CN116645514A - Improved U 2 Ceramic tile surface defect segmentation method of Net - Google Patents

Abstract

The invention relates to an improved U ² ‑Net tile surface defect segmentation method, which belongs to the technical field of salient object segmentation. In order to overcome the deficiencies in the prior art, the present invention aims to provide a method for segmenting surface defects of ceramic tiles that improves ^U2 -Net, including obtaining a data set for detecting surface defects of ceramic tiles; constructing a network for segmenting surface defects of ceramic tiles based on ^U2 -Net Model; through the training set, the tile surface defect segmentation network model is continuously iteratively trained until the network finally converges, and the trained tile surface defect segmentation network model is obtained; the image to be processed is input into the trained tile surface defect segmentation network model , to get the segmented target. The present invention constructs a deep learning network model based on codec structure and multi-scale feature fusion for target segmentation of tile surface defects, so as to improve the segmentation effect of tile surface defects.

Description

An Improved U2-Net Method for Tile Surface Defect Segmentation

technical field

The invention relates to an improved U ² -Net tile surface defect segmentation method, which belongs to the technical field of salient object segmentation.

Background technique

Salient object segmentation is also called salient object detection. The main goal is to distinguish the most obvious areas in the image. Analysis is performed by extracting the target area in the image. At present, it is widely used in the fields of intelligent security, unmanned driving, security monitoring and other fields such as scene object segmentation, human body foreground segmentation, face and human body parsing, and 3D reconstruction technology.

Traditional salient object detection algorithms are generally based on low-level visual features, including center bias, contrast prior, and background prior. Achanata et al. respectively processed two low-level features of brightness and color, and used Gaussian difference function to perform frequency domain filtering, so as to calculate the contrast between the current pixel and the pixels in the surrounding domains of different sizes, so as to determine the salient value of the image pixel. Klein et al. use the K-L divergence in information theory to measure the difference between the center of the feature space of the image and its surrounding features. These algorithms are based on the consideration of local pixel contrast, which can perform well in detecting the edge information of the target, but it is difficult to detect the target as a whole. The detection algorithm based on global contrast is to calculate the contrast of the pixel area relative to all pixels in the image, so that the specific position of the salient target can be detected. For example, Wei et al. proposed the concept of background connectivity. In their method, the global contrast is used to obtain the effective range and calculate the saliency of the image. In Gao et al.’s method, the image after superpixel segmentation is obtained first, then the texture details and the global contrast of the image are calculated in the CIELab color space, and the saliency map is obtained by fusing these two features with the help of multi-core reinforcement learning, and using the filter The prediction map obtained by optimizing the post-processing method. Perazzi et al. first decomposed the image into different blocks. The picture relationship between these blocks is compact and has visual smoothness, and then calculates and evaluates the uniqueness and spatial distribution of these blocks, and generates A saliency map of an image. In addition, some researchers calculate the saliency value of image pixels by constructing a graph model. Based on this, Yang et al. proposed a SOD algorithm based on flow sorting. Their approach is to first use superpixels to segment the image, and then select a suitable background seed point. The similarity between other nodes and the seed point can be calculated. Calculate the similarity and then sort to get the result saliency map. But whether this algorithm can detect the saliency map well in the end is very related to the selection of seed points. Yuan et al. proposed a saliency inversion correction process to remove foreground superpixels near the boundary and prevent saliency inversion, thereby improving the accuracy and robustness of saliency prediction based on boundary priors. Shan et al. use background weight maps to seed manifold sorting, and use a third-order smoothness framework to improve the performance of manifold sorting. In Wu et al., addressing the problem that the previously existing methods may miss target regions when detecting some images with low background contrast, a new propagation model is proposed, which takes into account the use of deformation smoothness constraints. The model locally regularizes the nodes of the image and its surrounding points, and then uses deformable smoothness constraints to prevent potentially erroneous results from propagating.

In short, most of the features extracted by traditional target detection algorithms are relatively simple, and the training process for the saliency value of the picture is not very complicated, and it currently has relatively good performance in many scenarios. However, in a large number of data sets and actual scene images, the environment and picture conditions that need to be detected become more and more complex, and the requirements for detection performance and detection rate are increasing. Traditional methods are gradually difficult to meet these requirements.

Contents of the invention

In order to overcome the defects in the prior art, the present invention aims to provide an improved U ² -Net tile surface defect segmentation method.

