A method for crack identification in electro-imaging images based on an improved FCN semantic segmentation model

Abstract

This invention discloses a crack identification method for electro-imaging images based on an improved FCN semantic segmentation model. The method includes: constructing a pixel-level labeled dataset for deep learning training using a superpixel segmentation algorithm; performing data augmentation on the pixel-level labeled dataset using image rotation; constructing an improved FCN semantic segmentation model; training the improved FCN semantic segmentation model using the pixel-level labeled dataset; and using the trained improved FCN semantic segmentation model to identify cracks in the electro-imaging image to be identified. This invention integrates channel and spatial cross-attention modules into the semantic segmentation model to obtain semantic dependencies in the image. The spatial cross-attention reduces the computational cost of the model. Secondly, ResNet-50 is used as the backbone extraction network, which has sufficient depth to fully extract contextual information from the FMI image. Finally, superpixel segmentation is used to assist manual image calibration, improving accuracy and efficiency.

Description

Technical Field

[0001] This invention relates to a crack identification method for electrical imaging images based on an improved FCN semantic segmentation model, belonging to the field of oil and gas exploration and development. Background Technology

[0002] The exploration and development of oil and gas reserves are crucial for maintaining national energy security and alleviating global energy shortages. Fractures in oil and gas reservoirs are important storage spaces and seepage channels, significantly impacting the design of oil and gas accumulation and extraction schemes, single-well productivity, and the overall development efficiency of oil and gas fields. Fracture detection technology plays a vital role in oil and gas exploration and development, providing scientific evidence for oil and gas field exploitation and assessing the potential and economic value of oil and gas resources. Electromagnetic imaging logging images are most widely used in fracture identification; they are a high-precision and intuitive method for displaying wellbore and formation distribution, characterizing fracture geometry, and distinguishing various geological features.

[0003] Traditional fracture identification typically involves manually picking key points and fitting a sine curve. In 1990, D. Torres et al. proposed using the Hough transform to automatically identify fractures, dip angles, and other information in FMI images. The principle is that well logging images can be unfolded based on cylindrical geometry, and the shape of fractures in well logging images is considered to be a sine curve. The two-dimensional Hough transform can be used to fit the fracture curve of a fixed pattern in the image. However, the Hough transform has the disadvantages of large computational cost and large space consumption.

[0004] Introducing deep learning-based semantic image segmentation into fracture identification in electrical imaging logging images is beneficial for the accurate acquisition of fracture features. However, while these networks reduce computational cost, they all employ ordinary convolutions, resulting in a small receptive field that limits contextual information and consequently reduces accuracy. Furthermore, they merely utilize the aggregation of contextual information without leveraging the semantic relationships between image pixels, thus failing to identify complex fracture information in logging images. Summary of the Invention

[0005] In order to overcome the shortcomings of the existing technology, the present invention aims to provide a method for crack recognition in electro-imaging images based on an improved FCN semantic segmentation model.

[0006] The technical solution provided by this invention to solve the above-mentioned technical problems is: a method for crack recognition in electro-imaging images based on an improved FCN semantic segmentation model, comprising the following steps:

[0007] S1. Construct a pixel-level labeled dataset for deep learning training using the superpixel segmentation algorithm;

[0008] S2. Use image rotation to augment the pixel-level labeled dataset;

[0009] S3. Construct an improved FCN semantic segmentation model;

[0010] S4. The improved FCN semantic segmentation model is trained using a pixel-level labeled dataset;

[0011] S5. Use the trained and improved FCN semantic segmentation model to identify cracks in the electro-imaging image to be identified.

[0012] A further technical solution is that the specific process of step S1 is as follows:

[0013] S11. Based on the pixel size and the training requirements of the deep learning model, produce several electrical imaging images with typical crack features, each with a pixel size of 256×256.

[0014] S12. Use the superpixel segmentation algorithm to preprocess the electro-imaging image and obtain the boundary of the original image;

[0015] S13. Finally, the cracks are color-coded, and pixel-level labeled data is generated by manual labeling, which is the pixel-level labeled dataset used for deep learning training.

[0016] A further technical solution is that the specific process of step S12 is as follows: using a simple linear iterative clustering algorithm to cluster pixels based on color similarity and proximity to generate superpixels, and segmenting the image to obtain the boundary of the original image.

