A lightweight target recognition method for UAV inspection

Abstract

The present invention discloses a lightweight method for identifying targets during drone inspections, comprising: preprocessing of sample images, including: removing the color information of the images, ensuring that the images are of the same size, and expanding the amount of sample data; calculating texture feature vectors and contour feature vectors, and splicing them to obtain feature vectors corresponding to the sample images; marking the types of the sample images, and using the feature vectors corresponding to the sample images as input, and the types as results to be input into a convolutional neural network for training; when preset training conditions are met, pruning the convolutional neural network and the contour feature vectors; for real-time images collected during drone inspections, calculating the feature vectors corresponding to the pruned real-time images, and substituting them into the pruned convolutional neural network to predict their types. The present invention reduces the number of parameters and computational complexity of the neural network. While maintaining recognition accuracy, it enables efficient operation on resource-constrained edge devices.

Description

Technical Field

[0001] The present invention belongs to the field of image processing target recognition, and more specifically, relates to a lightweight unmanned aerial vehicle inspection target recognition method. Background Art

[0002] With the continuous advancement of science and technology and the rapid development of drone technology, drones have been widely used in a variety of fields, including power inspection, agricultural monitoring, and disaster response. In the field of power inspection, drones, with their flexibility and efficiency, can quickly acquire image and video data from power grid equipment, greatly improving inspection efficiency and quality. However, in practical applications, achieving efficient and accurate identification of inspection targets still faces many challenges, especially with limited edge computing resources. Currently, drone inspection target recognition relies primarily on deep learning technology, particularly the application of convolutional neural networks (CNNs) in image recognition. By training large amounts of power grid equipment image data and building high-performance deep learning models, drones can automatically identify and classify inspection targets during flight. These models are typically trained in high-performance computing environments and deployed on drones for real-time inference.

[0003] To improve recognition accuracy and speed, researchers are continuously exploring new network structures and optimization algorithms, such as ResNet, YOLO, and EfficientNet, to enhance model detection performance and computational efficiency. Simultaneously, model compression techniques, such as pruning, quantization, and distillation, are increasingly being applied to drone inspection target recognition to reduce model size and computational overhead. However, existing technologies still suffer from several major drawbacks: High model complexity: Current deep learning models typically have a large number of parameters and complex structures, resulting in significant consumption of computational and storage resources. This makes efficient and real-time target recognition difficult for drones with limited computing power. Slow inference speed: The complex model structure and numerous computational operations result in slow inference speed, making it unsuitable for real-time inspections in practical applications, particularly in high-speed flight or large-scale inspection scenarios. High energy consumption: Complex models require high computational resources, increasing drone energy consumption, shortening flight time, and limiting the sustainability and scope of inspection missions.

[0004] Therefore, a lightweight target recognition method for drone inspections is urgently needed. By reducing the number of model parameters and computational complexity, it can achieve efficient and accurate target recognition while meeting the requirements of drone edge computing environments. This method can not only improve the efficiency and quality of drone inspections, but also reduce energy consumption and extend flight time, providing more reliable technical support for power grid inspections. Summary of the Invention

[0005] In order to address the deficiencies in the prior art, the purpose of the present invention is to solve the above-mentioned defects and further propose a lightweight UAV inspection target recognition method.

[0006] The present invention adopts the following technical solutions.

[0007] The first aspect of the present invention discloses a lightweight UAV inspection target recognition method, comprising steps 1 to 5;

[0008] Step 1: Preprocessing of sample images, including: removing color information from the images, ensuring image size consistency, and expanding the sample data volume;

[0009] Step 2: Calculate the texture feature vector and the contour feature vector, and concatenate them to obtain the feature vector corresponding to the sample image;

[0010] Step 3: Mark the type of the sample image, and use the feature vector corresponding to the sample image as input, and the type as the result to input into the convolutional neural network for training;

[0011] Step 4: When the preset training conditions are met, the convolutional neural network and the contour feature vector are pruned;

[0012] Step 5: For the real-time images collected during the drone inspection process, calculate the feature vector corresponding to the pruned real-time image and substitute it into the pruned convolutional neural network to predict its type.

