Image segmentation network and image segmentation method

Abstract

The invention discloses an image segmentation network and an image segmentation method; in the image segmentation method, a welding pool image is input into the image segmentation network for image segmentation; wherein the image segmentation network is based on a Unet structure, the image segmentation network comprises a multi-layer feature extraction structure and a multi-layer feature fusion structure, the feature extraction structure is jump-connected through an attention module and a feature fusion structure, the attention module focuses on the target features of the image and filters the redundant information generated in the jump connection; the feature extraction structure comprises a plurality of branches and a feature splicing module; the plurality of branches extract features from different scales according to the input information to enrich the semantic information; this enables the image segmentation network to have a better image segmentation effect.

Description

Technical Field

[0001] The invention relates to the field of image processing, in particular to an image segmentation network and an image segmentation method. Background Art

[0002] As one of the key issues in computer vision, semantic segmentation of images is an important research direction. During the welding process, molten metal drips onto the base material to form a molten pool. The contour is the most basic visual morphological feature of the molten shape. Image recognition technology can be used to extract the molten pool features and predict the welding quality based on the molten pool features. However, the features of the weld pool image are complex. The leading edge of the weld pool contour is covered by the welding arc, and the brightness saturation area on the surface of the weld pool impacts the trailing edge of the contour, and the boundary between the edge of the weld pool and the surrounding area is unclear. The weld pool has various shapes and its edges are irregular. This makes it difficult to extract the contour features of the weld pool image, and most image segmentation algorithms or image segmentation networks cannot achieve high-precision and good robustness image segmentation effects for weld pool images. Summary of the invention

[0003] The purpose of the present invention is to solve at least one of the technical problems existing in the prior art and to provide an image segmentation network and an image segmentation method.

[0004] The technical solution adopted by the present invention to solve the problem is:

[0005] The first aspect of the present invention is an image segmentation network, which is based on a Unet neural network structure. The image segmentation network includes a multi-layer feature extraction structure and a multi-layer feature fusion structure. The feature extraction structure is jump-connected with the feature fusion structure in a one-to-one correspondence through an attention module, and the attention module is used to filter the redundant information generated in the jump connection; the feature extraction structure includes multiple branches and a feature splicing module; the multiple branches extract features from different scales according to input information; the feature splicing module is connected to the multiple branches, and the feature splicing module is used to splice and fuse features of different scales extracted by the multiple branches.

[0006] According to the first aspect of the present invention, the branch of the feature extraction structure includes a first branch, and the first branch includes a first convolutional layer.

[0007] According to the first aspect of the present invention, the second branch includes a second convolutional layer, a first asymmetric convolutional layer, a second asymmetric convolutional layer, a first asymmetric hole convolutional layer, a second asymmetric hole convolutional layer, a third convolutional layer and a first feature fusion layer; the second convolutional layer is used to compress the number of feature channels, the input end of the first asymmetric convolutional layer and the input end of the first asymmetric hole convolutional layer are both connected to the output end of the second convolutional layer, the input end of the second asymmetric convolutional layer is connected to the output end of the first asymmetric convolutional layer, the input end of the second asymmetric hole convolutional layer is connected to the output end of the first asymmetric hole convolutional layer, the output end of the second asymmetric convolutional layer and the output end of the second asymmetric hole convolutional layer are both connected to the input end of the third convolutional layer, the third convolutional layer is used to expand the number of feature channels, and the first feature fusion layer is used to fuse the output of the third convolutional layer with the input information.

[0008] According to the first aspect of the present invention, the third branch includes a fourth convolutional layer, multiple fifth convolutional layers connected in sequence, a feature splicing layer and a second feature fusion layer, multiple fifth convolutional layers are connected to the feature splicing layer, the feature splicing layer is used to splice the outputs of multiple fifth convolutional layers, the feature splicing layer and the fourth convolutional layer are both connected to the second feature fusion layer, and the second feature fusion layer is used to perform feature fusion on the output of the feature splicing layer and the output of the fourth convolutional layer.

