Safety rope semantic segmentation method and system based on codec network structure

Abstract

The invention relates to the technical field of image processing, and discloses a safety rope semantic segmentation method and system based on a codec network structure, and the method comprises the steps: firstly obtaining and preprocessing a to-be-segmented safety rope image, inputting the preprocessed image into a pre-training codec network, and obtaining a pixel-level semantic segmentation result through forward propagation. The codec network comprises an encoder, a multi-scale feature extraction module, a feature calibration module and a decoder which are connected in sequence; the encoder samples and extracts multi-level features by using a pooling-free convolutional layer; the multi-scale feature extraction module processes advanced features through parallel convolution branches of different receptive fields, and fuses the advanced features to generate a multi-scale aggregation feature map; the feature calibration module optimizes features by using a double-attention mechanism of local interactive attention and channel weight calibration; and the decoder combines up-sampling and encoder jump connection to recover the resolution, and outputs a safety rope semantic segmentation map. The safety rope can be accurately segmented, and the method is suitable for safety monitoring of high-altitude operation, industrial construction and other scenes.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically, to a method and system for semantic segmentation of safety ropes based on a codec network structure. Background Technology

[0002] Safety ropes are core safety equipment in high-altitude operations and industrial construction, and their proper wearing and condition monitoring are directly related to the lives of workers. With the development of computer vision technology, semantic segmentation technology can achieve pixel-level division of target areas in images, providing key technical support for the automated detection of safety ropes and becoming an important means of ensuring safety in high-altitude operations and industrial production.

[0003] Existing semantic segmentation methods are mainly based on encoder-decoder architectures such as U-Net and DoveNet, but they have significant technical shortcomings in safety rope segmentation tasks, making it difficult to meet the needs of practical applications. First, safety ropes have a wide scale range, with details as small as 1-2 pixels and long distances extending to the entire image screen, making it difficult for existing networks to accurately capture multi-scale features simultaneously. Second, safety ropes have a single color and sparse texture, easily blending with backgrounds such as building structures, equipment pipelines, and the natural environment. Redundant information interference exists during cross-scale feature fusion, leading to decreased segmentation accuracy. Third, safety ropes have blurred edges and strong continuity, making existing methods prone to segmentation breaks and loss of details, failing to fully reconstruct the shape of the safety rope. Fourth, existing models have a large number of parameters and high computational complexity, making it difficult to meet the deployment requirements of real-time safety monitoring, thus limiting their widespread application in practical industrial scenarios.

[0004] In addition, in high-altitude operations and industrial construction scenarios, safety ropes often face complex situations such as obstruction and sudden changes in lighting, which further exacerbates the problems of sudden changes in local scale and weak edge features, making it difficult to guarantee the segmentation accuracy and continuity of traditional semantic segmentation methods. Summary of the Invention

[0005] In response, this invention provides a method and system for semantic segmentation of safety ropes based on an encoder-decoder network structure, in order to at least partially solve the above-mentioned technical problems.

[0006] This invention provides a semantic segmentation method for safety ropes based on an encoder-decoder network structure, comprising the following steps:

[0007] S1. Obtain the safety rope image to be segmented, and preprocess the safety rope image;

[0008] S2. Input the preprocessed safety rope image into a pre-trained encoder-decoder network, and obtain pixel-level semantic segmentation results through the forward propagation of the encoder-decoder network;

[0009] The encoder-decoder network comprises an encoder, a multi-scale feature extraction module, a feature calibration module, and a decoder connected in sequence. The encoder downsamples the input image and extracts multi-level features using unpooled convolutional layers. The multi-scale feature extraction module performs convolution operations on the high-level features output by the encoder using parallel convolutional branches with different receptive fields, and fuses the feature maps output by each branch to generate a multi-scale aggregated feature map. The feature calibration module optimizes the multi-scale aggregated feature map using a dual attention mechanism that includes local interactive attention and channel weight calibration. The decoder gradually restores the image resolution through upsampling operations combined with skip connections from the encoder, ultimately outputting a semantic segmentation map of the safety rope.

[0010] In one possible embodiment, a dynamic scale adaptation and edge enhancement attention module is further embedded between the multi-scale feature extraction module and the feature calibration module; the dynamic scale adaptation and edge enhancement attention module is used to receive the multi-scale aggregated feature map, and sequentially perform dynamic scale perception and edge enhancement interactive processing, and output a scale-edge co-optimized feature map to the feature calibration module.

[0011] In one possible embodiment, the dynamic scale-aware processing specifically includes:

[0012] Based on the multi-scale aggregated feature map, multiple scale probability maps corresponding to different scale attributes are generated.

[0013] Set up multiple parallel convolution branches corresponding to the scale attribute, and each branch uses convolution kernels or dilated convolutions of different sizes;

[0014] For each spatial location in the feature map, based on its probability value on the probability map at each scale, the feature vectors of the corresponding parallel convolutional branches at that location are weighted and fused to generate a scale-adapted feature map.

[0015] In one possible embodiment, the edge-enhanced interaction processing specifically includes:

[0016] Edge responses are extracted from the scale-adapted feature map, and an edge attention weight map is generated;

[0017] For pixel positions in the edge attention weight map with a weight higher than a preset threshold, calculate the feature similarity between them and their neighboring pixels on the scale-adaptive feature map.

[0018] The feature vectors of neighboring pixels are weighted and aggregated based on the feature similarity, and the aggregated features are fused with the original scale-adapted feature map through residual connections.

[0019] In one possible embodiment, the loss function used in the pre-training of the codec network includes an edge alignment loss; the edge alignment loss is calculated based on edge information extracted from the ground truth labeled mask and an edge attention weight map obtained from the edge enhancement interaction process, and is used to supervise the model's recognition of the safety rope edge.

[0020] In one possible embodiment, in the dynamic scale-aware processing,

[0021] For convolutional branches used to capture long-range continuous features, the dilation rate of their dilated convolutions is dynamically adjusted according to the probability value of the corresponding scale, so that the dilation rate increases with the increase of the confidence of that scale.

