CN113034512B - Weld joint tracking method based on feature segmentation - Google Patents

Abstract

The invention relates to a weld joint tracking method based on feature segmentation, which comprises the following steps: 1. and collecting a molten pool image, and performing ROI selection on the molten pool image. 2. An Encoder-Decoder segmentation network structure of ERFNet is adopted to carry out molten pool image segmentation, a cross entropy loss function commonly used for semantic segmentation is used, global feature information is fused, and a pyramid pooling module is added into a backbone network. The invention improves the network structure and the Loss function of the existing ERFNet, and carries out multi-scale feature fusion on the high-level features and the low-level features by referring to UNet on the network structure, and Focal local is used for replacing the cross entropy Loss function on the Loss function. The problems of line breakage of the central line of the laser stripe, overlarge deviation of welding characteristic points and the like are avoided. The performance of the feature extraction algorithm is effectively improved, the laser stripe center line and the weld joint feature points of each weld joint of each layer are accurately extracted, the efficiency of the algorithm is not lost, and the requirement of weld joint tracking on real-time performance is guaranteed.

Description

Seam Tracking Method Based on Feature Segmentation

technical field

The invention relates to a seam tracking method based on feature segmentation, which belongs to the technical field of welding automation.

Background technique

Additive welding is not only used in the task of groove filling, but also more widely used in the rapid prototyping of solid parts through layer-by-layer welding methods. Taking advantage of its advantages of short processing cycle, fast forming rate and high material utilization rate, it can realize Rapid manufacturing of parts with complex geometries and structures. With the increasing application of additive manufacturing technology in important fields such as aerospace, national defense, automobile manufacturing, electronic products, and biomedicine, the demand for rapid prototyping of metal parts is also increasing, so the welding seam tracking technology It is of great strategic significance to be applied to the task of rapid prototyping of metal parts.

It can be known from the large-groove additive weld seam tracking experiment that only the laser stripe contour line is valid information in the laser stripe image collected by the weld seam tracking system, so there will be an imbalance of positive and negative samples in the ERFNet training process At the same time, since the number of welding layers and welding passes in flat plate additive welding is much larger than that of large groove filling experiments, the laser stripe characteristics of high-rise welds will be less obvious, so ERFNet is directly applied to flat plate additive welds. In the feature extraction experiment, problems such as broken lines in the center line of laser stripes and excessive deviation of welding feature points will occur. It is necessary to improve the network structure and loss function of ERFNet. In the network structure, refer to UNet to perform multi-scale feature fusion of high-level features and low-level features, and replace the cross-entropy loss function with Focal Loss in the loss function, these problems can be avoided.

In the deep convolutional neural network, according to the theory of the receptive field, the low-level feature map has higher resolution and contains more position and detail information, but due to fewer convolution operations, the semantic information is lower and the noise is more , while the high-level feature map has stronger semantic information, but due to more convolution operations, the resolution is very low and the perception of image details is poor. According to this feature of the feature map, multi-scale feature fusion can be used to improve the ability of the network to obtain feature information in the image.

Multi-scale feature fusion is to fuse the low-level feature map with rich spatial features and the high-level feature map with rich semantic information to obtain a new feature map, so that the new feature map has the characteristics of high resolution and high semantic information at the same time. The method is widely used in the fields of object detection and semantic segmentation.

The present invention improves the network structure of ERFNet with reference to the UNet network, and adopts a multi-scale feature fusion method to improve the network's effect of extracting laser stripe centerlines and welding feature points. The network structure of UNet is the same as that of ERFNet, and it is also an algorithm based on the Encoder-Decoder structure, but in UNet, the Concat operation is used to splice the feature maps of the same size in the encoder and decoder, so that the network can be upsampled. Obtain more spatial information and semantic information, thereby improving the accuracy of segmentation.

The present invention directly adds the corresponding feature maps to achieve multi-scale feature fusion. The reason is that using the Concat operation for feature superposition will increase the number of feature channels, resulting in an increase in the amount of calculation and slowing down the speed of the algorithm to extract the laser centerline and welding feature points. Therefore, in order to ensure the real-time performance and reliability of the algorithm of this method, the multi-scale feature fusion is realized by directly adding the corresponding feature maps.

