform card splicing identification on the mask feature maps of different feature resolutions, and then obtain the card identification result. Among them, since the card identification result integrates the local details and global overview of the mask feature map, the card identification result can more accurately reflect the splicing of the target card image, and in financial business, it is also more helpful to quickly and accurately verify the validity of financial cards.

[0073] In some embodiments, step S101 is to obtain a target card image. It should be noted that the target card image refers to the image of the card to be identified. In the embodiments of the present application, the target card image is obtained to identify it and prevent abnormal editing or splicing of the card image or digital copy to ensure the authenticity and integrity of the card.

[0074] Reference Figure 2According to some embodiments of the present application, step S102 obtains a pre-trained target card identification model, which may include, but is not limited to, the following steps S201 to S204.

[0075] Step S201, obtaining a training sample set; wherein the training sample set includes a plurality of card sample images and a sample image annotation corresponding to each card sample image;

[0076] Step S202, inputting the card sample image into the preset original card identification model for current round identification processing to obtain the current round identification result;

[0077] Step S203, comparing the sample image annotation corresponding to the card sample image with the current round identification result to obtain comparison deviation data;

[0078] Step S204, based on the comparison deviation data, the model parameters of the original card identification model are updated, and the card sample image is input into the original card identification model for this round of identification processing until the original card identification model meets the preset training termination condition, and the original card identification model is determined as the pre-trained target card identification model.

[0079] In step S201 of some embodiments, a training sample set is obtained; wherein the training sample set includes a plurality of card sample images and sample image annotations corresponding to each card sample image. It should be noted that, as the starting point of the original card identification model training process, this step involves the acquisition and preprocessing of the training sample set, as well as the preparation of the original card identification model. In this step, it is first necessary to collect a plurality of card sample images, which refer to various types of card samples that the original card identification model needs to identify during the pre-training process. The source of the card sample images can be actual photographs or digital images extracted from a database. In addition to the card sample images themselves, corresponding annotation information, i.e., sample image annotations, needs to be prepared for each card sample image. Sample image annotations are descriptions of card features in card sample images, which may include but are not limited to key information such as whether the card sample image is a spliced ​​image and spliced ​​contours. These annotations will serve as the true value or standard answer of the original card identification model during the training process.

[0080] In some embodiments, in order to ensure the effectiveness of the training process and the generalization ability of the model, the training sample set needs to be diverse and representative, covering a variety of card types, backgrounds, lighting conditions, and possible occlusions. In addition, the sample images and their annotation information need to be preprocessed to meet the input requirements of the model, which may include operations such as cropping, scaling, rotation correction, grayscale, and formatting of the annotation information.

[0081] In some embodiments, the original card authentication model can be an off-the-shelf deep learning model or a custom model designed specifically for the card authentication task. After selecting the model architecture, the parameters of the model need to be initialized, which may be randomly initialized or fine-tuned based on a pre-trained model.

[0082] It should be understood that step S201 lays the foundation for the training and optimization of the original card identification model, ensuring the quality of the model training input and the initial state of the model. Through this step, the original card identification model can be trained to gradually improve its ability to identify card samples.

[0083] Step S202, input the card sample image into the original card identification model for this round of identification processing to obtain the identification result of this round. It should be noted that inputting the card sample image into the original card identification model for this round of identification processing involves inputting the prepared card sample image into the original card identification model for actual identification processing. In this step, the model will use its current parameters and learning algorithm to analyze and classify the input card sample image, with the purpose of identifying the card features in the image and predicting the corresponding sample image annotation.

[0084] In some more specific embodiments, the card sample images can be sent to the original card identification model in batches or one by one. The original card identification model receives these image data and extracts the features of the card sample images through a series of mathematical transformations. These transformations may include convolution operations to capture local features, activation functions to introduce nonlinearity, pooling operations to reduce the spatial dimensions of features, etc. As the data processing process deepens, the original card identification model gradually abstracts useful feature representations from the original images, and then the original card identification model gives one or more categories of prediction results for each card sample image based on the learned knowledge and feature representations, which is the current round of identification results. These results constitute the output of this round of identification, that is, the model's understanding and judgment of the current input sample. After step S202 is completed, a data set of identification results will be obtained, which contains the current round of identification results corresponding to each card sample image.

[0085] Step S203, compare the sample image annotation corresponding to the card sample image with the current round of identification results to obtain comparison deviation data. It should be noted that the current round of identification results are used to compare with the sample image annotation to evaluate the performance of the model and calculate the comparison deviation data. The comparison deviation data will reflect the accuracy of the model prediction and provide a basis for model optimization and parameter update.

[0086] In some embodiments, the sample image annotation corresponding to the card sample image is compared with the current round of identification results, which may specifically involve the difference between the current round of identification results and the true annotation, which may be a classification error, a mismatch in the splicing contour, etc. In order to quantify these deviations, the embodiments of the present application may use specific error metrics or loss functions, such as cross entropy loss, mean square error, IoU (intersection over union), etc. These metrics can reflect the gap between the predicted results and the actual situation.