[0013] Specifically, in step 1, removing the color information of the image specifically includes:

[0014] Grayscale information p of the sample image ij As shown below:

[0015] p ij =(0.299R+0.587G+0.114B) / p max

[0016] Among them, p ij Represents the grayscale information at position (i, j) in the sample image, p max Indicates the maximum value in the grayscale information.

[0017] Specifically, the calculation method of the texture feature vector includes:

[0018] Calculate the Gabor function G k (x,y), as shown below:

[0019]

[0020] The value range of x and y is [1, s], where s is the window size; λ, φ, σ, and γ represent the wavelength, phase offset, standard deviation, and aspect ratio, respectively, and their value ranges are {0.5, 1, 1.5, 2}, {0°, 90°}, {0.5, 1, 1.5, 2}, and {1, 2}.

[0021] Calculate the texture feature matrix T k , as shown below:

[0022]

[0023] Among them, T k (x,y) represents the texture feature matrix T k The element in row x and column y in ij is the grayscale information of the sample image;

[0024] The texture feature vector T is calculated based on the texture feature matrix, as shown in the following formula:

[0025]

[0026] Where T[k] represents the kth element in the texture feature vector T.

[0027] Specifically, the calculation method of the contour feature vector includes:

[0028] Calculate the gradient matrix D as shown below:

[0029] D(i,j)=[abs(p ij -p (i-1)j )+abs(p ij -p i(j-1) )] / D max

[0030] Where D(i,j) represents the element in the i-th row and j-th column of the gradient matrix D. max is the maximum value among all D(i,j), p ij is the grayscale information of the sample image;

[0031] Binarize the gradient matrix to obtain the edge matrix B, as shown below:

[0032]

[0033] Where B(i,j) represents the element in the i-th row and j-th column of the edge matrix B, T is the preset binarization threshold, and when i or j is 1, B(i,j) is equal to 0;

[0034] Extract the geometry matrix M based on the edge matrix k The formula is as follows:

[0035]

[0036] Among them, M k (x,y) represents the geometric matrix M k The element in the xth row and yth column, K is an even number, and K is greater than or equal to 32, and s is the window size;

[0037] Calculate the contour feature matrix C k , as shown below:

[0038]

[0039] Among them, C k (x,y) represents the contour feature matrix C k The element at row x and column y in ;

[0040] Based on the contour feature matrix, the contour feature vector C is calculated as shown below:

[0041]

[0042] Where C[k] represents the kth element in the contour feature vector C.

[0043] Specifically, the feature vector TC corresponding to the sample image is a vector with a dimension of 2K+64, as shown in the following formula:

[0044]

[0045] Among them, TC[k] represents the kth element in the feature vector TC corresponding to the sample image, T[k] and C[k-64] are the texture feature vector and contour feature vector respectively, and 2K represents the dimension of the contour feature vector.

[0046] Specifically, the preset training condition is that when the number of training sample images increases to a certain value, when the increase or decrease in the value of each element in the weight matrix is ​​less than a preset training threshold, it is determined that the preset training condition is met.

[0047] Specifically, the influence factor A(k) of the kth element in the contour feature vector C is calculated as shown in the following formula:

[0048]

[0049] Where R and R′ represent the classification vector of the final output of the convolutional neural network when TC is input, and the classification vector of the final output of the convolutional neural network when TC[64+k] is equal to 0;

[0050] If the impact factor is less than the preset deletion threshold, the kth element in the contour feature vector C is deleted, otherwise it is not deleted.

[0051] The second aspect of the present invention discloses a lightweight UAV inspection target recognition system for executing the method described in the first aspect, the system comprising: an image processing module, a logic calculation module, and a convolutional neural network module;

[0052] The image processing module is used to pre-process the sample images, including: removing the color information of the image, ensuring the image size is consistent, and expanding the sample data volume;

[0053] The logic calculation module is used to calculate the texture feature vector and the contour feature vector, and to concatenate them to obtain the feature vector corresponding to the sample image; and when the preset training conditions are met, the convolutional neural network and the contour feature vector are pruned;

[0054] The convolutional neural network module is used to mark the type of sample images, and take the feature vector corresponding to the sample image as input, and the type as the result into the convolutional neural network for training; and for the real-time images collected during the drone inspection process, the feature vector corresponding to the pruned real-time image is calculated and substituted into the pruned convolutional neural network to predict its type.

