Rock mass fracture identification method based on boundary detection neural network model

Abstract

The invention relates to the technical field of rock mass fracture recognition, in particular to a rock mass fracture recognition method based on a boundary detection neural network model. According to the method, an innovative deep learning framework is provided for boundary recognition, the characteristics of a refinement strategy and sub-pixel convolution are combined, and a leading result is obtained on a data set. The research shows that the CED model can improve the performance of middle-level and high-level visual tasks including optical flow estimation, target suggestion generation, semantic segmentation and the like. According to the method, a novel architecture is provided to solve the precision problem of a CED algorithm, a reverse refining path is supplemented for an HED network to improve the recognition precision, and the path uses efficient sub-pixel convolution to merge features step by step; the CED provided by the invention can generate an edge image which is well aligned with an image boundary.

Description

Technical Field

[0001] The invention relates to the technical field of rock mass crack identification, and in particular to a rock mass crack identification method based on a boundary detection neural network model. Background Art

[0002] Traditional rock crack image recognition mainly includes methods such as raster image processing and image vectorization. Raster image processing can distinguish cracks from rock mass by graying or binarizing color images. However, this image processing method has many defects. On the one hand, the results after image processing still need to be recognized by the naked eye, lacking quantitative indicators and methods; on the other hand, when the image background is complex, the rock mass is dark, and the illumination is limited, the effect of raster image processing is poor and cannot meet the actual engineering needs. The image vectorization method also has poor results when the rock mass structure is complex and the boundaries are blurred. Most traditional image recognition methods cannot meet the needs of crack recognition under complex conditions such as limited lighting, complex and different types of rock mass, and large background interference in tunnel engineering. Therefore, it is necessary to develop new image recognition methods to accurately identify rock wall cracks in tunnel environments.

[0003] There are two main methods in the research of crack detection based on machine learning. The first type is to extract features from images and identify cracks through classifiers of machine learning methods. In the research of this method, features are extracted through methods such as adaptive filters, deformation and morphological operations, and then these features are transferred to the machine learning classifier to classify crack images and non-crack images. This image extraction method has been widely used in fields such as pavement crack detection and pipeline crack identification. The second type of method is a deep learning method represented by convolutional neural networks, and its feature extraction stage is carried out in a black box algorithm. In the deep learning algorithm, the original image is used as input data, and there is no need to manually set features. The algorithm automatically finds the mapping relationship and rules between data to complete the target task. Since the features used to classify cracks are automatically generated by the model, human bias or errors in feature extraction are avoided, and only systematic errors exist. It can be considered that crack identification based on deep learning is a data-driven method rather than a knowledge-driven method. At present, there are results that can be used as an example study on the application of convolutional neural networks in road crack detection. This method uses a convolutional neural network with 6 convolutional layers to perform binary crack detection tasks on roads. A total of 600,000 images were collected for training, 200,000 of which were used for testing and achieved an F1 score of 0.8965. Another model used 11 convolutional layers and used 4.9 million images as a training set, 1.2 million of which were used as a test set, and finally achieved an F1 score of 0.7246 on the dataset provided by itself.

[0004] The above studies have achieved good results. However, the existing research cannot meet the needs of rock crack identification in tunnels. The main reasons are as follows: (1) The image background is different. Existing crack identification methods are mainly aimed at crack identification on relatively flat surfaces such as cement pavements and walls, while crack identification on tunnel rock walls is more complicated. On the one hand, compared with crack identification on pavements and walls, tunnel walls are usually curved, rock walls are often uneven, and there are interference objects such as steel bars, which increases the difficulty of crack identification.

[0005] (2) Different identification objects. Although both are crack identification, cracks on the road surface and wall are quite different from cracks on the tunnel wall in this study. Different cracks in the same tunnel have quite different appearances. Some cracks are narrow and difficult to detect, while some surrounding rock cracks have fuzzy boundaries and complex shapes. Surrounding rock cracks are usually accompanied by water seepage, and the color of the seepage area is darker, covering the direction of the crack itself. The color of the crack is often close to that of the surrounding rock. At the same time, the scratches of the TBM edge knife are similar in appearance to some cracks, which can easily lead to identification errors.

