Line flexible body segmentation method based on improved DeepLabv3+

Abstract

The application provides a kind of linear flexible body segmentation method based on improved DeepLabv3+, improved light-weight MobileNetv2 network is used instead of original Xception network, and the activation function ReLU6 after 3x3 convolution and 1x1 convolution in reverse residual module is replaced by Sigmoid, the first layer convolution layer and the first seventeen Bottleneck in MobileNetv2 network are used, and the average pooling between the subsequent two 1x1 convolutions is replaced by maximum pooling, which reduces the parameter quantity in MobileNetv2.Dual attention mechanism DAMM is applied before shallow feature enters decoder 1x1 convolution, after channel stacking of improved sampling rate ASPP, and after deep and shallow feature information fusion.The problem that the existing linear flexible body segmentation technology has poor wire segmentation effect in complex scene and poor robustness is solved, the accuracy of linear flexible body semantic segmentation can be effectively improved, and the overall segmentation prediction performance of the network is improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of linear flexible body segmentation, and particularly relates to a linear flexible body segmentation method based on improved DeepLabv3+. BACKGROUND

[0002] With the increasing automation in industrial production assembly, the demand for flexible body autonomous operation is increasing. Linear flexible objects such as power cords, ropes, industrial hoses and the like are often encountered in real industrial production, but there are currently few systems that can correctly perceive the wire bundle. Work related to such materials can currently only rely on manual labor. The perception and manipulation of linear flexible bodies is still an open problem. Although the geometric structure of linear flexible bodies is very simple, they can exhibit complex and variable winding states. Different winding states may have the same topological structure, and there is no specific shape, color or feature. Moreover, the wires are too long and crisscrossed, making it difficult to distinguish each wire.

[0003] Common linear flexible body recognition techniques include threshold-based segmentation, curvature enhancement, edge detection, object detection, and point cloud-based methods. Threshold-based segmentation methods are prone to segmentation errors due to uneven lighting, noise interference, and other issues. Curvature enhancement algorithms can handle uneven lighting, but require appropriate parameter settings and can produce errors for a large number of interfering pixels. Common edge detection algorithms include Sobel and Canny operators, but their segmentation performance is unstable in complex scenarios. Object detection can be prone to errors when linear flexible bodies are close together, and cannot be recognized when they cross. Using point cloud methods to recognize linear flexible bodies has the problems of complex processing and environmental restrictions.

[0004] In view of the above, in view of the deficiencies of the prior art, a method is needed to accurately identify complex linear flexible bodies to address the limitations of current methods. SUMMARY

[0005] Therefore, the purpose of the present application is to provide a linear flexible body semantic segmentation method based on DeepLabv3+ improvement, to solve the problem of poor wire segmentation effect in complex scenes and poor robustness in existing linear flexible body segmentation technology, effectively improve the accuracy of linear flexible body semantic segmentation, and improve the overall segmentation prediction performance of the network.

[0006] In a first aspect, the present application provides a linear flexible body semantic segmentation method based on DeepLabv3+ improvement, comprising the following steps:

[0007] Step 1, collect linear flexible body image data and construct a linear flexible body image dataset;

[0008] Step 2, build an improved DeepLabV3+ semantic segmentation model, including an encoder module and a decoder module, the input encoder module uses a lightweight improved MobileNetv2 network as a backbone feature extraction network, fuses a lightweight improved DAMM double attention mechanism and an improved sampling rate ASPP module, obtains deep semantic feature information, and transmits it to the decoder module for processing;

[0009] Step 3, train the improved DeepLabV3+ semantic segmentation model in step 2 using the data set in step 1 to obtain a trained improved DeepLabV3+ network;

[0010] Step 4, based on the obtained linear flexible body image, input it into the improved DeepLabV3+ semantic segmentation model for image semantic feature processing in the encoder module, after feature extraction, perform feature decoding structure, output high semantic information to the decoder module, to segment the linear flexible body and its intersection in real time.

[0011] Further, the encoder module is used for image semantic feature processing, including a lightweight improved MobileNetv2 network as a feature extraction network, an improved sampling rate ASPP module, and an improved lightweight DAMM double attention mechanism; the decoder module is used for feature information restoration.