[0009] According to the first aspect of the present invention, the attention module includes a feature extraction layer, a maximum pooling layer, a mapping layer, a weight calculation layer and a multiplication layer, the maximum pooling layer, the mapping layer and the weight calculation layer are connected in sequence, and the output end of the feature extraction layer and the output end of the weight calculation layer are connected to the input end of the multiplication layer; the feature extraction layer is used to extract features from the encoded feature map to obtain mapping features; the maximum pooling layer is used to spatially compress the decoded feature map to obtain compressed features; the mapping layer is used to map the compressed features so that the length of the compressed features is restored to obtain restored features; the weight calculation layer is used to calculate the attention weight based on the restored features; the multiplication layer is used to multiply the attention weight and the mapping features to obtain the target attention features.

[0010] According to the first aspect of the present invention, the mapping layer adopts a ReLU activation function.

[0011] According to the first aspect of the present invention, the weight calculation layer adopts a Sigmoid activation function.

[0012] According to the first aspect of the present invention, for two adjacent feature extraction structures, an output end of the feature extraction structure of an upper layer is connected to an input end of the feature extraction structure of a lower layer through a downsampling layer.

[0013] According to the first aspect of the present invention, for two adjacent feature fusion structures, the output end of the feature fusion structure of the lower layer is connected to the input end of the feature fusion structure of the upper layer through an upsampling layer.

[0014] A second aspect of the present invention is an image segmentation method, which uses the image segmentation network as described in the first aspect of the present invention, and the image segmentation method comprises:

[0015] Acquire weld pool image;

[0016] The weld pool image is input into the image segmentation network for image segmentation.

[0017] The above scheme has at least the following beneficial effects: extracting features from different scales through multiple branches of the feature extraction structure to obtain detailed information of the input image; setting an attention module in the jump connection between the feature extraction structure and the feature fusion structure, focusing on the target features of the image through the attention module, filtering the background features of the image, and filtering the redundant information generated in the jump connection; this can make the image segmentation network have a better image segmentation effect.

[0018] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The present invention will be further described below in conjunction with the accompanying drawings and examples.

[0020] Figure 1 is a structural diagram of an image segmentation network according to an embodiment of the present invention;

[0021] Figure 2 It is a structural diagram of the feature extraction structure;

[0022] Figure 3 It is the structural diagram of the attention module. DETAILED DESCRIPTION

[0023] This section will describe in detail the specific embodiments of the present invention. The preferred embodiments of the present invention are shown in the accompanying drawings. The purpose of the accompanying drawings is to supplement the description of the text part of the specification with graphics, so that people can intuitively and vividly understand each technical feature and the overall technical solution of the present invention, but it cannot be understood as a limitation on the scope of protection of the present invention.

[0024] In the description of the present invention, it should be understood that descriptions involving orientations, such as up, down, front, back, left, right, etc., and orientations or positional relationships indicated are based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present invention.

[0025] In the description of the present invention, "several" means one or more, "more" means more than two, "greater than", "less than", "exceed" etc. are understood as not including the number itself, and "above", "below", "within" etc. are understood as including the number itself. If there is a description of "first" or "second", it is only used for the purpose of distinguishing the technical features, and cannot be understood as indicating or implying the relative importance or implicitly indicating the number of the indicated technical features or implicitly indicating the order of the indicated technical features.

[0026] In the description of the present invention, unless otherwise clearly defined, terms such as setting, installing, connecting, etc. should be understood in a broad sense, and technicians in the relevant technical field can reasonably determine the specific meanings of the above terms in the present invention based on the specific content of the technical solution.

[0027] An embodiment of the first aspect of the present invention provides an image segmentation network.

[0028] The image segmentation network is based on the Unet neural network structure. The image segmentation network includes a multi-layer feature extraction structure 100 and a multi-layer feature fusion structure 200. The feature extraction structure 100 is jump-connected one-to-one through the attention module 300 and the feature fusion structure 200. The attention module 300 is used to filter the redundant information generated in the jump connection; the feature extraction structure 100 includes multiple branches and a feature splicing module 113; the multiple branches extract features from different scales according to the input information; the feature splicing module 113 is connected to the multiple branches, and the feature splicing module 113 is used to splice and fuse the features of different scales extracted by the multiple branches.