The technical solution provided by the present invention to solve the above-mentioned technical problems is: an improved U ² -Net tile surface defect segmentation method, comprising the following steps:

Step S1, obtaining the tile surface defect detection data set, and dividing it into a training set and a testing set;

Step S2, constructing a tile surface defect segmentation network model based on U ² -Net;

The tile surface defect segmentation network model is a six-layer U-shaped structure, including a 6-level encoder, a 5-level decoder, a 2-level multi-scale feature fusion module and a saliency map fusion module;

Among them, the first 4 encoders and the corresponding 4 decoders are composed of the feature extraction structure DCRSU, and the number of layers of each DCRSU decreases as the number of layers of the encoder and decoder increases, that is, the first 4 encoders use DCRSU respectively. -7, DCRSU-6, DCRSU-5, DCRSU-4, the same is true for the first 4 decoders;

The fifth encoder and corresponding decoder use RSU-4F;

The sixth encoder introduces SKnet as the deepest encoder;

Introduce two improved attention gate modules at the input of the 5th decoder and the 4th decoder respectively, fuse the features output by the 6th encoder and the 5th encoder into the 5th decoder, and use the The 4 encoders are fused with the features output by the 5th decoder and input to the 4th decoder; then use 3×3 convolution and sigmoid function from the 6th encoder, the 5th decoder, the 4th decoder , the 3rd decoder, the 2nd decoder and the 1st decoder generate 6 output saliency probability maps; then the logic map of the output saliency map is up-sampled to the same size as the input image, and are fused by cascade operation ;Finally, through the 1×1 convolutional layer and the sigmoid function to generate the final significance probability map;

Step S3, continuously iteratively train the tile surface defect segmentation network model through the training set, until the network finally converges, and obtain the trained tile surface defect segmentation network model;

Step S4: Input the image to be processed into the trained tile surface defect segmentation network model to obtain the segmentation target.

A further technical solution is that in the step S10, the training set and the testing set are divided by 4:1.

A further technical solution is that the DCRSU consists of four parts: an input convolutional layer, an encoder, a decoder, and a residual structure;

The input convolution layer is used to extract local features and transform channels;

In the encoding stage, the last encoder adopts convolution + batch normalization + ReLU activation function structure, and the penultimate layer adopts depth separable convolution + batch normalization + ReLU activation function structure; the remaining encoders use the residual structure to The features extracted by the depth-separable convolution and the input features processed by the attention mechanism module are added and then input to the next feature extraction layer for feature extraction, so that the output features of each level can be focused on channels with more effective feature information and strengthened. The ability to extract effective features at each level and obtain multi-scale feature information;

The residual structure fuses the input layer and the middle layer, and splices features of two different scales;

In the decoding stage, the decoder module takes the concatenated feature map, passes through a 3×3 convolution, a batch normalization layer and a Relu activation function to gradually repair the details and spatial dimensions of the segmented object through upsampling; the final decoder output The feature map is added and fused with the feature map of the input convolutional layer to obtain the final feature map processed by the DCRSU module.

A further technical solution is that the RSU-4F replaces downsampling and upsampling with dilation convolution, and the input feature CxHxW first passes through two modules consisting of convolution + batch normalization layer + Relu, and then goes through dilation volume The products are 1, 2, 4, and 8 in turn, and the size of the feature map remains unchanged throughout the process.

The further technical solution is that SKnet is used as the deepest encoder, and the operation of extracting multi-scale features is to generate two feature maps through a 3×3 grouping/depth convolution and a 3×3 hole convolution of the original feature map. : and /> Then add these two feature maps to generate U;

The generated U generates a 1×1×C feature map through global average pooling. The feature map generates a d×1 vector z through a fully connected layer, and the vector z is converted back to length C through 2 FC layers. For 2 A vector is calculated for softmax in the channel dimension to obtain its respective weight vector, and the weight is multiplied by the two outputs of stage 1 to obtain a new feature map, and the two new feature maps are summed to obtain the final output, which is sent to the next decoder.

A further technical solution is that the two inputs of the improved attention gate module are the current layer xl of the encoder and the next layer g of the decoder, and the input features are CxHxW. They first go through element-by-element addition, and then pass Relu Activation function, get the feature map of CxHxW, then reduce the number of channels to 1 through 1×1 convolution, then normalize the sigmoid activation function to get the attention coefficient, and then restore the size through a 1×1 module, Get the coefficient of CxHxW, and finally use the obtained attention coefficient to multiply the two input feature maps, and then splicing, and send the final feature map to the next decoder module.

