Image visual semantic segmentation method for insulators in power lines

Abstract

An image visual semantic segmentation method of insulators in a power line comprises the following steps: 1, collecting multiple insulator images, constructing an insulator label data set by using the multiple insulator images, and dividing the insulator label data set into a training set and a verification set; 2, constructing a backbone feature extraction network, wherein the backbone feature extraction network comprises an encoder and a decoder; 3, constructing an insulator feature enhancement module, and optimizing the encoder by using the insulator feature enhancement module; 4, adding an attention mechanism in the decoder, fusing the optimized encoder and the decoder, and obtaining an improved U-Net model; 5, training the improved U-Net model by using the training set; and 6, verifying the trained U-Net model by using the verification set. Through ablation experiments and comparison experiments, the effectiveness and superiority of the insulator segmentation method based on the residual attention mechanism are verified.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a method for visual semantic segmentation of images of insulators in power lines. Background Art

[0002] In power electronics transmission lines, insulators effectively isolate the ground from the current-carrying towers. They are critical components for stable inter-tower connections, electrical insulation, and power transmission. Therefore, insulator inspection during power line inspections is essential. Due to the complex environments in which power lines are installed, drones are currently being used to replace manual inspections in complex field environments. Cameras mounted on drones serve as sensors for environmental perception and defect detection, capturing images of insulators. Clearly and accurately segmenting insulators in aerial images makes it easier to analyze and perceive insulator status and conduct timely remote fault diagnosis. This not only ensures stable and reliable power line operation, but also saves significant manpower and material resources and significantly reduces the risk of accidents. To achieve intelligent image processing, traditional technologies primarily rely on infrared image recognition and texture information extraction and classification to automatically identify and segment insulators.

[0003] Although traditional image segmentation techniques based on threshold segmentation, edge segmentation, and region segmentation can achieve good processing results when the segmentation task is small, with the development and prevalence of power lines, the insulator segmentation task faces more complex image forms and larger sample data sets. At this time, the use of traditional insulator segmentation is not only costly and time-consuming, but also cannot achieve stable and accurate segmentation results. Therefore, with the development of deep learning in recent years, the improvement of equipment and the improvement of technology level, the insulator segmentation task has gradually shifted from traditional segmentation based on various characteristics of the target insulator to the use of semantic segmentation models based on deep learning.

[0004] To better process large amounts of image data and simplify complex problems, most deep learning-based insulator segmentation methods use convolutional neural networks as their underlying structure. Researchers achieve higher accuracy by optimizing the network structure and the encoder and decoder paths. They also use fully convolutional neural networks and encoder-decoders with asymmetric convolution kernels to extract feature information and perform discrimination and screening, and use improved conditional generative adversarial networks for pixel-level segmentation of insulators. However, these network models often suffer from complex structures and large number of model parameters. Furthermore, they are unable to effectively distinguish between foreground and background and complete segmentation when insulators are significantly disturbed by background noise.

[0005] At present, the insulator segmentation technology still has limitations that need to be improved and resolved:

[0006] 1) Different types of insulators correspond to different color textures, resolutions, and spectra. In addition, insulators have different positions and postures. These differences and randomness reduce the accuracy of insulator segmentation.

[0007] 2) The locations of power lines span a wide range, from cities to mountains and forests. Therefore, the insulator images are always in a cluttered background environment, which makes the target insulator segmentation more susceptible to background interference.

[0008] 3) Due to the jitter of the inspection drone during movement, the collected insulator images may be blurred, which requires the insulator segmentation model to be more robust. Summary of the Invention

[0009] The present invention provides a method for visual semantic segmentation of images of insulators in power lines, so as to solve the technical problems in the prior art of low accuracy of insulator image segmentation by traditional models and large background interference.