[0006] (3) Different recognition tasks. Common crack recognition methods usually split a photo into a large number of small pieces and perform a binary classification calculation on each small picture into two results: the presence of a crack or the absence of a crack. However, the cracks in the tunnel rock wall are large in scale. Evenly splitting the picture cannot include the entire crack in the picture. It is impossible to determine whether the cracks in different pictures belong to the same one. In order to calculate the integrity of the surrounding rock, it is necessary to distinguish different cracks and calculate the distance between the cracks.

[0007] Therefore, a new algorithm for identifying cracks in tunnel rock walls is needed to facilitate the observation of the trend and shape of the entire crack and to distinguish different cracks. Summary of the invention

[0008] In order to make up for the deficiencies of the prior art, the present invention provides a rock mass fissure identification method based on a boundary detection neural network model.

[0009] The present invention is achieved through the following technical solutions: A rock mass crack identification method based on a boundary detection neural network model comprises the following steps: Step 1: Obtaining the image of the rock mass crack to be detected; Step 2: preprocessing the rock mass crack image to be detected; Step 3: Build a fine boundary detection (CED) model and pre-train the model using open-source wall and pavement crack datasets. When building a CED model, you first need to build the output edge map of the overall edge detection algorithm - HED. HED is an edge detection algorithm based on a convolutional neural network. By calculating the confidence of each layer of the network for the edge of the image, a set of complete edge images is obtained. In the network, the image is continuously downsampled through continuous pooling operations, that is, the edge features of smaller images are continuously collected, and then the boundaries of the small images are synthesized upward through bilinear interpolation, that is, by fitting and averaging the resolution of the image, the edge map of the large image is finally portrayed; Step 4: After completing the CED model training, input the test image into the CED model and display some calculation results.

[0010] Furthermore, in step 1, a binocular camera is used to capture the surrounding rock image in real time. While the deep learning algorithm identifies the surrounding rock cracks, the distance between the cracks is automatically measured to achieve a complete new calculation of the surrounding rock. The radial distance calculation formula of the tunnel is as follows: ; Among them, P is a point on the object to be measured, O R and O T are the optical centers of the two cameras, and the imaging points of point P on the two camera photoreceptors are p and p' respectively. f is the focal length of the camera, B is the center distance of the camera, Z is the depth information, X R From point P to O R The vertical distance, X T From point P to O T The vertical distance between p and p' is dis, then: .

[0011] Furthermore, in step 3, the training data set is recorded as , where the sample Represents the original input image, and the binary edge map corresponding to the image is recorded as ; The set of all network hyperparameters of the model is recorded as W , j represents the serial number of each pixel. Assuming that the image has 10 pixels, the value of j is 1, 2, ..., 10; assuming that the number of sub-network layers is M , each layer is associated with a classifier, and the weights of each classifier are expressed as , where classifier represents a model for classifying input pixels. It can judge the category of input pixels according to their characteristics, such as conventional rock mass or fractured rock mass. The classification algorithm can be regarded as the comprehensive effect of multiple classifiers.

[0012] Furthermore, in step 4, the objective function of the model is recorded as: in, represents the loss function of the output image, which is calculated for all pixels of the training image. m represents the sequence number of the classifier. , , and and denote the set of labels belonging to the edge and not belonging to the edge, respectively. , To balance the weight represents the sigmoid activation function, represents the processing result of the j-th pixel and the m-th classifier, Indicates that the above processing results are calculated using the sigmoid activation function. ) also represents the above calculation result.

[0013] In addition, a weighted fusion layer is added to the network, and the activation function of this layer is: ; in ; w is the weight matrix of multiple classifiers in the algorithm, h is the "fusion weight" that divides the image into multiple levels, Dist(Y, Y fuse ) represents the distance between the predicted boundary and the true boundary. The stochastic gradient descent method is used to optimize the objective function and find its minimum value. Y is the true value, Y fuse is the predicted value; is the activation function, M is the total number of classifiers, m is the sequence number of the classifier, and h m is the “fusion weight” of the m-th classifier, represents the “boundary” position of the image predicted by the mth classifier, h is the weight, and its value is ; .