A further technical solution is that when training the model in step S3, set the batch size to 16, and use the AdamW optimizer for optimization; first use a learning rate of 1×10 ^-3 for initial training, and then use 1×10 ^-5 The learning rate fine-tuning model.

A further technical solution is that the loss formula in the step S3 is:

In the formula: l _fuse represents the loss of the final predicted probability map, l represents the binary cross-entropy loss, and w represents the weight of each loss.

The present invention has the following beneficial effects: the U ² -Net network structure of the present invention is deep and complex, and the intra-level and inter-level information of different sizes of pictures can be extracted through RSU and skip connections, but invalid features such as non-defect areas are prone to appear in the connection There is still a certain leakage of segmentation due to the retention of information such as the retention of defects and the loss of edge defects. In order to further improve the performance of network segmentation, the present invention mainly improves the network from two aspects: improving the extraction ability of effective features and reducing information loss;

Aiming at the problem of the retention of invalid features such as non-defect areas and the loss of information such as defect edges due to skip connections in the current salient object detection and decoding stage, the present invention can make full use of the context information of the image to reduce information loss and improve the segmentation of salient subjects Effect;

Aiming at the problem of low contrast between target and background and low detection accuracy in the detection of salient targets in the tile surface defect scene, the present invention can make full use of the backbone network's ability to extract effective features, so that the segmentation effect of defective targets in pictures is better.

Description of drawings

Fig. 1 flow chart of the present invention;

Fig. 2 is the network structural diagram of the inventive method;

Fig. 3 is the feature extraction DCRSU module in the method coder-decoder of the present invention;

Fig. 4 is the CA attention module used in the method coder-decoder of the present invention;

Fig. 5 is the feature extraction RSU4-F module in the method coder-decoder of the present invention;

Fig. 6 is the coder module SKnet of the inventive method;

Fig. 7 is the multi-scale feature fusion module AG+ module of the method of the present invention;

Fig. 8 is a schematic diagram of the effect of the method of the present invention.

Implementation

The technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Apparently, the described embodiments are part of the embodiments of the present invention, but not all of them. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

As shown in Figure 1, an improved U ² -Net tile surface defect segmentation method of the present invention includes the following steps:

Step S1, obtain the Magnetic-tile-defect-datasets data set in the field of surface defect detection, and divide it into a training set and a test set, and divide it by 4:1;

Step S2, constructing a tile surface defect segmentation network model based on U ² -Net;

The network is based on the encoding and decoding structure, and adopts the method of side fusion for salient target prediction. As shown in Figure 2, the model consists of 6-level encoder, 5-level decoder, 2-level multi-scale feature fusion module and saliency map fusion module.

The network structure is generally U-shaped. The first 4 encoders and the corresponding 4 decoders are composed of the feature extraction structure DCRSU (as shown in Figure 3). The number of layers of each DCRSU increases with the number of layers of the encoder and decoder. Increase and decrease, that is, the first encoder, the second encoder, the third encoder, and the fourth encoder use DCRSU-7, DCRSU-6, DCRSU-5, DCRSU-4, and the decoder The same is true.

The fifth encoder and the corresponding decoder use RSU-4F (as shown in Figure 4), and F means that the size will not change, that is, only feature extraction is performed.

The present invention introduces SKnet (as shown in FIG. 6 ) as the deepest encoder at the sixth encoder.

In addition, in order to solve the problem that only using skip connections to simulate the global multi-scale context may easily lead to the loss of spatial information, the present invention introduces two AG+ modules at the input of the fifth decoder and the fourth decoder respectively (as shown in Figure 5 shown), the features output by the sixth encoder and the fifth encoder are fused into the fifth decoder, and the features output by the fourth encoder and the fifth decoder are fused into the fourth decoder .

Then use 3×3 convolution and sigmoid function to generate 6 outputs from 6th encoder, 5th decoder, 4th decoder, 3rd decoder, 2nd decoder and 1st decoder Significant probability map Then the logic map of the output saliency map is up-sampled to the same size as the input image, and fused through cascading operations; finally, the final saliency probability map is generated through a 1×1 convolutional layer and a sigmoid function.