[0010] To achieve the above object, the technical solution of the present invention is achieved as follows:

[0011] The present invention provides a method for visual semantic segmentation of images of insulators in power lines, comprising the following steps:

[0012] Step S1: collecting multiple insulator images, constructing an insulator label dataset using the multiple insulator images, and dividing the insulator label dataset into a training set and a validation set;

[0013] Step S2: construct a backbone feature extraction network, which includes an encoder and a decoder;

[0014] Step S3: constructing multiple insulator feature enhancement modules, and optimizing the encoder using the multiple insulator feature enhancement modules to obtain an optimized encoder;

[0015] Step S4: construct multiple attention feature fusion modules, use the multiple attention feature fusion modules to optimize the decoder to obtain an optimized decoder, and fuse the optimized encoder and the optimized decoder to obtain an improved U-Net model;

[0016] Step S5: Using the training set to train the improved U-Net model to obtain a trained U-Net model;

[0017] Step S6: Use the validation set to verify the trained U-Net model.

[0018] Furthermore, the step S1 specifically includes the following steps:

[0019] Step S11: collecting multiple insulator images, and labeling the multiple insulator images to generate multiple corresponding insulator label images;

[0020] Step S12: performing data enhancement on the multiple insulator label images, and constructing an insulator label dataset using the multiple insulator label images after data enhancement;

[0021] Step S13: randomly divide the insulator label dataset into a training set and a validation set according to a set ratio.

[0022] Furthermore, the data enhancement in step S12 includes image rotation, image flipping and image parameter adjustment.

[0023] Furthermore, the encoder in step S2 includes 9 convolutional layers, 9 activation layers and 5 maximum pooling layers; the 9 convolutional layers are respectively the first convolutional layer to the ninth convolutional layer; the 9 activation layers are the first activation layer to the ninth activation layer, the 9 activation layers are respectively connected in sequence through the 9 convolutional layers, and the 5 maximum pooling layers are respectively connected to the input ends of the second convolutional layer, the third convolutional layer, the fifth convolutional layer, the seventh convolutional layer, and the ninth convolutional layer.

[0024] The convolutional layers all use a convolution kernel of size 3×3; the number of channels of the first to ninth convolutional layers are 64, 128, 128, 256, 256, 512, 512, 512, and 512, respectively.

[0025] Furthermore, the decoder in step S2 includes 5 transposed convolutional layers and 5 activation layers. The 5 transposed convolutional layers are connected in sequence through 5 activation layers. The 5 transposed convolutional layers are transposed convolutional layer 1 to transposed convolutional layer 5, respectively. Bilinear interpolation operation is performed between each transposed convolutional layer and the adjacent activation layer; the features output by each transposed convolutional layer after bilinear interpolation operation and activation of the corresponding activation layer are the decoding feature layer.

[0026] Furthermore, the number of the insulator feature enhancement modules in step S3 is five, namely, insulator feature enhancement module 1 to insulator feature enhancement module 5, and insulator feature enhancement module 1 to insulator feature enhancement module 5 are connected to the output ends of the first activation layer, the second activation layer, the fourth activation layer, the sixth activation layer, and the eighth activation layer, respectively;

[0027] Each of the five insulator feature enhancement modules includes three sequentially connected feature enhancement submodules, namely the first feature enhancement submodule to the third feature enhancement submodule; wherein the first feature enhancement submodule includes three tenth convolutional layers and three first filters, and the three tenth convolutional layers and the three first filters are connected in parallel; the second feature enhancement submodule includes two eleventh convolutional layers and two second filters, and the two eleventh convolutional layers and the two second filters are connected in parallel; the third feature enhancement submodule includes a twelfth convolutional layer and a third filter connected in parallel;

[0028] The scales of the tenth to twelfth convolutional layers are all 3×3; the sizes of the first to third filters are all 1×1;

[0029] The output feature layer obtained by connecting the three feature enhancement sub-modules is the encoding feature layer, which serves as the input layer of the next maximum pooling operation. This process is the downsampling process of the encoder.

[0030] Furthermore, the step S4 specifically includes the following steps:

[0031] Step S41: The coding feature layer output by the insulator feature enhancement module in step S3 and the decoding feature layer information extracted at the corresponding level are converted into one-dimensional vectors through linear transformation, and then connected and Activate the operation and obtain activation information;

[0032] Step S42: deriving an attention grid coefficient a by analyzing activation information and context information provided by the encoding feature layer in the spatial region;

[0033] Step S43: multiply the attention grid coefficient a by the input coding feature layer to obtain a new coding feature layer after the target area is highlighted. ;

[0034] Step S44: The decoding layer features are enhanced by the insulator feature enhancement module in step S3 to obtain a new decoding feature layer. ;

[0035] Step S45: New coding feature layer and the new decoded feature layer Perform skip connections to obtain the U-Net model.