[0022] In another aspect, this application also provides a safety rope semantic segmentation system based on a codec network structure, comprising:

[0023] An image preprocessing module is used to acquire the safety rope image to be segmented and to preprocess the safety rope image;

[0024] The semantic segmentation module is used to input the preprocessed safety rope image into a pre-trained encoder-decoder network, and obtain pixel-level semantic segmentation results through the forward propagation of the encoder-decoder network;

[0025] The encoder-decoder network comprises an encoder, a multi-scale feature extraction module, a feature calibration module, and a decoder connected in sequence. The encoder downsamples the input image and extracts multi-level features using unpooled convolutional layers. The multi-scale feature extraction module performs convolution operations on the high-level features output by the encoder using parallel convolutional branches with different receptive fields, and fuses the feature maps output by each branch to generate a multi-scale aggregated feature map. The feature calibration module optimizes the multi-scale aggregated feature map using a dual attention mechanism that includes local interactive attention and channel weight calibration. The decoder gradually restores the image resolution through upsampling operations combined with skip connections from the encoder, ultimately outputting a semantic segmentation map of the safety rope.

[0026] This application also provides an electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the safety rope semantic segmentation method based on a codec network structure as described above.

[0027] In another aspect, this application provides a computer-readable storage medium having stored thereon computer program instructions that can be executed by a processor to implement the safety rope semantic segmentation method based on a codec network structure as described above.

[0028] Another aspect of this application provides a computer program product, including a computer program that, when executed by a processor, implements the safety rope semantic segmentation method based on a codec network structure as described above.

[0029] This application addresses the segmentation fragmentation and detail loss issues in existing safety rope segmentation technologies by embedding a dynamic scale adaptation and edge enhancement attention module between the multi-scale feature extraction and feature calibration modules in the codec network. The module employs dynamic scale perception, dynamically adapting the receptive field based on pixel scale probability to accurately capture multi-scale features of the safety rope, ranging from fine details to the entire screen. Furthermore, the long-distance branch hole rate is dynamically adjusted with scale confidence, further improving the accuracy of long-scale feature capture. Combined with edge enhancement interaction, edge attention weights and neighborhood feature similarity aggregation strengthen the continuity of weak edges, preventing isolated edge segments. Simultaneously, a weighted loss function combining foreground normalized mean square error loss, Dice loss, and edge alignment loss addresses the issues of small foreground proportion and pixel imbalance, balancing high-precision segmentation with real-time deployment requirements, and effectively adapting to safety monitoring in scenarios such as high-altitude operations and industrial construction. Attached Figure Description

[0030] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0031] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0032] Figure 1 This is a schematic diagram of a semantic segmentation method for safety ropes based on a codec network structure, provided in an embodiment of the present invention.

[0033] Figure 2 This is a schematic diagram of the codec network structure provided in an embodiment of the present invention.

[0034] Figure 3 This is a schematic diagram of the dynamic scale processing flow provided in an embodiment of the present invention.

[0035] Figure 4 This is a schematic diagram of the edge enhancement processing flow provided in an embodiment of the present invention.

[0036] Figure 5 This is a schematic diagram of the feature calibration module processing provided in an embodiment of the present invention.

[0037] Figure 6 This is a schematic diagram of a safety rope semantic segmentation system based on a codec network structure provided in an embodiment of the present invention.

[0038] Figure 7 This is a schematic diagram of the structure of a device provided in an embodiment of the present invention. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0040] It should be noted that all user information (including but not limited to user device information, user personal information, object information corresponding to device usage data, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, device usage data, etc.) involved in all embodiments of this disclosure are information and data authorized by the user or fully authorized by all parties.

[0041] The implementation process of the safety rope semantic segmentation method based on the encoder-decoder network structure described in this invention will be explained in detail below with reference to specific embodiments. It should be noted that this embodiment is only used to explain this invention and is not intended to limit the scope of protection of this invention. Conventional adjustments or substitutions made by those skilled in the art to each step without departing from the concept of this invention should be included within the scope of protection of this invention.

[0042] like Figure 1 As shown in the figure, an embodiment of the present invention discloses a semantic segmentation method for safety ropes based on an encoder-decoder network structure, schematic diagram 100, which includes the following method steps:

[0043] S1. Obtain the safety rope image to be segmented, and preprocess the safety rope image;

[0044] S2. Input the preprocessed safety rope image into a pre-trained encoder-decoder network, and obtain pixel-level semantic segmentation results through the forward propagation of the encoder-decoder network;

[0045] The encoder-decoder network comprises an encoder, a multi-scale feature extraction module, a feature calibration module, and a decoder connected in sequence. The encoder downsamples the input image and extracts multi-level features using unpooled convolutional layers. The multi-scale feature extraction module performs convolution operations on the high-level features output by the encoder using parallel convolutional branches with different receptive fields, and fuses the feature maps output by each branch to generate a multi-scale aggregated feature map. The feature calibration module optimizes the multi-scale aggregated feature map using a dual attention mechanism that includes local interactive attention and channel weight calibration. The decoder gradually restores the image resolution through upsampling operations combined with skip connections from the encoder, ultimately outputting a semantic segmentation map of the safety rope.

[0046] In some embodiments, for step S1, the image preprocessing in step S1 aims to convert the original safety rope image to be segmented into standardized feature data that meets the network input requirements, while retaining the key feature information of the safety rope in the image, so as to provide high-quality input for subsequent network inference.

[0047] For example, the resolution of the original safety rope image to be segmented may include common specifications in practical applications such as 1080P (1920×1080) and 4K (3840×2160). Since each module of the network has a fixed requirement for the resolution of the input feature map, the image is first normalized in size.

[0048] Preferably, a bilinear interpolation algorithm is used to uniformly scale the original image to a resolution of 512×512. This resolution ensures that the fine-scale features (1-2 pixels) of the safety rope are not lost, while balancing computational complexity and inference efficiency. During the scaling process, the aspect ratio of the image is kept constant. For blank areas generated after scaling, black pixels (all RGB values ​​are 0) are used to fill them, ensuring that the size of the output image is strictly 512×512.

[0049] According to an embodiment of the present invention, after size normalization, the image needs to be standardized to eliminate the impact of channel distribution differences on network training and inference. Specifically, the RGB channel mean and standard deviation of the ImageNet dataset are used as standardization parameters, where the RGB mean is (0.485, 0.456, 0.406) and the standard deviation is (0.229, 0.224, 0.225). The calculation formula for the standardization process is as follows:

[0050]

[0051] in, Indicates the original image in channels Pixel position Pixel value at that location, This represents the mean of the corresponding channel. This represents the standard deviation of the corresponding channel. This represents the standardized pixel value. This formula maps the pixel values ​​of each channel of the image to the interval [-1, 1], accelerating the feature convergence speed during the network's forward propagation.