[0017] A further technical solution is that the simple linear iterative clustering algorithm includes the following steps:

[0018] 1. Initialization: First, some seed points need to be set in the image. Then, the image is divided into M superpixels of equal size, each superpixel being [size missing].

[0019] 2. Reselect seed point: First, calculate the gradient values ​​of all neighborhood points, and then move the seed point to the location with the smallest gradient among all neighborhood points;

[0020] 3. Simple linear iterative clustering optimization: First, aggregate similar pixels into a superpixel; then, update the center of each cluster region, and iteratively execute the above two steps until the cluster center no longer changes.

[0021] A further technical solution is that, in step 3, the distance D between each pixel and its surrounding seed points is calculated, and then the seed point corresponding to the minimum value is taken as the cluster center of that pixel.

[0022] The formula for calculating the distance D between each pixel and its surrounding seed points is:

[0023]

[0024]

[0025]

[0026] In the formula: L represents brightness; a represents the range from red to green; b represents the range from yellow to blue; d ω Indicates color distance; x and y represent pixel position, d λ represents the Euclidean distance between pixels; S represents the size of each superpixel.

[0027] A further technical solution is that, in step S2, the manually labeled images are rotated by 90°, 180°, and 270° to triple the size of the dataset. After being tripled, the dataset is combined with the unrotated images to form a new dataset.

[0028] A further technical solution is that the improved FCN semantic segmentation model comprises three parts: a backbone network, a fusion channel, a spatial cross-attention module, and an upsampling layer.

[0029] The backbone network is used to extract features and input these features into pooling layers, upsampling layers and convolutional layers to generate pixel-level segmentation output;

[0030] The fusion channel and spatial cross-attention module comprises two parts: a spatial cross-attention module and a channel attention module.

[0031] The channel attention module consists of a max pooling layer, an average pooling layer, and a 3-layer perceptron, wherein the 3-layer perceptron consists of two fully connected layers.

[0032] The spatial cross-attention module consists of a set of 1×1 convolutional layers and an association matrix layer.

[0033] A further technical solution is that, in step S4, 70% of the pixel-level labeled dataset is randomly selected as the training set, and the remaining 30% is used as the validation dataset.

[0034] A further technical solution is that, in step S4, the batch size of the training set is set to 16; the initial learning rate is set to 0.01, and a learning rate decay strategy is adopted to dynamically adjust the learning rate in order to better guide the training process of the model; the number of iterations is set to 100, and experiments are conducted using the above parameter settings to statistically analyze the evaluation metrics.

[0035] A further technical solution is that the evaluation metrics include training loss, accuracy, and intersection-over-union ratio.

[0036] The present invention has the following beneficial effects: First, it integrates the channel and spatial cross attention (FCSA) module into the semantic segmentation model, which can obtain the semantic dependencies in the image. The purpose of spatial cross attention can reduce the computational load of the model. Second, it uses ResNet-50 as the backbone extraction network, which has a certain depth and can fully extract the contextual information in the FMI image. Finally, it uses superpixel segmentation to assist manual image labeling, which can improve accuracy and efficiency. Attached Figure Description

[0037] Figure 1 The principle of electrical imaging logging images;

[0038] Figure 2 A crack recognition model diagram integrating channel and spatial intersection attention modules;

[0039] Figure 3 A schematic diagram of the fusion channel and spatial cross-attention module;

[0040] Figure 4 This is a typical FMI image dataset example;

[0041] Figure 5 This is a portion of the FMI image after processing using a superpixel segmentation algorithm;

[0042] Figure 6 The image is an FMI image with manual annotation;

[0043] Figure 7 The data graphs are after rotation by 90°, 180°, and 270°.

[0044] Figure 8 Training loss curves for different semantic segmentation models;

[0045] Figure 9 The training accuracy curves for different semantic segmentation models are shown.

[0046] Figure 10 The following is a comparison of experimental results. Detailed Implementation

[0047] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] The present invention provides a crack identification method for electro-imaging images based on an improved FCN semantic segmentation model, comprising the following steps:

[0049] S1. Construct a pixel-level labeled dataset for deep learning training using the superpixel segmentation algorithm;

[0050] S11. Based on the pixel size and the training requirements of the deep learning model, produce 504 electrical imaging images with typical crack features, each with a pixel size of 256×256.