Contents of the invention

In order to solve the above technical problems, the present invention provides a seam tracking method based on feature segmentation, and its specific technical scheme is as follows:

A seam tracking method based on feature segmentation, characterized in that: comprising the steps of:

Step S1: collecting the molten pool image, and performing ROI selection on the molten pool image;

式中α_t为正样本时，α_t＝α∈[0，1]，α_t为负样本时，α_t＝1-α，y_c表示样本c的标签，p_c表示样本c预测为正类的概率，M为类别数量；Step S2: Use the Encoder-Decoder segmentation network structure of ERFNet to segment the melt pool image, and use the cross-entropy loss function commonly used in semantic segmentation to fuse global feature information, and add a pyramid pooling module to the backbone network, which has a balance The formula for the cross-entropy loss function of factor α _t is:

In the formula, when α _t is a positive sample, α _t = α∈[0,1], when α _t is a negative sample, α _t =1-α, y _c represents the label of sample c, p _c represents that sample c is predicted to be positive The probability of the class, M is the number of categories;

Further, the ROI area in the step S1 is a fixed area including the contour of the molten pool, and the size of the ROI area is 512 pixels×512 pixels.

Further, the network structure in the step S2 is based on SegNet and ENet, and the whole model includes 23 layers, wherein layers 1-16 are Encoder, and layers 17-23 are Decoder.

Further, in the step S2, the network structure adopts multi-scale feature fusion of high-level features and low-level features, and fuses the feature maps of the 8th layer and the 16th layer in ERFNet, and fuses the feature maps of the 3rd layer and the 17th layer. Layer 2 is fused with the feature maps of layer 20.

Further, the frame number of ERFNet in the step S2 is above 120FPS.

Further, in the step S2, the mean value of RMSE and PDE of the high-level welds of the network structure is on the rise relative to the low-level welds, and the offset between the feature points of the welds and the manually marked feature points is 1-2 pixels.

The beneficial effects of the present invention are: the weld feature extraction algorithm is optimized and improved, the multi-scale feature fusion strategy and Focal Loss are introduced into ERFNet, and the plate additive weld feature extraction experiment is designed to verify the feasibility of the improved algorithm , not only can effectively improve the performance of the feature extraction algorithm, accurately extract the laser stripe centerline and weld feature points of each layer and each weld, but also can not lose the efficiency of the algorithm, and ensure the real-time requirements of weld tracking. The plate additive welding experiment was carried out to verify the adaptability of the improved seam tracking scheme under actual working conditions. By comparing and analyzing the manually welded cube and the system welded cube, the results of many experiments showed that the forming effects of the two are very different. Micro, it is proved that the improved seam tracking scheme has high reliability in the task of rapid prototyping of solid parts.

Description of drawings

Fig. 1 is the molding of additive manufacturing parts of the present invention;

Fig. 2 is a schematic diagram of the UNet network structure of the present invention;

Fig. 3 is the network structure figure of ERFNet after the improvement of the present invention;

Fig. 4 is the data set diagram of flat plate additive welding seam of the present invention,

In the figure, (a1) the third laser stripe pattern of the first layer; (a2) the characteristic map of the third weld seam of the first layer;

(b1) The second laser stripe pattern of the second layer; (b2) The characteristic map of the second layer of the second weld;

(c1) The fourth laser stripe pattern of the second layer; (c2) The characteristic map of the fourth weld seam of the second layer;

(d1) The third laser stripe pattern of the third layer; (d2) The characteristic map of the third layer and the third weld seam;

(e1) The second laser fringe pattern of the fourth layer; (e2) The characteristic map of the second weld seam of the fourth layer;

(f1) The second laser fringe pattern of the fifth layer; (f2) The second weld feature map of the fifth layer;

Fig. 5 is the segmentation result figure of each layer and each track of the present invention,

In the figure, a represents the input test graph; b represents the output result graph;

(1-1)～(1-4) The laser fringe pattern and feature segmentation map of the four-pass weld of the first layer;

(2-1)～(2-4) The laser fringe pattern and feature segmentation map of the second layer of four welds;

(3-1)～(3-4) The laser fringe pattern and feature segmentation map of the four-pass weld on the third layer;

(4-1)～(4-4) The laser fringe pattern and feature segmentation map of the four-pass weld on the fourth layer;

(5-1)～(5-4) Laser fringe pattern and feature segmentation map of the fifth layer four-pass weld

Fig. 6 is a flat plate additive welding experiment weld bead planning and weld bead numbering diagram of the present invention;

Fig. 7 is the comparison diagram of the cube surface topography after each layer of the present invention is welded,

In the figure, (a1) the surface topography of the first layer of the standard workpiece; (b1) the topography of the first layer of the experimental workpiece;

(a2) The second-layer surface topography of the standard workpiece; (b2) The second-layer surface topography of the experimental workpiece;

(a3) The surface topography of the third layer of the standard workpiece; (b3) The topography of the third layer of the experimental workpiece;

(a4) The topography of the fourth layer of the standard workpiece; (b4) The topography of the fourth layer of the experimental workpiece;

(a5) The fifth-layer surface topography of the standard workpiece; (b5) The fifth-layer surface topography of the experimental workpiece.