[0052] Optionally, the standardized image data is converted to a tensor format supported by the network. For network implementations based on the PyTorch framework, the image's H×W×C (height × width × number of channels) format is converted to C×H×W format, and a batch dimension is added, resulting in a final tensor format of [dimensionality missing]. The input tensor is denoted by (where 1 represents the batch size for inference on a single image, and 3 represents the RGB three channels). Understandably, if batch inference is performed, the batch dimension can be adjusted to an integer greater than 1 depending on the hardware's computing power, but the subsequent processing flow remains consistent.

[0053] In some embodiments, step S2 involves using a pre-trained encoder-decoder network to perform feature extraction, optimization, and segmentation inference on the preprocessed image, ultimately outputting pixel-level semantic segmentation results for the safety rope. See also... Figure 2 , Figure 2 This is a schematic diagram of the codec network structure provided in an embodiment of the present invention. The codec network includes an encoder, a multi-scale feature extraction module, a dynamic scale adaptation and edge enhancement attention module, a feature calibration module, and a decoder connected in sequence. The specific implementation process of each module is as follows:

[0054] First, regarding the encoder, according to an embodiment of the present invention, the encoder adopts a pooling-free layer design. Its core purpose is to avoid the loss of safety rope edge details caused by traditional pooling operations, while simultaneously achieving hierarchical feature extraction through progressive downsampling. Specifically, the encoder consists of four consecutive convolutional blocks, and the structure and function of each convolutional block are implemented as follows:

[0055] (1) Convolutional block structure design

[0056] Each convolutional block contains two 3×3 convolutional layers, a batch normalization layer, and a ReLU activation function. The implementation order of each component is as follows: 3×3 convolutional layer → batch normalization layer → ReLU activation function → 3×3 convolutional layer → batch normalization layer → ReLU activation function.

[0057] For example, the stride of the 3×3 convolutional layer is set to 1 and the padding is set to 1 to ensure that the resolution of the feature map does not change after the convolution operation; the momentum parameter of the batch normalization layer is set to 0.9 and the epsilon parameter is set to 1e-5 to normalize the distribution of the output features of the convolutional layer and alleviate the gradient vanishing problem; the ReLU activation function enhances the non-linear expressive power of the features by setting negative values ​​to zero.

[0058] (2) Downsampling and feature output

[0059] Between the four convolutional blocks, a downsampling operation is performed using a 3×3 convolutional layer with a stride of 2. The padding of this convolutional layer is set to 1, which halves the resolution of the feature map and doubles the number of channels after each downsampling. The specific downsampling and feature output process is as follows:

[0060] The first convolutional block receives the input tensor (1, 3, 512, 512), which is processed by two internal 3×3 convolutional layers and activation functions to output a feature map with 64 channels and a resolution of 512×512.

[0061] After downsampling by a 3×3 convolutional layer with a stride of 2, the feature map resolution becomes 256×256, while the number of channels remains unchanged at 64, and is then input into the second convolutional block.

[0062] The second convolutional block outputs a feature map with 128 channels and a resolution of 256×256. This map is then downsampled to 128×128 resolution by a convolutional layer with a stride of 2 and input into the third convolutional block.

[0063] The third convolutional block outputs a feature map with 256 channels and a resolution of 128×128. This map is then downsampled to a resolution of 64×64 by a convolutional layer with a stride of 2 and input into the fourth convolutional block.

[0064] After processing by the fourth convolutional block, the final output is a high-level semantic feature map with 256 channels and a resolution of 64×64.

[0065] Understandably, the encoder ultimately outputs four sets of feature maps at different levels, with resolutions of 512×512, 256×256, 128×128, and 64×64, and channel numbers of 64, 128, 256, and 256, respectively. These feature maps successively contain low-level features such as the edge and texture of the safety rope and high-level features such as semantics, providing multi-dimensional feature inputs for subsequent modules.

[0066] For the multi-scale feature extraction module, this module captures multi-scale information of the safety rope, from fine-scale texture to long-range continuous features, through parallel convolutional branches with different receptive fields. The specific implementation process is as follows:

[0067] First, (1) receive input features. The multi-scale feature extraction module receives the highest-level semantic feature map output by the encoder, that is, the feature map with 256 channels and a resolution of 64×64. This feature map contains the global semantic information of the safety rope and is the basis for multi-scale feature extraction.

[0068] (2) Design of parallel branches: Specifically, according to an embodiment of the present invention, the module is designed with three parallel convolutional branches. Each branch achieves differentiation of the receptive field through convolutional kernels of different sizes or dilated convolution, thereby specifically capturing safety rope features at different scales:

[0069] Fine-scale branch: Employs a 3×3 convolutional layer with a stride of 1, padding of 1, and 128 kernels. This branch has a small receptive field, enabling it to accurately capture texture features of 1-2 pixels in the finer details of the safety rope.

[0070] Mesoscale branch: Employs a 5×5 convolutional layer with a stride of 1, padding of 2, and 128 kernels. This branch has a medium receptive field and is used to capture local structural features of the safety rope, such as morphological features at bends.

[0071] Long-distance branch: Employs a 7×7 dilated convolutional layer with a stride of 1, padding of 6, a dilation rate of 3, and 128 kernels. Dilated convolution expands the receptive field without increasing the number of parameters by introducing holes into the kernels. This branch can effectively capture continuous features of a long-distance safety rope extension (up to the entire image screen).

[0072] (3) Branch feature processing and fusion: Specifically, the feature maps output by each parallel branch (all 128 channels, 64×64 resolution) need to be processed by instance normalization and LeakyReLU activation function, and then channel splicing and fusion are performed. Specifically:

[0073] The epsilon parameter of the instance normalization layer is set to 1e-5 to normalize the feature map of each sample, thereby enhancing the robustness of the features.

[0074] The LeakyReLU activation function has a negative slope of 0.01, which, compared to the ReLU function, can preserve negative numerical features and avoid loss of feature information.