[0051] S12. Use a simple linear iterative clustering algorithm to cluster pixels based on color similarity and proximity to generate superpixels, and segment the image to obtain the boundaries of the original image.

[0052] The simple linear iterative clustering algorithm includes the following steps:

[0053] 1. Initialization: First, some seed points need to be set in the image. Then, the image is divided into M superpixels of equal size, each superpixel being [size missing].

[0054] 2. Reselect seed point: First, calculate the gradient values ​​of all neighborhood points, and then move the seed point to the location with the smallest gradient among all neighborhood points;

[0055] 3. Simple linear iterative clustering optimization: First, similar pixels are aggregated into a superpixel. To improve the running speed, the aggregation operation is only performed on each seed point within a (2S·2S) region. Then, the center of each cluster region is updated. The above two steps are iteratively executed until the cluster center no longer changes.

[0056] In the SLIC algorithm, a key step is calculating the similarity between pixels, which includes color similarity and spatial similarity. For each pixel being searched, its similarity to the seed point is calculated separately, as follows:

[0057]

[0058]

[0059]

[0060] In the formula: L represents brightness; a represents the range from red to green, with a value range of [127, -128]; b represents the range from yellow to blue, with a value range of [127, -128]; d ω Indicates color distance; x and y represent pixel position, d λ S represents the Euclidean distance between pixels; S represents the size of each superpixel; D defines the distance between pixels, which can characterize the similarity between pixels;

[0061] S13. Finally, the cracks are color-marked, and pixel-level labeled data is generated by manual labeling, which is the pixel-level labeled dataset used for deep learning training.

[0062] S2. To increase the amount of training data, this invention employs image rotation for data augmentation. Manually labeled images are rotated by 90°, 180°, and 270°, tripling the size of the dataset. This tripled dataset, combined with the unrotated images, forms a new dataset of 2016 images.

[0063] S3. Construct an improved FCN semantic segmentation model;

[0064] The improved FCN semantic segmentation model consists of three parts: a backbone network, a fusion channel and a spatial cross-attention module, and an upsampling layer.

[0065] The backbone network is used to extract features and input these features into pooling layers, upsampling layers, and convolutional layers to generate pixel-level segmentation output. The backbone network uses ResNet-50, which increases network depth, improves the model's semantic segmentation performance, and reduces the number of parameters. Dilated convolutions are applied to increase the effective receptive field of the convolutional kernels, enabling the model to perceive broader contextual information without increasing the number of parameters or computational cost. Then, the fusion channel and spatial cross-attention module of this invention are introduced to construct semantic dependencies between FMI image pixels with a small increase in computation. The upsampling layer can restore the obtained high-dimensional features back to the original image size through deconvolution and bilinear interpolation operations, resulting in a semantically segmented image that includes cracks and background.

[0066] The fusion channel and spatial cross-attention module comprises two parts: a spatial cross-attention module and a channel attention module. The working process is as follows: first, the pre-encoder provides the input features F for channel attention, and then the intermediate features Fi are obtained. M The process involves spatial cross-attention, ultimately yielding the output feature F'.

[0067] The channel attention module consists of a max pooling layer, an average pooling layer, and a 3-layer perceptron. The 3-layer perceptron consists of two fully connected layers. The feature map F is first subjected to average pooling and global pooling along its spatial dimension, reducing its spatial dimension to a single channel attention vector. This vector is then fed into the 3-layer perceptron. The first fully connected layer projects it onto a lower-dimensional space for dimensionality reduction, and the second fully connected layer maps the lower-dimensional space back to the original channel dimension. This process generates two C×1×1 attention vectors, which are then weighted according to their importance for each channel of the feature map, where C represents the number of channels. Finally, a weighted sum is used to obtain the channel attention map, and the input feature F is multiplied element-wise by this sum to obtain the intermediate feature F.M ;

[0068] Channel attention M c The calculation formula is:

[0069] M c (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F)))

[0070] In the formula: F is the input feature map received by the channel attention module; σ is the Sigmoid function; MLP is a 3-layer perceptron with activation function R. e LU and AvgPool(F) are average pooling features, and MaxPool(F) is the max pooling feature;

[0071] The spatial cross-attention module consists of a set of 1×1 convolutional layers and an association matrix layer; the intermediate feature F M It can be represented as a tensor of shape (C, H, W), where C represents the number of feature channels, H and W represent the height and width of the feature map, and the intermediate feature F... M The features are mapped to new features Q, K, V through three independent 1×1 convolutional layers, with the shape (C', H, W), where C represents the number of channels of the new features, the query tensor Q represents the current position or region of interest in the input image, the key tensor K represents the features at all positions in the image, and the value tensor V is used to weight the importance of different positions in the input image.

