A seal recognition system and recognition method based on neural network
By annotating and data augmenting the customs seal data, using RA-CNN and APN networks to extract features, combined with MobileNetV2 classification network, the problem of low accuracy in customs seal recognition is solved, high accuracy and real-time recognition are achieved, and manpower consumption is reduced .
Patent Information
- Application Number
- CN202210512526.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-05-12
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2042-05-12
AI Technical Summary
The prior art has low recognition accuracy in customs seal recognition, and traditional algorithms cannot meet the real-time recognition requirements, and manual recognition is time-consuming and labor-intensive. Traditional algorithms such as VGG and LENET are insufficient in customs seal recognition.
By annotating and data augmenting the seal data, features are extracted using the RA-CNN network model, and attention area information is obtained through the APN network, and repeated training is used to improve recognition accuracy, combining the MobileNetV2 classification network and cross entropy loss function and sorting loss Functions make predictions.
It realizes high-accuracy seal recognition, with a recognition speed of within 7 to 8 milliseconds, meeting real-time detection needs and reducing manpower consumption. It has a higher accuracy rate than traditional networks and is more generalized.
Smart Images

Figure CN114998646B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image recognition, and is a neural network-based seal recognition system and recognition method. Background Art
[0002] Currently, image recognition technologies and algorithms are gradually improving, and are increasingly being applied to various scenarios in our lives, such as facial recognition and biomedical image recognition.
[0003] Customs inspections require significant seal recognition. Currently, seal recognition in customs is primarily performed manually. Compared to machine recognition, manual recognition is time-consuming and labor-intensive. Existing traditional algorithms cannot achieve ideal classification results, and image recognition technology is not widely used in customs seal recognition.
[0004] Customs seal recognition, like other image recognition tasks, is primarily text-based and involves a large number of possible results. Traditional algorithms, such as VGG and LENET, have low accuracy in this area and cannot meet real-world requirements. Furthermore, we also need to consider whether the algorithm's scale can achieve real-time recognition. Summary of the Invention
[0005] To overcome the shortcomings of existing technologies and address the low accuracy of automatic seal recognition, this invention labels seal data and uses data augmentation to increase the data volume, enabling the trained model to have stronger generalization capabilities. Furthermore, the data is used to train a RACNN network model to extract seal features. Finally, the features are classified to produce classification results. The seal classification model trained using this method can achieve high-accuracy seal recognition.
[0006] The present invention provides a seal recognition system and recognition method based on a neural network, and the present invention provides the following technical solutions:
[0007] A seal recognition method based on a neural network, the method comprising the following steps:
[0008] Step 1: Label the seal image data and increase the data volume through data augmentation to obtain sample image data;
[0009] Step 2: Based on the sample image data after data augmentation, training is performed to obtain training data, and the seal image features are extracted and classified;
[0010] Step 3: Based on the obtained seal image features, the APN network is used to obtain the attention area information;
[0011] Step 4: Crop the attention area to obtain the cropped image;
[0012] Step 5: Repeat steps 2 to 4 three times to obtain the output results of the three-scale networks. By adding the loss functions of the three-scale networks as the loss function of the result, matching it with the results in the training set, and predicting the image to obtain the seal recognition prediction result.
[0013] Preferably, the step 1 specifically comprises: annotating the seal image data and performing data augmentation, by making a series of random changes to the training image to generate similar but different training samples, thereby expanding the scale of the training data set;
[0014] Data augmentation specifically includes random rotation; random translation along the horizontal or vertical direction with a percentage of the image size as the variation range; random scaling of the image size proportionally; and a combination of the above methods.
[0015] Preferably, the step 2 is specifically as follows:
[0016] Use the MobileNetV2 classification network to extract features and classify sample images to obtain image features;
[0017] The above feature extraction operation is: Depth-wise Separable Convolution is used, that is, PW convolution, DW convolution, and PW convolution are used to extract features.