DCRSU (shown in Figure 3) consists of 4 parts: input convolutional layer, encoder, decoder and residual structure. And assume three parameters set in DCRSU: input channel number Cin output channel number Cout intermediate module output channel number Cmid.

The input convolutional layer is used to extract local features and convert channels. Taking the input picture 3×512×512 as an example, it convolutes a BN layer and Relu activation function through a 3×3, and the channel rises to become Cout×512×512 (This feature is set to a).

In the encoding stage, the last encoder adopts convolution + batch normalization + ReLU activation function structure, and the penultimate layer adopts depth separable convolution + batch normalization + ReLU activation function structure. The remaining encoders use the residual structure to add the features extracted by the depthwise separable convolution and the input features processed by the CA module, and then input them into the next feature extraction layer for feature extraction, so that the output features of each level can be focused on more The effective feature information channel strengthens the ability to extract effective features at each level and obtains multi-scale feature information.

Taking a as the input feature as an example, it becomes Cin×512×512 through Cout 3×3 convolutions; then through Cmid Cin×1×1 convolutions, a BN layer and Relu activation function, it becomes Cmid× 512×512.

At the same time, a is passed through a CA attention module and a BN layer as an input feature. The CA attention module (as shown in Figure 4) performs average pooling on the input feature map of C×H×W shape channel-by-channel, using (H, 1) and (1, W) pooling kernels on the X and Y axes, respectively Direction pooling encodes each channel, producing feature maps of C×H×1 and C×1×W shapes. The pair of method-aware feature maps generated in this way can enable CA attention to capture long-distance dependencies within one channel, and also help to preserve precise location information, enabling the network to localize objects more accurately. The feature maps extracted above are spliced according to the spatial dimension, and spliced into The feature map of the shape, where r is used to control the reduction rate of the block and has the same effect as in SE. Then the feature map is passed through the F1 convolution transformation function (1×1 convolution) and the nonlinear activation function to generate the f intermediate feature map. After splitting f into two tensors f _h and f _w according to the spatial dimension, the shapes are /> and /> Then perform F _h and F _w convolution transformation functions (1×1 convolution) and Sigmoid activation functions to obtain gh and gw coordinate attention. Then multiply gh and gw with the original input to get an output with the same shape as the input, and then go through a 1×1 convolution and batch normalization layer to make the shape of the output feature map, which is the same as the feature a at the same stage after the depth can be obtained. The shape of the separable convolution output is the same. Finally, the feature a processed by the two channels is added and input to the next feature extraction layer.

The residual structure fuses the input layer and the intermediate layer, and performs a concatenation of features of two different scales.

In the decoding stage, the decoder module takes the spliced feature map, passes through a 3×3 convolution, a batch normalization layer and a Relu activation function to gradually restore the details and spatial dimensions of the segmented object through upsampling. The feature map output by the last decoder is added and fused with the feature map input to the convolutional layer to obtain the final feature map processed by the DCRSU module.

In the RSU-4F module, due to several times of downsampling, the original image is already very small. In order to avoid information loss, RSU-4F no longer performs downsampling (as shown in Figure 5), and replaces downsampling and upsampling with dilation Convolution, the input feature C×H×W first passes through 2 modules composed of convolution + batch normalization layer + Relu, and then after expansion convolution is sequentially 1, 2, 4, 8, the size of the feature map in the whole process constant.

SKnet (Figure 6) is the deepest encoder. The multi-scale feature extraction operation is to pass the original feature map through a 3×3 group/depth convolution and a 3×3 hole convolution (the receptive field is 5×5 ) to generate two feature maps: (yellow in the picture) and /> (green in the picture). Then these two feature maps are added to generate U. The generated U generates a 1×1×C feature map (s in the figure) through global average pooling (Fgp), and the feature map generates a d×1 vector (z in the figure) through a fully connected layer (Ffc function) , change the vector z back to the length C through 2 FC layers respectively, and calculate the softmax of the 2 vectors in the channel dimension to obtain their respective weight vectors, and multiply the weights with the 2 outputs of stage 1 to obtain new features map, the two new feature maps are summed to get the final output, which is sent to the next decoder.