[0055] A third aspect of the present invention discloses a terminal, comprising a processor and a storage medium, characterized in that:

[0056] The storage medium is used to store instructions;

[0057] The processor is configured to operate according to the instructions to execute the steps of the method of the first aspect.

[0058] A fourth aspect of the present invention discloses a computer-readable storage medium having a computer program stored thereon, wherein the program implements the steps of the method described in the first aspect when executed by a processor.

[0059] The beneficial effects of the present invention are that, compared with the prior art, the present invention has the following advantages:

[0060] By extracting the texture feature vector and contour feature vector of the image as the input of the convolutional neural network, the computational complexity of real-time image feature extraction can be reduced. At the same time, when the convolutional neural network training is more complete, weight pruning can be used to prune invalid neurons and invalid elements in the contour feature vector, thereby making the model lightweight. This invention significantly reduces the number of parameters and computational complexity of the neural network. While maintaining recognition accuracy, the model volume is greatly compressed, enabling it to run efficiently on resource-constrained edge devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 This is a flow chart of a lightweight UAV inspection target recognition method. DETAILED DESCRIPTION

[0062] The present application will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present application.

[0063] As for the target recognition method of drone inspection, it is foreseeable that in the initial stage of drone inspection, a large number of on-site pictures can be collected with the help of drones. Of course, these pictures can be assisted by staff or machines and classified into sample pictures of various distribution equipment, including: normal sample pictures and defective sample pictures.

[0064] As the sample images expand, the backend server can continuously learn these sample images. In this process, the present invention proposes a lightweight drone inspection target recognition method. Its core idea is to improve the operation efficiency of the entire model by strengthening the feature extraction technology of the input data and continuously combining model pruning to perform model paper cutting on the deep learning network. Figure 1 As shown, the target recognition method includes steps 1 to 5.

[0065] Step 1: Preprocessing of sample images, including: removing color information of the images, ensuring the image size is consistent, and expanding the sample data volume.

[0066] Specifically, each pixel in the sample image obtained by the drone can be represented as [R, G, B]. After being converted into a grayscale image, the grayscale information p ij To express it, as follows:

[0067] p ij =(0.299R+0.587G+0.114B) / p max

[0068] Among them, p ij Represents the grayscale information at position (i, j) in the sample image, which is used to remove color information, p max Indicates the maximum value in the grayscale information and is used for normalization.

[0069] The size of all images can be adjusted to a uniform size, for example, 128*128. In this case, each sample image corresponds to a grayscale matrix P, as shown below:

[0070]

[0071] Here, n is a uniform size, that is, n is the size of the image. In the embodiment of the present invention, n may be equal to 128.

[0072] In some embodiments, a sample image can be converted into multiple sample images by random cropping and rotation.

[0073] Step 2: Calculate the texture feature vector and the contour feature vector, and concatenate them to obtain the feature vector corresponding to the sample image.

[0074] Texture features are used to extract texture information from sample images. In some embodiments, texture features can be extracted using a Gabor filter.

[0075] First, calculate the Gabor function G k (x,y), as shown below:

[0076]

[0077] The values ​​of x and y both range from [1, s], where s is the window size and must be divisible by n. λ, φ, σ, and γ represent the wavelength, phase offset, standard deviation, and aspect ratio, respectively, and range from {0.5, 1, 1.5, 2}, {0°, 90°}, {0.5, 1, 1.5, 2}, and {1, 2}. Therefore, there are 64 different combinations, i.e., k = 1, 2, …, 64.

[0078] Secondly, calculate the texture feature matrix T k , as shown below:

[0079]

[0080] Among them, T k (x,y) represents the texture feature matrix T k The element in row x and column y in uv Represents the grayscale information at the (u, v) position in the sample image.

[0081] In the present invention, s can be set to 16. Since the value of n is 128, each sample image is exactly divided into 8*8 windows, that is, the number of texture feature matrices is 64.

[0082] Finally, the texture feature vector T is calculated based on the texture feature matrix, as shown in the following formula:

[0083]

[0084] Among them, T[k] represents the kth element in the texture feature vector T, so the texture feature vector T is a vector with a dimension of 64.

[0085] Contour features are used to focus on the shape, size, and position information in an image.

