A red tide outbreak early warning method and system based on weak target detection

Abstract

The present application relates to the technical field of red tide early warning, in particular to a red tide outbreak early warning method and system based on weak target detection. The method comprises the following steps: constructing a weak target detection enhancement model based on a Yolov8 structure, and performing feature extraction on red tide data set images by using the weak target detection enhancement model based on the Yolov8 structure; performing feature fusion on the extracted features; performing target detection based on the fused features, incrementally training the weak target detection enhancement model based on the target detection result; and performing red tide early warning by using the trained weak target detection enhancement model. The model proposed in the present application has a recognition accuracy of 91.1% for red tide outbreak period images and a recognition accuracy of 84.3% for red tide early stage images, and can effectively recognize red tide images at different stages.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of red tide early warning, in particular to a red tide outbreak early warning method and system based on weak target detection. BACKGROUND

[0002] Red tide is a marine ecological disaster caused by the abnormal proliferation of phytoplankton, and its detection and early warning are crucial for fishery protection and marine ecological safety. Traditional red tide early warning methods usually use satellite remote sensing analysis, on-site sampling detection, and high-altitude visual methods. However, in the early stage of red tide, the distribution of algae is sparse and the color characteristics are weak, and the traditional methods are affected by signal strength, environmental interference, etc., making it difficult to quickly and effectively identify red tide and issue warnings.

[0003] Existing computer vision-based methods can quickly and effectively identify marine anomalies such as red tide, but for the early stage of red tide, the distribution of algae is small, and the existing computer vision-based identification methods need to be improved in terms of weak target recognition accuracy. Therefore, developing a system that can accurately detect weak algal targets in the early stage of red tide and issue warnings has important practical application value. SUMMARY

[0004] To solve the problem that existing computer vision-based methods cannot identify weak algal targets in the early stage of red tide, a small target enhanced target recognition model based on Yolov8 structure is proposed to accurately identify weak algal targets. The color gamut, density, and area of the identified weak algal targets are comprehensively judged, and automatic sample collection technology is used to automatically collect samples in the area for further analysis, and a warning is issued for cases that exceed the threshold.

[0005] In a first aspect, the present application provides a red tide outbreak early warning method based on weak target detection, which adopts the following technical solution:

[0006] A red tide outbreak early warning method based on weak target detection, comprising:

[0007] Obtain red tide dataset images and perform image preprocessing through Gaussian filtering;

[0008] Construct a weak target detection enhancement model based on the Yolov8 structure, and use the weak target detection enhancement model based on the Yolov8 structure to extract features from the red tide dataset images;

[0009] Fuse the extracted features;

[0010] Perform target detection based on the fused features, and perform incremental training of the weak target detection enhancement model based on the target detection results;

[0011] Use the trained weak target detection enhancement model to perform red tide early warning.

[0012] Further, the feature extraction of the red tide data set image by using the weak target detection enhanced model based on the Yolov8 structure comprises wherein C is the number of channels, W is the image width, and H is the image height, the input image is first input into the backbone network to extract feature information, the backbone network gradually reduces the feature map size and expands the channel number from shallow to deep.

[0013] Further, the feature fusion of the extracted features comprises fusing the high-order semantic information finally extracted by the backbone network through the neck network of the model and the low-order features by upsampling, wherein the entire feature map is first compressed to a single feature dimension, the original feature map is compressed to a single feature dimension is expressed as:

[0014]

[0015] wherein and are the length and width of the input feature map.

[0016] Further, the compression of the entire feature map to a single feature dimension comprises using a fully connected layer to first reduce the dimension of the entire feature map times, restoring it to by using a fully connected layer after a nonlinear function, and finally obtaining the inter-channel attention by weighting the original input map again through a Sigmoid function, which is expressed as:

[0017]

[0018] wherein, is a Sigmoid function, and are two weight matrices, is a Relu function.

[0019] Further, the feature fusion of the extracted features also comprises a channel attention mechanism LSE based on the addition of position information, after any convolutional transformation U, the original feature map is compressed from the high direction and the wide direction respectively, and the two compressed matrices are deformed and fused, the position information is extracted by using a convolutional layer, which is expressed as:

[0020]

[0021] wherein, is a ReLU activation function, and the weight matrix , is the reduction rate of the control block size for reducing the computational overhead, and then separated into two independent tensors and , the number of channels is changed back to C by using another two convolution layers, and the final position channel attention is expressed as:

[0022]

[0023] wherein, is a Sigmod activation function, is a segmentation operation, are two weight matrices.

