Weld joint segmentation method based on stripe feature enhanced extraction

Abstract

The invention provides a welding seam segmentation method based on stripe feature enhanced extraction. The welding seam segmentation method comprises the following steps: carrying out efficient encoding and decoding learning on welding seam stripe information in an encoder and a decoder by using transfer learning and cascade information learning thoughts respectively; through the two structures, richer low-layer and high-layer semantic information can be extracted, and the extraction capability of welding seam stripes of the segmentation model is ensured; besides, the weld stripe image contains complex noise such as arc light, smoke dust and molten drop splashing, the defects of an existing jump connection mode are analyzed, a multi-layer feature fusion mechanism fusing a global attention mechanism is designed, information association in different coding layers can be learned through the model, and the fusion precision is improved. And the influence of noise can be greatly weakened, so that effective fusion of shallow-layer semantic information and high-layer semantic information is realized, and the segmentation capability of the model to weld seam stripe features is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of weld image segmentation and deep learning in welding automation, and in particular to a weld segmentation method based on stripe feature enhancement extraction. Background Art

[0002] The quality of pipeline welds plays a vital role in the safety of oil and gas pipeline transportation. Nondestructive testing can detect the type and nature of internal defects in welds without damaging the weld structure of oil and gas pipelines. It has become the main method for pipeline weld detection. X-ray digital imaging detection is one of the important methods for nondestructive testing of oil and gas pipeline welds. This method first uses an X-ray machine to irradiate the pipeline weld, and then uses X-ray film to view the defects of the weld. Traditional manual film evaluation is inefficient and greatly affected by human subjective factors. It is difficult to meet the needs of industrial applications. Therefore, it is urgent to use computers to realize automatic detection and identification of pipeline weld image defects. The background interference information of X-ray weld images is large, and it is difficult to directly detect and identify weld defects. Therefore, scholars at home and abroad use the image segmentation method in computer vision to extract the accurate weld position, so as to more effectively assist the identification and detection of weld defects.

[0003] In recent years, weld image segmentation algorithms based on deep learning have made breakthrough progress. For example, a pipeline weld image segmentation method based on improved Unet, CN116740088A, includes the following steps: Step 1: Process the pipeline weld image using methods such as image enhancement and image denoising; Step 2: The processed weld data relies on professionals to complete data annotation, and then divides the training set and the verification set; Step 3: Improve the Unet network using dense connection ideas and parallel attention mechanisms; Step 4: Use the improved model to train the data set, and optimize the training process with mixed cross entropy and Dice loss functions; Step 5: Use the optimal weights in the training process to segment the prediction set. The present invention uses dense structures and deep separable convolutions to solve the problem of limited feature extraction capabilities of the original Unet, adds a parallel attention mechanism at the jump connection to improve the model's recognition ability of weld features, and adopts a mixed loss function to improve the imbalance between the pixels to be segmented and the background pixels. The weld segmentation ability of the improved model has been greatly improved. This segmentation method only uses the original jump connection method, which can easily cause information loss problems, resulting in a limited amount of useful information that can be learned in the subsequent decoding layer, and cannot meet the effective segmentation of the model.

[0004] For another example, a method for extracting the center of laser structured light stripes, CN116740088A, includes the following key steps: Step 1: Collect weld stripe images required for model training during welding; Step 2: Perform image preprocessing on the collected data to meet the experimental data requirements; Step 3: Construct a network model that meets the requirements for weld stripe segmentation and train the weld stripe data set; Step 4: Use the trained weights to automatically segment the weld stripes for the weld prediction data. The present invention adopts the idea of ​​interactive learning of semantic information and spatial information to increase the model's learning ability for weld stripe features, and enhances the network's anti-interference ability by adding hierarchical information fusion units and detail enhancement units, thereby improving the model's robustness for segmenting weld laser stripes. Compared with ordinary convolution, this extraction method is slow, requires high parameters, and has insufficient accuracy. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention provides a weld segmentation method based on stripe feature enhanced extraction, which is characterized by: including weld data annotation, accurately annotating and processing the obtained weld image raw data;

[0006] Weld data division: the processed weld stripes are divided into training set and test set according to the preset ratio;

[0007] Build an improved model, build a network model used for weld stripe segmentation, adopt the improved mode of Unet model, improve the encoding structure and decoding structure in Unet model, and introduce transfer learning strategy in data training;

[0008] Data training and recognition: the divided data set is sent to the improved model for training, and the weights that meet the evaluation requirements are selected as the subsequent training weights. The training weights are verified to obtain the weld results and segmentation performance indicators.

