A method for extracting information about benthic materials in coral reefs

Abstract

The present invention relates to a method for extracting information about benthic materials on coral reefs, comprising: constructing an initial improved U-Net network model, the initial improved U-Net network model comprising: an encoder, a decoder, a far-hop connection, and an underlying feature map, wherein the encoder and the decoder are both provided with an attention residual block and a depthwise separable attention residual block to classify objects contained in an image and extract features thereof respectively, the far-hop connection is provided with a CBAM attention mechanism module to configure weights for features of the image, and the underlying feature map is provided with at least one dilated spatial pyramid pooling module to enable the initial improved U-Net network model to mark features of smaller areas in the image. The present invention can effectively improve the ability to extract information about benthic materials on coral reefs through the target U-Net network model, and further achieve high-precision automatic extraction.

Description

Technical Field

[0001] The present invention relates to the fields of remote sensing image processing and deep learning, and in particular to a method for extracting information of coral reef benthic materials. Background Art

[0002] Coral reefs are among the world's most diverse ecosystems, boasting a rich diversity of species. They not only generate economic and social benefits, but also mitigate wave energy and protect coastal zones. However, coral reef systems are fragile and susceptible to disturbances from both natural and human activities, leading to coral bleaching. Prolonged bleaching can lead to the death of reef-building corals. Therefore, monitoring and protecting coral reefs is crucial. To better monitor coral reefs, mapping the distribution of benthic materials is essential.

[0003] Currently, the most commonly used methods for mapping the distribution of benthic materials on coral reefs are traditional machine learning and object-oriented methods. However, the complex environment of benthic materials on coral reefs means that the accuracy of information extracted by traditional machine learning methods is generally low. While object-oriented methods are more sophisticated than traditional machine learning methods, they require more human intervention and are less efficient. Therefore, improving the accuracy and efficiency of information extraction is of great significance.

[0004] Using deep learning technology for image segmentation involves building a network model, feeding it a training set and a validation set to train the network, and then using the learned prior knowledge to achieve refined image segmentation. This can produce better results than traditional methods. Deep learning technology is currently widely used in the field of remote sensing information extraction.

[0005] The U-Net network is a classic semantic segmentation network built on FCN. Its two most prominent features are its U-shaped structure and long-hop connections. Feature extraction is performed through convolutional layers and max-pooling layers in the network encoder, while feature map size is restored through convolutional layers and upsampling operations in the network decoder. Long-hop connections strengthen the connection between low-level and high-level features. While the U-Net network has demonstrated good results in some remote sensing image segmentation tasks, its relatively simple structure, with feature extraction consisting of simple convolution and downsampling, does not yield good results for complex segmentation tasks such as coral reef benthic material. Therefore, improvements to the U-Net network are needed. Summary of the Invention

[0006] The purpose of the present invention is to solve the above problems existing in the prior art.

[0007] To achieve the above-mentioned object, the present invention provides a method for extracting information of benthic materials of coral reefs, comprising: constructing an initial improved U-Net network model, wherein the initial improved U-Net network model comprises: an encoder, a decoder, a far jump connection and an underlying feature map, wherein the encoder and the decoder are both provided with an attention residual block and a depth-separable attention residual block to classify the targets contained in the image and extract features respectively, the far jump connection is provided with a CBAM attention mechanism module to configure the weights of the features of the image, and the underlying feature map is provided with at least one void space pyramid pooling module to enable the initial improved U-Net network model to mark the features of smaller areas in the image; the initial improved U-Net network model is trained by a training data set to obtain a target U-Net network model, wherein the training The training data set includes: multiple training images that have been classified and labeled for various areas of the training area; obtaining the target image of the target area, the target image including the image of the coral reef benthic material in the target area, inputting the target image into the target U-Net network model, encoding the image of the coral reef benthic material through the attention residual block and the depth-separable attention residual block of the encoder, and extracting feature data to obtain an initial feature data set; weighting the feature data in the initial feature data set through the CBAM attention mechanism module of the far jump connection, assigning a higher weight coefficient to feature data with higher importance and a lower weight coefficient to feature data with lower importance, to form a target data set; performing data feature analysis on the target data set through the decoder to obtain target analysis data.