[0024] Further, the target detection based on the fused features comprises constructing a decoupled feature head by using a Yolov8 Header according to the fused features, and respectively calculating a target frame loss and a classification loss, wherein an Anchor-Free structure is adopted, a target center point is used to replace an anchor frame, and a center point and distances between the target frame and the center point are predicted to output a final target frame, a loss function of the target frame loss uses a sum of a CIOU and a DFL loss function, and the classification loss uses a binary cross-entropy function for optimization.

[0025] Further, the incremental training of the weak target detection enhancement model based on the target detection result comprises adopting an incremental training mode to input collected data to a previously trained model for further incremental training, and adopting an EWC mode to alleviate the problem that the model is too inclined to the distribution of a new data set and forgets the distribution of an old data set, wherein an additional EWC loss function is added when the incremental training is performed, and the loss function is expressed as:

[0026]

[0027] wherein, is a loss function trained on a new data set, is a parameter of a new model, is an optimal parameter of an old model, is a Fisher information matrix of the parameter , used to measure the importance of the parameter, and λ is a regularization coefficient, controlling the protection strength of the old task knowledge.

[0028] In a second aspect, a red tide outbreak early warning system based on weak target detection comprises:

[0029] A data acquisition module configured to acquire red tide data set images and perform image preprocessing by Gaussian filtering;

[0030] The feature extraction module is configured to construct a weak target detection enhancement model based on a Yolov8 structure, and perform feature extraction on red tide data set images by using the weak target detection enhancement model based on the Yolov8 structure.

[0031] The feature fusion module is configured to perform feature fusion on the extracted features.

[0032] The target detection module is configured to perform target detection based on the fused features, and perform incremental training on the weak target detection enhancement model based on a target detection result.

[0033] The early warning module is configured to perform red tide early warning by using the trained weak target detection enhancement model.

[0034] In a third aspect, the present application provides a computer readable storage medium, wherein a plurality of instructions are stored, the instructions being adapted to be loaded by a processor of a terminal device and to execute the method.

[0035] In a fourth aspect, the present application provides a terminal device, comprising a processor and a computer readable storage medium, the processor being used to implement instructions, and the computer readable storage medium being used to store a plurality of instructions, the instructions being adapted to be loaded by the processor and to execute the method.

[0036] To sum up, the present application has the following beneficial technical effects:

[0037] 1. The model proposed by the present application has a recognition accuracy of 91.1% for red tide outbreak period images and a recognition accuracy of 84.3% for red tide early stage images, effectively recognizing red tide images at different stages.

[0038] 2. The present application realizes an automatic red tide early warning and monitoring technology, and uses an improved channel attention and an additional Neck link method to solve the problem of small early stage red tide algal targets being difficult to identify, effectively identifying the early stage of red tide. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 FIG. 1 is a red tide early warning process schematic diagram of embodiment 1 of the present application;

[0040] Figure 2 FIG. 2 is a neural network overall structure schematic diagram of embodiment 1 of the present application;

[0041] Figure 3 FIG. 3 is an EPM module structure schematic diagram of embodiment 1 of the present application;

[0042] Figure 4 FIG. 4 is an LSE attention structure schematic diagram of embodiment 1 of the present application. DETAILED DESCRIPTION

[0043] The application will be further described in detail below with reference to the accompanying drawings.

[0044] Embodiment 1

[0045] With reference to Figure 1 , the red tide outbreak early warning method based on weak target detection of the embodiment comprises:

[0046] obtaining a red tide data set image;

[0047] constructing a weak target detection enhancement model based on a Yolov8 structure, and performing feature extraction on the red tide data set image by using the weak target detection enhancement model based on the Yolov8 structure;

[0048] performing feature fusion on the extracted features;

[0049] performing target detection based on the fused features, and performing incremental training on the weak target detection enhancement model based on the target detection result;

[0050] performing red tide early warning by using the trained weak target detection enhancement model.

[0051] Specifically,

[0052] (1) a weak target detection enhancement model based on a Yolov8 structure.