The two inputs of the AG+ module (as shown in Figure 7) are the current layer xl of the encoder and the next layer g of the decoder, and the input features are C×H×W. They are first added element by element and passed through the Relu activation function , get the feature map of C×H×W, then reduce the number of channels to 1 through 1×1 convolution, then normalize the sigmoid activation function to get the attention coefficient, and then restore the size through a 1×1 module Come back, get the coefficient of C×H×W, and finally use the obtained attention coefficient to multiply the two input feature maps, and then splicing, and send the final feature map to the next decoder module.

Step S3, continuously iteratively train the tile surface defect segmentation network model through the training set, until the network finally converges, and obtain the trained tile surface defect segmentation network model;

The experimental equipment uses NVIDIA V100, and the entire model is implemented with PyTorch.

When training the model, set the batch size to 16 and use the AdamW optimizer for optimization. We first use a learning rate of 1× ^10-3 for initial training, and then use a learning rate of 1× ^10-5 to fine-tune the model.

The formula is the calculated loss formula as follows, where M=6, which means the 1st decoder, the 2nd decoder, the 3rd decoder, the 4th decoder, the 5th decoder, and the 6th encoder There are six outputs, l _fuse represents the loss of the final predicted probability map, l represents the binary cross entropy loss, and w represents the weight of each loss.

The binary cross-entropy loss is shown below, where (r,c) are pixel coordinates and (H,W) are image dimensions: height and width. PG(r,c) and Ps(r,c) denote pixel ground truth values and predicted saliency probability-power maps.

Step S4: Input the image to be processed into the trained tile surface defect segmentation network model to obtain the segmentation target.

In order to verify the effectiveness of the algorithm, the Magnetic-tile-defect-datasets data set is used in the experiment to compare the algorithm proposed by the present invention with the U-Net and U ² -net algorithms.

Analysis: Experiments are carried out on the Magnetic-tile-defect-datasets dataset, and the generated results are compared quantitatively. There are mainly two evaluation indicators, maxF _β and MAE are shown in the table below. Compared with other methods, our method has achieved the best results in both indicators.

Method MAE maxF _β U2NET 0.002 0.781 U-Net 0.004 0.733 Ours 0.001 0.792

MAE is mean absolute error, which is one of the indicators to measure image quality. The calculation principle is the absolute value of the difference between the real value and the predicted value and then summed and then averaged. The smaller the MAE value, the better the image quality. The formula is as follows:

where P and G are the probabilistic map detections of salient objects and the corresponding ground truth, and (H, W) and (r, c) are the (height, width) and pixel coordinates.

Accuracy = (TP+TN)/total samples. The definition is: for a given test data set, the ratio of the number of samples correctly classified by the classifier to the total number of samples.

Precision (Precision) = TP/(TP+FP). It means: How many of the samples predicted to be positive are true positive samples, which is for our prediction results.

F _β is used to comprehensively evaluate precision and recall as follows:

Set β ² to 0.3 and report the maximum F _β (maxF _β ).

The above description does not limit the present invention in any form. Although the present invention has been disclosed by the above-mentioned embodiments, it is not intended to limit the present invention. The technical content disclosed above can be used to make some changes or be modified into equivalent embodiments of equivalent changes, but any simple modifications and equivalent changes made to the above embodiments according to the technical essence of the present invention will not deviate from the content of the technical solution of the present invention and modifications, all still belong to the scope of the technical solution of the present invention.

Claims (8)