[0008] The present invention can effectively improve the ability to extract coral reef benthic material information through the target U-Net network model, and further achieve high-precision automatic extraction. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 1 is a flow chart of a method for extracting information of coral reef benthic materials according to an embodiment of the present invention;

[0010] Figure 2 Schematic diagram of the target U-Net network model according to an embodiment of the present invention;

[0011] Figure 3 is a schematic structural diagram of an encoder according to an embodiment of the present invention;

[0012] Figure 4 is a schematic diagram of the attention residual block structure of an embodiment of the present invention;

[0013] Figure 5 2 is a schematic diagram of the structure of a depth-separable residual attention block according to an embodiment of the present invention;

[0014] Figure 6 is a schematic diagram of the decoder structure of an embodiment of the present invention;

[0015] Figure 7 is a schematic diagram of the underlying feature mapping structure of an embodiment of the present invention;

[0016] Figure 8 is a schematic diagram of a far-hop connection structure according to an embodiment of the present invention;

[0017] Figure 9 is an example flow chart of preprocessing of a training data set according to an embodiment of the present invention;

[0018] Figure 10 is a comparison chart of test results of an embodiment of the present invention;

[0019] Figure 11 1 is a graph showing the F1-Score evaluation index results of the test results of an embodiment of the present invention. DETAILED DESCRIPTION

[0020] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments.

[0021] like Figure 1 FIG. 1 is a method for extracting information about benthic materials in coral reefs according to an embodiment of the present invention, comprising:

[0022] S1: Construct an initial improved U-Net network model, which includes an encoder, a decoder, a far-hop connection, and an underlying feature map. An attention residual block and a depthwise separable attention residual block are provided between the encoder and the decoder to classify the objects contained in the image and extract features respectively. A CBAM attention mechanism module is provided in the far-hop connection to configure the weights of the image features. At least one dilated spatial pyramid pooling module is provided in the underlying feature map to enable the initial improved U-Net network model to mark the features of smaller areas in the image.

[0023] Specifically, in this step, it is necessary to construct an initial improved U-Net network model, where the initial improved U-Net network model includes: an encoder, a decoder, a long-hop connection and an underlying feature map, and both the encoder and the decoder are provided with an attention residual block and a deep attention residual block.

[0024] It will be understood by those skilled in the art that the U-Net network model in the prior art uses convolutional layers, while in the embodiments of the present invention, the initial improved U-Net network model uses attention residual blocks and deep attention residual blocks to replace the ordinary convolutional layers of the U-Net network model in the prior art, which can effectively improve the extraction of features of targets in images, make the extracted features richer, and thus improve the accuracy of subsequent feature classification.

[0025] A CBAM attention mechanism is set in the long-hop connection, which can assign weights to the features in the image, that is, effectively control the attention to the features in the image.

[0026] In other words, it is possible to effectively pay more attention to features with higher importance, and suppress attention to features with lower importance, thereby achieving efficient recognition and marking of features.

[0027] In the initial improved U-Net network model, an underlying feature map is provided, and a void space pyramid pooling module is provided in the underlying feature map, which can improve the image recognition accuracy of a small area in the image, thereby improving the accuracy and precision of the initial improved U-Net network model in image recognition.

[0028] S2: training the initial improved U-Net network model using a training data set to obtain a target U-Net network model, wherein the training data set includes: a plurality of training images in which various regions of the training area have been classified and labeled;

[0029] Specifically, in this step, the purpose of training is to obtain training weights. The initial improved U-Net network model just built has no weights, so it must be trained with a training data set. After training and verification, the obtained target U-Net network model can be used to improve the efficiency and accuracy of information extraction from the target image.

[0030] The training data set consists of multiple training images. It can be understood that the training image is an image of the training area. The training image includes multiple areas of the training area, and the multiple areas have been classified and marked.

[0031] Among them, the target U-Net network model includes: four operating layers, each operating layer is equipped with an encoder, a decoder and a long-jump connection, and the encoder and decoder are both equipped with attention residual blocks and depth separation attention residual blocks.

[0032] like Figure 2 The following is a schematic diagram of the target U-Net network model. Figure 2It can be understood that in the target U-Net network model of the embodiment of the present invention, there are four layers, the layers are as follows: Figure 2 In the up and down direction shown, the levels increase sequentially, namely the first level, the second level, the third level and the fourth level.