[0036] Furthermore, the calculation formula of the attention grid coefficient a in step S42 is:

[0037] ;

[0038] in, 、 Represent the encoding feature layer and the decoding feature layer respectively; 、 is the vector group corresponding to its linear transformation; Computes the linear transformation of the input tensor using channel-level 1x1x1 convolution; and is the corresponding activation function.

[0039] Beneficial effects of the present invention:

[0040] The present invention provides an insulator segmentation method based on an improved U-Net using attention and residual connections. Its characteristic is that the U-Net is improved by combining the attention mechanism and residual connections to address the problem of low insulator segmentation accuracy caused by complex backgrounds, unclear contrast, and unguaranteed image quality in the insulator images of current transmission line inspection aerial photography. For the acquired insulator image, VGG11 is selected as the backbone feature extraction network to construct a U-shaped feature encoder and decoder for insulator feature extraction. In order to reduce the semantic gap in feature transfer, an insulator enhancement module is designed in each layer of the encoder. This module mainly uses residual connections for network convolution; in order to better integrate high-level features and low-level features, an attention mechanism is added to each layer of the decoder to effectively reduce the error between high- and low-level semantics and improve segmentation accuracy. The present invention confirms the effectiveness and superiority of the insulator segmentation method based on the residual attention mechanism through ablation experiments and comparative experiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a flow chart of the present invention.

[0042] Figure 2 It is a model diagram of the present invention.

[0043] Figure 3 It is a connection diagram of the encoder and decoder of the present invention.

[0044] Figure 4 It is a structural diagram of the insulator feature enhancement module of the present invention.

[0045] Figure 5 It is a flow chart of the insulator feature enhancement module of the present invention.

[0046] Figure 6 It is a structural diagram of the attention feature fusion module of the present invention.

[0047] Figure 7 It is a segmentation effect diagram of the present invention. DETAILED DESCRIPTION

[0048] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0049] In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "mounted," "connected," and "connected" should be understood in a broad sense. For example, they may refer to fixed connections, detachable connections, or integral connections; mechanical connections or electrical connections; direct connections or indirect connections through an intermediate medium; and internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0050] In addition, the terms "first," "second," and so on in the present invention are for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number or order of the technical features indicated. Therefore, a feature specified as "first" or "second" may explicitly or implicitly include at least one of such features. In the description of the present invention, "plurality" means at least two, such as two or three, unless otherwise specifically defined.

[0051] Reference Figure 1 The embodiment of the present application provides a method for visual semantic segmentation of images of insulators in power lines, comprising the following steps:

[0052] Step S1: collecting multiple insulator images, constructing an insulator label dataset using the multiple insulator images, and dividing the insulator label dataset into a training set and a validation set;

[0053] In this embodiment, step S1 specifically includes the following steps:

[0054] Step S11: collecting multiple insulator images, and labeling the multiple insulator images to generate multiple corresponding insulator label images;

[0055] Step S12: performing data enhancement on the multiple insulator label images, and constructing an insulator label dataset using the multiple insulator label images after data enhancement;

[0056] In this example, we collected 750 images of insulators with varying background complexity, contrast, and clarity. We annotated these images to generate corresponding insulator label images. Then, through rotation, flipping, image parameter adjustment, and data augmentation, we expanded the dataset to 5,320 insulator images and their corresponding labels. The label images contain two categories: pixels in the insulator are labeled with a value of 1, while pixels in the background are labeled with a value of 0.

[0057] Step S13: randomly divide the insulator label dataset into a training set and a validation set according to a set ratio.

[0058] The present invention randomly extracts insulator label data set images at a ratio of 9:1 and divides them into a training set and a verification set. The training set is used for network learning and training, and the verification set is used to verify and optimize network model parameters.

[0059] Step S2: construct a backbone feature extraction network, which includes an encoder and a decoder;

[0060] Specifically, we first pre-train the parameters of the traditional VGG11 network using images from ImageNet (a foreign image database). The pre-trained VGG11 network is used as an encoder, and a corresponding decoder is constructed. The encoder and decoder form the backbone feature extraction network.