[0012] Further, the feature extraction network processing step is:

[0013] Step 21, compress the input feature matrix by 1x1 convolution to reduce the size of the channel;

[0014] Step 22, perform convolution processing through 3x3 DW convolution and;

[0015] Step 33, perform dimension lifting through 1x1 convolution to form an inverted residual structure, the lead picture is input into the 1x1 convolution to obtain a feature map F1, F1 passes through seventeen Bottleneck to obtain a feature map F2, 1x1 convolution is performed to generate a feature map F3, and then maximum pooling is performed to obtain a feature map F4, which reduces the parameter amount, and then 1x1 convolution is performed to generate a feature map F5, completing feature extraction.

[0016] Wherein, the activation function ReLU6 after 3x3 convolution and 1x1 convolution in the inverted residual structure is replaced by Sigmoid, the first convolution layer and the first seventeen Bottleneck in the MobileNetv2 network are adopted, and the average pooling between the subsequent two 1x1 convolutions is replaced by maximum pooling.

[0017] Further, the lightweight improved DAMM double attention mechanism is:

[0018] The max pooling in the channel attention mechanism is replaced by softpooling, the tanh activation function is used to replace the ReLU activation function in the original CAM for nonlinear activation, the sigmoid is replaced by LeakyReLU, and the output weight is obtained by LeakyReLU activation;

[0019] The spatial attention mechanism uses four 3x3 convolutions to replace the 7x7 convolution kernel in the original SAM, and a LeakyReLU activation function is added after each 3x3 convolution layer for nonlinear mapping, then a 1x1 convolution is used to adjust the channel number, and a softmax is used for activation to obtain a weight matrix

[0020] Further, the double attention mechanism DAMM is applied before the shallow feature enters the decoder 1x1 convolution, after the improved channel superposition of ASPP, and after the fusion of deep and shallow feature information.

[0021] Further, the ASPP module with improved sampling rate includes a 1x1 convolution layer, six parallel groups of dilated convolution and pooling layers;

[0022] The implementation method of the ASPP module is: on the basis of the original sampling rate, increase the sampling rate of 3, 6, and 9 dilated convolutions, and delete the sampling rate of 24 dilated convolution, reduce the number of dilated convolution kernels from 256 to 128, which is used to reduce the consumption of model size and inference time;

[0023] The ASPP formula is as follows:

[0024]

[0025] Where i represents the size of the input dilated convolution, s represents the step, k represents the original convolution kernel size, d is the number of padding spaces, and O represents the size of the feature map after dilated convolution.

[0026] Further, the steps of the feature decoding structure are:

[0027] Step 41, the shallow feature is processed by the DAMM double attention mechanism, and then a 1x1 convolution is used to adjust the dimension to obtain the shallow feature information Low_LevelFeature;

[0028] Step 42, the deep feature information High_LevelFeature obtained by processing the improved ASPP module is restored after 4 times of bilinear interpolation upsampling, and is stacked with the shallow feature information Low_LevelFeature to obtain the fused feature information concat Feature;

[0029] Step 43, after 3x3 convolution operation, enter DAMM double attention network, and perform bilinear interpolation four times upsampling to obtain a prediction map.

[0030] The embodiment of the present application brings the following beneficial effects:

[0031] The method provided by the present application first collects the lead picture data, uses a light-weight improved MobileNetv2 network as the backbone network of the semantic segmentation model, greatly reduces the model parameter quantity, so that the network convergence becomes faster. In view of the problems of large calculation amount and relatively high inference time, the ASPP uses the dilated convolution with expansion rates of 3, 6, 9, 12, 15 and 18, and uses the depth separable convolution to replace the ordinary convolution in the ASPP, which effectively reduces the parameter quantity and meets the real-time requirement of the model for segmenting the lead. In view of the problem that the model is prone to lose feature information, the light-weight DAMM double attention mechanism is added to the model, so that the extraction of the edge features of the linear flexible body is more accurate and efficient.

[0032] Other features and advantages of the present application will be described in the following specification, and some features will become apparent from the specification, or will be understood by those skilled in the art through implementation of the present application. The purpose and other advantages of the present application are achieved and obtained by the structure specifically pointed out in the specification, claims and drawings.

[0033] In order to make the above-mentioned purpose, features and advantages of the present application more obvious and easy to understand, the following preferred embodiments are specifically described below, and the accompanying drawings are described as follows. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 The line-shaped flexible body semantic segmentation method provided by the embodiment of the present application is shown in the flowchart.