[0075] Channel splicing and fusion involves superimposing the feature maps processed by the three branches along the channel dimension, ultimately generating a multi-scale aggregated feature map with 128×3=384 channels and a resolution of 64×64.

[0076] According to a preferred embodiment of the present invention, the codec network further includes a dynamic scale adaptation and edge enhancement attention module, which is embedded between the multi-scale feature extraction module and the feature calibration module to solve the problems of local scale abrupt changes and weak edges caused by occlusion and sudden changes in illumination of the safety rope.

[0077] Specifically, the core objective of dynamic scale-aware processing is to enable each pixel in a multi-scale aggregated feature map to adaptively match the optimal receptive field based on its own local scale attributes, ensuring consistency between feature representation and scale attributes. (See also: [link to relevant documentation]). Figure 3 , Figure 3 This is a schematic diagram of the dynamic scale processing flow provided in an embodiment of the present invention. The specific implementation process is as follows:

[0078] In S301, based on the multi-scale aggregated feature map (384 channels, 64×64 resolution), multiple scale probability maps corresponding to different scale attributes are generated through three sets of parallel 1×1 convolutional layers.

[0079] Specifically, firstly, three sets of parallel 1×1 convolutional layers are set up, each containing 128 convolutional kernels. The stride of the convolutional layers is 1, with no padding. The three sets of convolutional layers simultaneously receive multi-scale aggregated feature maps with 384 channels and output intermediate feature maps with 128 channels and 64×64 resolution, corresponding to three scale attributes: fine scale, medium scale, and long scale.

[0080] Secondly, for each group of 128-channel intermediate feature maps, channel compression is performed using a single 1×1 convolutional kernel (3 in total), converting the 128-channel feature maps into single-channel feature maps, resulting in three original scale response maps with a resolution of 64×64, denoted as follows: (Detailed scale) (Mesoscale) (Long-scale);

[0081] Finally, the three original scale response maps were subjected to Softmax normalization to obtain the scale probability map. , , ,in ( ; ; ) represents a pixel Belongs to the The probability value at this scale satisfies The probability value ranges from [0,1]. The formula for calculating Softmax normalization is as follows:

[0082]

[0083] This step provides a basis for the confidence distribution of each pixel across three scales, enabling subsequent dynamic receptive field adaptation.

[0084] In S302, three parallel convolutional branches corresponding to the scale attribute are set up. The kernel size and type of each branch correspond to the branches of the multi-scale feature extraction module to ensure the consistency of the scale attribute.

[0085] Specifically, the fine-scale corresponding branch uses a 3×3 convolutional layer with a stride of 1, padding of 1, and 512 kernels to enhance feature capture of fine-scale edge segments.

[0086] Mesoscale corresponding branch: A 5×5 convolutional layer with a stride of 1, padding of 2, and 512 kernels is used to match the expression of local structural features;

[0087] Long-scale corresponding branch: 7×7 dilated convolutional layers with stride of 1, padding of 6, and 512 kernels are used to capture long-range continuous features.

[0088] In S303, for each spatial location in the feature map Based on the probability values ​​of each scale probability map, the feature vectors of the corresponding parallel convolution branches at that position are weighted and fused to generate a scale-adapted feature map.

[0089] Specifically, first, extract each pixel. Probability values ​​on 3-scale probability maps , , , and use them as weighting coefficients for the output features of the corresponding convolutional branches;

[0090] Secondly, the three parallel convolutional branches perform convolution operations on the multi-scale aggregated feature map, outputting three branch feature maps with 512 channels and a resolution of 64×64, denoted as follows: , , ;

[0091] Finally, pixel-level weighted summation is performed on the three branch feature maps to generate a scale-adapted feature map. The calculation formula is as follows:

[0092]

[0093] in, This represents the channel index of the feature map. Through this weighted fusion process, the feature representation of each pixel is dynamically adjusted to best match its local scale attributes. For example, the features of pixels in fine-scale edge segments are mainly contributed by 3×3 convolutional branches, while the features of pixels in long-distance continuous regions are mainly contributed by 7×7 dilated convolutional branches.

[0094] Preferably, for the 7×7 dilated convolution of the corresponding branch at the long scale, its dilation rate is determined based on the long-scale probability map. The probability value is dynamically adjusted so that the void ratio increases with the increase of long-scale confidence.

[0095] Specifically, the adjustment range for the hole ratio is set to 1-3, with an initial hole ratio of 1; for each pixel... Calculate the dynamic void ratio The calculation formula is as follows:

[0096]

[0097] in, This indicates a round-down operation. When hour, ;when hour, ;when When the value is between 0 and 1, the porosity dynamically adapts between 1 and 3. Through this dynamic adjustment, the receptive field of the long-scale branch can flexibly change according to the long-scale confidence of the pixel, further improving the capture accuracy of long-distance continuous features.

[0098] In S304, scale-adapted feature maps Instance normalization is performed, with the epsilon parameter set to 1e-5. Then, the LeakyReLU activation function (negative slope 0.01) is used to enhance the non-linear expressive power of the features, avoid the gradient vanishing problem, and obtain the final scale-adapted feature map.

[0099] According to embodiments of the present invention, the core objective of edge enhancement interactive processing is to mine the correlation between edge features and features of adjacent regions, enhance the continuity of weak edges, and avoid segmentation breaks caused by isolated edge fragments. See also... Figure 4 , Figure 4 This is a schematic diagram of the edge enhancement processing flow provided in an embodiment of the present invention, specifically including:

[0100] In S401, edge response extraction and edge attention weight map generation, specifically, firstly, from the scale-adapted feature map... Extract edge responses and generate an edge attention weight map.

[0101] Specifically, a differentiable Sobel operator is embedded in the network layer, and horizontal and vertical edge responses of the feature map are extracted through convolution operations. The convolution kernel of the Sobel operator in the horizontal direction is... The convolution kernel of the Sobel operator in the vertical direction is ;

[0102] The Sobel operators in the two directions respectively Perform a convolution operation to obtain the horizontal edge response map. Vertical edge response map Both have a resolution of 64×64 and the number of channels. Consistent (512 channels);

[0103] right and Perform pixel-level square root operations to obtain the edge intensity map. The calculation formula is as follows:

[0104]

[0105] in, For channel indexing;

[0106] Edge intensity map Perform global mean calculation to obtain the global mean. Then, the Sigmoid activation function is used to activate... Convert to edge attention weight map The calculation formula is as follows:

[0107]

[0108] in, This is the weighted discrimination enhancement coefficient. for The global mean (adaptively updated during training). The pixel value range is [0,1]. The weight value of the safety rope edge region is close to 1, and the weight value of the background region is close to 0, thus highlighting the importance of edge features.