[0087] According to the error measurement method used, step S203 will generate a series of comparison deviation data, which provide feedback for further optimization of the model. If the model performs poorly on certain types of cards, the deviation data will reveal this, suggesting that more training is needed for these samples or adjustments to the model structure. In addition, step S203 can also involve statistical analysis of model performance, such as the calculation of indicators such as accuracy, recall, and F1 score, which can comprehensively evaluate the performance of the model in different aspects. By analyzing these performance indicators, the advantages and disadvantages of the model can be identified, providing direction for model improvement. The comparison deviation data not only indicates the current performance level of the original card identification model, but also provides a basis for subsequent parameter updates and model optimization. In this way, the original card identification model can be gradually iterated during the training process and continuously develop towards higher accuracy.

[0088] Step S204, based on the comparison deviation data, the model parameters of the original card identification model are updated, and the card sample image is input into the original card identification model for this round of identification processing, until the original card identification model meets the preset training termination condition, and the original card identification model is determined as the pre-trained target card identification model. It should be noted that step S204 is the optimization and parameter update stage in the original card identification model training process. In this step, based on the comparison deviation data obtained in step S203, the original card identification model will adjust itself to reduce the prediction error and improve the identification accuracy. Specifically, the comparison deviation data will drive the update of the model parameters, and this process involves gradient descent or its variant algorithm.

[0089] First, the comparison deviation data will be used to calculate the gradient of the loss function, which indicates how to adjust the model parameters to reduce the prediction error. In deep learning, this process is implemented in the back-propagation algorithm, which starts from the output layer and traverses each layer of the network in reverse, calculating the gradient of each parameter.

[0090] Next, the model parameters will be updated according to the calculated gradient and the predetermined learning rate. The learning rate is a hyperparameter that controls the magnitude of parameter updates. If the learning rate is set too high, it may cause the model training to be unstable or miss the optimal solution; if the learning rate is too low, the training process will be slow or even fall into a local optimum.

[0091] After the parameters are updated, the model will perform another round of identification processing, that is, return to step S202, and use the updated model to perform a new round of prediction on the training sample set. This process will be repeated, and each round of iteration includes model prediction, error evaluation and parameter update until the model meets the preset training termination conditions. Among them, the preset training termination conditions may include that the performance of the model on the validation set is no longer significantly improved, the predetermined number of training rounds is reached, or the training time is too long. When the original card identification model meets these conditions, it means that the original card identification model has been fully trained and further iterative updates are stopped. In the end, the pre-trained target card identification model is obtained, which can show high accuracy when processing new card sample images. In this way, the original card identification model can gradually adapt to the distribution characteristics of the training samples, improve the ability to identify cards, and finally reach practical standards.

[0092] Reference Figure 3 According to some embodiments of the present application, the original card identification model includes an original backbone network, an original classification network, and a contour extraction network, and the current round of identification results includes the current round of splicing contour information and the current round of card category information. After the original card identification model is pre-trained, the original backbone network is immediately transformed into a target backbone network, and the original classification network is immediately transformed into a target classification network. Step S202 inputs the card sample image into the original card identification model for this round of identification processing to obtain the current round of identification results, which may include, but are not limited to, the following steps S301 to S303.

[0093] Step S301, downsampling each card sample image through the original backbone network to obtain at least two sample feature maps; wherein each sample feature map corresponds to a feature resolution;

[0094] Step S302, extracting the current round of splicing contour information from at least two sample feature maps through a contour extraction network;

[0095] Step S303: extract the card category information of this round from at least two sample feature maps through the original classification network.

[0096] In some embodiments of the present application, the original card identification model adopts a multi-network structure, including an original backbone network, an original classification network, and a contour extraction network. This design aims to fully extract the features of the card sample image and perform effective identification processing.

[0097] In step S301 of some embodiments, it should be noted that the original backbone network performs feature extraction and downsampling processing on the input card sample image. The original backbone network receives the original high-resolution card image and generates at least two sample feature maps of different resolutions. These feature maps capture the features of the image at different scales, providing a basis for subsequent contour information extraction and category information extraction. Downsampling not only reduces the spatial size of the data, but also helps to extract more abstract features in the image, so that the original card identification model can extract the key features of the image, thereby forming a sample feature map.

[0098] Reference Figure 4In some more specific embodiments, the original backbone network can be composed of feature extraction layer 1 to feature extraction layer 4. The core principle of each feature extraction layer is to extract features based on the convolution layer (Convolutional Layer, Conv), which is a network architecture particularly suitable for image analysis. The convolution layer is the basic building block in the convolutional neural network (CNN) and is used to extract local features of input data (such as images). It slides on the input data using a learnable convolution kernel (or filter), calculates the dot product of the convolution kernel and the local area of ​​the input data, and generates feature maps (FeatureMaps). The convolution layer can capture local patterns in the image, such as edges, textures, etc., and has translation invariance. In addition, in addition to the convolution layer, each feature extraction layer can also include a batch normalization layer (Batch Normalization, BN) and a rectified linear activation layer (Rectified Linear Unity, ReLU). Among them, batch normalization is a technology used to improve the training process, which can accelerate the convergence speed and improve the stability and generalization ability of the model. The BN layer normalizes the data of each mini-batch by subtracting the mean and dividing by the standard deviation, and then scaling and translating it through learnable parameters. This helps to reduce internal covariate shift, that is, the problem that the input distribution of the previous layer in the network changes during the training process; the rectified linear activation layer is a commonly used activation function used to introduce nonlinearity in neural networks. ReLU can increase the expressive power of the network, and due to its linear characteristics, it has a fast calculation speed and helps to alleviate the gradient vanishing problem. It should be noted that convolutional layers, batch normalization layers, and rectified linear activation layers can be used in combination inside the feature extraction layer to achieve effective feature extraction and nonlinear transformation. For example, a convolutional block of a feature extraction layer may contain a convolutional layer, followed by a batch normalization layer, followed by a rectified linear activation layer.

