Dam disaster point identification method based on improved DeepLabV < 3 + >

Abstract

The invention relates to a dam disaster point identification method based on improved DeepLabV < 3 + >, and the method comprises the following steps: Step 1: data preprocessing: changing a data set composed of dam images into a size which can be trained by a network according to the demands; step2, establishing a dam disaster point identification model based on the improved DeepLabV < 3 + >, applying a lightweight network MobileNetV2 in the model, optimizing the expansion rate of a cavity space pyramid pooling module, and introducing a channel attention mechanism and a space attention mechanism, and Step3, training by using the dam disaster point identification model based on the improved DeepLabV < 3 + >, and obtaining a dam disaster point image segmentation result. According to the algorithm, the recognition precision of dam disaster points is improved by introducing an adaptive attention mechanism, a feature extraction module in the algorithm adopts specific cavity convolution configuration, the capture capability of small-scale disaster features is enhanced, and a decoder structure is improved, so that more efficient feature fusion is realized, and the accuracy of semantic segmentation is improved.

Description

Technical Field

[0001] The present invention relates to an identification method, in particular to a dam disaster point identification method based on improved DeepLabV3+, and belongs to the technical field of artificial intelligence and water conservancy project safety monitoring. Background Art

[0002] With the rapid development of water conservancy projects, dams, as important flood control facilities, are of vital importance to the safety of people's lives and property. At present, the identification of dam disaster points mainly relies on manual inspections, which is not only inefficient but also easily affected by human factors, resulting in inaccurate identification or missing important disaster information.

[0003] In order to improve the efficiency and accuracy of identifying dam disaster points, some studies have begun to explore automated identification methods. Deep learning technology, especially convolutional neural networks (CNNs), has attracted widespread attention due to its outstanding performance in the field of image recognition. DeepLabV3+ is an advanced semantic segmentation network that improves the accuracy and robustness of segmentation by introducing dilated convolution and feature fusion strategies. However, the existing DeepLabV3+ algorithm still has certain limitations in identifying dam disaster points, such as insufficient recognition of small targets and complex backgrounds, and high demand for computing resources in practical applications. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention proposes an algorithm based on improved DeepLabV3+, aiming to improve the recognition accuracy and efficiency of dam disaster points. The present invention effectively improves the algorithm's recognition ability of dam disaster points by introducing specific image preprocessing technology, optimizing network structure and adjusting training strategy, especially in small target detection and application performance in complex environments. In addition, the present invention also takes into account the practicality of the algorithm, and makes it more suitable for actual dam monitoring scenarios by reducing the amount of calculation and increasing the running speed.

[0005] In order to achieve the above object, the technical solution of the present invention is as follows: a method for identifying dam disaster points based on improved DeepLabV3+, the method is as follows:

[0006] Step 1: Data preprocessing: convert the dataset consisting of dam images into a size that can be trained by the network according to the requirements;

[0007] Step 2: Establish a dam disaster point recognition model based on the improved DeepLabV3+, apply the lightweight network MobileNetV2 in the model, optimize the expansion rate of the void space pyramid pooling module, and quote the channel attention mechanism and spatial attention mechanism.

[0008] Step 3: Use the dam disaster point recognition model based on the improved DeepLabV3+ for training to obtain the dam disaster point image segmentation results.

[0009] The specific process in Step 1 is as follows:

[0010] Step 1.1 Remove highly similar disaster point images;

[0011] Step 1.2 Use the labeling tool labelme to complete the labeling of the disaster point image;

[0012] Step 1.3 divides the labeled dam images into training samples, verification samples and test samples in a ratio of 7:2:1.

[0013] The specific process in Step 2 is as follows:

[0014] Step 2.1 replaces the feature extraction network Xception in the original DeepLabV3+ network with a lightweight network MobileNetV2. MobileNetV2 is composed of multiple bottleneck layers, each of which contains three parts: expansion, convolution, and compression.

[0015] Step 2.1.1 expansion is to expand the number of channels of the dam feature map by using 1×1 convolution;

[0016] Step 2.1.2 convolution uses 3×3 depth-separable convolution to identify the characteristics of each channel, thereby reducing the parameters of the model;

[0017] Step 2.1.3 compression uses 1×1 convolution to reduce the number of channels of the dam feature map.

[0018] Step 2.1.4 In addition, MobileNetv2 also enhances each bottleneck layer through residual connections, thereby improving the training accuracy of the network.