[0072] Next, the attention score is calculated using the dot product between the query tensor and the key tensor, resulting in a shape of (H+W-1,H,W); where a vector Q of size (C',1,1) can be obtained at position l in the space dimension of the query tensor Q. l In the key tensor K, H+W-1 vectors can be extracted from position l and the remaining positions in the same row and column, resulting in a set K of size (C', H+W-1). l The attention score at position l can then be expressed as:

[0073]

[0074] Among them, A l If the size is (H+W-1,1,1), then the total attention score for all positions in the same row and column is A, with a size of (H+W-1,H,W). Then, by normalizing the attention scores along the channel dimension using the softmax function, the spatial attention map can be obtained. Therefore, the spatial cross-attention between position l of the query tensor Q and position i in the same row and column of the key tensor K is:

[0075]

[0076] Finally, obtain the importance set Vi of the value tensor Vi in the same row and column. l The size is (C', H+W-1). Multiplying it by the correlation matrix and summing the results yields the attention feature vector at position l. This vector is then compared with the original feature A at the corresponding position. l Attention is calibrated in the spatial dimension.

[0077] In summary, the spatial cross-attention module M s The calculation formula is:

[0078]

[0079] S4. The improved FCN semantic segmentation model is trained using a pixel-level labeled dataset;

[0080] The dataset contains 2016 FMI images, labeled with VOC (Visual Object Classification Format), which includes two object classes: black for background and white for cracks. During each training iteration, 70% of the dataset is randomly selected as the training set, and the remaining 30% is used as the validation set.

[0081] During network training, setting the batch size of the training set to 16 strikes a good balance between computer performance and model training effectiveness. The initial learning rate was set to 0.01, and a learning rate decay strategy was used to dynamically adjust the learning rate to better guide the model's training process. The number of iterations was set to 100. The choice of the number of iterations requires a trade-off between model convergence and training time; too few iterations may result in the model failing to fully learn the features of the data, while too many iterations may lead to overfitting. Experiments were conducted using the above parameter settings, and the evaluation metrics were statistically analyzed.

[0082] This invention uses training loss, accuracy, and the intersection-union ratio (IU), commonly used in semantic segmentation, as evaluation metrics. The training loss is used to determine how close the actual output value is to the expected output value, while the accuracy reflects the training precision of the model.

[0083] S5. Use the trained and improved FCN semantic segmentation model to identify cracks in the electro-imaging image to be identified.

[0084] Example

[0085] To verify the effectiveness of the model in this invention for crack segmentation in FMI images, it was compared with existing classic semantic segmentation models, including U-Net, PSPNet, SegNet, and DeepLabv3. All models were trained and tested using the same experimental settings. Figure 8The training loss curves for various semantic segmentation models on the dataset are shown, with each model trained for 100 epochs. The losses of all five models initially showed a rapid decrease, but then stabilized. The model from this invention exhibited the smoothest curve and the lowest loss, with a final value of 0.08. PSPNet and DeepLabv3 followed, with relatively smooth curves and final losses of 0.11 and 0.14, respectively. U-Net and SegNet showed more volatile curves and higher final losses of 0.23 and 0.25, respectively. The loss curves demonstrate that, compared to other semantic segmentation models, the actual output value of this invention matches the expected value more closely.

[0086] like Figure 9 The training accuracy curves of various semantic segmentation models on the FMI crack recognition dataset are shown. The accuracy variation curves reflect the differences in accuracy among different semantic segmentation models with the same number of training samples. The results show that the accuracy of each model initially rises rapidly and then stabilizes, with the U-Net model achieving the lowest final accuracy of 93.93%. The accuracy of the other models is relatively similar, with DeepLabv3, SegNet, and PSPNet achieving final accuracies of 96.83%, 96.83%, and 96.59%, respectively. The model of this invention achieves the highest accuracy, reaching 97.02%. This indicates that, with the same number of training samples, the model of this invention stabilizes first and achieves the highest accuracy relative to other semantic segmentation models.