[0099] For feature extraction layer 1, which is the first layer of the original backbone network, it is used to extract the most basic image features of the card sample image, such as edges and color changes. The features corresponding to this layer have higher resolution and smaller size, which can capture details and provide original visual information for subsequent layers.

[0100] For feature extraction layers 2 to 4, as the network level goes deeper, the features processed by each feature extraction layer become more abstract and complex, that is, the card sample image is downsampled by half, one quarter, and one eighth respectively to obtain the sample features. Figure 1 , sample characteristics Figure 2 , sample characteristics Figure 3 and sample characteristics Figure 4For example, feature extraction layer 2 may focus on shapes, and feature extraction layer 3 may be more focused on recognizing specific textures or patterns. By feature extraction layer 4, the original backbone network is already able to recognize highly abstract features that are critical to the final classification or recognition task.

[0101] For each feature extraction layer from feature extraction layer 1 to feature extraction layer 4, the original backbone network can continuously extract and synthesize higher-level features through continuous operations of convolution, activation, and pooling. The output of each layer is a representation of a deeper understanding of the input image, which will ultimately serve as features from the global scale to each local scale, forming sample feature maps corresponding to different feature resolutions. Through this hierarchical approach, feature extraction layers 1 to 4 effectively extract and construct increasingly complex features from the card sample images, allowing the network to efficiently and accurately analyze the images.

[0102] In step S302 of some embodiments, the task of the contour extraction network is to extract the splicing contour information from the sample feature map generated by the original backbone network. The contour extraction network is used to identify and extract possible splicing traces or contours in the card image. Extracting this information can improve the accuracy and sensitivity of the identification model in identifying splicing operations.

[0103] Reference Figure 5 According to some embodiments of the present application, step S302 extracts the contour information of this round of stitching from at least two sample feature maps through a contour extraction network, which may include, but is not limited to, the following steps S501 to S502.

[0104] Step S501, determining the resolution ranking based on the high and low relationship of the feature resolutions corresponding to each sample feature map;

[0105] Step S502: Based on the resolution sorting, the sample feature maps corresponding to each feature resolution are merged to obtain the contour information of this round of stitching.

[0106] It should be emphasized that in the original card identification model, the role of the contour extraction network is to identify and extract the splicing contour information in the card image.

[0107] In step S501 of some embodiments, the contour extraction network needs to determine the relationship between the feature resolutions of each sample feature map and sort them accordingly. It should be noted that resolution sorting refers to the sorting of the feature resolutions, rather than the sorting of the sample feature maps. This is because in the downsampling process of the original backbone network, the generated feature maps have different feature resolutions. For example, one sample feature map may correspond to a higher feature resolution of the original image, while another sample feature map corresponds to a lower feature resolution. By comparing the feature resolutions of these sample feature maps, the contour extraction network can determine their position in the feature hierarchy, thereby providing a reference for subsequent processing.

[0108] In step S502 of some embodiments, the contour extraction network will fuse the sample feature maps of different resolutions based on the resolution sorting. The purpose of this step is to integrate the feature information at different scales to obtain more comprehensive splicing contour information. The fusion process can involve a variety of technologies, such as feature pyramid, feature weighting, feature superposition, etc. In this way, the contour extraction network can utilize the characteristics of each sample feature map to enhance the recognition ability of the splicing contour. The fused current round of splicing contour information will contain rich information from different feature resolution levels. For example, a high-resolution feature map may provide more detailed detail information, while a low-resolution feature map may provide more extensive context information. In this way, the current round of splicing contour information is extracted through the contour extraction network, aiming to reflect the features of the splicing area through the current round of splicing contour information, such as edge discontinuity, texture inconsistency, etc., thereby providing more comprehensive and accurate contour features for the pre-training of the original card identification model and enhancing the performance of the model in detecting image splicing. .

[0109] Reference Figure 6 According to some embodiments of the present application, step S502 is based on resolution sorting, and the sample feature maps corresponding to each feature resolution are merged to obtain the current round of stitching contour information, which may include, but is not limited to, the following steps S601 to S603.

[0110] Step S601, for a sample feature map with the lowest feature resolution in the resolution sorting, input the sample feature map with the lowest feature resolution into a contour extraction network to extract contour features, and obtain contour sub-feature information corresponding to the sample feature map;

[0111] Step S602: for the sample feature map with non-lowest feature resolution in the resolution sorting, obtain the sample feature map with low feature resolution Figure 1 The contour sub-feature information of the first level is obtained, and the contour sub-feature information is merged with the sample feature map of non-lowest feature resolution to obtain the contour sub-feature information corresponding to the sample feature map;

[0112] Step S603, according to the sample feature map with the highest feature resolution in the resolution sorting, the contour sub-feature information corresponding to the sample feature map with the highest feature resolution is determined as the contour information of this round of splicing.