[0019] Step 2.2 optimizes and adjusts the expansion rate of the dilated space pyramid pooling module in the original network;

[0020] The expansion coefficients of the hollow space pyramid pooling in the DeepLabv3+ network structure are 6, 12, and 18. This setting focuses mainly on the receptive field of larger scales and may ignore some small-scale detail information. In the identification of dam disaster points, some minor changes or anomalies may be critical for accurate identification. In addition, setting a larger expansion rate will increase the computational complexity, especially in resource-constrained environments, such as mobile devices, which will affect the speed of identifying dam disaster points. In response to the above problems, the present invention adjusts the expansion coefficient of the hollow space pyramid pooling. The adjusted expansion coefficients are 1, 3, and 5. This coefficient can reduce the loss of detail information and speed up the recognition speed and accuracy.

[0021] Step 2.3 introduces channel attention mechanism and spatial attention mechanism;

[0022] Step 2.3.1 After adjusting the number of channels through 1×1 convolution, the channel attention mechanism module is used to weight the channels to further enhance the feature expression and obtain the information of key dam disaster points;

[0023] Step 2.3.2 In the decoder, adjust the number of low-order feature channels with rich detail information extracted by the backbone extraction network with 4 times downsampling through 1×1 convolution. The high-order features obtained in the encoder are upsampled by 4 times to make the feature map size the same as the low-order feature map size;

[0024] Step 2.3.3 stacks the low-order features and high-order features, and then adjusts the obtained feature map to the number of channels through 3×3 convolution;

[0025] Step 2.3.4 Use the spatial attention mechanism to process the feature map after adjusting the channel, reduce the calculation of irrelevant areas, strengthen the attention to the target area, and ignore the areas irrelevant to the dam disaster point;

[0026] Step 2.3.5 Finally, upsample by 4 times to make the feature map resolution the same as the input dam image resolution.

[0027] The specific process in Step 3 is as follows:

[0028] Step 3.1 Select ReLU as the activation function;

[0029] Step 3.2 Select Adam as the optimizer;

[0030] Step 3.3 uses the cross entropy loss function;

[0031] Step 3.4 selected mean Inetersection over Union (mIoU), mean pixel accuracy (MPA), and Dice coefficient to evaluate the training model.

[0032] Among them, Step 3.4 is as follows:

[0033] Step 3.4.1 Mean Intersection over Union (mIoU),

[0034] In semantic segmentation models, average intersection-over-union is widely considered to be a key evaluation criterion. Its core concept is to calculate the ratio of the intersection and union of all recognition result sets in two categories. In particular, in the dam hazard point recognition task in this paper, the intersection and union represent the original label and predicted label of the pixel. If the intersection is defined as TP and the union is represented as the sum of TP, FP and FN, then the intersection-over-union ratio of each type can be calculated and averaged to obtain the average intersection-over-union ratio. The calculation of the average intersection-over-union ratio is shown in formula (1).

[0035]

[0036] Step 3.4.2 Average pixel accuracy (MPA),

[0037] PA is the ratio of the number of correctly predicted pixels to the total number of pixels, and the calculation formula is as follows:

[0038]

[0039] Among them, k is the number of target categories, p ii is originally class i and is also predicted to be class i, p ij It means that the original class is i and is predicted to be class j.

[0040] MPA is the sum of all PA categories and then the average value. The calculation formula is:

[0041]

[0042] Step3.4.3Dice coefficient,

[0043] The Dice coefficient is a statistical indicator used to evaluate the similarity of two data sets, usually used to measure the degree of overlap between two sample sets. The value of this coefficient ranges from 0 to 1. The closer the value is to 1, the higher the similarity between the two data sets, while the closer it is to 0, the lower the similarity between the two data sets.

[0044]

[0045] A computer-readable storage medium having computer instructions stored thereon, characterized in that: when the computer instructions are executed by a processor, the steps of the dam disaster point identification method based on the improved DeepLabV3+ are implemented.