[0053] (1) model structure,

[0054] The model is composed of three parts, namely Backbone, Neck and Detector, which are respectively responsible for feature extraction, feature fusion and target detection.

[0055] 1. Backbone structure,

[0056] As shown in Figure 2 , for an input image , wherein C is the number of channels, W is the image width, and H is the image height, the input image is first input into the backbone network to extract feature information. The backbone network gradually reduces the size of the feature map and expands the number of channels from shallow to deep. The input image size used in this paper is , the feature map size is reduced and the channel number is expanded through the ConV layer and the C2f layer. Finally, the features are input into the SPPF module for pyramid pooling operation. Each ConV layer includes a BatchNorm layer, a standard convolution layer and a ReLU activation function layer. The BatchNorm layer is used to normalize the batch input data to improve the overall convergence speed of the model and alleviate the gradient descent and gradient explosion. Its specific definition is that for each Batch containing m data of the model input, the mean and variance are calculated as follows:

[0057]

[0058] wherein, is the number of all samples in the Batch, is the mean of all samples in the Batch, is the variance. Then each sample in the Batch is normalized as follows:

[0059]

[0060] wherein, is a very small number to prevent division by zero.

[0061] The standard convolution layer performs sliding window calculation on the input feature map, extracts features in the local area, and generates a new feature map, which is used for feature extraction and feature map dimension reduction. The definition of the standard convolution operation is:

[0062]

[0063] wherein, is the convolution kernel, is the height and width of the convolution kernel, is the pixel value of a certain point on the feature map, which is achieved by traversing all pixel values on the feature map and performing weighted sliding window to extract features of surrounding elements. In addition, the convolution operation can increase or decrease the number of kernel windows to increase or decrease the channel number.

[0064] In the convolution operation, the step controls the sampling frequency. The default convolution step is 1, that is, all points of the feature map are traversed. When the step is 2, the convolution operation samples every other point, and the output feature degree is reduced to half of the input feature map.

[0065] The activation function is used to perform nonlinear transformation on the features. The SiLU activation function is used in this model. The definition of SiLU is:

[0066]

[0067] The C2f layer is composed of different ConV layers, and the number of input and output channels of the feature map is unchanged when passing through the C2f layer. Specifically, the input is first passed through a ConV layer with a convolution kernel of 1 and the number of input channels is doubled. Then the feature map is divided by the number of channels to generate two feature maps with the same number of input channels for the final connection. The purpose of this is to increase the expression ability of the model. Then pass through 3 Bottleneck layers, which are composed of two ConV layers with a convolution kernel of 1 and a step size of 1, which are used to extract features. Finally, the input itself, the feature map separated after the first ConV layer, and the feature map output after the Bottleneck layer are spliced to obtain a feature map three times the original channel number. The feature map is transformed to the channel number at the input time through a ConV layer with a convolution kernel of 1 to ensure that the number of input and output channels is unchanged.

[0068] The order of the size change of the feature map of the backbone network is , , , , , The change of the number of channels and the size is carried out through the ConV layer, which is marked as P1-P5 in Figure 2 . The step size of these ConV layers is 2, and the output channel number is twice the input channel number.

[0069] 2. Neck structure,

[0070] The feature pyramid structure (FPN) and the path aggregation network (PAN) have been proved to be an effective network for aggregating feature information extracted by the backbone network at multiple scales. The structure of the neck network is shown in Figure 2 . The neck network fuses the high-order semantic information extracted by the backbone network through up-sampling with low-order features, improving the detection ability of different target sizes. Weak targets usually only account for 5-10% of the original image size, and feature maps located at a deeper position in the backbone network are too small to capture the position information of these small targets. On the other hand, feature maps that are too large are located at a shallow position in the backbone network, with narrow channels and poor feature expression ability, so it is difficult to effectively identify the feature information of small targets. In order to solve these problems, the model improves the Neck structure, integrates early large feature maps into the detection head, and uses a feature extraction structure with channel attention to enhance the semantic information of early feature maps.

[0071] Channel attention methods represented by SE attention mechanism effectively improve the performance of CNN network for image processing tasks. The essence of SE attention is to compress the entire feature map to a single feature dimension after any convolutional transformation U, which is expressed as:

[0072]

[0073] where and are the length and width of the input feature map. Through this squeezing operation, the original feature map is compressed to a single feature dimension . In order to utilize the feature information aggregated by the squeezing operation, a fully connected layer is used to reduce its dimension times, and then restored to by a fully connected layer after a nonlinear function. Finally, the inter-channel attention is obtained by the Sigmoid function, and this excitation method is expressed as:

[0074]

[0075] where, is the Sigmoid function, and are two weight matrices, is the Relu function.