[0086] First, the gradient matrix D is calculated and then normalized as shown below:

[0087] D(i,j)=[abs(p ij -p (i-1)j )+abs(p ij -p i(j-1) )] / D max

[0088] Where D(i,j) represents the element in the i-th row and j-th column of the gradient matrix D. max is the maximum value among all D(i,j). It is understandable that in D(i,j), the value range of i,j is [2,n].

[0089] It should be noted that, like the formula for calculating grayscale information, for convenience, the process of calculating grayscale information or gradient matrix is ​​combined with the normalization process. However, in fact, there is a sequence relationship, and the formula is only for simplifying the description.

[0090] Secondly, the gradient matrix is ​​binarized to obtain the edge matrix B, as shown below:

[0091]

[0092] Where B(i,j) represents the element in the i-th row and j-th column of the edge matrix B, and T is the preset binarization threshold, which can be 0.5. When i or j is 1, B(i,j) is assumed to be 0.

[0093] It is understandable that the edge matrix is ​​used to obtain the outline of the image.

[0094] Furthermore, the geometric matrix M is extracted according to the edge matrix k The formula is as follows:

[0095]

[0096]

[0097] Among them, M k (x,y) represents the geometric matrix M k The element in row x and column y in M ​​is understandable. k The value range of i and j in (i,j) is also [1,s]. In addition, f k (x, y) includes K straight line functions and K circular functions, which can be understood as follows:

[0098] y=x·tanθ

[0099] x 2 +y 2 =r 2

[0100] Therefore, f k (x, y) essentially takes K equally spaced values ​​of the parameter θ in the linear function, and takes K equally spaced values ​​of the parameter r in the circular function. The ranges of θ and r are (-90°, 90°) and (0, s / 2), respectively. In the embodiment of the present invention, K should be an even number, and K should be greater than or equal to 32.

[0101] 2K represents the dimension of the contour feature vector. The larger the value, the richer the information, but the slower the training time.

[0102] Next, calculate the contour feature matrix C k , as shown below:

[0103]

[0104] Among them, C k (x,y) represents the contour feature matrix C k The element in the xth row and yth column in , B(u,v) represents the element in the uth row and vth column in the edge matrix B.

[0105] Finally, based on the contour feature matrix, the contour feature vector C is calculated as shown below:

[0106]

[0107] Among them, C[k] represents the kth element in the contour feature vector C. Therefore, the contour feature vector C is a vector with a dimension of 2K.

[0108] Therefore, the feature vector TC corresponding to the sample image is a vector with a dimension of 2K+64, as shown in the following formula:

[0109]

[0110] Among them, TC[k] represents the kth element in the feature vector TC corresponding to the sample image, and C[k-64] represents the k-64th elements in the contour feature vector C.

[0111] Step 3: Mark the type of the sample image, and use the feature vector corresponding to the sample image as input, and the type as the result to input into the convolutional neural network for training.

[0112] In embodiments of the present invention, convolutional neural networks can be selected from Vision Transformer, DenseNet, ResNet, AlexNet, and others. In fact, any convolutional neural network should generally include: a convolutional layer, a feature extraction layer, a pooling layer, and a fully connected layer. The convolutional layer is used to convolve the feature vector; the feature extraction layer is used to extract feature data from the convolution result; the pooling layer is used to calculate the feature information corresponding to each sample image from the feature data; and the fully connected layer outputs a classification vector based on the feature information.

[0113] In the embodiment of the present invention, since the sample data will continue to expand during the inspection process, a Vision Transformer or DenseNet with complex neurons should be selected as the convolutional neural network.

[0114] Step 4: When the preset training conditions are met, the convolutional neural network and the contour feature vector are pruned.

[0115] For convolutional neural networks, model pruning is a technique that reduces the number of neural network parameters and computational complexity by removing redundant neurons and connections. The core idea is to reduce model complexity by removing unimportant components while minimizing performance. Common model pruning methods include weight pruning and structural pruning.

[0116] Weight pruning simplifies the model by removing connections with smaller weights. The pruning strategy can be summarized as follows: sort the weights of all neurons according to their importance and select the weights with the lowest importance (e.g., 0.5-0.7) for pruning.