[0009] Furthermore, in the annotation of the weld data, the grayscale ratio, color gamut change, image flipping and size normalization are performed on the annotated data and graphics, and the image pixels are adjusted to a preset size.

[0010] Furthermore, the image pixel size is adjusted to 512*512.

[0011] Furthermore, when dividing the weld data, the processed weld stripe data is divided into a training set and a validation set in a ratio of 8:2.

[0012] Furthermore, in the improved model, the coding structure of the Unet model is improved by using the first 13 layers of VGG16 as the improved model, the coding blocks of the first two layers are composed of two convolution modules and a maximum pooling operation, and the convolution modules of the last three layers are increased from the original two layers to three layers, and then a small amount of training parameters are added to improve the receptive field of the model; the specific calculation results are as follows:

[0013] x 1 =R(B(Conv(x in ))

[0014] x 2 =R(B(Conv(x 1 ))

[0015] x 3 =R(B(Conv(x 2 ))

[0016] x 4 =R(B(Conv(x 3 ))

[0017] x out =R(B(Conv(x 4 ))

[0018] Where x in The image initial input value representing the mode, x 1 、x 2 、x 3 、x 4 Represents the output of the convolution operation of each of the first 13 layers of Vgg16, x out It is the result of using Vgg16 as the final output of the encoder, Conv(·) is the convolution operation, B(·) is the normalization operation, and R is the Relu linear operation.

[0019] Furthermore, in the improved model, a multi-scale feature fusion strategy is used to fuse low-level information at different levels into the decoder to improve the model's ability to obtain weld detail features. The specific method is as follows:

[0020] Feed the information learned by each layer of encoder into the GA attention mechanism;

[0021] The results output by the GA attention mechanism are then subjected to a jump connection operation. The jump connections in the same layer are directly transmitted, and the different layers are downsampled in turn to keep the image scale and channel information consistent with the image information in the connected jump connections.

[0022] Furthermore, the GA attention mechanism includes two parts of attention mechanism: channel and space.

[0023] Furthermore, the specific operation method of applying the multi-scale feature fusion strategy to the decoder is as follows:

[0024] The information learned by the GA attention mechanism in the first layer is directly transmitted to the decoding layer through the same-layer skip connection;

[0025] The skip connection of the second layer needs to go through Maxpool downsampling with a step size of 2, conv1

[0026] ×1 and Relu linear operations, and then fused with the information of the second layer of skip connection and sent to the second layer of decoding layer;

[0027] The skip connection of the third layer needs to undergo Maxpool downsampling with a step size of 4, conv1×1 and Relu linear operations, and then be fused with the information of the skip connection of the third layer and sent to the third decoding layer;

[0028] The skip connection of the fourth layer needs to undergo Maxpool downsampling with a step size of 8, conv1×1 and Relu linear operations, and then be fused with the information of the skip connection of the fourth layer and sent to the fourth decoding layer. The operations of the remaining layers are similar;

[0029] Taking the multi-scale feature fusion operation of the first-layer encoder as the input information as an example, the calculation result structure is described as follows:

[0030] x T1 =GA(x D1 )

[0031] x T2 =R(Conv1×1(MaxPool 2 (GA(x D1 ))))

[0032] x T3 =R(Conv1×1(MaxPool 4 (GA(x D1 ))))

[0033] x T4 =R(Conv1×1(MaxPool 8 (GA(x D1 ))))

[0034] Where x T1 、x T2 、x T3 、x T4are the results of the first layer encoder after the multi-scale feature fusion operation at the first, second, third, and fourth skip connection layers; R is the Relu linear operation, GA(·) is the GA attention mechanism operation, conv1×1 is the convolution operation with a size of 1, and MaxPool 2 (·), MaxPool 4 (·), MaxPool 8 (·) is a downsampling operation with a step size of 2, 4, and 8; after the multi-scale feature fusion operation, it is fused with the second, third, and fourth layer skip connection information and transmitted to the calculation structure of the corresponding level decoding layer:

[0035] U T1 =x T1

[0036]

[0037] Where U T1 , U T2 , U T3 , U T4 They are the multi-scale feature fusion operation results and the information fused with the current skip connection layer and transmitted to the corresponding decoding layer. is the fusion operation, x D2 、x D3 、x D4 are the information of the current skip connection layer respectively.