[0113] In step S601 of some embodiments, the starting point for fusion of each sample feature map is the sample feature map with the lowest feature resolution in the resolution sorting. This sample feature map is input into the contour extraction network, and its purpose is to extract the contour sub-feature information corresponding to the sample feature map with the lowest feature resolution in the resolution sorting. Through this step, a feature representation with low resolution but possibly containing important contour information, i.e., contour sub-feature information, can be obtained.

[0114] In step S602 of some embodiments, for sample feature maps that are not of the lowest feature resolution in the resolution sorting, it is first necessary to obtain the resolution sorting, and find the contour sub-feature information with a feature resolution one level lower than these sample feature maps from the resolution sorting. Then, this part of the contour sub-feature information is fused with the current sample feature map. The fusion process may involve superposition, weighted averaging or other fusion techniques of the contour sub-feature information and the current sample feature map, the purpose of which is to integrate feature information at different feature resolutions and enhance the recognition ability of the spliced ​​contour. In this way, richer and more detailed contour sub-feature information can be obtained.

[0115] In step S603 of some embodiments, according to the sample feature map with the highest feature resolution in the resolution sorting, the contour sub-feature information corresponding to the sample feature map with the highest feature resolution is determined as the contour information of this round of splicing. This step is based on the highest sample feature map in the resolution sorting, and the contour sub-feature information corresponding to the sample feature map is used as the contour information of this round of splicing. The contour sub-feature information corresponding to the sample feature map with the highest resolution is selected as the final output because the high-resolution feature map usually contains finer detail information, which is beneficial for accurately identifying the splicing contour.

[0116] Through the embodiment of the present application shown in steps S601 to S603, the contour extraction network can extract the current round of stitching contour information from sample feature maps of different resolutions and effectively integrate the information.

[0117] The final stitching contour information not only contains the global contour information in the low-resolution feature map, but also integrates the detailed contour information in the high-resolution feature map, thereby providing more comprehensive and accurate contour features for the pre-training of the original card identification model and enhancing the model's performance in detecting image stitching.

[0118] It is worth mentioning that since the stitching contour information of this round obtained in step S603 is the contour sub-feature information corresponding to the sample feature map with the highest feature resolution. Therefore, under the guidance of the stitching contour position by the stitching contour information of this round, the original classification network can better classify the sample feature map based on the stitching contour information of this round to determine the corresponding category, for example, to determine whether the card sample image is a non-stitched image, a stitched image, or a suspected stitched image. In this embodiment, the original classification network can determine more accurate card category information of this round in combination with the guidance of the stitching contour position by the stitching contour information of this round.

[0119] In step S303 of some embodiments, the original classification network is responsible for extracting the category information of the card from the sample feature graph. The original classification network can identify the type or attribute of the card, such as spliced ​​card, non-spliced ​​card, suspected spliced ​​card, etc. by analyzing the sample feature graph.

[0120] In the embodiment of the present application, when the card sample image is input into the original card identification model for this round of identification processing, the original card identification model first processes the card sample image through the original backbone network to generate a sample feature map. Then, the contour extraction network and the original classification network respectively operate on these sample feature maps to extract the splicing contour information and the card category information of this round. The splicing contour information and the card category information of this round together constitute the identification result of this round.

[0121] Reference Figure 7 According to some embodiments of the present application, the original card identification model includes an original backbone network, an original classification network, and a contour extraction network, and the current round of identification results includes the current round of spliced ​​contour information and the current round of card category information. Specifically, the present application embodiment can downsample each card sample image through the original backbone network to obtain the sample feature Figure 1 , sample characteristics Figure 2 , sample characteristics Figure 3 etc., wherein each sample feature map corresponds to a feature resolution. Further, the contour extraction network extracts the current round of splicing contour information from at least two sample feature maps, and the original classification network extracts the current round of card category information from at least two sample feature maps.

[0122] Reference Figure 8 According to some embodiments of the present application, the sample image annotation includes the splicing contour annotation and the card category annotation of the card sample image, and the comparison deviation data includes the contour deviation data and the category deviation data. Step S203 compares the sample image annotation corresponding to the card sample image with the current round identification result to obtain the comparison deviation data, which may include, but not limited to, the following steps S801 to S802.

[0123] Step S801, performing contour deviation comparison based on the splicing contour annotation and the current round of splicing contour information to obtain contour deviation data;

[0124] Step S802: performing a category deviation comparison based on the card category label and the card category information of this round to obtain category deviation data.

[0125] In some embodiments of the present application, the sample image annotation may specifically include two parts: the splicing contour annotation of the card sample image and the card category annotation. These two parts of annotation correspond to the structural information and type information in the card image, respectively, and provide a standard for training and evaluation of the model.

[0126] In step S801 of some embodiments, contour deviation comparison is performed based on the splicing contour annotation and the current round of splicing contour information to obtain contour deviation data. It should be noted that the current round of splicing contour information generated by the contour extraction network is compared with the actual splicing contour annotation to identify the difference between the two. This comparison can reveal the accuracy of the contour extraction network in identifying the image splicing contour, thereby obtaining contour deviation data. Contour deviation data is an important indicator for measuring model performance, which reflects the degree of deviation between the current round of splicing contour information predicted by the contour extraction network and the actual annotation.

[0127] Reference Fig. 9 According to some embodiments of the present application, step S801 performs contour deviation comparison based on the stitching contour annotation and the current round of stitching contour information to obtain contour deviation data, which may include, but is not limited to, the following steps S901 to S902.