1. A tile surface defect segmentation method for improving U ² -Net, characterized in that, comprising the following steps: Step S1, obtaining the tile surface defect detection data set, and dividing it into a training set and a testing set; Step S2, constructing a tile surface defect segmentation network model based on U ² -Net; The tile surface defect segmentation network model is a six-layer U-shaped structure, including a 6-level encoder, a 5-level decoder, a 2-level multi-scale feature fusion module and a saliency map fusion module; Among them, the first 4 encoders and the corresponding 4 decoders are composed of the feature extraction structure DCRSU, and the number of layers of each DCRSU decreases as the number of layers of the encoder and decoder increases, that is, the first 4 encoders use DCRSU respectively. -7, DCRSU-6, DCRSU-5, DCRSU-4, the same is true for the first 4 decoders; The fifth encoder and corresponding decoder use RSU-4F; The sixth encoder introduces SKnet as the deepest encoder; Introduce two improved attention gate modules at the input of the 5th decoder and the 4th decoder respectively, fuse the features output by the 6th encoder and the 5th encoder into the 5th decoder, and use the The 4 encoders are fused with the features output by the 5th decoder and input to the 4th decoder; then use 3×3 convolution and sigmoid function from the 6th encoder, the 5th decoder, the 4th decoder , the 3rd decoder, the 2nd decoder and the 1st decoder generate 6 output saliency probability maps; then the logic map of the output saliency map is up-sampled to the same size as the input image, and are fused by cascade operation ;Finally, through the 1×1 convolutional layer and the sigmoid function to generate the final significance probability map; Step S3, continuously iteratively train the tile surface defect segmentation network model through the training set, until the network finally converges, and obtain the trained tile surface defect segmentation network model; Step S4: Input the image to be processed into the trained tile surface defect segmentation network model to obtain the segmentation target. 2 . The improved U ² -Net tile surface defect segmentation method according to claim 1 , wherein the training set and test set are divided by 4:1 in the step S10 . 3. A kind of tile surface defect segmentation method of improving U ² -Net according to claim 1, is characterized in that, described DCRSU is made up of 4 parts of input convolutional layer, encoder, decoder and residual structure; The input convolution layer is used to extract local features and transform channels; In the encoding stage, the last encoder adopts convolution + batch normalization + ReLU activation function structure, and the penultimate layer adopts depth separable convolution + batch normalization + ReLU activation function structure; the remaining encoders use the residual structure to The features extracted by the depth-separable convolution and the input features processed by the attention mechanism module are added and then input to the next feature extraction layer for feature extraction, so that the output features of each level can be focused on channels with more effective feature information and strengthened. The ability to extract effective features at each level and obtain multi-scale feature information; The residual structure fuses the input layer and the middle layer, and splices features of two different scales; In the decoding stage, the decoder module takes the concatenated feature map, passes through a 3×3 convolution, a batch normalization layer and a Relu activation function to gradually repair the details and spatial dimensions of the segmented object through upsampling; the final decoder output The feature map is added and fused with the feature map of the input convolutional layer to obtain the final feature map processed by the DCRSU module. 4. A tile surface defect segmentation method for improved U ² -Net according to claim 1, characterized in that, the RSU-4F replaces downsampling and upsampling with dilated convolution, and the input features CxHxW are first passed through 2 modules composed of convolution + batch normalization layer + Relu, and then after expansion convolution are sequentially 1, 2, 4, 8, the size of the feature map remains unchanged throughout the process. 5. A kind of improved U ² -Net tile surface defect segmentation method according to claim 1, characterized in that, SKnet is used as the deepest encoder, and the operation of extracting multi-scale features is to pass the original feature map through a 3 The ×3 group/depth convolution and the 3×3 atrous convolution generate two feature maps: and /> Then add these two feature maps to generate U; The generated U generates a 1×1×C feature map through global average pooling. The feature map generates a d×1 vector z through a fully connected layer, and the vector z is converted back to length C through 2 FC layers. For 2 A vector is calculated for softmax in the channel dimension to obtain its respective weight vector, and the weight is multiplied by the two outputs of stage 1 to obtain a new feature map, and the two new feature maps are summed to obtain the final output, which is sent to the next decoder. 6. A kind of improved ^U2 -Net tile surface defect segmentation method according to claim 1, is characterized in that, two inputs of described improved attention gate module are the current layer x1 of encoder and decoder respectively The next layer g, the input feature is CxHxW, they first go through the element-by-element addition and the Relu activation function to get the feature map of CxHxW, and then reduce the number of channels to 1 through 1×1 convolution, and then the sigmoid activation function performs Normalize to get the attention coefficient, and then restore the size through a 1×1 module to get the coefficient of CxHxW, and finally use the obtained attention coefficient to multiply the two input feature maps, and then stitch them together, and finally get The feature map of is sent to the next decoder module. 7. A kind of improved U ² -Net tile surface defect segmentation method according to claim 1, it is characterized in that, when training model in the described step S3, setting batch size is 16, uses AdamW optimizer to optimize; We first use a learning rate of 1× ^10-3 for initial training, and then use a learning rate of 1× ^10-5 to fine-tune the model. 8. A method for segmenting surface defects of ceramic tiles by improving ^U2 -Net according to claim 1 or 7, wherein the loss formula in the step S3 is: In the formula: l _fuse represents the loss of the final predicted probability map, l represents the binary cross-entropy loss, and w represents the weight of each loss.