[0014] Furthermore, in step 4, the CED model includes two components: a forward propagation path and a reverse refinement path. The forward propagation path generates a high-dimensional low-resolution feature map with rich semantic information, and the reverse refinement path fuses the feature map with the intermediate features along the forward propagation path.

[0015] Furthermore, each reverse refinement module fuses the top-down feature map of the reverse path with the feature map of the current layer in the forward path, and further merges the map by a small factor (2x) before passing it down the path; this module has two core components, namely fusion and upsampling. Furthermore, the fusion process is as follows: the dimensionality of the two feature maps is reduced by additional convolutions, and the two low-dimensional feature maps with equal channels are concatenated. Specifically, the number of channels of the input forward path feature map isk h After convolution and ReLU operations, the channel is reduced to, which is much smaller than k h , perform the same operation on the feature map of the previous refinement module to obtain k u generate ; Connect the above feature maps into a The new feature map of the channel is reduced to a 3× 3 convolutional layer with The feature map of the channel.

[0016] Furthermore, the upsampling process is as follows: sub-pixel convolution is used to upsample the fused feature map; assuming that the network has i input channels and o The desired output channels, the kernel size of the convolutional layer is recorded as ( o , i , r , c ),in r and c Represent the kernel width and kernel height respectively; the resolution of the output feature map is the input feature map k times, the kernel size of the convolutional layer is ( o × k 2 , i , r , c ), thus generating o × k 2 The feature maps of feature channels have the same resolution; sub-pixel convolution is then used to assemble the output feature maps into o feature channels but high resolution k times the feature map.

[0017] Compared with the prior art, the present invention has the following advantages: (1) This paper proposes an innovative deep learning framework for boundary recognition, combining the features of refinement strategy and sub-pixel convolution, and achieves leading results on datasets. The study shows that the CED model can improve the performance of mid-level and high-level vision tasks including optical flow estimation, object proposal generation, and semantic segmentation.

[0018] (2) The present invention solves the accuracy problem of the CED algorithm by proposing a novel architecture, and supplements the HED network with a reverse refinement path to improve recognition accuracy. This path uses efficient sub-pixel convolution to gradually merge features; the CED proposed in the present invention can generate an edge map that is well aligned with the image boundary. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0020] Figure 1 This is a diagram of the CED architecture of the present invention; Figure 2 This is a basic principle diagram of the binocular vision camera depth measurement of the present invention; Figure 3 The surrounding rock crack image of the tunnel of the Pearl River Delta water resources allocation project in Example 3 of the present invention; Figure 4 is the crack identification result obtained after training with the CED model of the present invention in Example 3 of the present invention; Figure 5 This is a binocular vision ranging diagram in Example 3 of the present invention. DETAILED DESCRIPTION

[0021] It should be noted that the following detailed descriptions are illustrative and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used in the present invention have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0022] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0023] Example 1 A rock mass crack identification method based on a boundary detection neural network model is characterized by comprising the following steps: Step 1: Obtaining the image of the rock mass crack to be detected; Step 2: preprocessing the rock mass crack image to be detected; Step 3: Build a fine boundary detection (CED) model and pre-train the model using open-source wall and pavement crack datasets. When building a CED model, you first need to build the output edge map of the overall edge detection algorithm - HED. HED is an edge detection algorithm based on a convolutional neural network. By calculating the confidence of each layer of the network for the edge of the image, a set of complete edge images is obtained. In the network, the image is continuously downsampled through continuous pooling operations, that is, the edge features of smaller images are continuously collected, and then the boundaries of the small images are synthesized upward through bilinear interpolation, that is, by fitting and averaging the resolution of the image, the edge map of the large image is finally portrayed; Step 4: After completing the CED model training, input the test image into the CED model and calculate the result.

[0024] like Figure 1 As shown in Figure 1, there are two main components of CED: the forward propagation path and the reverse refinement path with a refinement module. The forward propagation path is similar to HED and generates high-dimensional low-resolution feature maps with rich semantic information. The reverse refinement path fuses the feature maps with the intermediate features along the forward propagation path. This refinement is done multiple times by the refinement module. In each iteration, sub-pixel convolution increases the feature resolution by a small factor (2x). Figure 1 Pooling is pooling, which can also be called sampling. It extracts the pixel statistical features of a sub-domain of a certain area and judges the pixel features of the entire area. Convolution is convolution, which divides the image into layers of different scales, such as w×h, w / 2×h / 2, etc., and performs mining calculations. Relu and Up-sampling are activation functions, and Period shuffling is periodically cleaning pixel data.