[0128] Step S901, performing scale matching processing on the splicing contour annotation to obtain a labeling feature map; wherein the feature resolution of the labeling feature map matches the current round of splicing contour information;

[0129] Step S902 , performing contour deviation comparison based on the marked feature map and the contour information of this round of splicing to obtain contour deviation data.

[0130] It should be noted that the purpose of step S801 is to obtain contour deviation data through accurate comparison, and the contour deviation data helps to evaluate and improve the performance of the original card identification model in splicing contour recognition. The key to achieving this goal is to ensure that the splicing contour annotation is consistent with the splicing contour information output by the model in terms of scale and resolution, so that an effective comparison can be performed.

[0131] In step S901 of some embodiments, the spliced ​​contour annotation is scale-matched to obtain a labeled feature map. It should be noted that this process is to adjust the scale of the spliced ​​contour annotation so that it matches the current round of spliced ​​contour information output by the contour extraction network in terms of feature resolution. The scale matching process may include scaling, rotation or other geometric transformations to ensure that the size, direction and position of the labeled feature map are consistent with the current round of spliced ​​contour information. The labeled feature map is a scaled annotation that can directly correspond to the output of the contour extraction network in terms of scale.

[0132] In step S902 of some embodiments, a contour deviation comparison is performed based on the marked feature map and the current round of splicing contour information to obtain contour deviation data. It should be noted that this comparison process involves quantitative analysis of the difference between the marked feature map and the current round of splicing contour information, and a variety of different methods may be used to measure the deviation, such as calculating the distance, angle difference or shape dissimilarity between the marked feature map and the current round of splicing contour information. Through this comparison, contour deviation data can be obtained, which reflects the difference between the current round of splicing contour information output by the contour extraction network and the marked feature map.

[0133] Profile deviation data are crucial for model training because they provide direct feedback on model performance. If the deviation is large, it may indicate that the model has problems recognizing certain features and needs further optimization. By analyzing the profile deviation data, the weaknesses of the model can be identified, so that model parameters can be adjusted or feature extraction methods can be improved in a targeted manner.

[0134] In addition, the contour deviation data can also be used to guide the refined training of the original card identification model, such as by adding samples with large differences between the labeled and predicted contours, or adjusting the loss function to pay more attention to these areas with large deviations. This feedback-based iterative training process helps to gradually improve the accuracy of the original card identification model, and ultimately achieve more reliable card identification.

[0135] In step S802 of some embodiments, a category deviation comparison is performed based on the card category annotation and the card category information of this round to obtain category deviation data. It should be noted that the category deviation comparison is to compare the card category annotation with the card category information of this round. The purpose of this step is to evaluate the performance of the original classification network on the card classification task. By comparing the card category information of this round predicted by the original classification network with the actual card category annotation, the category deviation data can be obtained. The category deviation data reveals the accuracy of the original classification network in card classification, including any classification errors or uncertainties.

[0136] Through step S801 and step S802, the comparison deviation data obtained include contour deviation data and category deviation data, which provide comprehensive feedback information for the model. These data not only indicate the performance of the model in the current training stage, but also provide a basis for further model optimization and parameter adjustment. For example, if the contour deviation data is large, it may indicate that the model is insufficient in splicing contour recognition, and the parameters or structure of the contour extraction network need to be further adjusted. Similarly, if the category deviation data is large, it may be necessary to improve the original classification network to improve the recognition accuracy of the card category. It should be understood that the generation and analysis of comparison deviation data is an iterative and gradual improvement process in the model training process. Through continuous evaluation and adjustment, the original card identification can gradually learn more accurate feature representations, improve its performance in the card identification task, and finally reach the preset training termination condition to form a pre-trained target card identification model.

[0137] Reference Fig.10 In some more specific embodiments, the original backbone network may be composed of feature extraction layer 1 to feature extraction layer 4.

[0138] For feature extraction layer 1, which is the first layer of the original backbone network, it is used to extract the most basic image features of the card sample image, such as edges and color changes. The features corresponding to this layer have higher resolution and smaller size, which can capture details and provide original visual information for subsequent layers.

[0139] For feature extraction layers 2 to 4, as the network goes deeper, the features processed by each feature extraction layer become more abstract and complex. For example, feature extraction layer 2 may focus on shapes, and feature extraction layer 3 may recognize more specific textures or patterns. By feature extraction layer 4, the original backbone network is already able to recognize highly abstract features that are crucial for the final classification or recognition task.

[0140] For each feature extraction layer from feature extraction layer 1 to feature extraction layer 4, the original backbone network can continuously extract and synthesize higher-level features through continuous operations of convolution, activation, and pooling. The output of each layer is a representation of a deeper understanding of the input image, which will ultimately serve as features from the global scale to each local scale, forming sample feature maps corresponding to different feature resolutions. Through this hierarchical approach, feature extraction layers 1 to 4 effectively extract and construct increasingly complex features from the card sample images, allowing the network to efficiently and accurately analyze the images.