Detailed ways

The present invention is described in further detail now in conjunction with accompanying drawing. These drawings are all simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, so they only show the configurations related to the present invention.

As shown in Figures 1 to 3, the seam tracking method based on feature segmentation of the present invention is specifically as follows:

A seam tracking method based on feature segmentation, comprising the following steps:

Step S1: collect the molten pool image, and perform ROI selection on the molten pool image; the ROI area is a fixed area including the outline of the molten pool, and the size of the ROI area is 512 pixels×512 pixels.

式中α_t为正样本时，α_t＝α∈[0，1]，α_t为负样本时，α_t＝1-α，y_c表示样本c的标签，p_c表示样本c预测为正类的概率，M为类别数量；网络结构基于SegNet与ENet，整个模型包含23层，其中1-16层为Encoder，17-23层为Decoder。网络结构采用高层特征和底层特征的多尺度特征融合，将ERFNet中第8层与第16层的特征图融合，第3层与第17层的特征图融合，第2层与第20层的特征图融合。ERFNet的帧数为120FPS以上。网络结构的高层焊缝相对于低层焊缝的RMSE与PDE均值呈上升趋势，焊缝特征点与人工标记特征点之间的偏移量为1～2个像素。Step S2: Use ErfNet's Encoder-Decoder segmentation network structure to segment the melt pool image, and use the cross-entropy loss function commonly used in semantic segmentation to fuse global feature information, and add a pyramid pooling module to the backbone network, which has a balance The formula for the cross-entropy loss function of factor α _t is:

In the formula, when α _t is a positive sample, α _t = α∈[0,1], when α _t is a negative sample, α _t =1-α, y _c represents the label of sample c, p _c represents that sample c is predicted to be positive The probability of the class, M is the number of categories; the network structure is based on SegNet and ENet, the whole model contains 23 layers, of which 1-16 layers are Encoder, and 17-23 layers are Decoder. The network structure adopts multi-scale feature fusion of high-level features and low-level features. The feature maps of the 8th and 16th layers in ERFNet are fused, the feature maps of the 3rd and 17th layers are fused, and the features of the 2nd and 20th layers are fused. Graph Fusion. The frame rate of ERFNet is above 120FPS. The average value of RMSE and PDE of the high-level welds in the network structure is on the rise compared with the low-level welds, and the offset between the weld feature points and the manually marked feature points is 1 to 2 pixels.

The technical effect of the present invention is verified by the following examples:

(1) Feature extraction algorithm of plate additive weld seam based on improved ERFNet

It can be known from the large-groove additive weld seam tracking experiment that only the laser stripe contour line is valid information in the laser stripe image collected by the weld seam tracking system, so there will be an imbalance of positive and negative samples in the ERFNet training process At the same time, since the number of welding layers and welding passes in flat plate additive welding is much larger than that of large groove filling experiments, the laser stripe characteristics of high-rise welds will be less obvious, so ERFNet is directly applied to flat plate additive welds. In the feature extraction experiment, problems such as broken lines in the center line of laser stripes and excessive deviation of welding feature points will occur. In order to avoid these problems, the present invention improves the network structure and loss function of ERFNet: the network structure refers to UNet to perform multi-scale feature fusion of high-level features and low-level features, and the loss function uses Focal Loss instead of cross-entropy loss function.

(1) Multi-scale feature fusion

In the deep convolutional neural network, according to the theory of the receptive field, the low-level feature map has higher resolution and contains more position and detail information, but due to fewer convolution operations, the semantic information is lower and the noise is more , while the high-level feature map has stronger semantic information, but due to more convolution operations, the resolution is very low and the perception of image details is poor. According to this feature of the feature map, multi-scale feature fusion can be used to improve the ability of the network to obtain feature information in the image.