[0018] Preferably, the step 3 is specifically:
[0019] Based on the extracted features, the APN network is used to obtain the attention area information. The extracted image is input into a pre-trained convolutional layer to extract region-based deep features. The network is modeled as a multi-task with two outputs at each scale. The first task is to generate a probability distribution over fine-grained categories, and the second task is to predict the coordinates of a set of participating regions for the next scale.
[0020] Preferably, the step 4 is specifically as follows:
[0021] Cropping is achieved by multiplying the original image with the attention mask at a coarse scale. The function of the cropping method in RA-CNN is a differentiable logistic function, and backpropagation is performed to complete end-to-end training;
[0022] The cropped area is enlarged to a larger size using an upsampling method based on bilinear interpolation and implemented using the torch.nn.functional function in the pytorch framework.
[0023] Preferably, the step 5 is specifically as follows:
[0024] The loss function is determined by the following formula:
[0025]
[0026] Among them, L cls is the commonly used cross entropy loss function, that is
[0027] L cls (Y (s) ,Y * )=CE(Y (s) ,Y * )
[0028] Used to optimize parameters at each scale;
[0029] Part II rank is the ranking loss function,
[0030]
[0031] The role is that the latter scale network uses the previous scale network as a reference, forcing the latter scale network to locate the most discriminative area and produce a higher prediction value.
[0032] Preferably, the neural network model in the jetson nano device predicts the acquired image to obtain a seal recognition prediction result, and displays the nationality name and the national flag image corresponding to the result in the display unit; the processed information is transmitted to the storage unit for storage and used to record data.
[0033] A seal recognition system based on neural network, the system comprises:
[0034] a marking module, which marks the seal image data, increases the data volume through a data augmentation method, and obtains sample image data;
[0035] A classification module is used to train the sample image data after data augmentation to obtain training data, extract seal image features and classify them;
[0036] APN network module, the labeling module uses the APN network to obtain attention area information according to the obtained seal image features;
[0037] a cropping module, wherein the cropping module crops the attention area to obtain a cropped image;
[0038] The prediction module obtains the output results of the three scale networks, adds the loss functions of the three scale networks as the loss function of the result, matches it with the results in the training set, and predicts the image to obtain the seal recognition prediction result.
[0039] A computer-readable storage medium stores a computer program, which is executed by a processor to implement a seal recognition method based on a neural network.
[0040] A computer device includes a memory and a processor, wherein a computer program is stored in the memory. When the processor runs the computer program stored in the memory, the processor executes a seal recognition method based on a neural network.
[0041] The present invention has the following beneficial effects:
[0042] This invention, based on deep learning methods, which have recently gained significant traction in the field of image recognition, classifies customs stamps of different nationalities. By optimizing and improving the neural network structure, classification accuracy is improved. Finally, the seal recognition results are displayed on the instrument display, enabling real-time, rapid, and highly accurate seal recognition, saving time and effort while also reducing human labor.
[0043] Based on the RA-CNN network, this paper proposes a fast, real-time seal detection method. The overall training process of this method is as follows: the input image is extracted and classified using the MobileNetV2 classification network. The APN (attention proposal network) is then trained based on the extracted features to obtain attention region information. This attention region is then cropped and amplified, and then used as input to a second scale network. This process is repeated three times to obtain the output results of the three scale networks. By fusing the results of the different scale networks, better results are achieved. Using this network has the following advantages:
[0044] 1. The image processing technology based on the RA-CNN neural network proposed in this invention has an overall detection speed fluctuating between 7 and 8 milliseconds per image, which means that the network can be used for real-time detection (<4s), meeting the needs of customs seal recognition in real life.
[0045] 2. Compared to other neural network methods, the image processing technology based on the RA-CNN neural network proposed in this paper has strong generalization, high accuracy, and fast detection speed. Compared to traditional networks such as VGG, it has higher accuracy. Compared to the YOLO series, while its recognition speed is slower, it has higher generalization and accuracy. Compared to some sound processing networks such as Transform, it has greater applicability. Therefore, this technology can process seals while maintaining accuracy.