[0061] The encoder in step S2 consists of nine convolutional layers, nine activation layers, and five max pooling layers. The nine convolutional layers are from the first to the ninth convolutional layers, and the seven activation layers are connected sequentially through nine convolutional layers. The five max pooling layers are connected to the inputs of the second, third, fifth, seventh, and ninth convolutional layers, respectively. All convolutional layers use 3×3 convolution kernels. The first convolutional layer has 64 channels, and as the network deepens, the number of channels doubles after each max pooling operation, until it reaches 512.

[0062] In the encoder, the fully connected layer of the traditional VGG11 network is removed, and a single convolutional layer with a channel number of 512 is used as the central part of the backbone feature extraction network. The central part of the backbone feature extraction network refers to the seventh convolutional layer.

[0063] The decoder consists of five transposed convolutional layers and five activation layers. The five transposed convolutional layers are connected sequentially through five activation layers. The five transposed convolutional layers are transposed convolutional layers 1 to 5, and bilinear interpolation is performed between each transposed convolutional layer and the adjacent activation layer. The output features of each transposed convolutional layer after bilinear interpolation and activation with the corresponding activation layer are the decoded feature layer. The transposed convolutional layers double the size of the feature maps of adjacent transposed convolutional layers while halving the number of channels. This process, known as upsampling, is repeated five times, corresponding to the five maximum pooling layers of the encoder. Finally, the outputs of the five transposed convolutional layers are concatenated with the outputs of the corresponding parts of the decoder. The resulting feature maps are processed through convolution operations to maintain the same number of channels as the symmetric encoder layers.

[0064] Step S3: constructing multiple insulator feature enhancement modules, and optimizing the encoder using the multiple insulator feature enhancement modules;

[0065] In step S3, there are five insulator feature enhancement modules, namely, insulator feature enhancement module 1 to insulator feature enhancement module 5, and insulator feature enhancement module 1 to insulator feature enhancement module 5 are connected to the output terminals of the first activation layer, the second activation layer, the fourth activation layer, the sixth activation layer, and the eighth activation layer, respectively.

[0066] Each of the five insulator feature enhancement modules includes three sequentially connected feature enhancement submodules, namely the first feature enhancement submodule to the third feature enhancement submodule; wherein the first feature enhancement submodule includes three tenth convolutional layers and three first filters, and the three tenth convolutional layers and the three first filters are connected in parallel; the second feature enhancement submodule includes two eleventh convolutional layers and two second filters, and the two eleventh convolutional layers and the two second filters are connected in parallel; the third feature enhancement submodule includes a twelfth convolutional layer and a third filter connected in parallel;

[0067] The scales of the tenth to twelfth convolutional layers are all 3×3; the sizes of the first to third filters are all 1×1.

[0068] In the insulator feature enhancement module, instead of using multi-layer convolution with single-line connection, residual connection is introduced to perform feature enhancement.

[0069] The output feature layer obtained by connecting the three feature enhancement sub-modules is the encoding feature layer, which serves as the input layer of the next maximum pooling operation. This process is the downsampling process of the encoder.

[0070] The insulator feature enhancement method involves adding an insulator feature enhancement module before each downsampling layer. The input features of each layer are first subjected to a simple feature extraction process using a 3x3 convolution and the corresponding activation function. This is then input into the insulator feature enhancement module for feature enhancement, minimizing the accumulated error in insulator features as they are transmitted deeper into the network. The feature layer output by this module serves as the encoding feature input layer for each level of skip connections and subsequent max pooling. This enhances insulator texture feature extraction and eliminates background interference.

[0071] Specifically, such as Figure 5 As shown, where i is the current convolution layer count, j is the number of convolution layers set for each connection in the insulator feature enhancement module, and k is the number of residual connections. If k=1, it corresponds to the first feature enhancement submodule; if k=2, it corresponds to the second feature enhancement submodule; and if k=3, it corresponds to the third feature enhancement submodule.