[0046] Compared with the prior art, the present invention has the following advantages: 1) The technical solution proposes an improved DeepLabV3+ disaster point recognition algorithm. First, the feature extraction network Xception in the original DeepLabV3+ network is replaced by the lightweight network MobileNetV2, which not only reduces the parameters required for the model, but also improves the segmentation efficiency of the model. Then, the expansion rate of the atrous spatial pyramid pooling module (ASPP) in the original network is optimized and adjusted. After the adjustment, the number of parameters of the model is significantly reduced, and it is also ensured that the convolution layer can complete the extraction of dense features under a wider perspective. Finally, after 1×1 convolution in the encoding stage and 3×3 convolution in the decoding stage, the channel attention mechanism (CA) and spatial attention mechanism (SA) are introduced to enhance the model's feature expression ability; 2) The network model proposed in the present invention is superior to PSPNet and DeepLabV3+ in terms of MPA, mIoU, and Dice coefficient indicators. The indicators of the UNet method and the method in this paper are close, but the UNet method has the phenomenon of misidentifying the correct area as the disaster point area. In comparison, the network of the present invention can identify the disaster point more accurately, so it is shown that the improved DeepLabV3+ network model is more effective; 3) The algorithm improves the recognition accuracy of the dam disaster point by introducing an adaptive attention mechanism. The feature extraction module in the algorithm adopts a specific hole convolution configuration to enhance the ability to capture small-scale disaster features. The improved decoder structure achieves more efficient feature fusion and improves the accuracy of semantic segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a schematic diagram of the dam disaster point identification model based on the improved DeepLabV3+;

[0048] Figure 2 Schematic diagram of the bottleneck layer;

[0049] Figure 3 Schematic diagram of the process of marking dam hazard points;

[0050] Figure 4 Schematic diagram of experimental comparison visualization results. DETAILED DESCRIPTION

[0051] In order to deepen the understanding of the present invention, the present embodiment is described in detail below with reference to the accompanying drawings.

[0052] Example 1: See Figure 1-Figure 4 , a dam disaster point identification method based on improved DeepLabV3+, the method comprises the following steps:

[0053] Step 1: Data preprocessing: convert the dataset consisting of dam images into a size that can be trained by the network according to the requirements;

[0054] Step 2: Establish a dam disaster point recognition model based on the improved DeepLabV3+, apply the lightweight network MobileNetV2 in the model, optimize the expansion rate of the void space pyramid pooling module, and quote the channel attention mechanism and spatial attention mechanism, such as Figure 1 As shown;

[0055] Step 3: Use the dam disaster point recognition model based on the improved DeepLabV3+ for training to obtain the dam disaster point image segmentation results.

[0056] The specific process in Step 1 is as follows:

[0057] Step 1.1 Remove highly similar disaster point images;

[0058] Step 1.2 Use the labeling tool labelme to complete the labeling of the disaster point image;

[0059] Step 1.3 divides the labeled dam images into training samples, verification samples and test samples in a ratio of 7:2:1.

[0060] The specific process in Step 2 is as follows:

[0061] Step 2.1 replaces the feature extraction network Xception in the original DeepLabV3+ network with a lightweight network MobileNetV2. MobileNetv2 is composed of multiple bottleneck layers, such as Figure 2 As shown, each bottleneck layer contains three parts: expansion, convolution and compression;

[0062] Step 2.1.1 expansion is to expand the number of channels of the dam feature map by using 1×1 convolution;

[0063] Step 2.1.2 convolution uses 3×3 depth-separable convolution to identify the characteristics of each channel, thereby reducing the parameters of the model;

[0064] Step 2.1.3 compression uses 1×1 convolution to reduce the number of channels of the dam feature map.

[0065] Step 2.1.4 In addition, MobileNetv2 also enhances each bottleneck layer through residual connections, thereby improving the training accuracy of the network.

[0066] Step 2.2 optimizes and adjusts the expansion rate of the dilated space pyramid pooling module in the original network;

[0067] The expansion coefficients of the hollow space pyramid pooling in the DeepLabv3+ network structure are 6, 12, and 18. This setting focuses mainly on the receptive field of larger scales and may ignore some small-scale detail information. In the identification of dam disaster points, some minor changes or anomalies may be critical for accurate identification. In addition, setting a larger expansion rate will increase the computational complexity, especially in resource-constrained environments, such as mobile devices, which will affect the speed of identifying dam disaster points. In response to the above problems, the present invention adjusts the expansion coefficient of the hollow space pyramid pooling. The adjusted expansion coefficients are 1, 3, and 5. This coefficient can reduce the loss of detail information and speed up the recognition speed and accuracy.