[0109] In S402, for the edge attention weight map For pixels with a weight higher than a preset threshold, their feature similarity to neighboring pixels is calculated, and feature aggregation is performed based on this similarity. Specifically, this includes...

[0110] Preset threshold setting: The initial threshold is set to 0.3. This threshold can be adaptively optimized during training to balance the sensitivity and accuracy of edge detection.

[0111] Neighborhood range setting: 8-neighborhood interaction range is used, that is, for each pixel... Its neighboring pixels are ( and (8 adjacent pixels, which conform to the continuous distribution pattern of the safety rope edge);

[0112] Feature similarity calculation: For For edge pixels, calculate their scale-fit feature maps with their 8 neighboring pixels. cosine similarity of features The calculation formula is as follows:

[0113]

[0114] in, To avoid the minimum value where the denominator is 0, This represents the vector dot product operation. The L2 norm of a vector;

[0115] Neighborhood feature weighted aggregation: For each edge pixel, the features of its 8 neighboring pixels are summed according to similarity weights to obtain the neighborhood enhancement features. The calculation formula is as follows:

[0116]

[0117] The summation range is 8 neighboring pixels. .

[0118] In S403, neighborhood enhancement features are used. Feature maps adapted to the original scale Fusion is performed using residual connections to enhance the continuity and contextual information of edge regions.

[0119] Specifically, the fusion formula is as follows:

[0120]

[0121] in, This is the fusion coefficient, initially set to 0.5. It can be adaptively learned during training to balance the contribution ratio of the original features and the neighborhood enhanced features.

[0122] For the fused feature map A 1×1 convolutional compression is performed with 512 kernels to ensure the number of output channels matches the input. Then, a batch normalization layer (momentum 0.9, epsilon = 1e-5) is used to stabilize the feature distribution, ultimately outputting a 512-channel, 64×64 resolution scale-edge co-optimized feature map. The feature map is then directly input into the subsequent feature calibration module.

[0123] The feature calibration module optimizes the scale-edge co-optimized feature map using a dual attention mechanism that includes local interactive attention and channel weight calibration. This suppresses redundant features and enhances the representation of key safety rope features. Please refer to [link to relevant documentation]. Figure 5 , Figure 5 This is a schematic diagram of the feature calibration module processing provided in an embodiment of the present invention. The specific implementation process is as follows:

[0124] In S501, the local interactive attention submodule is implemented. The core purpose of this submodule is to enhance the effective information interaction between adjacent channels and improve the correlation of local features.

[0125] Specifically, firstly, the scale-edge co-optimization feature map of the input is performed. (512 channels, 64×64 resolution) Perform a channel-dimensional reshape operation to convert it into a 512×(64×64) two-dimensional feature matrix. ;

[0126] Next, calculate the interaction relationship between each channel and its 7 neighboring channels: for the ... The first channel, select the... to Each channel (with circular padding when the channel index exceeds the range) is used as an adjacent channel to construct a 7×(64×64) local channel feature matrix. ;

[0127] Then, through a 1×1 convolutional layer... Dimensional compression is performed to obtain a 1×(64×64) interaction response vector, which is then used to generate a local attention weight vector via a sigmoid activation function. Its dimensions are (64×64);

[0128] Finally, the local attention weight vector With the The feature vectors of each channel are multiplied element-wise to obtain the calibrated local channel features. The calibrated features of all channels are then concatenated to form a local interactive optimization feature map. (512 channels, 64×64 resolution).

[0129] In S502, the channel weight calibration submodule is implemented. The core purpose of this submodule is to dynamically learn the importance weight of the channels, suppress redundant channel features, and highlight the key features of the safety rope.

[0130] Specifically, firstly, optimize the feature map of local interactions. A global average pooling operation is performed, averaging the (64×64) pixel values ​​for each channel to obtain a 512-dimensional channel statistical vector. ,in ;

[0131] Secondly, the channel statistical vector The input is fed into two fully connected layers. The output dimension of the first fully connected layer is 128, and the output dimension of the second fully connected layer is 512. The GELU activation function is inserted between the two fully connected layers to enhance the non-linear expressive power.

[0132] Then, the output of the second fully connected layer is normalized using the Sigmoid activation function to obtain a 512-dimensional channel weight vector. Its element value range is [0,1], which indicates the importance of the corresponding channel;

[0133] Finally, the channel weight vector Optimize feature maps with local interactions Perform channel-level weighted multiplication, i.e. The final feature map after feature calibration is obtained. (512 channels, 64×64 resolution).

[0134] According to an embodiment of the present invention, the core function of the decoder is to gradually restore the image resolution through upsampling operations, and supplement detailed information by combining the skip connections of the encoder, and finally output the semantic segmentation map of the safety rope.

[0135] Specifically, the decoder consists of four consecutive deconvolutional blocks, each containing a transposed convolutional layer, a batch normalization layer, and a ReLU activation function. It also fuses the feature maps of the corresponding layers of the encoder through skip connections.

[0136] For the first deconvolution block: Receive the feature map after feature calibration. (512 channels, 64×64 resolution) Upsampling is performed using a 3×3 transposed convolutional layer with a stride of 2 and padding set to 1, resulting in 256 output channels and a resolution restored to 128×128. A 128×128 resolution, 256-channel feature map from the encoder output is introduced through skip connections. This feature map is compressed by channels using a 1×1 convolutional layer (256 kernels) and then added pixel-by-pixel to the upsampled feature map to supplement the semantic details of the middle layer.