[0072] Step S4: construct multiple attention feature fusion modules, use the multiple attention feature fusion modules to optimize the decoder to obtain an optimized decoder, and fuse the optimized encoder and the optimized decoder to obtain a U-Net model;

[0073] Step S41: The coding feature layer output by the insulator feature enhancement module in step S3 and the decoding feature layer information extracted at the corresponding level are converted into one-dimensional vectors through linear transformation, and then connected and Activate the operation and obtain activation information;

[0074] The corresponding layers refer to the corresponding coding feature layer output by the insulator feature enhancement module 5 and the corresponding decoding feature layer output by the transposed convolution layer 1; the corresponding coding feature layer output by the insulator feature enhancement module 4 and the corresponding decoding feature layer output by the transposed convolution layer 2; and so on.

[0075] Step S42: Analyze the activation information and context information provided by the encoding feature layer in the spatial region to obtain the attention grid coefficient a that combines the high-level and low-level features. The attention grid coefficient a is expressed as:

[0076] ;

[0077] in, 、 They represent the encoding feature layer and the decoding feature layer respectively; they represent the encoding feature layer and the decoding feature layer respectively. 、 is the vector group corresponding to its linear transformation. Computes a linear transformation of the input tensor using channel-wise 1x1x1 convolution. and is the corresponding activation function.

[0078] Step S43: Multiply the attention grid coefficient a with the input coding feature layer to obtain a new coding feature layer after the target area is highlighted ;

[0079] Step S44: The decoding layer features are enhanced by the insulator feature enhancement module in step S3 to obtain a new decoding feature layer. ;

[0080] Step S45: Finally, the encoding feature layer and decoding feature layer The jump connections of the network are completed by connecting them to obtain the improved U-Net model.

[0081] In order to reduce the possible errors caused by connecting the feature maps from the encoder stage to the decoder stage, further improve the imbalance ratio of local information and global information, enhance the recognition of insulator areas and the refinement of edges, the attention connection mechanism is used as the feature fusion path between the encoding layer and the decoding layer, reduce the semantic errors between feature layers at different stages when the insulator features are fused, and alleviate the serious problem of gradient dissipation caused by the deepening of the network layers.

[0082] Step S5: Using the training set to train the improved U-Net model to obtain a trained U-Net model;

[0083] Step S6: Use the validation set to verify the trained U-Net model.

[0084] In order to verify the effectiveness and superiority of the present invention, a series of comparative experiments were conducted in this case. The specific implementation process is as follows:

[0085] The experiments in this case study used a deep learning framework based on the PyTorch environment, relying on the GPU computing resources provided by the Colaboratory platform developed by the Google team. To ensure consistency in experimental conditions, we debugged the validation model methods to uniformly set the initial learning rate to 0.00001, the momentum to 0.9, and the weight decay to 0.0005. The number of training rounds was recorded until the loss of each model was minimized and there was virtually no fluctuation.

[0086] In this case, the main indicator for judging the insulator segmentation performance is Dice score ( ), which represents the sum of twice the overlap between the predicted result and the true mask, and the formula is as follows:

[0087] ;

[0088] in and Represent the predicted segmentation results and the true segmentation labels respectively. A higher value indicates more accurate segmentation results and better model performance.

[0089] In order to verify the effectiveness of the method of the present invention, this case compares the model verification results before and after the main improvement steps through ablation experiments. After adding the insulator feature enhancement module and the attention feature fusion mechanism, the model verification results are And it has achieved good results in both recall rate and precision rate.