[0029] In this embodiment, features are extracted from different scales through multiple branches of the feature extraction structure 100 to obtain detailed information of the input image; an attention module 300 is set through a jump connection between the feature extraction structure 100 and the feature fusion structure 200, and the attention module 300 focuses on the target features of the image, filters the background features of the image, and filters the redundant information generated in the jump connection; this enables the image segmentation network to have a better image segmentation effect.

[0030] Reference Figure 1In this specific embodiment, the feature extraction structure 100 of the first layer extracts a feature map with 64 channels from the input image with 1 channel, and downsamples the feature map with 64 channels to the feature extraction structure 100 of the second layer; the feature extraction structure 100 of the second layer extracts a feature map with 128 channels from the feature map with 64 channels, and downsamples the feature map with 128 channels to the feature extraction structure 100 of the third layer; the feature extraction structure 100 of the third layer extracts a feature map with 256 channels from the feature map with 128 channels, and downsamples the feature map with 128 channels to the feature extraction structure 100 of the third layer. The feature map with 256 channels is downsampled to the feature extraction structure 100 of the fourth layer; the feature extraction structure 100 of the fourth layer extracts a feature map with 512 channels from the feature map with 256 channels, and downsamples the feature map with 512 channels to the first feature extraction structure 100 of the fifth layer; the first feature extraction structure 100 of the fifth layer extracts a feature map with 1024 channels from the feature map with 512 channels, and the second feature extraction structure 100 of the fifth layer restores the feature map with 512 channels from the feature map with 1024 channels.

[0031] The feature map with the number of channels of 512 output by the second feature extraction structure 100 of the fifth layer is upsampled to the feature fusion structure 200 of the fourth layer as the first input feature map with the number of channels of 512. The feature map with the number of channels of 512 output by the feature extraction structure 100 of the fourth layer is jump-connected to the feature fusion structure 200 of the fourth layer. During the jump connection process, it is processed by the attention module 300 of the fourth layer as the second input feature map with the number of channels of 512. The feature fusion structure 200 of the fourth layer fuses the two feature maps with the number of channels of 512 and outputs a new feature map with the number of channels of 512. The feature map with the number of channels of 512 output by the feature fusion structure 200 of the fourth layer is upsampled to the feature fusion structure 200 of the third layer as the first input feature map with the number of channels of 256; the feature map with the number of channels of 256 output by the feature extraction structure 100 of the third layer is jump-connected to the feature fusion structure 200 of the third layer, and is processed by the attention module 300 of the third layer during the jump connection process as the second input feature map with the number of channels of 256. The feature fusion structure 200 of the third layer fuses the two feature maps with the number of channels of 256 to output a new feature map with the number of channels of 256. The feature map with 128 channels output by the feature fusion structure 200 of the third layer is upsampled to the feature fusion structure 200 of the second layer as the first input feature map with 128 channels. The feature map with 128 channels output by the feature extraction structure 100 of the second layer is jump-connected to the feature fusion structure 200 of the second layer, and is processed by the attention module 300 of the second layer during the jump connection process as the second input feature map with 128 channels. The feature fusion structure 200 of the second layer fuses the two feature maps with 128 channels and outputs a new feature map with 128 channels. The feature map with 128 channels output by the feature fusion structure 200 of the second layer is upsampled to the feature fusion structure 200 of the first layer as the first input feature map with 64 channels. The feature map with 64 channels output by the feature extraction structure 100 of the first layer is jump-connected to the feature fusion structure 200 of the first layer, and is processed by the attention module 300 of the first layer during the jump connection process as the second input feature map with 64 channels. The feature fusion structure 200 of the first layer fuses the two feature maps with 64 channels through convolution calculation and outputs a new feature map with 1 channel as the output feature.

[0032] The feature fusion structures 200 of the fourth layer, the third layer, and the second layer are all multi-scale convolution modules, and the feature fusion structure 200 of the first layer is a convolution layer with a convolution kernel of 1x1.

[0033] It should be noted that the above is only a specific example of an image segmentation network. In other embodiments, the structure of the image segmentation network can be expanded on this basis.

[0034] Reference Figure 2 In certain embodiments of the first aspect of the present invention, for the feature extraction structure 100, the feature extraction structure 100 includes three branches and a feature splicing module 113, and the three branches are respectively a first branch, a second branch and a third branch.