Multi-scale feature fusion is to fuse the low-level feature map with rich spatial features and the high-level feature map with rich semantic information to obtain a new feature map, so that the new feature map has the characteristics of high resolution and high semantic information at the same time. The method is widely used in the fields of object detection and semantic segmentation.

This method improves the network structure of ERFNet with reference to the UNet network, and uses multi-scale feature fusion to improve the effect of network extraction of laser stripe centerlines and welding feature points. The network structure of UNet is shown in Figure 2. Same as ERFNet, UNet is also an algorithm based on the Encoder-Decoder structure, but in UNet, the Concat operation is used to splice the feature maps of the same size in the encoder and decoder, making the network In the process of upsampling, more spatial information and semantic information can be obtained, thereby improving the accuracy of segmentation.

Referring to the feature fusion method of corresponding size feature map splicing in UNet, the feature maps of the 8th layer and the 16th layer in ERFNet are fused, the feature maps of the 3rd layer and the 17th layer are fused, and the feature maps of the 2nd layer and the 20th layer are fused. Fusion, the network structure is shown in Figure 3. Different from using Concat operation to fuse feature maps in UNet, the present invention directly adds corresponding feature maps to achieve multi-scale feature fusion. The reason is that using Concat operation for feature superposition will increase the number of feature channels, resulting in increased calculation and slowing down the algorithm. The speed of the laser centerline and welding feature points is extracted. Therefore, in order to ensure the real-time performance and reliability of the algorithm in this paper, the method of directly adding the corresponding feature maps is used to achieve multi-scale feature fusion.

(2)Focal Loss

The proposal of Focal Loss comes from the field of target detection, mainly to solve the problem of extreme imbalance between positive and negative samples in target detection. In the seam tracking task of this method, only the laser stripes belong to the positive samples in the images collected by the laser vision system, and there is also the problem of imbalance between positive and negative samples. Therefore, in order to improve the algorithm's ability to extract laser stripe features, FocalLoss can be used instead of the cross-entropy loss function used in ERFNet.

平衡因子α_t可以较好的平衡正负样本比例不均，但是无法解决简单样本与困难样本的问题，所以在式平衡因子α_t的交叉熵损失函数的基础上引入参数γ，使模型更加集中于困难的、错误分类的样本。因此，FocalLoss的数学表达式如下：

式中参数γ为超参数，用于调节简单样本权重降低的速率，当γ＝0时，FocalLoss会转换为传统的交叉熵损失函数，当γ值增加时，调制因子的影响也在增加，使得易分类样本的损失减少。Focal Loss is improved on the basis of the traditional cross-entropy loss function. In the mathematical expression of the traditional cross-entropy loss function, for positive samples, the greater the output probability, the smaller the loss. For negative samples, the smaller the output probability, the smaller the loss. However, in the case of extreme imbalance between positive and negative samples Under this condition, the cross-entropy loss function may not be optimized to the optimum in the iterative process. In order to solve this problem, first introduce the balance factor α _t , for positive samples, α _t = α∈[0,1], for negative samples, α _t =1-α, then cross-entropy loss function with balance factor α _t as follows:

The balance factor α _t can better balance the uneven proportion of positive and negative samples, but it cannot solve the problem of simple samples and difficult samples, so the parameter γ is introduced on the basis of the cross-entropy loss function of the balance factor α _t to make the model more concentrated for difficult, misclassified samples. Therefore, the mathematical expression of FocalLoss is as follows:

In the formula, the parameter γ is a hyperparameter, which is used to adjust the rate of simple sample weight reduction. When γ=0, FocalLoss will be converted into a traditional cross-entropy loss function. When the value of γ increases, the influence of the modulation factor also increases, making The loss of easy-to-classify samples is reduced.

Using FocalLoss to train ERFNet can reduce the weight of a large number of simple negative samples in training, which is very important for the application scenario of the present invention.

(2) Feature extraction experiment of flat plate additive welding seam

In this experiment, the equipment and scheme of the laser vision weld seam tracking system based on the ERFNet network are used to collect and make a flat plate additive weld laser data set. Some dataset images are shown in Figure 4.

Comparing the flat-plate additive weld data set with the large-groove additive weld data set, it can be seen that the line shape of the laser stripes in the flat-plate additive welding is smoother, and the intensity of the laser stripes increases with the increase in the number of welding layers and welding passes. The edge position of the image will be weakened, the number of weld feature points will be less, and the position will be less obvious, so higher requirements are put forward for the feature extraction algorithm. In order to test the performance of the improved ERFNet, the laser fringe pattern of each layer and each weld is verified. The environment configuration of the network test is also the same as that of the laser vision seam tracking system based on the ERFNet network. Part of the test set test results are shown in Figure 5.