[0087] Although the training accuracy and training loss are similar to classic semantic segmentation models such as DeepLabv3, the model of this invention strikes a good balance between network accuracy, training loss, and Intersection over Union (MIoU). The MIoU directly determines the quality of the final prediction result. The MIoU was used to evaluate each model, and the results are shown in Table 1. It can be seen that the MIoU of the model of this invention is higher than other semantic segmentation models, reaching 74.61%, which is 3.15 percentage points higher than DeepLabv3. A higher MIoU indicates better model segmentation performance; therefore, the model of this invention can effectively identify crack information in FMI images.

[0088] Table 1

[0089] Model Intersection and Union U-Net 65.38 SegNet 65.15 DeepLabv3 67.32 PSPNet 71.46 Ours 74.61

[0090] Comparative experimental results are as follows Figure 10As shown in the figure, DeepLabv3 and U-Net exhibit misclassification when the original FMI image contains noise similar to cracks. On the other hand, some FMI images have unclear cracks, and DeepLabv3, PSPNet, SegNet, and U-Net show inconsistent crack prediction performance. Figure 10 The third row of images is particularly noticeable. In contrast, the model of this invention can effectively avoid noise similar to cracks and the phenomenon of discontinuous cracks in the images, and has the least amount of noise and interference information. It can improve the accuracy of crack recognition based on intuitive judgment alone, proving the effectiveness of the model in this paper.

[0091] The above description is not intended to limit the present invention in any way. Although the present invention has been disclosed through the above embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the present invention shall still fall within the scope of the present invention.

Claims

1. An electrical imaging image crack identification method based on an improved FCN semantic segmentation model, characterized in that, The method comprises the following steps: S1, constructing a pixel-level labeling dataset for deep learning training by using a superpixel segmentation algorithm; S2, performing data enhancement on the pixel-level labeling dataset by using an image rotation method; S3, constructing an improved FCN semantic segmentation model; The improved FCN semantic segmentation model comprises a backbone network, a fusion channel, a spatial cross-attention module and an upsampling layer; The backbone network is used for extracting features and inputting the features into a pooling layer, an upsampling layer and a convolutional layer to generate a pixel-level segmentation output; the backbone network adopts ResNet-50, which can deepen the network depth, improve the semantic segmentation performance of the model and reduce the number of parameters of the model; and an dilated convolution is applied to increase the effective receptive field of a convolution kernel, so that the model can perceive wider context information without increasing the number of parameters and the calculation cost; Then, the fusion channel and the spatial cross-attention module are introduced to construct the semantic dependency relationship between FMI image pixels with a small amount of calculation; The upsampling layer restores the obtained high-dimensional features to the original image size through deconvolution and bilinear interpolation operation to obtain a semantic segmentation picture including cracks and backgrounds; The fusion channel and space cross attention module comprises two parts, namely a space cross attention module and a channel attention module; the working process is that first, the input feature of the channel attention is given by the pre-encoder , then the intermediate feature is obtained ; the final output feature is obtained by entering the space cross attention The channel attention module is composed of a max-pooling layer, an average pooling, and a 3-layer perceptron composed of two fully connected layers; the feature map First, average pooling and global pooling are performed along the spatial dimension, reducing the spatial dimension of the feature map to a single channel attention vector; then it is sent to a 3-layer perceptron, the first fully connected layer projects it to a lower dimensional space for dimension reduction, and the second fully connected layer maps the low-dimensional space back to the original channel dimension, generating two ×1×1 attention vectors, which assign weights to each channel of the feature map according to their importance, where represents the number of channels; finally, the channel attention map is obtained by weighted summation, and the input feature is multiplied element by element with it to obtain the intermediate feature ; where channel attention The calculation formula is: In the formula: is the input feature map accepted by the channel attention module; is a Sigmoid function; is a 3-layer perceptron and the activation function is R e LU , is the average-pooled feature, is the max-pooled feature; The spatial cross-attention module is composed of a group of 1x1 convolutional layers, an association matrix layer; the intermediate feature is represented as a tensor with shape , where represents the number of feature channels, and represent the height and width of the feature map, the intermediate feature is mapped into a new feature by three independent 1x1 convolutional layers, with shape , where represents the number of new feature channels, the query tensor represents the current position or the region of interest in the input image, while the key tensor represents the features of all positions in the image, and the value tensor is used to weight the importance of different positions in the input image; S4, training the improved FCN semantic segmentation model by using the pixel-level labeling dataset; S5, identifying cracks in an electric imaging image to be identified by using the trained improved FCN semantic segmentation model.