[0035] Figure 2 The improved bottleneck structure diagram of the line-shaped flexible body semantic segmentation method provided by the embodiment of the present application is shown in the flowchart.

[0036] Figure 3 The improved ASPP structure of the line-shaped flexible body semantic segmentation method provided by the embodiment of the present application is shown in the flowchart.

[0037] Figure 4 The double attention mechanism diagram of the line-shaped flexible body semantic segmentation method provided by the embodiment of the present application is shown in the flowchart.

[0038] Figure 5 The improved DeepLabv3+ model diagram of the line-shaped flexible body semantic segmentation method provided by the embodiment of the present application is shown in the flowchart.

[0039] Figure 6 The segmentation result comparison diagram of the line-shaped flexible body semantic segmentation method provided by the embodiment of the present application is shown in the flowchart. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] To facilitate understanding of this embodiment, a detailed description of the linear flexible body semantic segmentation method based on DeepLabv3+ improvement disclosed in this embodiment of the invention will be provided first. The DeepLabv3+ network includes a backbone feature extraction network, an ASPP module, and a feature decoding structure. The encoder module is used for image semantic feature processing, including a lightweight improved MobileNetv2 network as the feature extraction network, an ASPP module with improved sampling rate, and an improved lightweight DAMM dual attention mechanism. The decoder module is used for feature information restoration.

[0042] Step 1: Collect image data of the linear flexible body and construct the original linear flexible body image dataset;

[0043] Step 2: Construct an improved DeepLabV3+ semantic segmentation model. Use a lightweight improved MobileNetv2 network as the backbone feature extraction network, integrate a lightweight improved DAMM dual attention mechanism and an ASPP module with improved sampling rate. After passing through the shallow effective feature layer obtained by the feature extraction network, obtain the deep semantic feature information of the wires, and transmit it to the decoder module for processing.

[0044] Furthermore, an improved lightweight MobileNetv2 network is used instead of the original Xception network, significantly reducing computational costs. The ReLU6 activation function after the 3x3 and 1x1 convolutions in the inverse residual module is replaced with Sigmoid. The improved Bottleneck architecture is shown in the diagram below. Figure 2 As shown, the first convolutional layer and the first seventeen Bottlenecks of the MobileNetv2 network are adopted, and the average pooling between the subsequent two 1x1 convolutions is replaced with max pooling. This reduces the number of parameters in MobileNetv2. The overall structure of the improved MobileNetv2 is shown in Table 1.

[0045] Table 1

[0046] Input Operation t c n s 512 X 512 X 3 Conv2d - 32 1 2 256 X 256 X 32 Bottleneck 1 16 1 1 128 X 128 X 24 Bottleneck 6 24 2 2 64 X 64 X 32 Bottleneck 6 32 3 2 32 X 32 X 64 Bottleneck 6 64 4 2 32 X 32 X 96 Bottleneck 6 96 3 1 16 X 16 X 96 Bottleneck 6 160 3 2 16 X 16 X 160 Bottleneck 6 320 1 1 16 X 16 X 320 Conv2d 1 x 1 - 1680 1 1 16 X 16 X 1280 Maxpool 16 X 16 - - 1 - 1 X 1 X 1280 Conv2d 1 x 1 - - -

[0047] Furthermore, a lightweight improved DAMM dual attention mechanism structure is as follows: Figure 4As shown, the double attention mechanism DAMM is introduced to process the high-level feature layer and the low-level feature layer, so that the model gives different weights and attention to different parts of the input image, and improves the sensitivity and accuracy of the feature extraction of the semantic segmentation network.

[0048] In the embodiment, the max pooling in the channel attention mechanism is replaced by softpooling, the tanh replaces the ReLU activation function in the original CAM for nonlinear activation, and the sigmoid is replaced by LeakyReLU, which has a fast convergence speed and achieves the effect of resource allocation.

[0049] In the embodiment, the spatial attention mechanism uses 4 3x3 convolutions instead of the 7x7 convolution kernel in the original SAM, which can reduce the parameter amount without changing the receptive field, and increases the LeakyReLU activation function after each 3x3 convolution layer to solve the problem of insufficient feature information expression in linear mapping. After adjusting the channel number to 1 by 1x1 convolution, the soft-max function is activated to obtain the weight matrix. The matrix is multiplied by the input to obtain the output of the spatial attention mechanism, which can effectively improve the semantic segmentation capability.