[0141] The main task of the original classification network is to classify the input image data into preset categories. It is usually located at the end of the deep learning model and receives the feature map output by the previous feature extraction layer. In the classification task, a fully connected layer (or dense layer) is usually used to implement it. The fully connected layer integrates and weights all the features extracted previously and outputs the predicted probability of each category. The most commonly used activation function is Softmax, which can convert the output into a probability format so that the final output can be intuitively expressed as the probability of each category. The original classification network is trained through the back propagation algorithm. According to the error between the card category label and the current card category information, the weights in the original classification network are continuously adjusted to minimize the classification error.

[0142] The goal of the contour extraction network is to identify and extract the contour or boundary of the target object from the image. This is crucial for tasks such as image segmentation and target detection. The contour extraction network of the embodiment of the present application can handle the task of contour extraction based on edge detection. Specifically, the edge information in the image can be enhanced by a specific filter, such as Sobel or Canny edge detector. In addition to traditional filters, the contour extraction of the embodiment of the present application can also use deep features in the convolutional neural network, combined with specific filters, to improve the accuracy and robustness of edge detection.

[0143] In the contour extraction network, the spliced ​​contour annotation needs to be scale-matched, that is, the spliced ​​contour annotation is downsampled by half, one quarter, and one eighth respectively to obtain the annotation features Figure 2 , annotation features Figure 3 , annotation features Figure 4 It should be noted that this processing is to adjust the scale of the spliced ​​contour annotation so that it matches the feature resolution of the current round of spliced ​​contour information output by the contour extraction network.

[0144] The contour deviation comparison is performed based on the annotated feature map and the current round of splicing contour information to obtain the contour deviation data. It should be noted that this comparison process involves quantitative analysis of the difference between the annotated feature map and the current round of splicing contour information. A variety of different methods may be used to measure the deviation, such as calculating the distance, angle difference or shape dissimilarity between the annotated feature map and the current round of splicing contour information. Through this comparison, contour deviation data can be obtained, which reflects the difference between the current round of splicing contour information output by the contour extraction network and the annotated feature map. Contour deviation data are crucial for model training because they provide direct feedback on model performance. If the deviation is large, it may indicate that the model has problems in identifying certain features and needs further optimization. By analyzing the contour deviation data, the weaknesses of the model can be identified, so that the model parameters can be adjusted or the feature extraction method can be improved in a targeted manner.

[0145] against Fig. 9 The original card identification model shown in the figure includes an original backbone network, an original classification network and a contour extraction network. The original backbone network extracts features from the global scale to each local scale for the card sample image. The original classification network performs classification by learning and identifying different category features, while the contour extraction network focuses on the spatial attributes and boundaries of objects. The two complement each other and jointly improve the overall performance and accuracy of the system.

[0146] Reference Fig.11 According to some embodiments of the present application, until the original card identification model meets the preset training termination condition in step S204, the original card identification model is determined as the pre-trained target card identification model, which may include, but is not limited to, the following steps S1101 to S1102:

[0147] Step S1101, in response to the original card identification model satisfying the training termination condition, removing the contour extraction network from the original card identification model;

[0148] Step S1102: determining the original card identification model with the contour extraction network removed as the pre-trained target card identification model.

[0149] It should be noted that in some embodiments provided in the present application, the pre-training process of the original card identification model may include multiple stages until the model meets the preset training termination conditions. The training termination conditions may include but are not limited to the improvement of model performance tending to be stable, reaching a predetermined accuracy rate, the end of the training cycle, or the limitation of computing resources. Once these conditions are met, the original card identification model will be identified as the pre-trained target card identification model, and the training process will enter the subsequent stage.

[0150] In step S1101 of some embodiments, in response to the original card identification model satisfying the training termination condition, the contour extraction network is removed from the original card identification model. It should be noted that the contour extraction network is necessary in the early stage of training because it can help the original card identification model learn how to identify and extract the spliced ​​contour information in the card image. However, once the original card identification model achieves satisfactory performance in spliced ​​contour recognition and the training termination condition is met, the contour extraction network is no longer needed. Removing the contour extraction network can simplify the model structure, reduce computational complexity, and may also help avoid overfitting and ensure the generalization ability of the model.

[0151] In step S1102 of some embodiments, after the contour extraction network is removed, the original card identification model is finally adjusted and verified and is formally determined as the pre-trained target card identification model. At this point, the model is ready for actual card identification tasks, and it includes the original backbone network and the original classification network, which have been optimized during the training process to ensure the model's efficiency and accuracy in actual applications.

[0152] In the embodiment of the present application shown by steps S1101 to S1102, after removing the contour extraction network, the original card identification model will retain its core identification capabilities and may obtain higher operating efficiency and better generalization performance. This optimized model is more suitable for deployment in actual scenarios, and can provide reliable card identification services in both resource-constrained environments and applications with high real-time requirements.

[0153] In step S103 of some embodiments, the target card image is downsampled by the target backbone network to obtain at least two mask feature maps; wherein each mask feature map has a corresponding feature resolution, and the feature resolutions of any two mask feature maps are different. It should be noted that since the target card identification model can be used to identify whether there is splicing in the target card image, the target card identification model has been pre-trained and has the ability to identify splicing. After the original card identification model is pre-trained, the original backbone network corresponding to the original card identification model is immediately transformed into the target backbone network, and the target backbone network corresponding to the target card identification model is immediately transformed into the target backbone network.