[0068] Step 2.3 introduces channel attention mechanism and spatial attention mechanism;

[0069] Step 2.3.1 After adjusting the number of channels through 1×1 convolution, the channel attention mechanism module is used to weight the channels to further enhance the feature expression and obtain the information of key dam disaster points;

[0070] Step 2.3.2 In the decoder, adjust the number of low-order feature channels with rich detail information extracted by the backbone extraction network with 4 times downsampling through 1×1 convolution. The high-order features obtained in the encoder are upsampled by 4 times to make the feature map size the same as the low-order feature map size;

[0071] Step 2.3.3 stacks the low-order features and high-order features, and then adjusts the obtained feature map to the number of channels through 3×3 convolution;

[0072] Step 2.3.4 Use the spatial attention mechanism to process the feature map after adjusting the channel, reduce the calculation of irrelevant areas, strengthen the attention to the target area, and ignore the areas irrelevant to the dam disaster point;

[0073] Step 2.3.5 Finally, upsample by 4 times to make the feature map resolution the same as the input dam image resolution.

[0074] The specific process in Step 3 is as follows:

[0075] Step 3.1 Select ReLU as the activation function;

[0076] Step 3.2 Select Adam as the optimizer;

[0077] Step 3.3 uses the cross entropy loss function;

[0078] Step 3.4 selected mean Inetersection over Union (mIoU), mean pixel accuracy (MPA), and Dice coefficient to evaluate the training model.

[0079] Among them, Step 3.4 is as follows:

[0080] Step 3.4.1 Mean Intersection over Union (mIoU),

[0081] In semantic segmentation models, average intersection-over-union is widely considered to be a key evaluation criterion. Its core concept is to calculate the ratio of the intersection and union of all recognition result sets in two categories. In particular, in the dam hazard point recognition task in this paper, the intersection and union represent the original label and predicted label of the pixel. If the intersection is defined as TP and the union is represented as the sum of TP, FP and FN, then the intersection-over-union ratio of each type can be calculated and averaged to obtain the average intersection-over-union ratio. The calculation of the average intersection-over-union ratio is shown in formula (1).

[0082]

[0083] Step 3.4.2 Average pixel accuracy (MPA),

[0084] PA is the ratio of the number of correctly predicted pixels to the total number of pixels, and the calculation formula is as follows:

[0085]

[0086] Among them, k is the number of target categories, p ii is originally class i and is also predicted to be class i, p ij It means that the original class is i and is predicted to be class j.

[0087] MPA is the sum of all PA categories and then the average value. The calculation formula is:

[0088]

[0089] Step3.4.3Dice coefficient,

[0090] The Dice coefficient is a statistical indicator used to evaluate the similarity of two data sets, usually used to measure the degree of overlap between two sample sets. The value of this coefficient ranges from 0 to 1. The closer the value is to 1, the higher the similarity between the two data sets, while the closer it is to 0, the lower the similarity between the two data sets.

[0091]

[0092] Example 2: See Figure 1 , Figure 2 , a dam disaster point identification method based on improved DeepLabV3+,

[0093] The details are as follows:

[0094] (1) Dataset: Describes the dataset used to train and test the algorithm, including the source, size, and annotation method of the data.

[0095] According to the actual situation of the dam and the collection requirements, taking into account factors such as obstacle avoidance, coverage, and safety, a reasonable collection route is planned, and the drone and unmanned tracked vehicle are controlled to collect data according to the planned route. During the collection process, the equipment status is monitored in real time, and an alarm will be issued for abnormal conditions that occur during the work of the collection equipment, such as obstruction of the unmanned tracked vehicle and insufficient power of the drone, to ensure the safety and stability of the collection process. This module transmits the collected images to the database in real time through the wireless network, including information such as the shooting location and time of the image. Taking a specific collection task as an example, the manager first determines the collection area of ​​the equipment and plans the specific collection route. While the drone and unmanned tracked vehicle are collecting images, the manager who controls the equipment on the ground monitors the operating status and collection of the equipment in real time. In emergencies such as weather deterioration and dense fog, the manager adjusts the equipment to ensure the safe return of the equipment. After the collection equipment completes the image collection, it will first return to the planned location, and then the manager will check the collection equipment and upload the collected dam image data to the database.

[0096] After completing data collection through the acquisition module, due to factors such as uneven lighting, bad weather conditions, poor shooting angles or equipment failure, the collected images may have incomplete information, blurred images and other problems. These poor quality images will affect the subsequent identification of disaster points, so the images need to be preprocessed. The main functions of the image preprocessing module include image specification adjustment, tilt correction, image denoising, etc. After uploading the collected dam image data to the database, the key step is to denoise the image. The technology used is to call the above-mentioned guided filtering and multi-branch residual attention denoising algorithm for infrared images, and call the above-mentioned residual dense network denoising algorithm for visible light images.