[0035] The first branch includes the first convolution layer 101, which is a convolution layer with a convolution kernel of 3x3. The first convolution layer 101 performs feature extraction on the input feature map to extract local context information. The first branch can be expressed as: Output1 = Conv 3,3 (Input).

[0036] The second branch includes a second convolutional layer 102, a first asymmetric convolutional layer 103, a second asymmetric convolutional layer 104, a first asymmetric hole convolutional layer 105, a second asymmetric hole convolutional layer 106, a third convolutional layer 107 and a first feature fusion layer 108; the second convolutional layer 102 is used to compress the number of feature channels, the input end of the first asymmetric convolutional layer 103 and the input end of the first asymmetric hole convolutional layer 105 are both connected to the output end of the second convolutional layer 102, and the second asymmetric convolutional layer The input end of 104 is connected to the output end of the first asymmetric convolution layer 103, the input end of the second asymmetric dilated convolution layer 106 is connected to the output end of the first asymmetric dilated convolution layer 105, the output end of the second asymmetric convolution layer 104 and the output end of the second asymmetric dilated convolution layer 106 are both connected to the input end of the third convolution layer 107, the third convolution layer 107 is used to expand the number of feature channels, and the first feature fusion layer 108 is used to fuse the output of the third convolution layer 107 with the input information.

[0037] Among them, the second convolution layer 102 is a convolution layer with a convolution kernel of 3x3, the first asymmetric convolution layer 103 is a convolution layer with a convolution kernel of 3x1, the second asymmetric convolution layer 104 is a convolution layer with a convolution kernel of 1x3, the first asymmetric hole convolution layer 105 is a convolution layer with a convolution kernel of 3x1, the second asymmetric hole convolution layer 106 is a 1x3 convolution layer, and the third convolution layer 107 is a convolution layer with a convolution kernel of 1x1.

[0038] The first convolution layer 101 performs convolution calculation on the input feature map, so that the number of channels w of the output feature map is reduced to w / 2 compared with the original input feature map. Then the compressed feature map passes through the asymmetric convolution branch including the first asymmetric convolution layer 103 and the second asymmetric convolution layer 104 and the asymmetric dilated convolution branch including the first asymmetric dilated convolution layer 105 and the second asymmetric dilated convolution layer 106 to obtain features of different receptive fields. The feature maps of different receptive fields output by the asymmetric convolution branch and the asymmetric dilated convolution branch are added to the corresponding points, and then the number of channels w / 2 of the feature map is expanded to w through the third convolution layer 107. The output of the third convolution layer 107 is fused with the input information through the first feature fusion layer 108. The receptive field is expanded and more contextual information is captured by introducing dilated convolution, and rich semantic information is obtained by fusing the left and right branches.

[0039] Then the operation of the second branch is expressed as:

[0040] The third branch includes a fourth convolutional layer 110, a plurality of fifth convolutional layers 109 connected in sequence, a feature concatenation layer 111, and a second feature fusion layer 112. The plurality of fifth convolutional layers 109 are all connected to the feature concatenation layer 111. The feature concatenation layer 111 is used to concatenate the outputs of the plurality of fifth convolutional layers 109. The feature concatenation layer 111 and the fourth convolutional layer 110 are all connected to the second feature fusion layer 112. The second feature fusion layer 112 is used to perform feature fusion on the output of the feature concatenation layer 111 and the output of the fourth convolutional layer 110. Specifically, there are three fifth convolutional layers 109.

[0041] Then the operation of the third branch can be expressed as:

[0042] Then, the features of different scales extracted by multiple branches are spliced ​​and fused through the feature splicing module 113 as the output result of the feature extraction structure 100, which enriches the semantic information.

[0043] The operation of the feature concatenation module 113 is expressed as: Output = Concate (Output1, Output2, Output3).

[0044] Reference Figure 3In some embodiments of the first aspect of the present invention, the attention module 300 includes a feature extraction layer 301, a maximum pooling layer 302, a mapping layer 303, a weight calculation layer 304 and a multiplication layer 305, the maximum pooling layer 302, the mapping layer 303 and the weight calculation layer 304 are connected in sequence, and the output end of the feature extraction layer 301 and the output end of the weight calculation layer 304 are connected to the input end of the multiplication layer 305; the feature extraction layer 301 is used to extract features from the encoded feature map to obtain mapping features; the maximum pooling layer 302 is used to compress the decoded feature map in space to obtain compressed features; the mapping layer 303 is used to map the compressed features to restore the length of the compressed features to obtain restored features; the weight calculation layer 304 is used to calculate the attention weight according to the restored features; the multiplication layer 305 is used to multiply the attention weight and the mapping feature to obtain the target attention feature.