[0154] Downsampling is an operation that reduces the size of an image or feature map. It can be used to reduce the spatial resolution of the data, thereby reducing the amount of computation and preventing overfitting. In deep learning, downsampling can be used in convolutional neural networks (CNNs), especially when designing network structures, to increase the depth of the network by gradually reducing the spatial dimensions of the feature map.

[0155] The downsampling process mentioned in the embodiments of the present application refers to a technique for reducing the spatial size of data during image processing or feature extraction. Specifically for the downsampling process of the target card image, downsampling can be used to reduce the resolution of the target card image. Downsampling can be implemented in a variety of ways, including pixel discarding, average pooling, maximum pooling, or other more complex interpolation methods. In some cases, downsampling also involves adjusting the dimensions of the feature map to accommodate the input requirements of subsequent network layers. For example, if the size of the original target card image is 256x256, downsampling may reduce it to 128x128 or smaller, which can reduce the amount of computation and may help the network focus on more important global features.

[0156] In the embodiments of the present application, the downsampling process can be used to generate target card images of different scales at different stages, which are then used as input for further processing. For example, in a multi-scale feature extraction network, the downsampled target card image can be combined with the original size target card image to provide richer contextual information, thereby helping the network to better understand the image content and structure. In addition, the downsampled target card image can also be used as prior information to guide the subsequent upsampling or feature refinement process, achieving a step-by-step optimization from coarse to fine.

[0157] In practical applications, downsampling processing helps to build a multi-scale analysis framework, so that the original card identification model of the embodiment of the present application can capture local details and global information at the same time, which is particularly important for tasks such as image segmentation and target detection. In this way, the original card identification model of the embodiment of the present application can more effectively process objects of different sizes, improve the understanding of image structure, and thus improve the performance of the final task. It should be understood that the embodiment of the present application inputs the target card image into the target card image for downsampling processing to obtain at least two mask feature maps, aiming to obtain target card images at different scales, so as to extract mask feature maps from the global scale to each local scale for the target card image using the original card identification model. Among them, at least two mask feature maps are used to reflect the image features of the target card image from the global scale to each local scale, and each mask feature map corresponds to a feature resolution, which is also the resolution corresponding to the mask feature map.

[0158] In step S104 of some embodiments, the classification network of the target card identification model performs card splicing identification on the mask feature map corresponding to each feature resolution to obtain the card identification result. It should be noted that after the original card identification model is pre-trained, the original backbone network corresponding to the original card identification model is immediately transformed into the target backbone network, and the target backbone network corresponding to the target card identification model is immediately transformed into the target backbone network. After the target card image is input into the target card image for downsampling processing to obtain at least two mask feature maps, in order to further identify whether there is splicing in the target card image, it is necessary to perform card splicing identification on the mask feature map corresponding to each feature resolution through the classification network of the target card identification model to obtain the card identification result. Since the mask feature map is used to reflect the image features of the target card image from the global scale to each local scale, and the original card identification model has the ability to identify splicing after pre-training. Therefore, the original card identification model can identify card splicing for each mask feature map in combination with the image features from the global scale to each local scale, thereby obtaining a more accurate card identification result.

[0159] Reference Fig.12, Fig.12 The hardware structure of an electronic device of another embodiment is illustrated. The electronic device may include, but is not limited to:

[0160] The processor 1201 may be implemented by a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present application;

[0161] The memory 1202 can be implemented in the form of a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 1202 can store an operating system and other application programs. When the technical solution provided in the embodiment of this specification is implemented by software or firmware, the relevant program code is stored in the memory 1202, and the processor 1201 calls and executes the card splicing identification method of the embodiment of this application;

[0162] Input / output interface 1203, used to implement information input and output;

[0163] The communication interface 1204 is used to realize the communication interaction between the device and other devices. The communication can be realized through a wired manner (such as USB, network cable, etc.) or a wireless manner (such as mobile network, WI FI, Bluetooth, etc.);

[0164] A bus 1205 that transmits information between various components of the device (e.g., the processor 1201, the memory 1202, the input / output interface 1203, and the communication interface 1204);

[0165] The processor 1201 , the memory 1202 , the input / output interface 1203 and the communication interface 1204 are connected to each other in communication within the device via the bus 1205 .

[0166] The embodiment of the present application also provides a computer program product, which includes a computer program. The processor of a computer device reads and executes the computer program, so that the computer device executes and implements the above-mentioned card splicing identification method.

[0167] The terms "first", "second", "third", "fourth", etc. (if any) in the specification of the present disclosure and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present disclosure described herein can, for example, be implemented in an order other than those illustrated or described herein. In addition, the terms "comprises" and "comprising" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0168] It should be understood that in the present disclosure, "at least one (item)" means one or more, and "plurality" means two or more. "And / or" is used to describe the association relationship of associated objects, indicating that three relationships may exist. For example, "A and / or B" can mean: only A exists, only B exists, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally indicates that the objects associated before and after are in an "or" relationship. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, at least one of a, b or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.

[0169] It should be understood that in the description of the embodiments of the present application, the meaning of multiple (or multiple items) is more than two, greater than, less than, exceed, etc. are understood to not include the number, and above, below, within, etc. are understood to include the number.

[0170] In the several embodiments provided in the present disclosure, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0171] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0172] In addition, each functional unit in each embodiment of the present disclosure may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0173] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present disclosure, in essence, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the various embodiments of the present disclosure. The aforementioned storage medium may include, but is not limited to: U disk, mobile hard disk, read-only memory (Read-OnlyMemory, referred to as ROM), random access memory (Random Access Memory, referred to as RAM), disk or optical disk and other media that can store program codes.