[0033] The size of the feature image to be processed in each layer is different. As the number of layers increases, the size of the image to be processed becomes smaller, but the number of channels to be processed increases.

[0034] For example Figure 2 As shown, the target image is output in the first level, and the image features to be processed in the first level may be 100. In the second level, the image features of the target image need to be further refined, that is, the influence of the 100 image features is further decomposed, which may be 200 image features. However, compared with the image in the first level, the image size in the second level becomes smaller, but the number of image features increases.

[0035] In some embodiments of the present invention, the sizes of feature maps from the first level to the fourth level are 1024, 512, 256, 128, and 64, respectively.

[0036] like Figure 2 As shown, there is a corresponding encoder and decoder in each layer. The encoder can extract features of the input target image, and the decoder can restore the data to the form of an image for output.

[0037] Further, if Figure 3 and Figure 6 As shown, both the encoder and decoder are equipped with an attention residual block and a depth-separated attention residual block. The attention residual block can improve the target U-Net network model's ability to extract internal image information of the target image. The depth-separable attention residual block further replaces the ordinary convolution used in the U-Net network model in the prior art, and replacing it with a depth-separable attention residual block can further improve efficiency. This achieves the target U-Net network model of the embodiment of the present invention to extract information efficiency and accuracy from the target image.

[0038] like Figure 3 The figure shows a schematic diagram of the structure of the encoder. The encoders in the first and second layers of the four-layer hierarchy are both provided with attention residual blocks, the encoders in the third and fourth layers of the four-layer hierarchy are both provided with depth separation attention residual blocks, and the fourth layer is provided with at least one Dropout layer.

[0039] Specifically, the encoder uses the structure of the residual part of ResNet-34, whose first level consists of three layers such as Figure 4The attention residual block shown in the figure is composed of four Figure 4 The third level consists of six attention residual blocks as shown in Figure 5 The fourth level consists of three blocks such as Figure 5 The shown depthwise separable attention residual block composition.

[0040] Among them, such as Figure 4 As shown in , the convolution layer in the attention residual block is a convolution layer with a 3×3 kernel, as shown in Figure 5 As shown in the figure, the depthwise separable convolutional layer in the depthwise separable attention residual block is a depthwise separable convolutional layer with a 3×3 kernel. A batch normalization layer is added after each convolutional layer and depthwise separable convolutional layer. The addition of the batch normalization layer here serves as a planning layer to prevent gradient vanishing during the construction of the target U-Net network model and speed up training. A dropout layer with a dropout probability of 0.5 is added after the last depthwise separable attention residual block in the fourth level.

[0041] like Figure 7 The figure shows a schematic diagram of the structure of the decoder. The decoders in the first and second layers of the four-layer hierarchy are both provided with attention residual blocks, the decoders in the third and fourth layers of the four-layer hierarchy are both provided with depth separation attention residual blocks, and the fourth layer is provided with at least one Dropout layer.

[0042] Specifically, the first and second levels of the decoder are composed of two Figure 4 The third and fourth levels are composed of two attention residual blocks as shown in Figure 5 The network consists of a depthwise separable attention residual block as shown. The convolutional layers in the attention residual block are 3×3 kernel convolutional layers. A batch normalization layer is added after each convolutional layer and depthwise separable convolutional layer. The purpose of adding a batch normalization layer here is to prevent gradient vanishing during the construction of the target U-Net network model and speed up training. A dropout layer with a dropout probability of 0.5 is added after the last depthwise separable attention residual block in the fourth level.

[0043] like Figure 2 As shown in FIG, the U-Net network model also includes: an underlying feature map, and both ends of the underlying feature map are connected to the encoder and decoder of the upper layer.

[0044] Specifically, if Figure 2 As shown, under the four-layer hierarchy, there is an underlying feature map, which can be used as the fifth layer of the target U-Net network model. Figure 2 As can be seen in Figure 2, the two ends of the underlying feature map of the fifth layer are connected to the encoder and decoder of the fourth layer respectively.

[0045] Further, if Figure 7 As shown in Figure 3, at least one depth-wise separable attention residual block is set in the underlying feature map, and a Dropout layer is set after each depth-wise separable attention residual block.