The segmentation effect of the network on the center line of the laser stripe can be seen from Figure 5, the improved ERFNet can achieve an ideal segmentation effect on the laser stripes of each layer and each weld seam, and there are no problems such as broken lines at key positions . The second is the core task in weld tracking: the segmentation of weld feature points. Although compared with the large groove experiment, the number of feature points in this experiment is less, and the distribution of positive and negative samples is more unbalanced. However, through the improvement of ERFNet, For each layer and each weld seam, a good segmentation effect can be achieved, and the extraction of the location of the feature point area is consistent with the actual result. This shows that the improvement of the ERFNet network structure and loss function can improve the performance of the feature extraction algorithm. At the same time, the number of frames of the improved ERFNet can reach more than 120FPS, and there is no loss of efficiency due to the introduction of multi-scale feature fusion, which still meets the requirements of real-time seam tracking. sexual demands. Therefore, the improved feature extraction algorithm fully meets the precision and time requirements of intelligent welding.

In order to accurately evaluate the performance of the improved ERFNet, RMSE is used to evaluate the extraction effect of the laser stripe centerline, and PDE is used to evaluate the accuracy of the weld feature points. The error values of each weld of each layer are shown in Table 1.

Table 1 RMES and PDE between the segmentation map and the labeling map of ERFNet after improvement

It can be concluded from Table 1 that for each layer of welds, the mean values of RMSE are 0.6181, 0.6273, 0.6199, 0.6623 and 0.6557, and the mean values of PDE are 2.2035, 2.6789, 2.2414, 2.5516 and 2.4439. Compared with the first layer of welds, the mean values of RMSE and PDE of high-level welds will increase to varying degrees, indicating that as the number of welding layers increases, the segmentation error of the laser fringe pattern will increase. The average PDE value of the first layer of welds is 2.2035, indicating that the offset between the weld feature points obtained by the improved algorithm and the manually marked feature points is about two pixels. It can be considered that the improved feature extraction algorithm can be accurate The feature points of the first-layer weld are obtained accurately. At the same time, it can be concluded from the above table that the increase of RMSE and PDE of high-rise welds compared with the first-layer welds are lower than 0.0442 and 0.4754 respectively, indicating that with the increase of welding layers, although the segmentation accuracy decreases, the welding The offset of seam feature points is less than one pixel. It can be considered that the improved algorithm can achieve ideal segmentation accuracy for high-level welds. In summary, the improvement of the present invention to ERFNet can improve the performance of the feature extraction algorithm.

In order to verify the feasibility of the improved seam tracking scheme in the field of rapid prototyping solid parts, the plate additive welding experiment is carried out below, and the deviation between the actual welding point and the artificial welding point is analyzed to judge the final forming effect. Does the seam tracking solution meet the requirements in the field of intelligent additive manufacturing.

(3) Flat plate additive welding seam tracking experiment

The experimental platform of this experiment is based on the seam tracking system and seam tracking software, and the material of the plate is Q235 stainless steel. Due to the need to verify the feasibility of the seam tracking system in this paper in the field of rapid prototyping solid parts, the part forming will be simulated by forming a cube through additive welding. In this experiment, a stainless steel flat plate will be used as the bottom plate, the number of welding layers is 5 layers, and the number of welding passes for each layer is 5, making a total of 25 welding passes. The specific weld bead planning and weld bead numbers are shown in Figure 6.

According to the above experimental plan, combined with the seam tracking system and seam tracking software, the plate additive seam tracking experiment was carried out. Similar to the large groove welding experiment, this experiment will also verify the reliability of the improved algorithm and system in the task of rapid prototyping of parts by comparing the standard workpiece with the experimental workpiece.

First, focus on the error between the taught welding point and the welding point calculated by the welding system. According to the welding scheme shown in Figure 5, the three-dimensional coordinates of the starting welding point and the ending welding point of part of the weld bead are selected and error analysis is carried out. The measurement results are shown in Tables 2-3.