[0050] Further, the improved ASPP module structure after the sampling rate is as shown in Figure 3 As shown, the depth separable convolution is used to replace the ordinary convolution in the ASPP, which effectively reduces the parameter amount, improves the model performance, and further meets the real-time performance of the model for linear flexible body segmentation. Reducing the number of convolution kernels of the ASPP dilated convolution from 256 to 128 almost does not affect the segmentation performance of the model, but can reduce the consumption of model size and inference time, so the number of convolution kernels of the ASPP module is set to 128. All levels of the ASPP are stacked, and the output result is subjected to 1x1 convolution by the DAMM double attention mechanism, and high semantic information is output to the decoder.

[0051] Step 3, training the improved DeepLabV3+ semantic segmentation model in step 2 by using the data set in step 1;

[0052] Step 4, based on the obtained linear flexible body image, inputting the image into the improved DeepLabV3+ semantic segmentation model for image semantic feature processing of the encoder module, obtaining two feature maps, and inputting the two feature maps into the improved sampling rate ASPP module and the decoder module, respectively, and inputting the shallow layer features output by the decoder module into the lightweight improved DAMM double attention mechanism for 1x1 convolution, and outputting high semantic information to the decoder module.

[0053] In this embodiment, the specific implementation method in step 2 is: the feature extraction network processing process is first to compress the input feature matrix by 1x1 convolution to reduce the size of the channel, then to perform convolution processing through 3x3 DW convolution and, and finally to perform dimensionality reduction through 1x1 convolution to form an inverted residual structure. The difference between the improved Mobilenetv2 network and the original network is that the inverted residual structure adopts a Sigmoid activation function, which can avoid information loss. In the 1x1 convolution of the wire picture input, the feature map F1 is obtained, F1 passes through seventeen Bottleneck structures to obtain the feature map F2, and then 1x1 convolution is performed to generate the feature map F3. After that, the maximum pooling is performed to obtain the feature map F4, which reduces the parameter amount. Then, 1x1 convolution is performed to generate the feature map F5, and the feature extraction is completed.

[0054] Further, the improved ASPP module (Atrous Spatial Pyramid Pooling) obtains information of different scales by adding convolution kernels of different sampling rates for corresponding convolution operations. The original ASPP structure extracts a coefficient of features, thereby losing part of the local information of the image. The improved ASPP increases the sampling rate to 3, 6, and 9 on the basis of the original sampling rate, and deletes the sampling rate of 24, thereby reducing the loss rate of edge information. The number of dilated convolution kernels is reduced from 256 to 128, which almost does not affect the segmentation performance of the model, but can reduce the consumption of model size and inference time. The ASPP formula is as follows:

[0055]

[0056] wherein i represents the size of the input dilated convolution, s represents the step, k represents the original size of the convolution kernel, d represents the number of padding spaces, and O represents the size of the feature map after dilated convolution.

[0057] Then, the channel superposition of the dilated convolution features of different scales is performed, and then the DAMM double attention network is passed, and then 1x1 convolution, pooling, and Relu activation function are performed. Finally, the dimensionality is reduced to a specific channel number through Dropout to establish the connection between different channels and obtain the High_LevelFeature.

[0058] Further, the specific implementation method in step 4 is: the feature decoding structure mainly includes two upsampling, both of which are four times upsampling using bilinear interpolation to restore the image size. First, the shallow features are processed by DAMM double attention mechanism, and then the dimension is adjusted by 1x1 convolution to obtain Low_Level Feature. The High_Level Feature obtained by improved ASPP processing is restored by four times upsampling using bilinear interpolation, and then stacked with Low_Level Feature by concat to obtain concat Feature. After 3x3 convolution operation, it enters DAMM double attention network, and is four times upsampling using bilinear interpolation to obtain the prediction map.

[0059] Further, in the channel attention mechanism of DAMM double attention mechanism, the input group of feature channels is first averaged and soft pooled to extract the overall features and details of the feature map respectively. After each pooling, 3x3 convolution kernel is used for dimension reduction, and then tanh activation function is used for nonlinear activation to avoid neuron inactivation in the back propagation process. Then, 3x3 convolution is used for dimension increase, the processed results are stacked, and finally LeakyReLU activation is used to obtain the output weight.