[0045] Specifically, the mapping layer 303 uses the ReLU activation function, and the weight calculation layer 304 uses the Sigmoid activation function.

[0046] In this embodiment, the attention module 300 weights each pixel in the feature image so that the target area is multiplied with a larger weight and the irrelevant background area is multiplied with a smaller weight, thereby effectively suppressing the interference of irrelevant information and reducing most of the redundant information.

[0047] The feature extraction layer 301 extracts features from the encoded feature map to generate mapping features. The decoded feature map can be regarded as a channel combination G = {g1, g2, g3, ..., g n}, each decoded feature map is spatially compressed through the maximum pooling layer 302 so that it converges to one point to obtain the compressed feature q k ∈R 1×1×C The processing of the maximum pooling layer 302 can be expressed as In order to obtain the recalibrated feature gain, the compressed feature q k The compressed features are mapped through the ReLU activation function of the mapping layer 303 to restore them to their original length, thereby obtaining the restored features. The restored features are passed through the weight calculation layer 304 through the Sigmoid function to obtain the adjusted feature weight α limited to the interval [0, 1] k The process of obtaining feature weights is expressed as: α k =[δ(F c (q1)),δ(F c (q2)),......,δ(F c (q k ))]. The multiplication layer 305 is used to convert the attention weight α k Multiply it with the mapping feature P to get the target attention feature U. The target attention feature can be expressed as U = αk ×P(k=C,p∈R H×W×C ).

[0048] During the welding process, molten metal drips onto the base material to form a molten pool. The contour is the most basic visual morphological feature of the melt shape, and the molten pool feature can be used to predict the welding quality.

[0049] However, the features of the weld pool image are complex. The leading edge of the weld pool contour is covered by the welding arc, and the brightness saturation area on the weld pool surface impacts the trailing edge of the contour, making the boundary between the weld pool edge and the surrounding unclear. The weld pool has various shapes, and its edges are irregular. This makes it difficult to extract the contour features of the weld pool image, and most image segmentation algorithms cannot achieve high-precision and robust image segmentation effects for weld pool images.

[0050] The second aspect of the present invention provides an image segmentation method, which uses the image segmentation network of the first aspect of the present invention.

[0051] Image segmentation methods include:

[0052] Acquire weld pool image;

[0053] The weld pool image is input into the image segmentation network for image segmentation.

[0054] In this embodiment, in order to solve the problems of unclear boundaries between the edges and peripheries of the weld pool image, large size differences, and irregular edges, the image segmentation network can extract features from different scales through multiple branches of the feature extraction structure 100 to obtain detailed information of the weld pool image, obtain feature maps of receptive fields of different sizes, adapt to the complex segmentation background of the weld pool image, and enrich semantic information; by setting an attention module 300 through a jump connection between the feature extraction structure 100 and the feature fusion structure 200, the segmented area of ​​the weld pool image is highlighted through the attention module 300, the features of irrelevant areas are filtered, and the redundant information generated in the jump connection is filtered to reduce the occurrence of mis-segmentation; this can enable the image segmentation network to have a better image segmentation effect on the weld pool image.

[0055] The third aspect of the present invention provides a storage medium storing a computer program for executing the image segmentation method according to the second aspect of the present invention.

[0056] It will be appreciated by those skilled in the art that all or some of the steps and systems in the methods disclosed above may be implemented as software, firmware, hardware, and appropriate combinations thereof. Some or all physical components may be implemented as software executed by a processor, such as a central processing unit, a digital signal processor, or a microprocessor, or as hardware, or as an integrated circuit, such as an application specific integrated circuit. As known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information. Computer storage media include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disk (DVD) or other optical disk storage, magnetic cassettes, magnetic tapes, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to store the desired information and can be accessed by a computer. In addition, it is known to those skilled in the art that communication media typically contain computer-readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transmission mechanism, and may include any information delivery medium.