[0038] Furthermore, in the improved model, the encoding structure of the Unet model uses a cascade information learning structure to replace a single convolutional layer, and uses a 1×1 convolution and GSConv convolution with fewer parameters for cascade construction inside the structure. The specific operation method is as follows:

[0039] The input information is passed to two paths. Each path first uses conv1×1 convolution operation to increase the dimension, enrich the information dimension, and enhance the feature expression ability.

[0040] The result is then sent to the GSConv convolution of the two paths to form a cascade mode, and then the results of the two paths are fused and sent to the 1×1 convolution for dimensionality reduction to obtain the final result; the specific calculation structure is as follows:

[0041] y i =Gelu(B(C(2*G(2*C(y i-1 )))))

[0042] Where C(·) is a 1×1 convolution operation, G(·) is a GSConv convolution operation, B is a BN normalization operation, Gelu activation function, 2* represents that the information is doubled through cascade fusion, and yi-1 is the information of the previous layer decoder, y i The result is obtained for the cascade information learning module, which is also the information value to be learned by the decoder of the current layer.

[0043] Furthermore, in the construction of the improved model, the specific operation method of introducing the transfer learning strategy in data training is: using the transfer learning strategy to introduce the pre-trained encoder weights into the improved segmentation network training process, so as to reduce the training cost and improve the training speed of the model.

[0044] The present invention provides a weld segmentation method based on stripe feature enhanced extraction, which has the following advantages: using the transfer learning method in the encoder to reduce the training of the model while improving high-efficiency learning; introducing a multi-scale feature fusion mechanism at the jump connection to improve the model's learning ability for information in different coding layers while reducing the influence of noise such as arc light, smoke and molten droplet splashing; in addition, the two structures are used in series to extract richer detail information from different coding layers, and the cross-dimensional interaction is amplified through global attention to capture more important features; using a cascaded fusion learning module that integrates GSConv convolution in the decoder to enhance the model's expression of weld features in high-level semantic information, and enhance the model's weld feature segmentation ability. Compared with ordinary convolution, it has faster speed, fewer parameters and higher accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The above and other objects, features and advantages of the present invention will become more apparent through the following description of the embodiments of the present invention with reference to the accompanying drawings. The accompanying drawings are only used for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. In addition, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0046] Figure 1 A schematic flow chart of a disclosed weld segmentation method provided in an embodiment of the present invention;

[0047] Figure 2 A schematic diagram of a weld segmentation model based on stripe feature enhancement extraction provided by an embodiment of the present invention;

[0048] Figure 3 A schematic diagram of the structure of a multi-scale feature fusion mechanism provided by an embodiment of the present invention;

[0049] Figure 4 A schematic diagram of a global attention mechanism provided by an embodiment of the present invention;

[0050] Figure 5 A schematic diagram of a cascade informatics module provided by an embodiment of the present invention;

[0051] Figure 6A schematic diagram showing the segmentation effect comparison between the improved model provided by the segmentation method according to an embodiment of the present invention and the prior art. DETAILED DESCRIPTION

[0052] The present invention is further described in detail below in conjunction with the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art. It should be noted that the embodiments of the present invention and the features in the embodiments can be combined with each other without conflict.

[0053] Reference Figure 1-6 As shown, a weld segmentation method based on stripe feature enhancement extraction disclosed in an embodiment of the present invention includes the following operation steps.

[0054] The first step is to accurately annotate the original data of the weld image, and perform data processing operations such as image grayscale, color gamut conversion, image flipping, and size normalization on the annotated data and the original image, and adjust the image pixel size to 512*512.

[0055] In the second step, the processed weld stripe data is divided into training set and validation set in a ratio of 8:2, and then the divided data set is sent to the model learning.

[0056] The third step is to build the network model used for weld stripe segmentation. First, use the first 13 layers of VGG16 as the encoding structure in the improved model. The encoding blocks of the first two layers are composed of two convolution modules and maximum pooling operations. The convolution modules of the last three layers are increased from two layers to three layers. Then, a small amount of training parameters are added to improve the receptive field of the model. The specific calculation results are as follows:

[0057] x 1 =R(B(Conv(x in ))

[0058] x 2 =R(B(Conv(x 1 ))

[0059] x 3 =R(B(Conv(x 2 ))

[0060] x 4 =R(B(Conv(x 3 ))

[0061] x out =R(B(Conv(x 4))

[0062] Where x in The image initial input value representing the mode, x 1 、x 2 、x 3 、x 4 represents the output of the convolution operation of each of the first 13 layers of Vgg16, x out It is the result of using Vgg16 as the final output of the encoder, Conv(·) is the convolution operation, B(·) is the normalization operation, and R is the Relu linear operation.