[0117] In convolutional neural networks, convolutional layers, pooling layers, and fully connected layers all contain a large number of neurons. More generally, the neural network should be viewed as a layer L containing m neurons. m With a layer L containing n neurons n Fully connected, layer L m The generated output is input to layer L through the weight matrix V n As you can see, the weight matrix V is an m×n matrix. Therefore, weight pruning essentially aims to directly set smaller elements in the weight matrix V to zero whenever possible, thereby speeding up the computation. From a computational perspective, the feature vector corresponding to the sample image can itself be considered a neuron in the first layer.

[0118] In some embodiments, the preset training condition may be that when the number of sample images for training increases to a certain value, the increase or decrease in the value of each element in the weight matrix is ​​less than a preset training threshold, and it is determined that the preset training condition is met.

[0119] When a large number of elements in the weight matrix are 0, for example, when the elements in the k-th column of the weight matrix of the first layer are almost all 0, it can be understood that this means that TC[k] has almost no effect. Therefore, it can also be used to reduce the dimension of the feature vector corresponding to the sample image.

[0120] When the kth element in the contour feature vector C is deleted (i.e., TC[64+k] or C[k]), its corresponding impact factor A(k) needs to be calculated through forward propagation, as shown in the following formula:

[0121]

[0122] Here, R and R′ represent the classification vector of the final output of the convolutional neural network when TC is input, and the classification vector of the final output of the convolutional neural network when TC[64+k] is equal to 0; TC[64+k] represents the 64+kth element in the feature vector TC corresponding to the sample image.

[0123] It is understandable that if the impact factor is less than the preset deletion threshold, the kth element in the contour feature vector C is deleted, otherwise it is not deleted. The deletion threshold can be set to be less than or equal to 0.01×A m , where A m is the average impact factor of all elements.

[0124] However, it should be noted that in the feature vector corresponding to the sample image, the dimension of the texture feature vector is usually not reduced, but the dimension of the contour feature vector is more reduced.

[0125] Step 5: For the real-time images collected during the drone inspection process, calculate the feature vector corresponding to the pruned real-time image and substitute it into the pruned convolutional neural network to predict its type.

[0126] Correspondingly, the present invention also discloses a lightweight UAV inspection target recognition system, comprising: an image processing module, a logic calculation module, and a convolutional neural network module;

[0127] The image processing module is used to pre-process the sample images, including: removing the color information of the image, ensuring the image size is consistent, and expanding the sample data volume;

[0128] The logic calculation module is used to calculate the texture feature vector and the contour feature vector, and to concatenate them to obtain the feature vector corresponding to the sample image; and when the preset training conditions are met, the convolutional neural network and the contour feature vector are pruned;

[0129] The convolutional neural network module is used to mark the type of sample images, and take the feature vector corresponding to the sample image as input, and the type as the result into the convolutional neural network for training; and for the real-time images collected during the drone inspection process, the feature vector corresponding to the pruned real-time image is calculated and substituted into the pruned convolutional neural network to predict its type.

[0130] The applicant of the present invention has made a detailed explanation and description of the implementation examples of the present invention in conjunction with the drawings in the specification. However, those skilled in the art should understand that the above implementation examples are only preferred implementation plans of the present invention, and the detailed description is only to help readers better understand the spirit of the present invention, and is not a limitation on the scope of protection of the present invention. On the contrary, any improvements or modifications based on the inventive spirit of the present invention should fall within the scope of protection of the present invention.

Claims

1. A lightweight UAV inspection target recognition method, characterized in that: The method comprises steps 1 to 5; Step 1: Preprocessing of sample images, including: removing color information from the images, ensuring image size consistency, and expanding the sample data volume; Step 2: Calculate the texture feature vector and the contour feature vector, and concatenate them to obtain the feature vector corresponding to the sample image; Step 3: Mark the type of the sample image, and use the feature vector corresponding to the sample image as input, and the type as the result to input into the convolutional neural network for training; Step 4: When the preset training conditions are met, the convolutional neural network and the contour feature vector are pruned; Step 5: For the real-time images collected during the drone inspection process, calculate the feature vector corresponding to the pruned real-time image and substitute it into the pruned convolutional neural network to predict its type; Calculate the influence factor A(k) of the kth element in the contour feature vector C as shown below: Where R and R′ represent the classification vector of the final output of the convolutional neural network when TC is input, and the classification vector of the final output of the convolutional neural network when TC[64+k] is equal to 0, respectively. C[k] represents the kth element in the contour feature vector C. Among them, the feature vector TC corresponding to the sample image is a vector with a dimension of 2K+64, and TC is shown as follows: Wherein, TC[k] represents the kth element in the feature vector TC corresponding to the sample image, TC[64+k] represents the 64+kth element in the feature vector TC corresponding to the sample image, wherein, T[k] represents the kth element in the texture feature vector T, C[k-64] represents the k-64th elements in the contour feature vector C, and 2K represents the dimension of the contour feature vector; If the impact factor is less than the preset deletion threshold, the kth element in the contour feature vector C is deleted, otherwise it is not deleted.