[0046] At least one depth-wise separable attention residual block is provided in the underlying feature map. As mentioned above, it can effectively improve the efficiency of information extraction, and there is at least one, that is, the number of depth-wise separable attention residual blocks set is at least one, and can also be more than one, which can greatly improve the extraction efficiency.

[0047] In some embodiments of the present invention, Figure 7 The underlying feature map shown in the figure includes two Figure 5 The depthwise separable attention residual block shown in the figure. The depthwise separable convolution layer in the depthwise separable attention residual block is a depthwise separable convolution layer with a 3×3 kernel. A dilated spatial pyramid pooling module is added between the two depthwise separable attention residual blocks. This module sets up 4 layers of dilated convolution connected in parallel, with dilation rates set to 1, 6, 12, and 18 respectively, and an automatic zero-padding strategy is also set.

[0048] In addition, a BN layer is added after each depth-wise separable convolutional layer, and a Dropout layer with a dropout probability of 0.5 is added after each depth-wise separable attention residual block. Setting the Dropout layer can effectively improve the learning efficiency of the target U-Net network model during the initial construction learning process.

[0049] The BN layer and Dropout layer mentioned above are explained again here. The BN layer is added to the target U-Net network model because it can keep the input of each layer of the deep neural network in the same distribution during training. It can speed up the training and convergence of the network to a certain extent, and can also control gradient explosion and prevent gradient disappearance. The Dropout layer is added to the target U-Net network model because it can remove the neural network training unit from the network according to a certain probability during training. This is a regularization method that can prevent overfitting to a certain extent.

[0050] Before step S2, the method further includes:

[0051] S10: Acquire a training image of a training area having coral reef benthic materials, and perform preprocessing to obtain a training dataset having characteristics of the coral reef benthic materials.

[0052] In this step, after the training images of the training area are acquired, the training images need to be preprocessed. The preprocessing may include performing atmospheric correction and radiation calibration on the training images, and generating a label map of the training images based on the first algorithm.

[0053] The first algorithm may be an object-oriented algorithm, through which each area in the training image may be marked to create a label map. For example, the coral reef area and the bleached area in the training image may be marked with different colors or images.

[0054] The label image and the training image are randomly cropped at the same time to obtain the training image set. That is, the original images of the label image and the training image are cropped into a uniform 128×128 small image, and data enhancement is performed on all images in the training image set to obtain the training dataset.

[0055] It should be noted that the number of images in the training picture set is limited after all. Therefore, the color, brightness or size of the training images can be adjusted, and the training images can be mirrored or flipped, so as to enhance the learning ability of the initial improved U-Net network model.

[0056] Specifically, such as Figure 9 The figure shows an example flow chart of the preprocessing of a training dataset. In the process of obtaining training images containing nearshore coral reef areas and using them to prepare the training dataset required for network training, data augmentation and dataset segmentation are performed. The obtained training images are labeled for seven categories, namely coral reefs, bleached coral reefs, algae mixtures, sand, waves, deep sea areas, and land, using an object-oriented algorithm. Three-level segmentation at different scales is performed to make the labels more refined. At the same time, the labels are manually adjusted based on the visual interpretation results of higher-resolution Google Earth images. The original images and manually adjusted labels are randomly cropped to 128×128 resolution images, and images containing only two non-target research areas, land and deep sea areas, are manually removed to prevent excessive class imbalance in the training process. Then, the original image dataset and the labeled image dataset are simultaneously rotated, offset, blurred, brightness changed, and noise added for data augmentation. Finally, all the processed data are divided into a training set and a validation set according to a 7:3 ratio.

[0057] Those skilled in the art will appreciate that the initial improved U-Net network model is trained using the training set, and the accuracy of the model is verified using the validation set, thereby obtaining a stable target U-Net network model. In other words, the training dataset can be divided into a training set and a validation set in a ratio of 7:3, with 70% used for training and 30% used for validation.

[0058] In step S2, the initial improved U-Net network model is trained using the training data set to obtain the target U-Net network model, including the following sub-steps:

[0059] S21: Input the training data set into the initial improved U-Net network model and set the maximum number of training times;

[0060] S22: Select the target optimization function and calculate the network loss;

[0061] S23: Update network parameters using the optimizer based on network loss;

[0062] S24: Repeat the above steps until the network is fitted or the maximum number of training times is reached;

[0063] S25: Save the trained target U-Net network model.