[0057] The above is only a preferred embodiment of the present invention. The present invention is not limited to the above implementation mode. As long as the technical effect of the present invention is achieved by the same means, it should belong to the protection scope of the present invention.

Claims

1. An image segmentation network, characterized in that: The image segmentation network is used to perform image segmentation on the welding pool image. The image segmentation network is based on the Unet neural network structure. The image segmentation network includes a multi-layer feature extraction structure and a multi-layer feature fusion structure. The feature extraction structure is jump-connected with the feature fusion structure in a one-to-one correspondence through an attention module. The attention module is used to filter the redundant information generated in the jump connection. The feature extraction structure includes multiple branches and a feature splicing module. The multiple branches extract features from different scales according to the input information. The feature splicing module is connected to the multiple branches, and is used to splice and fuse the features of different scales extracted by the multiple branches; Among them, the branch of the feature extraction structure includes a first branch, and the first branch of the feature extraction structure includes a first convolutional layer; the second branch of the feature extraction structure includes a second convolutional layer, a first asymmetric convolutional layer, a second asymmetric convolutional layer, a first asymmetric hole convolutional layer, a second asymmetric hole convolutional layer, a third convolutional layer and a first feature fusion layer; the second convolutional layer is used to compress the number of feature channels, the input end of the first asymmetric convolutional layer and the input end of the first asymmetric hole convolutional layer are both connected to the output end of the second convolutional layer, the input end of the second asymmetric convolutional layer is connected to the output end of the first asymmetric convolutional layer, the input end of the second asymmetric hole convolutional layer is connected to the output end of the first asymmetric hole convolutional layer, and the second asymmetric hole convolutional layer is connected to the output end of the first asymmetric hole convolutional layer. The output end of the convolution layer and the output end of the second asymmetric hole convolution layer are both connected to the input end of the third convolution layer, the third convolution layer is used to expand the number of feature channels, and the first feature fusion layer is used to fuse the output of the third convolution layer with the input information; the third branch of the feature extraction structure includes a fourth convolution layer, a plurality of sequentially connected fifth convolution layers, a feature splicing layer and a second feature fusion layer, a plurality of the fifth convolution layers are all connected to the feature splicing layer, the feature splicing layer is used to splice the outputs of a plurality of the fifth convolution layers, the feature splicing layer and the fourth convolution layer are both connected to the second feature fusion layer, and the second feature fusion layer is used to perform feature fusion on the output of the feature splicing layer and the output of the fourth convolution layer.

2. An image segmentation network according to claim 1, characterized in that: The attention module includes a feature extraction layer, a maximum pooling layer, a mapping layer, a weight calculation layer and a multiplication layer, wherein the maximum pooling layer, the mapping layer and the weight calculation layer are connected in sequence, and the output end of the feature extraction layer and the output end of the weight calculation layer are connected to the input end of the multiplication layer; the feature extraction layer is used to extract features from the encoding feature map to obtain mapping features; the maximum pooling layer is used to compress the decoding feature map in space to obtain compressed features; The mapping layer is used to map the compressed feature so that the length of the compressed feature is restored to obtain a restored feature; the weight calculation layer is used to calculate the attention weight according to the restored feature; the multiplication layer is used to multiply the attention weight and the mapping feature to obtain a target attention feature.

3. An image segmentation network according to claim 2, characterized in that: The mapping layer of the attention module adopts the ReLU activation function.

4. An image segmentation network according to claim 2, characterized in that: The weight calculation layer of the attention module adopts the Sigmoid activation function.

5. An image segmentation network according to claim 1, characterized in that: For two adjacent feature extraction structures, the output end of the feature extraction structure of the upper layer is connected to the input end of the feature extraction structure of the lower layer through the downsampling layer.

6. An image segmentation network according to claim 1, characterized in that: For two adjacent feature fusion structures, the output end of the feature fusion structure of the lower layer is connected to the input end of the feature fusion structure of the upper layer through an upsampling layer.

7. An image segmentation method, characterized in that: Applying the image segmentation network according to any one of claims 1 to 6, the image segmentation method comprises: Acquire weld pool image; The weld pool image is input into the image segmentation network for image segmentation.

Images