[0046] 3. The image processing technology based on the RA-CNN neural network proposed in this invention utilizes the MobileNetV2 classification network. It is a lightweight CNN neural network focused on mobile devices and embedded devices. Compared with traditional CNN networks, it uses depthwise separable convolution to greatly reduce the number of parameters, greatly reducing the model calculation amount while slightly reducing the accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0047] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.
[0048] Figure 1 This is a random rotation example image;
[0049] Figure 2 This is a translation example diagram;
[0050] Figure 3 This is a scaling example diagram;
[0051] Figure 4 For comprehensive use example diagram;
[0052] Figure 5 It is the RA-CNN network structure;
[0053] Figure 6 Flowchart for the implementation of the present invention. DETAILED DESCRIPTION
[0054] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.
[0055] In the description of the present invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings and are intended solely to facilitate and simplify the description of the present invention. They are not intended to indicate or imply that the devices or components referred to must have, be constructed, or operate in a specific orientation, and therefore should not be construed as limitations on the present invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.
[0056] In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "mounted," "connected," and "connected" should be understood in a broad sense. For example, they may refer to fixed, detachable, or integral connections; mechanical or electrical connections; direct or indirect connections through an intermediate medium; and internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on the specific circumstances.
[0057] In addition, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.
[0058] The present invention is described in detail below with reference to specific embodiments. Specific embodiment one:
[0060] according to Figures 1 to 6 As shown, the specific optimization technical solution adopted by the present invention to solve the above technical problems is: the present invention relates to a seal recognition system and recognition method based on neural network.
[0061] The present invention provides a seal recognition method based on a neural network, the method comprising the following steps:
[0062] Step 1: Label the seal image data and increase the data volume through data augmentation to obtain sample image data;
[0063] Step 2: Based on the sample image data after data augmentation, training is performed to obtain training data, and the seal image features are extracted and classified;
[0064] Step 3: Based on the obtained seal image features, the APN network is used to obtain the attention area information;
[0065] Step 4: Crop the attention area to obtain the cropped image;
[0066] Step 5: Repeat steps 2 to 4 three times to obtain the output results of the three-scale networks. By adding the loss functions of the three-scale networks as the loss function of the result, matching it with the results in the training set, and predicting the image to obtain the seal recognition prediction result. Specific embodiment two:
[0068] The difference between the second embodiment of the present application and the first embodiment is that:
[0069] The step 1 specifically includes: labeling the seal image data and performing data augmentation, by making a series of random changes to the training images to generate similar but different training samples, thereby expanding the scale of the training data set;
[0070] Data augmentation specifically includes random rotation; random translation along the horizontal or vertical direction with a percentage of the image size as the variation range; random scaling of the image size in proportion; and a combination of the above methods. The present invention uses four data augmentation methods, namely: random rotation (such as Figure 1 ); Randomly translate the image horizontally or vertically with a percentage of the image size as the range of variation (such as Figure 2 ); randomly scale the image size proportionally (e.g. Figure 3 ); Combine the above three methods (such as Figure 4 ). Specific embodiment three:
[0072] The only difference between the third embodiment of the present application and the second embodiment is that:
[0073] The step 2 is specifically as follows:
[0074] Use the MobileNetV2 classification network to extract features and classify sample images to obtain image features;
[0075] The above feature extraction operation is: using Depth-wise Separable Convolution, that is, using PW (Point-wise) convolution - DW (Depth-wise) convolution - PW (Point-wise) convolution to extract features. Specific embodiment four:
[0077] The only difference between the fourth embodiment of the present application and the third embodiment is that:
[0078] The step 3 is specifically as follows:
[0079] Based on the extracted features, the APN network is used to obtain the attention area information. The extracted image is input into a pre-trained convolutional layer to extract region-based deep features. The network is modeled as a multi-task with two outputs at each scale. The first task is to generate a probability distribution over fine-grained categories, and the second task is to predict the coordinates of a set of participating regions for the next scale. Specific embodiment five:
[0081] The only difference between the fifth embodiment of the present application and the fourth embodiment is that:
[0082] The step 4 is specifically as follows:
[0083] Cropping is achieved by multiplying the original image at a coarse scale with an attention mask. The cropping function in the RA-CNN is a differentiable logistic function, which is then backpropagated to complete end-to-end training. Traditional cropping methods directly extract the desired cropped area from the original image using the known coordinates of the top-left and bottom-right corners of the cropped area. While this method is simple and computationally inefficient, it cannot reversely compute gradients during training, making it difficult to train.