[0076] Although this channel attention method further improves the network performance by obtaining the attention between features, it ignores the position information of the features, which is very important for generating spatially selective attention maps. Therefore, this paper improves the current SE attention module and proposes a channel attention module LSE that adds position information. For any convolutional transformation U, the original feature map is compressed from the high and wide directions, respectively, and this method is expressed as:

[0077]

[0078]

[0079] where H, W are the height and width of the input feature map. After the above formula, the original input is separated into two different feature maps, with sizes , respectively. Then transpose it and concatenate it in the last dimension to get a new feature map with size . Finally, use the convolution layer of to extract the position information, which is expressed as:

[0080]

[0081] where, is the ReLU activation function, and , is the reduction rate to control the block size for reducing the computational overhead, and then is separated into two independent tensors and . The channel number is changed back to C by using another two convolutional layers, and the final position channel attention is represented as:

[0082]

[0083] where, is the Sigmod activation function, is the segmentation operation, and are two weight matrices.

[0084] As shown in Figure 4 , the entire improved channel attention module, referred to as the LSE model in this paper, is composed of SE attention mechanism and position attention mechanism, and finally all attention information is fed back to the original feature map, and a new feature map with added attention is obtained through convolution operation for the next layer training. The formula is represented as:

[0085]

[0086] where, is the output of the current layer, is the ReLU activation function, is the input of the current layer.

[0087] The Neck structure of the model is improved based on the Yolov8 Neck structure, as shown in the figure, the Yolov8 Neck structure is improved in this paper, and additional links are introduced into the originally Yolov8 Neck and size feature map in the Backbone and size feature extraction module. A feature extraction module is proposed in this paper for feature extraction on the link, referred to as the EPM module, and the module structure is convolutional layer, C2f module, LSE module and convolutional layer, which can effectively extract the features of large-scale feature maps, and the specific structure is shown in Figure 3 .

[0088] 3. Detection head,

[0089] The Yolov8 Header structure is adopted herein to construct a decoupled feature head, and target box loss and classification loss are calculated respectively. The specific structure is shown in the figure. In addition, the model adopts an Anchor-Free structure, that is, instead of determining anchor boxes in advance, the center point of the target is determined, and the distance between the center point and the top, bottom, left and right of the target box is predicted to output the final target box. The loss function of the target box loss uses the CIOU loss function, and the binary cross entropy function is used for classification loss optimization.

[0090] The specific calculation method of the Anchor-Free structure is as follows: for each feature map output by each detection head, each pixel point is assumed to be the center point of the target box, and the top, bottom, left and right (top, bottom, left, right) four distances are calculated. The final output of the network contains (4+target classification number+target confidence) elements, which are as follows:

[0091] [left offset, upper offset, right offset, lower offset, [algal species], [development stage], target confidence]

[0092] The target confidence represents the probability of the existence of the target in the predicted box. Assuming that each center point is x, the overall target box can be obtained by calculating the coordinate values of the top left corner and the bottom right corner of the target box, and the calculation method is as follows:

[0093]

[0094] These distances are optimized by CIOU, and the calculation method of CIOU loss is as follows:

[0095]

[0096] The calculation of CIOU consists of three parts, IOU, and . Among them, IOU is the general intersection over union loss, and its calculation method is as follows:

[0097]

[0098] Among them, is the position of the predicted box is the position of the real box, and through this term, the size of the predicted box covering the real box can be judged.

[0099] is the Euclidean distance between the center point of the predicted box and the center point of the real box , and its calculation method is as follows:

[0100]

[0101] wherein, is the horizontal and vertical coordinates of the center point of the prediction box, is the horizontal and vertical coordinates of the center point of the real box.

[0102] is the square of the diagonal length of the minimum bounding box enclosing the prediction box and the real box, used for normalizing the distance.

[0103] is the aspect ratio consistency term, which is calculated as:

[0104]

[0105] wherein, is the width and height of the prediction box, is the width and height of the real box.