[0174] It should also be understood that the various implementations provided in the embodiments of the present application can be combined arbitrarily to achieve different technical effects.

[0175] The above is a specific description of the implementation methods of the present disclosure, but the present disclosure is not limited to the above implementation methods. Technical personnel familiar with the art can also make various equivalent modifications or substitutions without violating the spirit of the present disclosure. These equivalent modifications or substitutions are all included in the scope defined by the claims of the present disclosure.

Claims

1. A card splicing identification method, characterized in that: include: Get the target card image; Obtain a pre-trained target card identification model; wherein the target card identification model includes a target backbone network and a target classification network; Downsampling the target card image through the target backbone network to obtain at least two mask feature maps; wherein each of the mask feature maps has a corresponding feature resolution, and the feature resolutions of any two mask feature maps are different; The target classification network performs card splicing identification on each of the mask feature maps to obtain a card identification result, and the card identification result is used to indicate whether the target card image is a spliced ​​image or a non-spliced ​​image.

2. The card splicing identification method according to claim 1, characterized in that: The step of obtaining a pre-trained target card identification model includes: Acquire a training sample set; wherein the training sample set includes a plurality of card sample images and a sample image annotation corresponding to each of the card sample images; Input the card sample image into a preset original card identification model for current round identification processing to obtain the current round identification result; Compare the sample image annotation corresponding to the card sample image with the current round identification result to obtain comparison deviation data; The model parameters of the original card identification model are updated based on the comparison deviation data, and the process returns to inputting the card sample image into the original card identification model for this round of identification processing until the original card identification model meets the preset training termination condition, and the original card identification model is determined as the pre-trained target card identification model.

3. The card splicing identification method according to claim 2, characterized in that: The original card identification model includes an original backbone network, an original classification network, and a contour extraction network, and the current round identification result includes the current round splicing contour information and the current round card category information; The step of inputting the card sample image into a preset original card identification model for current round identification processing to obtain the current round identification result includes: Downsampling each of the card sample images through the original backbone network to obtain at least two sample feature maps; wherein each of the sample feature maps corresponds to one of the feature resolutions; Extracting the current round of stitching contour information from at least two of the sample feature maps through the contour extraction network; The card category information of this round is extracted from at least two of the sample feature maps through the original classification network.

4. The card-document splicing identification method according to claim 3, characterized in that: The extracting the current round of stitching contour information from at least two of the sample feature maps by the contour extraction network includes: Determine a resolution ranking based on a high-low relationship between the feature resolutions of the sample feature graphs and the feature resolutions; Based on the resolution sorting, the sample feature maps corresponding to the feature resolutions are fused to obtain the current round of stitching contour information.

5. The card splicing identification method according to claim 4, characterized in that: The resolution-based sorting, fusing the sample feature maps corresponding to the feature resolutions to obtain the current round of stitching contour information, includes: For the sample feature map with the lowest feature resolution in the resolution sorting, input the sample feature map with the lowest feature resolution into the contour extraction network to extract contour features, so as to obtain contour sub-feature information corresponding to the sample feature map; For the sample feature map of the non-lowest feature resolution in the resolution sorting, obtain the contour sub-feature information of the feature resolution one level lower than the sample feature map, and fuse the contour sub-feature information with the sample feature map of the non-lowest feature resolution to obtain the contour sub-feature information corresponding to the sample feature map; According to the sample feature map with the highest feature resolution in the resolution sorting, the contour sub-feature information corresponding to the sample feature map with the highest feature resolution is determined as the current round of stitching contour information.

6. The card splicing identification method according to claim 3, characterized in that: The sample image annotation includes a splicing contour annotation and a card category annotation of the card sample image, and the comparison deviation data includes contour deviation data and category deviation data; The step of comparing the sample image annotation corresponding to the card sample image with the current round identification result to obtain comparison deviation data includes: Performing a contour deviation comparison based on the splicing contour annotation and the current round of splicing contour information to obtain the contour deviation data; A category deviation comparison is performed based on the card category label and the card category information of this round to obtain the category deviation data.

7. The card splicing identification method according to claim 6, characterized in that: The comparing the contour deviation based on the splicing contour annotation and the current round of splicing contour information to obtain the contour deviation data includes: Performing scale matching processing on the splicing contour annotation to obtain an annotation feature map; wherein the feature resolution of the annotation feature map matches the current round of splicing contour information; A contour deviation comparison is performed based on the marked feature map and the current round of splicing contour information to obtain the contour deviation data.

8. The card-certificate splicing identification method according to claim 3, characterized in that: The method of determining the original card identification model as the pre-trained target card identification model until the original card identification model satisfies a preset training termination condition comprises: In response to the original card authentication model satisfying the training termination condition, removing the contour extraction network from the original card authentication model; The original card authentication model with the contour extraction network removed is determined as the pre-trained target card authentication model.

9. An electronic device, characterized in that: include: A memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the card splicing identification method according to any one of claims 1 to 8 is implemented.

10. A computer-readable storage medium, characterized in that: The storage medium stores a program, and the program is executed by a processor to implement the card splicing identification method according to any one of claims 1 to 8.