[0084] Although the participating areas have been localized, it is sometimes still difficult to extract effective feature representations from highly localized areas. The cropped areas are enlarged to a larger size and an upsampling method with bilinear interpolation as the core is used to implement it using the torch.nn.functional function under the pytorch framework. Specific embodiment six:
[0086] The only difference between the sixth embodiment of the present application and the fifth embodiment is that:
[0087] The step 5 is specifically as follows:
[0088] The attention region is cropped and enlarged, and then used as the input of the second network;
[0089] Cropping is achieved by multiplying the original image with the attention mask at a coarse scale. The function of the cropping method in RA-CNN is a differentiable logistic function, and backpropagation is performed to complete end-to-end training;
[0090] The cropped area is enlarged to a larger size using an upsampling method based on bilinear interpolation and implemented using the torch.nn.functional function in the pytorch framework. Specific embodiment seven:
[0092] The only difference between the seventh embodiment of the present application and the sixth embodiment is that:
[0093] The neural network model in the Jetson Nano device predicts the acquired image to obtain a seal recognition prediction result, and displays the nationality name and the national flag image corresponding to the result on the display unit; the processed information is transmitted to the storage unit for storage and used to record data. Specific embodiment eight:
[0095] The only difference between the eighth embodiment of the present application and the seventh embodiment is that:
[0096] The present invention provides a seal recognition system based on a neural network, the system comprising:
[0097] a marking module, which marks the seal image data, increases the data volume through a data augmentation method, and obtains sample image data;
[0098] A classification module is used to train the sample image data after data augmentation to obtain training data, extract seal image features and classify them;
[0099] APN network module, the labeling module uses the APN network to obtain attention area information according to the obtained seal image features;
[0100] a cropping module, wherein the cropping module crops the attention area to obtain a cropped image;
[0101] The prediction module obtains the output results of the three scale networks, adds the loss functions of the three scale networks as the loss function of the result, matches it with the results in the training set, and predicts the image to obtain the seal recognition prediction result. Specific embodiment nine:
[0103] The only difference between the ninth embodiment of the present application and the eighth embodiment is that:
[0104] The present invention provides a computer-readable storage medium on which a computer program is stored. The program is executed by a processor to implement a seal recognition method based on a neural network. Specific embodiment ten:
[0106] The only difference between the tenth embodiment of the present application and the ninth embodiment is that:
[0107] The present invention provides a computer device, comprising a memory and a processor, wherein a computer program is stored in the memory. When the processor runs the computer program stored in the memory, the processor executes a seal recognition method based on a neural network. Specific embodiment eleven:
[0109] The only difference between the eleventh embodiment of the present application and the tenth embodiment is that:
[0110] The present invention provides a seal recognition device, which includes a data acquisition unit, a processing unit, a display unit and at least one storage unit. In the data acquisition unit, the required processing image is obtained through two methods and sent to the processing unit for processing, that is, an edge recognition device with trained network model weights is stored. After passing through the processing unit, the classification result is obtained and displayed on the display unit. This device specifically inputs image data through the following two methods: first, by directly uploading the seal image; second, by obtaining real-time image data through a camera. Appropriate methods are adopted in two different application scenarios. For example, during customs inspection, when checking a passport, the passport seal is placed under a fixed camera, and the camera obtains specific image data and then transmits the data to the neural network model.
[0111] During customs inspection, the passport stamp is placed under a fixed camera when checking the passport. The camera obtains the specific data of the image and then transmits the data to the neural network model.