[0106] is a weight factor, which is calculated as:

[0107]

[0108] The definition of the binary cross-entropy function is:

[0109]

[0110] wherein, N is the number of classifications, is the real value, is the predicted value.

[0111] (2) Model training data set collection and training method,

[0112] The model training data set is constructed by using a public data set and a collected data set. For the public data set, it is relabeled according to the algae species and the development period. The collected data set is continuously collected by a multi-rotor unmanned aerial vehicle equipped with an RGB camera at different angles in the sea area. Data under different light conditions in the morning, noon and evening are selected to ensure data diversity. For the collected data set, image cropping is performed first to remove the extra background part, and all pictures are adjusted to a resolution of 640*640. Since the background area of different pictures is different, the cropped picture may be lower than the preset resolution, which will cause noise generation. Therefore, Gaussian filtering is used to denoise all pictures, and the definition of Gaussian filtering is:

[0113]

[0114] wherein, x, y are pixel point coordinates, is the standard deviation of the Gaussian function, which controls the smoothing degree of the Gaussian filter. Through Gaussian filtering, salt and pepper noise and Gaussian noise in the picture can be removed while preserving the edges and details of the picture. For the collected pictures, a double labeling method is also used, and the pictures are labeled according to the algae species and the development period.

[0115] The overall data set is composed of 45% normal seawater, 30% early red tide, 15% red tide outbreak, and 10% red tide-like interference samples. The training set, validation set, and test set are divided in the ratio of 8:1:1. The model is trained for 300 iterations, the learning rate is 0.0001, the weight decay is set to 0.0005, and the random gradient descent method is used for optimization. Due to the seasonality and timeliness of red tide, there is not much data to train the network at the beginning of training, so the incremental training mode is adopted. This mode inputs the newly collected data into the previously trained model for further incremental training. To solve the problem that incremental learning may cause the model to be too inclined to the distribution of the new data set and forget the distribution of the old data set, the Elastic Weight Consolidation (EWC) method is used to alleviate this problem. An additional EWC loss function is added when incremental training is performed, and the loss function is expressed as:

[0116]

[0117] wherein, is the loss function trained on the new data set, is the parameter of the new model, is the optimal parameter of the old model, is the Fisher information matrix of the parameter , which is used to measure the importance of the parameter, and λ is the regularization coefficient, which controls the protection strength of the old task knowledge.

[0118] (2) Red tide monitoring and early warning method.

[0119] For a specific sea area, deploy an RGB monitoring camera with a distance of 10-30m from the sea surface. All RGB images of the camera are transmitted to the computing server in real time. These images are cropped to 640*640 and subjected to Gaussian filtering to eliminate noise on the computing server. The computing server uses the target detection model proposed above to detect the transmitted RGB images every 60 seconds. When the area of the algae target occupies more than 15% of the marine area in the image, the system will prompt the monitoring personnel that there is a risk of red tide. When the system prompts that there is a risk of red tide, use a multi-rotor unmanned aerial vehicle with an automatic sampling function to automatically sample seawater in the target monitoring sea area, and perform biological detection. For the warning condition, an early warning is issued.

[0120] Example 2

[0121] The embodiment provides a red tide outbreak early warning system based on weak target detection, which comprises the following steps:

[0122] A data acquisition module is configured to:

[0123] A computer readable storage medium, wherein a plurality of instructions are stored, the instructions are suitable for being loaded by a processor of a terminal device and performing the red tide outbreak early warning method based on weak target detection.

[0124] A terminal device comprises a processor and a computer readable storage medium, the processor is used for implementing instructions; the computer readable storage medium is used for storing a plurality of instructions, the instructions are suitable for being loaded by the processor and performing the red tide outbreak early warning method based on weak target detection.

[0125] The above are preferred embodiments of the present application, not limited to the protection scope of the present application, therefore: equivalent changes made according to the structure, shape, principle of the present application should be covered in the protection scope of the present application.