[0137] The second deconvolutional block receives the fused feature map (256 channels, 128×128 resolution) from the previous step, upsamples it to 256×256 resolution through a 3×3 transposed convolutional layer with a stride of 2, and outputs 128 channels; it then introduces the 256×256 resolution, 128-channel feature map output by the encoder through skip connections, compresses it through a 1×1 convolutional channel, and adds it pixel by pixel to supplement texture details;

[0138] The third deconvolutional block receives the fused feature map (128 channels, 256×256 resolution), upsamples it to 512×512 resolution through a 3×3 transposed convolutional layer with a stride of 2, and outputs 64 channels; it then introduces the 512×512 resolution, 64-channel feature map output by the encoder through skip connections, compresses it through a 1×1 convolutional channel, and adds it pixel by pixel to supplement edge details;

[0139] The fourth deconvolutional block receives the fused feature map (64 channels, 512×512 resolution), processes it through a 3×3 convolutional layer (stride 1, padding 1) and the ReLU activation function, and outputs a feature map with 64 channels and 512×512 resolution.

[0140] Finally, a 1×1 convolutional layer is used to compress the output feature map of the fourth deconvolution block. The number of convolutional kernels is set to 1, converting the 64-channel feature map into a single-channel feature map (512×512 resolution). Then, the pixel values ​​are mapped to the [0,1] interval by the Sigmoid activation function.

[0141] For example, a threshold of 0.5 is set. Regions with pixel values ​​greater than 0.5 are identified as safety ropes (mask value 1), and regions with pixel values ​​less than or equal to 0.5 are identified as background (mask value 0). The final output is a pixel-level semantic segmentation map of safety ropes.

[0142] According to a preferred embodiment of the present invention, the pre-training of the codec network includes dataset construction and preprocessing, loss function design, optimizer configuration, and training strategy implementation, as detailed below:

[0143] 1. Construction and Preprocessing of Safety Rope Image Dataset

[0144] According to embodiments of the present invention, safety rope images covering typical scenarios such as high-altitude operations, industrial workshops, outdoor construction, and severe weather are collected. The initial collection quantity is 5,000 images, with image resolutions including common specifications in practical applications such as 1080P and 4K. To ensure the validity of the dataset, the collected images are screened by professionals, removing blurry images, images without safety ropes, or images with unclear safety rope characteristics, ultimately retaining 4,800 valid images to ensure the scenario diversity and representativeness of the dataset.

[0145] Pixel-level annotations were performed on 4800 valid images using professional annotation tools (such as LabelMe). Annotators manually outlined the safety rope area and generated a binary mask (the safety rope area had a pixel value of 1, and the background area had a pixel value of 0). After annotation, another professional performed cross-checking to remove missing or incorrectly labeled images, ensuring the accuracy of the annotations.

[0146] The labeled dataset is divided into training, validation, and test sets in a 7:2:1 ratio:

[0147] Training set: 4800 × 70% = 3360 images, used for model parameter learning;

[0148] Validation set: 4800 × 20% = 960 images, used to monitor the training process and evaluate model performance;

[0149] Test set: 4800 × 10% = 480 images, used to evaluate the final segmentation accuracy of the model.

[0150] During the partitioning process, it is ensured that the images of each scene are evenly distributed across the three datasets to avoid data bias that could lead to a decrease in the model's generalization ability.

[0151] To improve the model's generalization ability, various data augmentation strategies are applied to the training set images, including:

[0152] Random cropping: The image is randomly cropped to a resolution of 512×512 to simulate safety rope images at different shooting distances;

[0153] Horizontal Flip: Flips the image horizontally with a probability of 0.5, increasing image diversity;

[0154] Brightness / contrast perturbation: Randomly adjust the brightness and contrast of the image within the range of 0.8-1.2 times to simulate different lighting conditions;

[0155] Gaussian noise addition: Gaussian noise with a standard deviation of 0.01 is added to the image to enhance the model's robustness to noise.

[0156] Images from all datasets (training set, validation set, and test set) need to be standardized. The standardization process is the same as that in step S1, which uses the RGB mean (0.485, 0.456, 0.406) and standard deviation (0.229, 0.224, 0.225) of the ImageNet dataset to normalize the pixel values ​​to the range [-1, 1].

[0157] 2. Loss Function Design and Implementation

[0158] According to an embodiment of the present invention, the loss function used for training is the foreground normalized mean square error loss (FN-MSE), the Dice loss, and the edge alignment loss (…). The weighted sum of () is used to address the issues of small foreground proportion of safety ropes, high requirements for segmentation continuity, and insufficient edge recognition accuracy, specifically including:

[0159] (1) Foreground Normalized Mean Squared Error Loss (FN-MSE)

[0160] FN-MSE addresses the training instability caused by an excessively small proportion of the safety rope by normalizing the error in the foreground region. Its calculation formula is as follows:

[0161]

[0162] in, This is the foreground mask (i.e., the pixel value of the safety rope area in the label mask, where 1 represents the foreground and 0 represents the background). The pixel values ​​representing the segmentation results output by the network. The pixel values ​​of the actual labeled mask. This is used to avoid excessive loss values ​​due to insufficient foreground pixels, thus ensuring stable training.

[0163] (2) Dice loss

[0164] Dice loss is used to alleviate the imbalance between foreground and background pixels and improve segmentation continuity. Its calculation formula is as follows:

[0165]

[0166] in, To avoid the minimum value where the denominator is 0, This represents pixel-level multiplication operations. Dice loss calculates the overlap between the predicted and actual regions, making it more sensitive to errors in the foreground region and effectively improving the continuity of safety rope segmentation.

[0167] (3) Edge alignment loss

[0168] The edge alignment loss is calculated based on the edge information extracted from the ground truth labeled mask and the edge attention weight map obtained from the edge enhancement interaction processing. It is used to supervise the model's recognition of the safety rope edge, and its calculation formula is as follows:

[0169]

[0170] in, This represents the total number of pixels in the image (512×512×512). The edge information is extracted from the real labeled mask (obtained by convolving the labeled mask with the Sobel operator, with a value of 1 for edge regions and 0 for non-edge regions). This is the edge attention weight map obtained in the edge enhancement interaction processing. The loss function supervises the alignment between the edge attention weight map and the real edge information through binary cross-entropy, forcing the model to accurately identify the edge of the safety rope.

[0171] The total loss function is a weighted sum of the three losses mentioned above. The weight ratios were determined through experimental optimization, and the specific calculation formula is as follows:

[0172]

[0173] The weight of FN-MSE is 1, the weight of Dice loss is 0.9, and the weight of edge alignment loss is 0.1. This weight ratio ensures both the accuracy and continuity of foreground segmentation and highlights the importance of edge recognition.