[0063] At the same time, a transfer learning strategy is used to introduce the pre-trained encoder weights into the improved segmentation network training process, reducing the training cost and greatly improving the training speed of the model.

[0064] Fourth, in the Unet model, the encoding layer can extract more low-level semantic information, which contains more position and useful detail information. The original segmentation model only fuses the single-layer encoding layer information with the decoding layer information, which reduces the efficiency of information use. The improved model in the present invention uses the idea of ​​multi-scale feature fusion to fuse low-level information of different levels into the decoder to improve the model's ability to obtain weld detail features;

[0065] The multi-scale feature fusion mechanism first sends the information learned by each layer of encoder to the GA attention mechanism, which includes two parts of channel and space attention. Then the result of the GA attention mechanism is skipped. The skip connection of the same layer is directly transmitted, and the different layers are downsampled in turn to keep the image scale and channel information consistent with the image information in the connected skip connection. The structure is as follows: Figure 2-4 As shown. For example, the information learned by the GA attention mechanism in the first layer is directly transmitted to the decoding layer in the same-layer skip connection; in the skip connection of the second layer, it needs to undergo Maxpool downsampling with a step size of 2, conv1×1 and Relu linear operations, and then be fused with the information of the second-layer skip connection and sent to the second-layer decoding layer; the skip connection of the third layer needs to undergo Maxpool downsampling with a step size of 4, conv1×1 and Relu linear operations, and then be fused with the information of the third-layer skip connection and sent to the third-layer decoding layer; the skip connection of the fourth layer needs to undergo Maxpool downsampling with a step size of 8, conv1×1 and Relu linear operations, and then be fused with the information of the fourth-layer skip connection and sent to the fourth-layer decoding layer. The operations of other layers are similar. The following describes the calculation structure based on the multi-scale feature fusion operation with the first-layer encoder result as input information:

[0066] The first layer encoder results are subjected to multi-scale feature fusion operation and then connected to different jump

[0067] The result structure of layer calculation is:

[0068] x T1 =GA(x D1 )

[0069] x T2 =R(Conv1×1(MaxPool 2 (GA(x D1 ))))

[0070] x T3 =R(Conv1×1(MaxPool 4 (GA(x D1 ))))

[0071] x T4 =R(Conv1×1(MaxPool 8 (GA(x D1 ))))

[0072] Where x T1 、x T2 、x T3 、x T4 are the results of the first layer encoder after the multi-scale feature fusion operation at the first, second, third, and fourth skip connection layers; R is the Relu linear operation, GA(·) is the GA attention mechanism operation, conv1×1 is the convolution operation with a size of 1, and MaxPool 2 (·), MaxPool 4 (·), MaxPool 8 (·) is the downsampling operation with a step size of 2, 4, and 8;

[0073] After the multi-scale feature fusion operation, the information of the second, third, and fourth layer skip connections is fused and transmitted to the calculation structure of the corresponding layer of the decoding layer:

[0074] U T1 =x T1

[0075]

[0076] Where U T1 , U T2 , U T3 , U T4 They are the multi-scale feature fusion operation results and the information fused with the current skip connection layer and transmitted to the corresponding decoding layer. is the fusion operation, x D2 、x D3 、x D4are the information of the current skip connection layer respectively.

[0077] In the fifth step, the semantic information learned in the decoder is a high-level feature expression with rich semantic information. However, the decoders of the existing models mostly use cascaded convolution operations. Such convolution operations have limited learning capabilities and cannot obtain more effective information from the feature map. The improved model in the example of the present invention uses a cascaded information learning structure to replace a single convolution layer. Figure 5 As shown in the figure, the effect of efficient extraction of weld stripe features has been achieved. At the same time, in order to reduce the training parameters of the model, a 1×1 convolution with fewer parameters and a GSConv convolution with higher efficiency and stronger learning ability are used in the structure for cascade construction;