Claims

1. A red tide outbreak early warning method based on weak target detection, characterized in that, The method comprises the following steps: An image of a red tide dataset is acquired, and image preprocessing is performed through Gaussian filtering; A weak target detection enhancement model based on a Yolov8 structure is constructed, and the weak target detection enhancement model based on the Yolov8 structure is used to perform feature extraction on the red tide dataset image; The extracted features are fused; Target detection is performed based on the fused features, and the weak target detection enhancement model is incrementally trained based on the target detection result; The trained weak target detection enhancement model is used for red tide early warning. The feature fusion on the extracted features comprises fusing the high-order semantic information finally extracted by the backbone network through upsampling with the low-order features by using a neck network of the model, wherein the whole feature map is first compressed to a single feature dimension, and the original feature map is compressed to a single feature dimension through the compression operation . is expressed as: ， wherein and are the length and width of the input feature map, C is the number of channels, and U denotes an arbitrary convolutional transformation; Also included is a channel attention mechanism LSE based on added position information, after any convolutional transformation U, the original feature map is compressed from the high direction and the wide direction respectively, and the two compressed matrices are deformed and fused, and the position information is extracted by the convolutional layer of , denoted as: ， where, is the ReLU activation function, and , is a reduction rate that controls the block size for reducing the computational overhead, and then is separated into two independent tensors and , and the channel number is changed back to C using another two convolutional layers. The final position-channel attention representation is given by: ， wherein, is a Sigmod activation function, is a split operation, are two weight matrices; The target detection based on the fused features comprises the following steps: a decoupled feature head is constructed using a Yolov8 Header according to the fused features, and target box loss and classification loss are calculated respectively, wherein an Anchor-Free structure is adopted, a target center point is used to replace an anchor box, and the center point and the distance between the target box and the center point are predicted to output the final target box, the loss function of the target box loss uses the sum of a CIOU and a DFL loss function, and the classification loss uses a binary cross-entropy function for optimization. The incremental training of the weak target detection enhancement model based on the target detection result comprises the following steps: based on the seasonality and timeliness of the red tide, the collected data is input into the previously trained model for further incremental training in an incremental training mode, and an EWC method is used to alleviate the problem that the model is too inclined to the distribution of the new dataset and forgets the distribution of the old dataset, wherein an additional EWC loss function is added during the incremental training, and the loss function is expressed as: ， where, is the loss function trained on the new dataset, is the parameter of the new model, is the optimal parameter of the old model, is the parameter Fisher information matrix of the parameter, used to measure the importance of the parameter, The method comprises the following steps: is the regularization coefficient, which controls the strength of protection of the old task knowledge.

2. The red tide outbreak early warning method based on weak target detection according to claim 1, characterized in that, The weak target detection enhancement model based on the Yolov8 structure is used for feature extraction on the red tide data set image, including inputting the input image into a backbone network to extract feature information Wherein C is the number of channels, W is the image width, and H is the image height, the input image is first input into a backbone network to extract feature information, and the backbone network gradually reduces the feature map size and expands the channel number from shallow to deep.

3. The red tide outbreak early warning method based on weak target detection according to claim 2, characterized in that, The compression of the entire feature map to a single feature dimension includes using a fully connected layer to first reduce the dimension of the entire feature map by a factor of 2, and then using a fully connected layer to restore to , and finally obtaining inter-channel attention by a Sigmoid function and re-weighting the original input map, expressed as: ， wherein Fexrepresents inter-channel weight distribution and feature enhancement, is a Sigmoid function, and are two weight matrices, is a Relu function.

4. A red tide outbreak early warning system based on weak target detection, which executes a red tide outbreak early warning method based on weak target detection according to claim 1, characterized in that, A data acquisition module is configured to acquire an image of a red tide dataset and perform image preprocessing through Gaussian filtering; A feature extraction module is configured to construct a weak target detection enhancement model based on a Yolov8 structure, and use the weak target detection enhancement model based on the Yolov8 structure to perform feature extraction on the red tide dataset image; A feature fusion module is configured to fuse the extracted features; A target detection module is configured to perform target detection based on the fused features, and incrementally train the weak target detection enhancement model based on the target detection result; An early warning module is configured to use the trained weak target detection enhancement model for red tide early warning. The instructions are suitable for being loaded and executed by a processor of a terminal device to perform the method of claim 1.

5. A computer readable storage medium having stored therein a plurality of instructions, wherein the instructions, when executed by a processor, cause the processor to perform the method of any one of claims 1-4. The instructions are suitable for being loaded and executed by a processor to perform the method of claim 1. 6.A terminal device, comprising a processor and a computer readable storage medium, the processor is configured to implement instructions; the computer readable storage medium is configured to store a plurality of instructions, characterized in that, ​

Images