[0097] (2) Network configuration: List the network parameters used in the algorithm, such as the type and number of convolutional layers, the configuration of pooling layers, the output dimensions of fully connected layers, etc.

[0098] 3000 highly similar disaster point images were removed, and the remaining 6400 images were trained. The disaster point images were labeled before training. The labeling tool used was labelme. The labeling process was as follows: Figure 3 shown.

[0099] The image of the dam disaster site is input into the network. In the encoder part, the disaster site image is first downsampled by the depthwise separable convolution (DSC) with different channel numbers in the backbone extraction network to obtain the feature information of the dam disaster site; the feature information is processed by the ASPP module, specifically, the results of the five branches of 1×1 convolution, dilated convolution with dilation rates of 1, 3, and 5, and global average pooling are stacked, and then the number of channels is adjusted by 1×1 convolution, and the channel attention mechanism module is used to weight the channels to further enhance the feature expression and obtain the key dam disaster site information; in the decoder part, the number of low-order feature channels with rich detail information extracted by 4 times downsampling of the backbone extraction network is adjusted by 1×1 convolution. After the high-order features obtained in the encoder are upsampled by 4 times, the feature map size is made the same as the low-order feature map size. The low-order features and high-order features are stacked, and the obtained feature map is adjusted to the number of channels through 3×3 convolution. Then, the spatial attention mechanism is used to process the feature map after adjusting the channels, reducing the calculation of irrelevant areas, strengthening the attention to the target area, and ignoring areas irrelevant to the dam disaster point. Finally, the feature map resolution is made the same as the input dam image resolution through 4 times upsampling.

[0100] (3) Training process: Detailed description of the model training process, including the training cycle, learning rate adjustment strategy, batch size, optimizer selection, etc.

[0101] The annotated dam images were divided into training samples, validation samples and test samples in a ratio of 7:2:1. The network was trained in batches of 100 epochs, with the batch size set to 8, the learning rate set to 0.0001, ReLU selected as the activation function, Adam as the optimizer, and the cross entropy loss function used, and the training lasted for about 20 hours.

[0102] (4) Experimental results: Display the experimental results, including quantitative evaluation indicators (such as accuracy, recall rate, F1 score, etc.) and qualitative analysis (such as confusion matrix, ROC curve, etc.).

[0103] To verify the accuracy and effectiveness of the proposed model, the prediction results of the improved network model were compared with those of the three network models: PSPNet, UNet, and DeepLabV3+. Figure 4 The prediction results of PSPNet, UNet, DeepLabV3+ and the proposed network in the test dataset are shown in Table 1. The performance indicators of disaster point identification of each network model are compared.

[0104] Table 1 Comparison of evaluation indicators of PSPNet, UNet, DeepLabV3+ and this network

[0105] method MPA(%) mIoU(%) Dice PSPNet 80.35 77.19 78.34 UNet 84.35 80.26 82.79 DeepLabV3+ 79.45 76.51 78.62 This article network 86.57 82.39 83.17

[0106] (5) Result analysis: Analyze the experimental results, explain the reasons for the improvement in model performance, and discuss the impact of various design choices on the results.

[0107] Compared with other network models, the network model proposed in this invention has the advantages of MPA, m1oU, Dic e The coefficient indexes reached 86.57%, 82.39%, and 83.17% respectively, which are better than PSPNet and DeepLabV3+. The indexes of the UNet method and the method of the present invention are close, but the UNet method has the phenomenon of misidentifying the correct area as the disaster point area. In comparison, the network in this paper can identify the disaster point more accurately, which shows that the improved DeepLabV3+ network model of the present invention has better effect.

[0108] (6) Optimization and improvement: Propose possible optimization directions and improvement measures, such as further network pruning, application of regularization technology, and integration of multi-task learning.

[0109] Compared with traditional image denoising algorithms, some effects have been achieved, but this is still a long way from the optimal noise reduction level. Therefore, more in-depth research on dam image denoising algorithms is needed to seek methods to achieve more superior results in this field, so as to support the subsequent improvement of the accuracy of dam disaster point identification.

[0110] In addition, a method is needed to classify the identified disaster points in a timely and efficient manner. Subsequently, training a dam disaster point classification model can be considered to achieve dam disaster point classification.

[0111] It should be noted that the above embodiments are not intended to limit the protection scope of the present invention, and equivalent changes or substitutions made on the basis of the above technical solutions all fall within the protection scope of the claims of the present invention.