[0174] The AdamW optimizer was selected for updating network parameters. This optimizer can effectively suppress overfitting and improve the model's generalization ability. Its specific parameter configuration is as follows: initial learning rate: 0.001; momentum parameter β1: 0.9; momentum parameter β2: 0.999; numerical stability parameter ε: 1e-8; weight decay coefficient: 0.0001.

[0175] During training, the optimizer calculates the gradient of the total loss function with respect to the parameters of each layer of the network through an automatic differentiation mechanism, and then updates the network weights according to the above parameter configuration to suppress the overfitting problem caused by excessively large parameters.

[0176] Batch Size: 32, determined based on the memory capacity of the NVIDIA GTX 2080 graphics card, to ensure efficient use of hardware resources in each iteration;

[0177] Training epochs: 150 epochs, sufficient training epochs to ensure that the model can fully learn the data features;

[0178] Parameter initialization: The weights of all convolutional and fully connected layers in the network are initialized using a He normal distribution, and the bias terms are initialized using a constant 0; the parameters of the dynamic scale adaptation and edge enhancement attention modules use the same initialization strategy as the original network parameters to ensure end-to-end training compatibility.

[0179] A segmented learning rate adjustment strategy is adopted to balance the model's convergence speed and accuracy: For the first 100 rounds, the initial learning rate is kept constant at 0.001 to ensure that the model converges quickly to a local optimum. For the last 50 rounds, a cosine annealing strategy is used to decay the learning rate.

[0180] To avoid model overfitting, an early stopping strategy is used to monitor the training process.

[0181] Evaluation frequency: After every 5 rounds of training, the model performance is evaluated using the validation set, and the total loss of the validation set is calculated. With F1 score;

[0182] Early stopping trigger condition: If the total loss of the validation set does not decrease for 15 consecutive rounds (the decrease is <1e-4), then stop training and save the model parameters of the current round as the optimal model;

[0183] Model saving: During training, the model parameter file (.pth format) is saved after each training round. The file contains information such as the weights of each layer of the network, the optimizer status, the current training round, and the learning rate, which facilitates subsequent inference and model reproduction.

[0184] After training, the performance of the optimal model is evaluated using a test set. Evaluation metrics include segmentation accuracy, F1 score, intersection-over-union (IoU) ratio, and inference time. For example, on an NVIDIA GTX 2080 graphics card, the inference time for a single 512×512 resolution image is approximately 15ms, with a frame rate of 65FPS, meeting the deployment requirements for real-time security monitoring; segmentation accuracy is ≥95%, F1 score ≥0.94, and IoU ≥0.89, ensuring high precision in safety rope segmentation.

[0185] It is understood that the present invention, through the design and implementation of a dynamic scale adaptation and edge enhancement attention module, addresses the segmentation breakage problem caused by local scale abrupt changes and weak edges in safety ropes through the synergistic effect of dynamic scale perception and edge enhancement interaction. It forms an organic synergy with the encoder, multi-scale feature extraction module, feature calibration module and decoder, and significantly improves the accuracy and continuity of semantic segmentation of safety ropes without significantly increasing computational complexity.

[0186] Please see Figure 6 , Figure 6 This is a schematic diagram of a safety rope semantic segmentation system 6 based on a codec network structure provided in an embodiment of this application. Figure 6 As shown, system 600 includes:

[0187] Image preprocessing module 601 is used to acquire the safety rope image to be segmented and to preprocess the safety rope image;

[0188] The semantic segmentation module 602 is used to input the preprocessed safety rope image into a pre-trained encoder-decoder network, and obtain pixel-level semantic segmentation results through the forward propagation of the encoder-decoder network;

[0189] The encoder-decoder network comprises an encoder, a multi-scale feature extraction module, a feature calibration module, and a decoder connected in sequence. The encoder downsamples the input image and extracts multi-level features using unpooled convolutional layers. The multi-scale feature extraction module performs convolution operations on the high-level features output by the encoder using parallel convolutional branches with different receptive fields, and fuses the feature maps output by each branch to generate a multi-scale aggregated feature map. The feature calibration module optimizes the multi-scale aggregated feature map using a dual attention mechanism that includes local interactive attention and channel weight calibration. The decoder gradually restores the image resolution through upsampling operations combined with skip connections from the encoder, ultimately outputting a semantic segmentation map of the safety rope.

[0190] Those skilled in the art will clearly understand that the technical solutions of the embodiments of this application can be implemented by means of software and / or hardware. In this specification, "unit" and "module" refer to software and / or hardware that can independently complete or cooperate with other components to complete a specific function, wherein the hardware may be, for example, a field-programmable gate array (FPGA), an integrated circuit (IC), etc.

[0191] Each processing unit and / or module in the embodiments of this application can be implemented by an analog circuit that implements the functions described in the embodiments of this application, or by software that executes the functions described in the embodiments of this application.

[0192] Please see Figure 7 It shows a schematic diagram of the structure of an electronic device according to an embodiment of this application, which can be used to implement... Figure 1 The method in the illustrated embodiment. (As shown) Figure 7 As shown, the electronic device 700 may include:

[0193] The system includes at least one processor 701, at least one network interface 704, a user interface 703, a memory 705, and at least one communication bus 702. The communication bus 702 is used to enable connection and communication between the components. The user interface 703 may include buttons, and optionally include a standard wired or wireless interface. The network interface 704 may include, but is not limited to, a Bluetooth module, an NFC module, a Wi-Fi module, etc.

[0194] The processor 701 may include one or more processing cores and connect to various parts within the device 700 via various interfaces and lines. It implements the various functions and data processing of the device 700 by running or executing instructions, programs, code sets, or instruction sets stored in the memory 705, and by accessing data in the memory 705. Optionally, the processor 701 may be implemented using at least one hardware form of DSP, FPGA, or PLA. The processor 701 may also integrate one or more combinations of CPU, GPU, and modem. The CPU is mainly used to handle the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content required for display on the screen; and the modem is used for wireless communication. It is understood that the modem may not be integrated into the processor 701, but may be implemented through a separate chip.