[0060] The spatial attention mechanism first performs maximum pooling and average pooling on the input image. Considering the real-time performance of the lead wire segmentation, 4 3x3 convolution kernels are used instead of the original 7x7 convolution kernel in the spatial attention mechanism, which increases the receptive field and reduces the parameter quantity. After each convolution kernel, a LeakyReLU activation function is added for nonlinear mapping to solve the problem of insufficient feature information in linear mapping. Then, 1x1 convolution is used to adjust the channel number, and softmax is used for activation to obtain the weight matrix.

[0061] Among them, the proposed double attention mechanism DAMM is applied to three positions: (1) before the shallow features enter the decoder 1x1 convolution; (2) after the channel stacking of the improved ASPP; (3) after the fusion of deep and shallow feature information.

[0062] The data set used by the present application is a public data set, the image size is [512, 512], including a line flexible body picture set and a corresponding labeled picture set, the sample label corresponding to the image is a 512x512 image obtained by binaryzation processing of a pure line flexible body image, the training set, the verification set and the test set are divided according to the ratio of 8:1:1, the pictures used for testing are divided into four categories, the first category of pictures is that only the line flexible body is located on a curved surface without other interference objects, the difficulty of this scene is that the line flexible body and the background may have possible chroma similarity; the second category of pictures is that the line flexible body is only on a high feature and a complex background without other interference objects, the challenge of this scene is to extract correct features belonging to the line flexible body in a chaotic scene; the third category of line flexible body is in a realistic industrial environment, and the difficulty may be that the metal surface reflects the wire and other interference objects; the fourth category of line flexible body picture is a combination of the difficulties in the previous three scenes.

[0063] In the training process, necessary parameters are set, the learning rate is set as Lr=0.005, the weight decay is set as weight_decay=0.0002, the batch_size is set as 16, only the wire object needs to be segmented, the total category is K+1, contains a background, so num_calsses=2, and 500 iterations of training are performed. The accuracy of the performance of the semantic segmentation model can be directly observed from the segmentation graph, and MPA and Miou can also be used for evaluation.

[0064] MioU calculates the ratio of the intersection and union value of the predicted value and the label of all categories in the image, and then calculates the average value of all category intersection and union ratios, and the formula is as follows:

[0065]

[0066] MPA is to first calculate the ratio of the number of pixels predicted as positive to the total number of pixels of each class, and then calculate the average of each class, and the calculation formula is as follows:

[0067]

[0068] Wherein, K+1 is the number of categories, P ii is the sum of the pixel points predicted as positive but actually positive, P ij is the sum of the pixel points predicted as positive but actually negative, P ji is the sum of the pixel points predicted as negative but actually positive.

[0069] In order to verify the reliability and practicability of the improved deeplabv3+ model, the original deeplabv3+ model and the improved deeplabv3+ model are compared, and the results of each model on the test set are shown in Table 2.

[0070] Table 2

[0071] Segmentation model Miou MPA Original deeplabv3+ 78.5％ 85.6％ Improved deeplabv3+ 82.2％ 88.4％

[0072] From the experimental results, the improved deeplabv3+ model is improved compared with the existing model, Miou is improved by 3.7%, reaching 82.2%, and MPA is improved by 2.8%, reaching 88.4%; the prediction results are shown in Table 2. Figure 6

[0073] From the segmentation results, the original deeplabv3+ model can roughly segment the linear flexible body, but the segmentation effect is not ideal, the intersection segmentation is not accurate, the edge details are not handled, and the segmentation range is too wide. The reason is that the image feature extraction of the linear flexible body is not perfect. The improved deeplabv3+ model has better segmentation effect than the original deeplabv3+ model, and can better identify and segment the linear flexible body and its intersection, and handle the edge details of the linear flexible body well. In summary, the improved deeplabv3+ linear flexible body segmentation model effectively improves the accuracy, robustness and generalization of the linear flexible body scene semantic segmentation.

[0074] It should be noted that in this paper, the term "includes", "contains" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or includes elements inherent to such process, method, article or equipment.