[0064] The network loss is the error rate in the verification, and the parameters are further optimized according to the error rate to obtain the final stable and accurate target U-Net network model.

[0065] The following example illustrates the training steps. Training the improved generative adversarial network model to obtain an optimized generative adversarial network model includes the following steps: setting the maximum training rounds to 40, setting the batch size to 2, setting the learning rate to 0.0001, calling the train function to iteratively train the initial U-Net model using the training set, and verifying the model after each round of training using the validation set. The network is evaluated using a new function obtained by adding the cross entropy loss function and the Dice Loss function. The Adam optimization algorithm is used to update the network weights. Training stops when the maximum training rounds are reached or the model has reached the optimal state.

[0066] S3: Obtain a target image of the target area, which includes an image of the coral reef benthic material in the target area. Input the target image into the target U-Net network model, encode the image of the coral reef benthic material through the encoder's attention residual block and depthwise separable attention residual block, and extract feature data to obtain an initial feature dataset.

[0067] In this step, the target area may be the area where coral reef benthic material information extraction is required, that is, the coral reef area near the coast. After acquiring the target image, the acquired target image may have a size problem, so the size of the target image needs to be adjusted.

[0068] In some embodiments of the present invention, the size of the target image can be cropped to reduce the size of the target image, and can be cropped into images of equal size with a resolution of 128×128, so as to adapt to the image size required by the target U-Net network model.

[0069] The target image is input into the target U-Net network model. As mentioned above, the encoder is equipped with an attention residual block and a depth-separable attention residual block, which can encode the image of the coral reef benthic material in the target area. After encoding the image, the features in the image can be extracted by the attention residual block and the depth-separable attention residual block to obtain an initial feature data set composed of feature data.

[0070] S4: The CBAM attention mechanism module with long jump connections is used to assign weights to the feature data in the initial feature dataset, assigning higher weight coefficients to feature data with higher importance and lower weight coefficients to feature data with lower importance, thus forming the target dataset.

[0071] In this step, if Figure 8 As shown in the figure, in the long-jump connection, the CBAM attention mechanism module can effectively distribute the weights of the initial feature data. As can be seen from the previous text, the feature data in the acquired feature data set are the feature data of all areas included in the target image of the target area. For some features with lower importance, the attention of the entire target U-Net network model to the feature data with higher importance can be increased by reducing the attention, thereby improving the processing speed of the feature data and thus improving efficiency.

[0072] S5: Perform data feature analysis on the target data set through the decoder to obtain target analysis data.

[0073] In this step, the decoder can re-decode the image encoded by the encoder. The target image with a resolution of 128×128 is input into the target U-Net network model, so that the target image can be analyzed by the target U-Net network model to extract coral reef benthic material information from the target image.

[0074] Specifically, the extracted information may be information such as the bleached area, the non-bleached area, the coast, and the water area of the target image of the target area.

[0075] In order to verify the method for extracting information about benthic materials in coral reefs according to the embodiment of the present invention, GF-2 images of West Daimao Island, Luhuitou Lane, and Xiaodonghai were first obtained. Then, operations such as labeling, cropping, and data enhancement were performed to create a training dataset. The improved U-Net model was trained based on the dataset. After the model training was completed, the GF-2 image of East Daimao Island was used for testing. Figure 10 The test results of the present invention based on the improved U-Net model and maximum likelihood method, SVM, FCN, U-Net, and Deeplabv3+ are shown as follows. Figure 11 The values of the F1-Score evaluation index for each category are shown. It can be seen that the F1-Score of the present invention is the best in each category, especially for the main research categories, namely coral reefs and bleached coral reefs. In addition, compared with the U-Net network, the F1-Score of the present invention in the coral reef and bleached coral reef categories is approximately 12% and 11% higher than that of the U-Net network, respectively, and the performance after improvement is greatly improved.