2. A lightweight UAV inspection target recognition method according to claim 1, characterized in that: In step 1, removing the color information of the image specifically includes: Grayscale information p of the sample image ij As shown below: p ij =(0.299R+0.587G+0.114B) / p max Among them, p ij Represents the grayscale information at position (i, j) in the sample image, p max Indicates the maximum value in the grayscale information.

3. A lightweight UAV inspection target recognition method according to claim 1, characterized in that: The calculation method of texture feature vector includes: Calculate the Gabor function G k (x,y), as shown below: The value range of x and y is [1, s], where s is the window size; λ, φ, σ, and γ represent the wavelength, phase offset, standard deviation, and aspect ratio, respectively, and their value ranges are {0.5, 1, 1.5, 2}, {0°, 90°}, {0.5, 1, 1.5, 2}, and {1, 2}. Calculate the texture feature matrix T k , as shown below: Among them, T k (x,y) represents the texture feature matrix T k The element in row x and column y in ij is the grayscale information of the sample image, p uv Represents the grayscale information at the (u, v) position in the sample image; The texture feature vector T is calculated based on the texture feature matrix, as shown in the following formula: Where T[k] represents the kth element in the texture feature vector T, and n is the size of the image.

4. A lightweight UAV inspection target recognition method according to claim 1, characterized in that: The calculation method of the contour feature vector includes: Calculate the gradient matrix D as shown below: D(i,j)=[abs(p ij -p (i-1)j )+abs(p ij -p i(j-1) )] / D max Where D(i,j) represents the element in the i-th row and j-th column of the gradient matrix D. max is the maximum value among all D(i,j), p ij is the grayscale information of the sample image; Binarize the gradient matrix to obtain the edge matrix B, as shown below: Where B(i,j) represents the element in the i-th row and j-th column of the edge matrix B, T is the preset binarization threshold, and when i or j is 1, B(i,j) is equal to 0; Extract the geometry matrix M based on the edge matrix k The formula is as follows: Among them, M k (x,y) represents the geometric matrix M k The element in the xth row and yth column, K is an even number, and K is greater than or equal to 32, and s is the window size; Calculate the contour feature matrix C k , as shown below: Among them, C k (x,y) represents the contour feature matrix C k The element in the xth row and yth column in the edge matrix B, B(u,v) represents the element in the uth row and vth column in the edge matrix B; Based on the contour feature matrix, the contour feature vector C is calculated as shown below: Where C[k] represents the kth element in the contour feature vector C, and n is the size of the image.

5. The lightweight UAV inspection target recognition method according to claim 1 is characterized in that: The preset training condition is that when the number of sample images for training increases to a certain value, and the increase or decrease in the value of each element in the weight matrix is ​​less than a preset training threshold, it is determined that the preset training condition is met.

6. A lightweight drone inspection target recognition system for executing the method according to any one of claims 1 to 5, characterized in that: The system includes: an image processing module, a logic calculation module, and a convolutional neural network module; The image processing module is used to pre-process the sample images, including: removing the color information of the image, ensuring the image size is consistent, and expanding the sample data volume; The logic calculation module is used to calculate the texture feature vector and the contour feature vector, and to concatenate them to obtain the feature vector corresponding to the sample image; and when the preset training conditions are met, the convolutional neural network and the contour feature vector are pruned; The convolutional neural network module is used to mark the type of sample images, and take the feature vector corresponding to the sample image as input, and the type as the result into the convolutional neural network for training; and for the real-time images collected during the drone inspection process, the feature vector corresponding to the pruned real-time image is calculated and substituted into the pruned convolutional neural network to predict its type.

7. A terminal comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images