[0112] The specific steps of the seal recognition process are:
[0113] Step 1: Data augmentation is performed by making a series of random changes to the training images to generate similar but different training samples, thereby expanding the size of the training data set. Randomly changing the training samples can reduce the model's dependence on certain attributes, thereby improving the model's generalization ability. The present invention uses four data augmentation methods, namely: random rotation (such as Figure 1 ); Randomly translate the image horizontally or vertically with a percentage of the image size as the range of variation (such as Figure 2 ); randomly scale the image size proportionally (e.g. Figure 3 ); Combine the above three methods (such as Figure 4 ).
[0114] Step 2: Input the sample image data after data augmentation;
[0115] Step 3: Use the MobileNetV2 classification network to extract features and classify the sample image to obtain image features;
[0116] The above feature extraction operation is: using Depth-wise Separable Convolution, that is, using PW (Point-wise) convolution - DW (Depth-wise) convolution - PW (Point-wise) convolution to extract features.
[0117] Step 4: Based on the extracted features, the APN (attention proposal network) network is used to obtain the attention area information;
[0118] Step 5: Crop the attention region and enlarge it, and then use it as the input of the second network;
[0119] Step 6: Repeat steps 3 to 5 three times to obtain the output results of the three-scale network. By adding the loss functions of the three-scale networks as the loss function of the result, the results of the networks of different scales achieve better results.
[0120] The APN network process in step 4 is as follows:
[0121] First, we extract region-based deep features by feeding the previously extracted image into a pre-trained convolutional layer. Furthermore, we model the network as a multi-task formulation with two outputs at each scale. The first task aims to generate a probability distribution over fine-grained categories, while the second task predicts the coordinates of a set of participating regions at the next finer scale.
[0122] The cropping operation described in step 5 is:
[0123] This is achieved by multiplying the original image with the attention mask at a coarse scale. Traditional cropping methods use known coordinates of the top-left and bottom-right corners of the cropped region to directly extract the desired cropped region from the original image. While this method is simple and computationally inefficient, it cannot back-calculate gradients during training, making it difficult to train. However, the cropping function in RA-CNN is a differentiable logistic function, allowing for backpropagation and end-to-end training.
[0124] Although the participating regions have been localized, it is sometimes still difficult to extract effective feature representations from highly localized regions. Therefore, we further adaptively enlarge the cropped regions to a larger size. This operation can be implemented using an upsampling method based on bilinear interpolation, using the torch.nn.functional function in the PyTorch framework.
[0125] Furthermore, the loss function in step six is designed as follows
[0126]
[0127] Among them L cls is the commonly used cross entropy loss function, that is
[0128] L cls (Y (s) ,Y * )=CE(Y (s) ,Y * )
[0129] This item is used to optimize the parameters at each scale. rankIt is the PairwiseRanking Loss function.
[0130]
[0131] The role of this item is to use the previous scale network as a reference for the next scale network, forcing the next scale network to locate the most discriminative area and produce a higher prediction value.
[0132] The neural network model in the Jetson Nano device predicts the seal recognition result based on the acquired image and displays the corresponding nationality name and national flag image on the display unit. On the other hand, the processed information is transmitted to the storage unit for storage and recording.
[0133] The method of obtaining images and their specific data is changed from obtaining through the camera to directly uploading, and then steps 1 to 6 are also performed.
[0134] The above is only a preferred embodiment of a neural network-based seal recognition system and recognition method. The scope of protection of a neural network-based seal recognition system and recognition method is not limited to the above embodiment. All technical solutions based on this concept fall within the scope of protection of the present invention. It should be pointed out that for those skilled in the art, various improvements and variations without departing from the principles of the present invention should also be considered as the scope of protection of the present invention.