[0090] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art who is familiar with the technical field can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. In addition, the technical solutions between the various embodiments of the present invention can be combined with each other, but it must be based on the ability of ordinary technicians in this field to implement. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the protection scope required by the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for visual semantic segmentation of images of insulators in power lines, characterized by: The steps include: Step S1: collecting multiple insulator images, constructing an insulator label dataset using the multiple insulator images, and dividing the insulator label dataset into a training set and a validation set; Step S2: construct a backbone feature extraction network, which includes an encoder and a decoder; Step S3: constructing multiple insulator feature enhancement modules, and optimizing the encoder using the multiple insulator feature enhancement modules to obtain an optimized encoder; Step S4: construct multiple attention feature fusion modules, and use the multiple attention feature fusion modules to optimize the decoder to obtain an optimized decoder, and fuse the optimized encoder and the optimized decoder to obtain an improved U-Net model; Step S5: Using the training set to train the improved U-Net model to obtain a trained U-Net model; Step S6: Use the validation set to validate the trained U-Net model; The encoder in step S2 includes 9 convolutional layers, 9 activation layers and 5 maximum pooling layers; the 9 convolutional layers are respectively the first convolutional layer to the ninth convolutional layer; the 9 activation layers are respectively the first activation layer to the ninth activation layer, the 9 activation layers are respectively connected in sequence through 9 convolutional layers, and the 5 maximum pooling layers are respectively connected to the input ends of the second convolutional layer, the third convolutional layer, the fifth convolutional layer, the seventh convolutional layer, and the ninth convolutional layer; The decoder in step S2 includes 5 transposed convolutional layers and 5 activation layers, the 5 transposed convolutional layers are connected in sequence through the 5 activation layers, the 5 transposed convolutional layers are transposed convolutional layer 1 to transposed convolutional layer 5, and bilinear interpolation operation is performed between each transposed convolutional layer and the adjacent activation layer; the features output by each transposed convolutional layer after the bilinear interpolation operation and the corresponding activation layer are activated are the decoding feature layer; The number of the insulator feature enhancement modules in step S3 is five, namely, insulator feature enhancement module 1 to insulator feature enhancement module 5, and insulator feature enhancement module 1 to insulator feature enhancement module 5 are connected to the output ends of the first activation layer, the second activation layer, the fourth activation layer, the sixth activation layer, and the eighth activation layer, respectively; Each of the five insulator feature enhancement modules includes three feature enhancer modules connected in sequence; they are the first feature enhancer module to the third feature enhancer module; among them, the first feature enhancer module includes three tenth convolutional layers and three first filters, and the three tenth convolutional layers and the three first filters are connected in parallel; the second feature enhancer module includes two eleventh convolutional layers and two second filters, and the two eleventh convolutional layers and the two second filters are connected in parallel; the third feature enhancer module includes a twelfth convolutional layer and a third filter connected in parallel.

2. The image visual semantic segmentation method according to claim 1, characterized in that The step S1 specifically includes the following steps: Step S11: collecting multiple insulator images, and labeling the multiple insulator images to generate multiple corresponding insulator label images; Step S12: performing data enhancement on the multiple insulator label images, and constructing an insulator label dataset using the multiple insulator label images after data enhancement; Step S13: randomly divide the insulator label dataset into a training set and a validation set according to a set ratio.

3. The image visual semantic segmentation method according to claim 2, characterized in that: The data enhancement in step S12 includes image rotation, image flipping and image parameter adjustment.

4. The image visual semantic segmentation method according to claim 3, characterized in that: The convolutional layers all use a convolution kernel of size 3×3; the number of channels of the first to ninth convolutional layers are 64, 128, 128, 256, 256, 512, 512, 512, and 512, respectively.

5. The image visual semantic segmentation method according to claim 4, characterized in that: The scales of the tenth to twelfth convolutional layers are all 3×3; the sizes of the first to third filters are all 1×1; The output feature layer obtained by connecting the three feature enhancement sub-modules is the encoding feature layer, which serves as the input layer of the next maximum pooling operation. This process is the downsampling process of the encoder.

6. The image visual semantic segmentation method according to claim 5, characterized in that: The step S4 specifically includes the following steps: Step S41: The coding feature layer output by the insulator feature enhancement module in step S3 and the decoding feature layer information extracted at the corresponding level are converted into one-dimensional vectors through linear transformation, and then connected and Activate the operation and obtain activation information; Step S42: deriving an attention grid coefficient a by analyzing activation information and context information provided by the encoding feature layer in the spatial region; Step S43: multiply the attention grid coefficient a by the input coding feature layer to obtain a new coding feature layer after the target area is highlighted. ; Step S44: The decoding layer features are enhanced by the insulator feature enhancement module in step S3 to obtain a new decoding feature layer. ; Step S45: New coding feature layer and the new decoded feature layer Perform skip connections to obtain the U-Net model.

7. The image visual semantic segmentation method according to claim 6, characterized in that: The calculation formula of the attention grid coefficient a in step S42 is: ; in, 、 Represent the encoding feature layer and the decoding feature layer respectively; 、 is the vector group corresponding to its linear transformation; Computes the linear transformation of the input tensor using channel-level 1x1x1 convolution; and is the corresponding activation function.

Images