Claims

1. A method for identifying dam disaster points based on improved DeepLabV3+, characterized in that: The method comprises the following steps: Step 1: Data preprocessing: convert the dataset consisting of dam images into a size that can be trained by the network according to the requirements; Step 2: Establish a dam disaster point recognition model based on the improved DeepLabV3+, apply the lightweight network MobileNetV2 in the model, optimize the expansion rate of the void space pyramid pooling module, and quote the channel attention mechanism and spatial attention mechanism. Step 3: Use the dam disaster point recognition model based on the improved DeepLabV3+ for training to obtain the dam disaster point image segmentation results.

2. The dam disaster point identification method based on improved DeepLabV3+ according to claim 1 is characterized in that: The specific process in Step 1 is as follows: Step 1.1 Remove highly similar disaster point images; Step 1.2 Use the labeling tool labelme to complete the labeling of the disaster point image; Step 1.3 divides the labeled dam images into training samples, verification samples and test samples in a ratio of 7:2:

1.

3. The dam disaster point identification method based on improved DeepLabV3+ according to claim 2 is characterized in that: The specific process in Step 2 is as follows: Step 2.1 replaces the feature extraction network Xception in the original DeepLabV3+ network with a lightweight network MobileNetV2. MobileNetV2 is composed of multiple bottleneck layers, each of which contains three parts: expansion, convolution, and compression. Step 2.1.1 expansion is to expand the number of channels of the dam feature map by using 1×1 convolution; Step 2.1.2 convolution uses 3×3 depth-separable convolution to identify the characteristics of each channel, thereby reducing the parameters of the model; Step 2.1.3 compression uses 1×1 convolution to reduce the number of channels of the dam feature map. Step 2.1.4 In addition, MobileNetv2 also enhances each bottleneck layer through residual connections, thereby improving the training accuracy of the network. Step 2.2 optimizes and adjusts the expansion rate of the dilated space pyramid pooling module in the original network; Step 2.3 introduces channel attention mechanism and spatial attention mechanism; Step 2.3.1 After adjusting the number of channels through 1×1 convolution, the channel attention mechanism module is used to weight the channels to further enhance the feature expression and obtain the information of key dam disaster points; Step 2.3.2 In the decoder part, the number of low-order feature channels with rich detail information extracted by 4-fold downsampling of the backbone extraction network is adjusted through 1×1 convolution, and the high-order features obtained in the encoder are upsampled by 4 times to make the feature map size the same as the low-order feature map size; Step 2.3.3 stacks the low-order features and high-order features, and then adjusts the obtained feature map to the number of channels through 3×3 convolution; Step 2.3.4 Use the spatial attention mechanism to process the feature map after adjusting the channel, reduce the calculation of irrelevant areas, strengthen the attention to the target area, and ignore the areas irrelevant to the dam disaster point; Step 2.3.5 Finally, upsample by 4 times to make the feature map resolution the same as the input dam image resolution.

4. The dam disaster point identification method based on improved DeepLabV3+ according to claim 3 is characterized in that: The specific process in Step 3 is as follows: Step 3.1 Select ReLU as the activation function; Step 3.2 Select Adam as the optimizer; Step 3.3 uses the cross entropy loss function; Step 3.4 selected mean Inetersection over Union (mIoU), mean pixel accuracy (MPA), and Dice coefficient to evaluate the training model.

5. The dam disaster point identification method based on improved DeepLabV3+ according to claim 4 is characterized in that: Step 3.4 is as follows: Step 3.4.1 Average Intersection over Union (mIoU), the intersection is defined as TP, and the union is represented as the sum of TP, FP, and FN. Then calculate each intersection over union ratio and average them to get the average intersection over union ratio. The calculation of the average intersection over union ratio is shown in formula (1). Step 3.4.2 Average pixel accuracy (MPA), PA is the ratio of the number of correctly predicted pixels to the total number of pixels, and the calculation formula is as follows: Among them, k is the number of target categories, p ii is originally class i and is also predicted to be class i, p ij It means that the original i class is predicted to be j class. MPA is the sum of all PA categories and then the average value. The calculation formula is: Step3.4.3Dice coefficient, The Dice coefficient ranges from 0 to 1. The closer the value is to 1, the higher the similarity between the two data sets. The closer to 0, the lower the similarity between the two.

6. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the computer instructions are executed by the processor, the steps of the dam disaster point identification method based on the improved DeepLabV3+ are implemented as described in any one of claims 1 to 5.

Images