[0076] The present invention discloses a method for extracting benthic material information of coral reefs based on an improved U-Net network. The improved U-Net network is used for extracting benthic material information of coral reefs. The present invention first improves the U-Net network. While retaining the U-shaped structure of the U-Net network consisting of an encoder, a decoder, an underlying feature map and a far-jump connection, the structure of the encoder, the decoder, the underlying feature map and the far-jump connection is improved. An attention residual block and a depth-separable attention residual block are used to replace the original ordinary convolution layer, so that the extracted features are richer. A CBAM attention mechanism module is added to the far-jump connection, which can distribute weights to features in the image, that is, effectively control the attention to the features in the image.

[0077] In other words, it is possible to effectively pay more attention to features with higher importance, and suppress attention to features with lower importance, thereby achieving efficient recognition and marking of features.

[0078] A dilated spatial pyramid pooling module is added to the underlying feature map to enable the network to have stronger segmentation capabilities for small objects. A batch normalization layer is added to the network to speed up network training and convergence. A dropout layer is added to the network to prevent network overfitting.

[0079] The above specific implementation methods further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above are only specific implementation methods of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for extracting information of coral reef benthic materials, characterized in that: include: Constructing an initial improved U-Net network model, the initial improved U-Net network model comprising: an encoder, a decoder, a far-jump connection, and an underlying feature map, wherein the encoder and the decoder are both provided with an attention residual block and a depthwise separable attention residual block to classify objects contained in the image and extract features respectively, the far-jump connection is provided with a CBAM attention mechanism module to configure weights for features of the image, and the underlying feature map is provided with at least one dilated spatial pyramid pooling module to enable the initial improved U-Net network model to mark features of smaller areas in the image; The initial improved U-Net network model is trained using a training data set to obtain a target U-Net network model, wherein the training data set includes: a plurality of training images in which various regions of a training area have been classified and labeled; Acquiring a target image of a target area, the target image including an image of coral reef benthic material in the target area, inputting the target image into the target U-Net network model, encoding the image of the coral reef benthic material using the attention residual block and the depthwise separable attention residual block of the encoder, and extracting feature data to obtain an initial feature dataset; The CBAM attention mechanism module connected by the far jump performs weight assignment on the feature data in the initial feature data set, assigning a higher weight coefficient to the feature data with higher importance and assigning a lower weight coefficient to the feature data with lower importance, so as to form a target data set; The decoder performs data feature analysis on the target data set to obtain target analysis data.

2. The method according to claim 1, characterized in that Before the initial improved U-Net network model is trained using a training data set to obtain a target U-Net network model, the method further includes: The training image of the training area having the coral reef benthic material is acquired and preprocessed to obtain a training data set having characteristics of the coral reef benthic material.

3. The method according to claim 1, characterized in that Both ends of the underlying feature map are connected to the encoder and the decoder.

4. The method according to claim 3, characterized in that At least one depth-wise separable attention residual block is provided in the underlying feature map, and a Dropout layer is provided after each depth-wise separable attention residual block.

5. The method according to claim 3, characterized in that The target U-Net network model includes: at least four levels, each level of the level is provided with an encoder, a decoder and a far jump connection, and the encoder and the decoder are both provided with the attention residual block and the depth-separable attention residual block.

6. The method according to claim 5, characterized in that The encoders in the first and second levels of the four levels are both provided with the attention residual block, the encoders in the third and fourth levels of the four levels are both provided with the depthwise separable attention residual block, and the fourth level is provided with at least one Dropout layer.

7. The method according to claim 5, characterized in that The decoders in the first and second layers of the four layers are both provided with the attention residual block, the decoders in the third and fourth layers of the four layers are both provided with the depth-separable attention residual block, and the fourth layer is provided with at least one Dropout layer.

8. The method according to claim 2, characterized in that The pretreatment includes: Performing atmospheric correction and radiometric calibration on the training image, and generating a label map of the training image based on an object-oriented algorithm; The label image and the training image are randomly cropped at the same time to obtain a training picture set, and data enhancement is performed on all pictures in the training picture set to obtain a training data set.

9. The method according to claim 1, characterized in that The initial improved U-Net network model is trained using a training data set to obtain a target U-Net network model, including: Inputting the training data set into the initial improved U-Net network model and setting a maximum number of training times; Select the target optimization function and calculate the network loss; updating network parameters using an optimizer based on the network loss; Repeat the above steps until the network is fitted or the maximum number of training times is reached; Save the trained target U-Net network model.

Images