[0078] Specific operation method: The structure of the cascade information learning module first passes the input information into two paths. Each path first uses conv1×1 convolution operation to increase the dimension, enrich the information dimension, and enhance the feature expression ability. Then the result is transmitted to the GSConv convolution of the two paths to form a cascade mode. Then the results of the two paths are fused and sent to 1×1 convolution for dimensionality reduction to obtain the final result. The specific calculation structure is as follows:

[0079] y i =Gelu(B(C(2*G(2*C(y i-1 )))))

[0080] Where C(·) is a 1×1 convolution operation, G(·) is a GSConv convolution operation, B is a BN normalization operation, Gelu activation function, 2* represents that the information is doubled through cascade fusion, and y i-1 is the information of the previous layer decoder, y i The result is obtained for the cascade information learning module, which is also the information value to be learned by the decoder of the current layer.

[0081] Step 6: Send the divided data set to the built improved model for training, and select the weight with the best evaluation value from the training process as the subsequent training weight;

[0082] Step 7: Finally, use the prepared prediction set to verify the training weights to obtain the final weld segmentation results and indicators for evaluating the model segmentation performance. The obtained weld segmentation effect comparison chart is shown in Figure 1. Figure 6 as shown in .

[0083] A weld segmentation method based on stripe feature enhanced extraction disclosed in an embodiment of the present invention: in order to solve the problem of serious imbalance in the ratio of background pixels to target pixels and limited learning ability of existing models, transfer learning and cascade information learning ideas are used in the encoder and decoder respectively to perform efficient encoding and decoding learning for weld stripe information; richer low-level and high-level semantic information can be extracted through the above two structures to ensure the weld stripe extraction ability of the segmentation model; in addition, the weld stripe image contains complex noise such as arc light, smoke and molten droplet splashing. The present invention analyzes the shortcomings of the existing jump connection method and designs a multi-layer feature fusion mechanism that integrates the global attention mechanism. The model can not only learn the information association in different coding layers, but also greatly weaken the influence of noise, thereby realizing the effective fusion of shallow and high-level semantic information and enhancing the model's segmentation ability for weld stripe features.

[0084] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0085] Finally, it should be noted that: Obviously, the above embodiments are only examples for clearly explaining the present invention, and are not intended to limit the implementation methods. For ordinary technicians in the relevant field, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived from this are still within the scope of protection of the present invention.

Claims

1. A weld segmentation method based on stripe feature enhancement extraction, characterized in that: include: Weld data annotation: accurate data annotation and processing of the original weld image data; Weld data division: the processed weld stripes are divided into training set and test set according to the preset ratio; Build an improved model, build a network model used for weld stripe segmentation, adopt the improved mode of Unet model, improve the encoding structure and decoding structure in Unet model, and introduce transfer learning strategy in data training; Data training and recognition: the divided data set is sent to the improved model for training, and the weights that meet the evaluation requirements are selected as the subsequent training weights. The training weights are verified to obtain the weld results and segmentation performance indicators.

2. The weld segmentation method based on stripe feature enhancement extraction according to claim 1 is characterized in that: In the annotation of the weld data, the grayscale ratio, color gamut change, image flipping and size normalization are performed on the annotated data and graphics, and the image pixels are adjusted to a preset size.

3. The weld segmentation method based on stripe feature enhancement extraction according to claim 2 is characterized in that: Resize the image pixel size to 512*512.

4. The weld segmentation method based on stripe feature enhancement extraction according to claim 1 is characterized in that: When dividing the weld data, the processed weld stripe data is divided into a training set and a verification set in a ratio of 8:

2.

5. The weld segmentation method based on stripe feature enhancement extraction according to claim 1 is characterized in that: In the improved model, the improvement method of the encoding structure of the Unet model is as follows: the first 13 layers of VGG16 are used as the improved model, the encoding blocks of the first two layers are composed of two convolution modules and maximum pooling operations, and the convolution modules of the last three layers are increased from the original two layers to three layers, and the receptive field of the model is improved by a small increase in the training parameters; the specific calculation results are as follows: x1=R(B(Conv(x in )) x2=R(B(Conv(x1)) x3=R(B(Conv(x2)) x4=R(B(Conv(x3)) x out =R(B(Conv(x4)) Where x in represents the initial input value of the image of the mode, x1, x2, x3, x4 represent the output results of the convolution operation of each of the first 13 layers of Vgg16, out It is the result of using Vgg16 as the final output of the encoder, Conv(·) is the convolution operation, B(·) is the normalization operation, and R is the Relu linear operation.