Claims
1. A seal recognition method based on neural network, characterized in that: The method comprises the following steps: Step 1: Annotate the seal image data, increase the data volume through data augmentation method, and obtain sample image data; Step 2: Perform training based on the sample image data after data augmentation to obtain a training set, extract seal image features and classify them; The step 2 is specifically as follows: Use MobileNetV2 classification network to extract features and classify sample images to obtain image features; The above feature extraction operation is: using Depth-wise Separable Convolution, that is, using PW convolution, DW convolution, PW convolution to extract features; Step 3: According to the obtained seal image features, the APN network is used to obtain the attention area information; The step 3 is specifically as follows: According to the extracted features, the APN network is used to obtain the attention area information. The extracted image is input into the pre-trained convolutional layer to extract the region-based deep features. The network is modeled as a multi-task with two outputs at each scale. The first task is to generate the probability distribution on the fine-grained category, and the second task is to predict the coordinates of a set of participating regions for the next scale. Step 4: Crop the attention area to obtain a cropped image; The step 4 is specifically as follows: Cropping is achieved by multiplying the original image with the attention mask at a coarse scale. The function of the cropping method in RA-CNN is a differentiable logistic function, and back propagation is performed to complete end-to-end training; the cropped area is enlarged to a larger size, and an upsampling method with bilinear interpolation as the core is used, which is implemented using the torch.nn.functional function under the pytorch framework; Step 5: Repeat steps 2 to 4 three times to obtain the output results of the three-scale networks. By adding the loss functions of the three-scale networks as the loss function of the result, matching it with the results in the training set, and predicting the image to obtain the seal recognition prediction result.
2. A seal recognition method based on neural network according to claim 1, characterized in that: The step 1 specifically includes: annotating the seal image data and performing data augmentation, by making a series of random changes to the training images to generate similar but different training samples, thereby expanding the scale of the training data set; Data augmentation specifically includes random rotation; random translation along the horizontal or vertical direction with a percentage of the image size as the variation range; random scaling of the image size in proportion; and a combination of the above methods.
3. A seal recognition method based on neural network according to claim 1, characterized in that: The step 5 is specifically as follows: Cropping is achieved by multiplying the original image with the attention mask at a coarse scale. The function of the cropping method in RA-CNN is a differentiable logistic function, which is back-propagated to complete end-to-end training; The cropped area is enlarged to a larger size, using an upsampling method with bilinear interpolation as the core, and implemented using the torch.nn.functional function under the pytorch framework.
4. A seal recognition method based on neural network according to claim 1, characterized in that: The neural network model in the jetsonNano device predicts the acquired image to obtain a seal recognition prediction result, and displays the nationality name and the national flag image corresponding to the result in the display unit; The processed information is transmitted to the storage unit for storage and used to record data.
5. A seal recognition system based on neural network, characterized by: The system comprises: A labeling module, which labels the seal image data, increases the data volume through a data augmentation method, and obtains sample image data; A classification module, wherein the classification module is trained according to the sample image data after data augmentation to obtain a training set, extract the seal image features and classify them; Use MobileNetV2 classification network to extract features and classify sample images to obtain image features; The above feature extraction operation is: using Depth-wise Separable Convolution, that is, using PW convolution, DW convolution, PW convolution to extract features; APN network module, the annotation module uses the APN network to obtain the attention area information according to the obtained seal image features; According to the extracted features, the APN network is used to obtain the attention area information. The extracted image is input into the pre-trained convolutional layer to extract the region-based deep features. The network is modeled as a multi-task with two outputs at each scale. The first task is to generate the probability distribution on the fine-grained category, and the second task is to predict the coordinates of a set of participating regions for the next scale. A cropping module, wherein the cropping module crops the attention area to obtain a cropped image; Cropping is achieved by multiplying the original image with the attention mask at a coarse scale. The function of the cropping method in RA-CNN is a differentiable logistic function, and back propagation is performed to complete end-to-end training; the cropped area is enlarged to a larger size, and an upsampling method with bilinear interpolation as the core is used, which is implemented using the torch.nn.functional function under the pytorch framework; The prediction module obtains the output results of the three-scale networks, adds the loss functions of the three-scale networks as the loss function of the result, matches it with the results in the training set, and predicts the picture to obtain the seal recognition prediction result.
6. A computer-readable storage medium having a computer program stored thereon, characterized in that: The program is executed by a processor to implement the neural network-based seal recognition method as described in any one of claims 1 to 4.
7. A computer device, characterized in that: It comprises a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the neural network-based seal recognition method according to any one of claims 1-4.