[0075] Although the embodiments of the present application have been shown and described, it can be understood by those skilled in the art that various changes, modifications, replacements and variations can be made to the embodiments without departing from the principles and spirits of the present application, and the scope of the present application is defined by the appended claims and their equivalents.​

Claims

1. A linear flexible volume segmentation method based on an improved DeepLabv3+ network, wherein the DeepLabv3+ network comprises a backbone feature extraction network, an ASPP module, and a feature decoding structure, characterized in that, The steps include the following: Step 1: Collect image data of the linear flexible body and construct an image dataset of the linear flexible body; Step 2: Construct an improved DeepLabV3+ semantic segmentation model, including an encoder module and a decoder module. The input encoder module uses a lightweight improved MobileNetv2 network as the backbone feature extraction network, integrates a lightweight improved DAMM dual attention mechanism and an ASPP module with improved sampling rate to obtain deep semantic feature information, and transmits it to the decoder module for processing. The feature extraction network processing steps are as follows: Step 21: Compress the input feature matrix using 1x1 convolution to reduce the channel size; Step 22: Perform convolution processing using a 3x3 DW convolution; Step 33: Upscale the structure by using 1x1 convolution to form an inverse residual structure. Input the wire image into the 1x1 convolution to obtain feature map F1. F1 passes through seventeen Bottlenecks to obtain feature map F2. Perform 1x1 convolution to generate feature map F3. Then, perform max pooling to obtain feature map F4, reducing the number of parameters. Finally, perform 1x1 convolution to generate feature map F5, completing feature extraction. In this process, the ReLU6 after the 3x3 convolution and 1x1 convolution in the inverted residual structure is replaced with Sigmoid, the first convolutional layer and the first seventeen Bottlenecks in the MobileNetv2 network are used, and the average pooling between the two subsequent 1x1 convolutions is replaced with max pooling. The lightweight improved DAMM dual attention mechanism is as follows: the max pooling in the channel attention mechanism is replaced with soft pooling, the tanh activation function is replaced with the ReLU activation function in the original CAM for nonlinear activation, the sigmoid is replaced with Leaky ReLU, and the output weights are obtained through Leaky ReLU activation. The spatial attention mechanism replaces the 7x7 convolution kernel in the original SAM with four 3x3 convolutions, and adds a LeakyReLU activation function after each 3×3 convolution layer for non-linear mapping, then uses 1x1 convolution to adjust the number of channels, and uses softmax for activation to obtain the weight matrix. Step 3: Train the improved DeepLabV3+ semantic segmentation model from Step 2 using the dataset from Step 1 to obtain the trained improved DeepLabV3+ network. Step 4: Based on the acquired linear flexible body image, input it into the improved DeepLabV3+ semantic segmentation model for image semantic feature processing of the encoder module. After feature extraction, perform feature decoding and output high semantic information to the decoder module to segment the linear flexible body and its intersections in real time.

2. The linear flexible body segmentation method according to claim 1, characterized in that, The encoder module is used for processing image semantic features, including a lightweight improved MobileNetv2 network as a feature extraction network, an ASPP module with improved sampling rate, and an improved lightweight DAMM dual attention mechanism. The decoder module is used to restore feature information.

3. The linear flexible body segmentation method according to claim 1, characterized in that, The dual attention mechanism DAMM is applied before shallow features enter the decoder's 1x1 convolution, after channel stacking in the improved ASPP, and after the fusion of deep and shallow feature information.

4. The linear flexible body segmentation method according to claim 1, wherein the ASPP module for improving the sampling rate in step 2 is characterized in that, The improved sampling rate ASPP module includes a 1x1 convolutional layer, six parallel groups of dilated convolutional and pooling layers; The ASPP module is implemented by adding dilated convolutions with sampling rates of 3, 6, and 9 on the basis of the original sampling rate, and deleting dilated convolutions with a sampling rate of 24, reducing the number of dilated convolution kernels from 256 to 128, in order to reduce the model size and inference time consumption. The ASPP formula is as follows: Where i represents the size of the input dilated convolution, s represents the stride, k represents the original size of the convolution kernel, d represents the number of spaces to fill, and O represents the size of the feature map after dilated convolution.

5. The method for segmenting a linear flexible body according to claim 1, characterized in that, The steps for feature decoding structure in step 4 are as follows: Step 41: The shallow features are processed by the DAMM dual attention network, and then the dimensions are adjusted by 1x1 convolution to obtain the shallow feature information Low_Level Feature. Step 42: The deep feature information High_Level Feature obtained by the improved ASPP module is restored by bilinear interpolation and upsampling by 4 times, and then concatted with the shallow feature information Low_Level Feature to obtain the fused deep and shallow feature information concatFeature. Step 43: After 3x3 convolution, the image is fed into the DAMM dual attention network and bilinear interpolation is performed to upsample it four times to obtain the prediction image.

Images