6. The weld segmentation method based on stripe feature enhancement extraction according to claim 1 is characterized in that: In the improved model, a multi-scale feature fusion strategy is used to fuse low-level information at different levels into the decoder to improve the model's ability to obtain weld detail features. The specific method is as follows: Feed the information learned by each layer of encoder into the GA attention mechanism; The results output by the GA attention mechanism are then subjected to a jump connection operation. The jump connections in the same layer are directly transmitted, and the different layers are downsampled in turn to keep the image scale and channel information consistent with the image information in the connected jump connections.

7. The weld segmentation method based on stripe feature enhancement extraction according to claim 6 is characterized in that: The GA attention mechanism includes two parts: channel and space attention mechanism.

8. The weld segmentation method based on stripe feature enhancement extraction according to claim 6 is characterized in that: The specific operation method of applying the multi-scale feature fusion strategy to the decoder is as follows: The information learned by the GA attention mechanism in the first layer is directly transmitted to the decoding layer through the same-layer skip connection; The skip connection of the second layer needs to undergo Maxpool downsampling with a step size of 2, conv1×1 and Relu linear operations, and then be fused with the information of the skip connection of the second layer and sent to the second decoding layer; The skip connection of the third layer needs to undergo Maxpool downsampling with a step size of 4, conv1×1 and Relu linear operations, and then be fused with the information of the skip connection of the third layer and sent to the third decoding layer; The skip connection of the fourth layer needs to undergo Maxpool downsampling with a step size of 8, conv1×1 and Relu linear operations, and then be fused with the information of the skip connection of the fourth layer and sent to the fourth decoding layer. The operations of the remaining layers are similar; Taking the multi-scale feature fusion operation of the first-layer encoder as the input information as an example, the calculation result structure is described as follows: x T1 =GA(x D1 ) x T2 =R(Conv1×1(MaxPool 2 (GA(x D1 )))) x T3 =R(Conv1×1(MaxPool 4 (GA(x D1 )))) x T4 =R(Conv1×1(MaxPool 8 (GA(x D1 )))) Where x T1 、x T2 、x T3 、x T4 are the results of the first layer encoder after the multi-scale feature fusion operation at the first, second, third, and fourth skip connection layers; R is the Relu linear operation, GA(·) is the GA attention mechanism operation, conv1×1 is the convolution operation with a size of 1, and MaxPool 2 (·), MaxPool 4 (·), MaxPool 8 (·) is the downsampling operation with a step size of 2, 4, and 8; After the multi-scale feature fusion operation, the information of the second, third, and fourth layer skip connections is fused and transmitted to the calculation structure of the corresponding layer of the decoding layer: U T1 =x T1 Where U T1 , U T2 , U T3 , U T4 They are the multi-scale feature fusion operation results and the information fused with the current skip connection layer and transmitted to the corresponding decoding layer. is the fusion operation, x D2 、x D3 、x D4 are the information of the current skip connection layer respectively.

9. The weld segmentation method based on stripe feature enhancement extraction according to claim 1 is characterized in that: In the improved model, the encoding structure of the Unet model uses a cascade information learning structure to replace a single convolutional layer, and uses 1×1 convolution and GSConv convolution with fewer parameters for cascade construction inside the structure. The specific operation method is as follows: The input information is passed to two paths. Each path first uses conv1×1 convolution operation to increase the dimension, enrich the information dimension, and enhance the feature expression ability. The result is then sent to the GSConv convolution of the two paths to form a cascade mode, and then the results of the two paths are fused and sent to the 1×1 convolution for dimensionality reduction to obtain the final result; the specific calculation structure is as follows: y i =Gelu(B(C(2*G(2*X(y i-1 ))))) Where C(·) is a 1×1 convolution operation, G(·) is a GSConv convolution operation, B is a BN normalization operation, Gelu activation function, 2* represents that the information is doubled through cascade fusion, and y i-1 is the information of the previous layer decoder, y i The result is obtained for the cascade information learning module, which is also the information value to be learned by the decoder of the current layer.

10. The weld segmentation method based on stripe feature enhancement extraction according to claim 1, characterized in that: In the construction of the improved model, the specific operation method of introducing the transfer learning strategy in data training is: using the transfer learning strategy to introduce the pre-trained encoder weights into the improved segmentation network training process to reduce the training cost and improve the training speed of the model.

Images