2. The method of claim 1, wherein the improved FCN semantic segmentation model is based on a U-Net model. The specific process of the step S1 is as follows: S11, according to the pixel size and the training requirement of the deep learning model, preparing a plurality of electric imaging images with typical crack features, and the pixels are 256*256; S12, performing image preprocessing on the electric imaging images by using a superpixel segmentation algorithm to obtain the boundary of the original image; S13, finally, performing color labeling on the cracks, and generating a pixel-level labeling dataset by using an artificial labeling method, that is, a pixel-level labeling dataset for deep learning training.

3. The method of claim 2, wherein the improved FCN semantic segmentation model is based on a U-Net model. The specific process of the step S12 is as follows: using a simple linear iterative clustering algorithm to produce superpixels by clustering pixels based on color similarity and proximity, and segmenting the image to obtain the boundary of the original image.

4. The method of claim 3, wherein the improved FCN semantic segmentation model is based on a U-Net model. The simple linear iterative clustering algorithm comprises the following steps: Step 1, initialization: first, some seed points need to be set in the image, then the image is divided into M superpixels of the same size, and the size of each superpixel is ; Step 2, reselecting seed points: first, calculating the gradient values of all field points, and then moving the seed points to the place with the smallest gradient among all field points; Step 3, simple linear iterative clustering optimization: first, aggregating similar pixels into a superpixel; then, updating the center of each cluster region, and iteratively executing the above two steps until the clustering center no longer changes.

5. The method of claim 4, wherein the improved FCN semantic segmentation model is based on a U-Net model. In the step 3, the distance D between each pixel point and the surrounding seed points is calculated, and then the seed point corresponding to the minimum value is taken as the clustering center of the pixel point; The calculation formula of the distance D between each pixel point and the surrounding seed points is as follows: In the formulae: represents luminance; represents a range from red to green; represents a range from yellow to blue; represents color distance; , represents pixel position, represents the Euclidean distance of the pixel; represents the size of each superpixel.

6. The method of claim 1, wherein the improved FCN semantic segmentation model-based electrical imaging image crack identification method is characterized by, In the step S2, the artificially labeled image is rotated by 90°, 180° and 270°, so that the dataset is expanded by three times, and the expanded three times and the image without rotation constitute a new dataset.

7. The method of claim 1, wherein the improved FCN semantic segmentation model is based on a U-Net model. The improved FCN semantic segmentation model comprises a backbone network, a fusion channel, a spatial cross-attention module and an up-sampling layer; The fusion channel and the spatial cross-attention module comprise two parts, namely a spatial cross-attention module and a channel attention module; The channel attention module is composed of a maximum pooling layer, an average pooling layer and a 3-layer perceptron, and the 3-layer perceptron is composed of two fully connected layers; The spatial cross-attention module is composed of a group of 1x1 convolution layers and a correlation matrix layer.

8. The method of claim 1, wherein the improved FCN semantic segmentation model is based on a U-Net model. In the step S4, 70% of the pixel-level labeled data set is randomly selected as a training set, and the remaining 30% is used as a validation data set.

9. The method of claim 8, wherein the improved FCN semantic segmentation model is based on a U-Net model. In the step S4, the batch size of the training set is set to 16; the initial learning rate is set to 0.01, and a learning rate decay strategy is used to dynamically adjust the learning rate to better guide the training process of the model; The number of iterations is set to 100 times, and the evaluation indexes are counted through the above parameter settings.

10. The method of claim 9, wherein the improved FCN semantic segmentation model is based on a U-Net model. The evaluation indexes include training loss, accuracy and intersection over union.

Images