[0195] The memory 705 may include random access memory (RAM) or read-only memory (ROM). Optionally, the memory 705 includes a non-transitory computer-readable medium for storing instructions, programs, code, code sets, or instruction sets. The memory 705 may be divided into a program storage area and a data storage area, wherein the program storage area may be used to store instructions for implementing an operating system, instructions for implementing at least one function (such as touch functionality, audio playback functionality, image playback functionality, etc.), and instructions for implementing the foregoing method embodiments; the data storage area may be used to store data involved in the relevant method embodiments. The memory 705 may also be at least one storage device located remotely from the processor 701. Figure 7 As shown, the memory 705, which serves as a computer storage medium, may contain an operating system, a network communication module, a user interface module, and program instructions.

[0196] In particular, the methods and / or embodiments in this application can be implemented as computer software programs. For example, the embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowchart. When the computer program is executed by processor 701, the functions defined in the methods of this application are performed.

[0197] Another embodiment of this application provides a computer-readable storage medium having computer program instructions stored thereon, which can be executed by a processor to implement the methods and / or technical solutions of any one or more embodiments of this application described above.

[0198] The computer-readable storage medium may include, but is not limited to, any type of disk, including floppy disks, optical disks, DVDs, CD-ROMs, microdrives, as well as magneto-optical disks, ROMs, RAMs, EPROMs, EEPROMs, DRAMs, VRAMs, flash memory devices, magnetic cards or optical cards, nanosystems (including molecular memory ICs), or any type of medium or device suitable for storing instructions and / or data.

[0199] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

Claims

1. A method for secure rope semantic segmentation based on codec network structure, characterized in that, The method comprises the following steps: S1, obtaining a safety rope image to be segmented, and preprocessing the safety rope image; S2, inputting the preprocessed safety rope image into a pre-trained encoder-decoder network, and obtaining a pixel-level semantic segmentation result through forward propagation of the encoder-decoder network; The encoder-decoder network comprises an encoder, a multi-scale feature extraction module, a feature calibration module and a decoder connected in sequence; the encoder is used for downsampling an input image and extracting multi-level features through a non-pooling convolutional layer; the multi-scale feature extraction module is used for performing convolutional operations on high-level features output by the encoder through parallel convolutional branches with different receptive fields, fusing feature maps output by each branch, and generating a multi-scale aggregated feature map; the feature calibration module is used for performing feature optimization on the multi-scale aggregated feature map through a double-attention mechanism comprising local interactive attention and channel weight calibration; and the decoder is used for gradually restoring image resolution through upsampling operations and combining with a skip connection from the encoder, and finally outputting a safety rope semantic segmentation map.

2. The safety rope semantic segmentation method based on the encoder-decoder network structure according to claim 1, wherein a dynamic scale adaptation and edge enhancement attention module is further embedded between the multi-scale feature extraction module and the feature calibration module; the dynamic scale adaptation and edge enhancement attention module is used for receiving the multi-scale aggregated feature map, sequentially performing dynamic scale perception and edge enhancement interaction processing, and outputting a scale-edge collaborative optimization feature map to the feature calibration module. The dynamic scale perception processing specifically comprises:

3. The method of claim 2, wherein the method is based on a codec network structure, and Based on the multi-scale aggregated feature map, a plurality of scale probability maps corresponding to different scale attributes are generated; A plurality of parallel convolutional branches corresponding to the scale attributes are set, and each branch uses a convolution kernel or a hollow convolution of different sizes; For each spatial position in the feature map, the feature vectors of the corresponding parallel convolutional branches at the position are weighted and fused according to the probability values of the spatial position on each scale probability map, to generate a scale adaptation feature map. The edge enhancement interaction processing specifically comprises:

4. The method of claim 3, wherein the method is based on a codec network structure, and Edge responses are extracted from the scale adaptation feature map, and an edge attention weight map is generated; For pixel positions with a weight higher than a preset threshold in the edge attention weight map, the feature similarity of the pixel position and neighbor pixels on the scale adaptation feature map is calculated; The feature vectors of the neighbor pixels are weighted and aggregated according to the feature similarity, and the aggregated features are fused with the original scale adaptation feature map through a residual connection.

5. The safety rope semantic segmentation method based on the encoder-decoder network structure according to claim 4, wherein the loss function used in the pre-training of the encoder-decoder network comprises an edge alignment loss; the edge alignment loss is calculated based on edge information extracted from a real label mask and an edge attention weight map obtained from the edge enhancement interaction processing, and is used to supervise the recognition of safety rope edges by the model. In the dynamic scale perception processing, ​ 6. The method of claim 2-5, wherein the method is a codec network structure based safety harness semantic segmentation method. ​ For the convolution branch used for capturing long-distance continuous features, the hole rate of the hole convolution is dynamically adjusted according to the probability value of the corresponding scale, so that the hole rate increases with the increase of the scale confidence.

7. A secure-rope semantic segmentation system based on a codec network structure, characterized by, The method comprises the steps of: An image preprocessing module is configured to acquire a safety rope image to be segmented and to preprocess the safety rope image. A semantic segmentation module is configured to input the preprocessed safety rope image into a pre-trained encoder-decoder network, and to obtain a pixel-level semantic segmentation result through forward propagation of the encoder-decoder network. The encoder-decoder network comprises, in sequence, an encoder, a multi-scale feature extraction module, a feature calibration module, and a decoder.

8. An electronic device, the electronic device comprising: The encoder is configured to down-sample an input image and extract multi-level features through a non-pooling convolution layer. The multi-scale feature extraction module is configured to perform convolution operations on high-level features output by the encoder through parallel convolution branches with different receptive fields, to fuse feature maps output by the branches, and to generate multi-scale aggregated feature maps. The feature calibration module is configured to perform feature optimization on the multi-scale aggregated feature maps through a double-attention mechanism comprising local interactive attention and channel weight calibration. The decoder is configured to gradually restore image resolution through up-sampling operations and in combination with a skip connection from the encoder, and to finally output a semantic segmentation map of the safety rope.

9. A computer-readable storage medium having stored thereon computer program instructions, wherein, The method comprises the steps of:

10. A computer program product comprising a computer program, characterized in that, At least one processor; and a memory connected in communication with the at least one processor; wherein The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-6. The computer program instructions can be executed by the processor to implement the method of any one of claims 1-6. The computer program is executed by the processor to implement the method of any one of claims 1-6. The computer program is executed by the processor to implement the method of any one of claims 1-6.

Images