Table 2 Three-dimensional coordinates and errors of the initial welding point of flat plate additive welding

Table 3 Three-dimensional coordinates and errors of the termination welding point of flat plate additive welding

From the experimental data recorded in the table, it can be seen that although the number of welding layers and the number of welding passes are more in flat plate additive welding, the deviation of each dimension coordinate between the teaching welding point and the welding point obtained by the algorithm does not exceed 1.00mm , which shows that the improved seam tracking scheme meets the high precision requirements of parts rapid prototyping tasks.

Secondly, pay attention to the shaping effect of the cube during the welding process. According to the welding scheme shown in Figure 5, after each layer is welded, the height of the cube is measured, and the height values of the experimental workpiece and the standard workpiece are compared. The measurement results are shown in Table 4.

Table 4 Height of the workpiece after welding of each layer in flat plate additive welding

It can be seen from the measurement data in the table that with the increase of the number of welding layers and welding passes, the error value will accumulate to a certain extent, but the maximum error does not exceed 0.90mm, indicating that the ideal forming effect can be achieved in the actual welding process.

While recording the height of the cube after each layer of welding is completed, compare the surface morphology of each layer of the experimental workpiece and the standard workpiece, as shown in Figure 7. It can be seen intuitively from the comparison chart of the appearance of each layer after welding that there is no obvious difference in the forming effect of each layer of the experimental workpiece and the standard workpiece. Combining the data and analysis in Table 4, it shows that this section has a good understanding of the welding seam tracking system. The improvement can meet the high-precision and real-time requirements in the field of rapid prototyping solid parts.

In order to verify the feasibility of the seam tracking scheme of the present invention in the field of rapid prototyping solid parts, a plate additive seam tracking experiment was carried out. Aiming at the problems of more complex laser stripe profile, unbalanced ratio of positive and negative samples, and more welding layers and passes in this experiment, the invention optimizes and improves the weld seam feature extraction algorithm, and introduces multi-scale features into ERFNet Fusion strategy and Focal Loss, and design experiments to verify the feasibility of the improved algorithm for flat plate additive weld feature extraction. The results show that the improved scheme adopted in the present invention can not only effectively improve the performance of the feature extraction algorithm, but also accurately extract The laser stripe centerline and weld feature points of the first-pass weld seam can not lose the efficiency of the algorithm at the same time, and ensure the real-time requirements of weld seam tracking. Through the follow-up experiment, the plate additive welding experiment was carried out to verify the adaptability of the improved seam tracking scheme under actual working conditions, and to compare and analyze the manually welded cube and the system welded cube. The results of multiple experiments showed that the forming effect of the two The difference is insignificant, which proves that the improved seam tracking scheme has high reliability in the task of rapid prototyping solid parts.

Inspired by the above-mentioned ideal embodiment according to the present invention, through the above-mentioned description content, relevant workers can make various changes and modifications within the scope of not departing from the technical idea of the present invention. The technical scope of the present invention is not limited to the content in the specification, but must be determined according to the scope of the claims.

Claims (3)

1. a seam tracking method based on feature segmentation, is characterized in that: comprise the steps: Step S1: collecting the molten pool image, and performing ROI selection on the molten pool image; Step S2: Use the Encoder-Decoder segmentation network structure of ERFNet to segment the molten pool image. The frame number of the ERFNet is above 120FPS. The network structure adopts multi-scale feature fusion of high-level features and bottom-level features. The high-level The average value of RMSE and PDE of welds relative to low-level welds is on the rise. The offset between weld feature points and manually marked feature points is 1~2 pixels, and it uses the cross-entropy loss function commonly used in semantic segmentation. In ERFNet, the feature maps of the 8th layer and the 16th layer are fused, the feature maps of the 3rd layer and the 17th layer are fused, the feature maps of the 2nd layer and the 20th layer are fused, and the pyramid pooling module is added to the backbone network. which has a balance factor

The formula of the cross entropy loss function is:

, where

When it is a positive sample,

,

When it is a negative sample,

,

Indicates the label of sample c,

Indicates the probability that sample c is predicted to be a positive class, and M is the number of categories. 2. The seam tracking method based on feature segmentation according to claim 1, characterized in that: in the step S1, the ROI area is a fixed area including the outline of the molten pool, and the size of the ROI area is 512 pixels×512 pixels. 3. The seam tracking method based on feature segmentation according to claim 1, characterized in that: the network structure in the step S2 is based on SegNet and ENet, and the whole model includes 23 layers, wherein 1-16 layers are Encoder, 17- The 23rd layer is Decoder.

Images