[0025] By iterating the reverse refinement module multiple times, the resolution of the feature map can be gradually improved. The present invention uses the edge evidence on the intermediate feature map aggregation path. The detailed structure of the module is as follows: Figure 1 As shown in the lower part, each backward refinement module fuses the top-down feature map of the backward path with the feature map of the current layer in the forward path, and further merges the map by a small factor (2x) before passing it down the path. This module has two core components, namely fusion and upsampling.

[0026] The detailed fusion process is as follows: The conventional strategy is to directly connect the two feature maps, which will significantly increase the variables of the model and further increase the computational cost required for modeling. However, the present invention reduces the dimension of the two feature maps through additional convolution and connects the two low-dimensional feature maps with equal channels.

[0027] The number of channels of the input forward path feature map is k h , after convolution and ReLU operations, the channels are reduced to , much smaller than k h , perform the same operation on the feature map of the previous refinement module to obtain k u generate .

[0028] The same operation is performed on the feature map of the back propagation path. The dimensions before and after the operation are ku and k'u, connect the above feature maps into a The new feature map of the channel is reduced to a 3 × 3 convolutional layer with The feature map of each channel. Therefore, the overall computational cost is reduced.

[0029] The detailed upsampling process is as follows: After fusion, the refinement module also needs to expand the resolution of the feature map. The present invention uses sub-pixel convolution to upsample the fused feature map. Sub-pixel convolution is different from bilinear interpolation for upsampling, it is a standard convolution. This method helps to eliminate block artifacts and maintain a low computational cost, which is of great significance for accurately locating edges.

[0030] Assume the network has i input channels and o The sub-pixel convolution does not directly output the enlarged feature map through a single deconvolution layer, but consists of a convolution layer and a subsequent phase shift layer. The kernel size of the convolution layer is ( o × k 2 , i , r , c ),in r and c Represent the kernel width and kernel height respectively, thus generating o × k 2 The feature maps of feature channels have the same resolution, and then sub-pixel convolution is used to assemble the output feature maps into o feature channels but high resolution k times the feature map.

[0031] The present invention also makes two improvements to the CED model. One is to use a better network skeleton (ResNet) for convolution training, and the other is to use a larger kernel for each sub-model output.

[0032] 1. Better network skeleton. The latest ResNet is used to replace the VGG skeleton used in conventional HED. Compared with VGG, ResNet shows better performance on many visual tasks. Since spatial details are crucial for clear boundary detection, the first pooling operation in the network is deleted, so the output dimension is reduced to 16.

[0033] 2. Larger sub-model kernels. CED is improved by adding two additional convolutional layers to each of the five sub-models of the network backbone. Large kernels help achieve accuracy of feature maps relative to the actual image, so the kernel sizes of the newly added layers of the five sub-models are set to 3x3, 3x3, 5x5, 5x5, and 7x7.

[0034] The training data set is denoted as , where the sample Represents the original input image, and the binary edge map corresponding to the image is recorded as .

[0035] The set of all network hyperparameters of the model is recorded as W . Assume that the number of sub-network layers is M , each layer is associated with a classifier, and the weights of each classifier are expressed as .

[0036] The objective function of the model is recorded as: ; in, Represents the loss function of the output image, which is calculated for all pixels of the training image. Specifically, the comparison sample and According to module "1", the boundaries of natural image recognition results are often fuzzy, and it is difficult to balance the accuracy and precision of recognition. In this regard, the present invention proposes a method for balancing the accuracy and precision of recognition, that is, introducing a balancing weight in the pixel parameters. Subsequently, the present invention uses the weight parameter to construct a new loss function, which is recorded as: ; in, , , and and denote the set of labels belonging to the edge and not belonging to the edge, respectively. , Represents the sigmoid activation function; the classification algorithm of the present invention can be understood as having multiple classifiers jointly judge the same target and finally obtain a bandit result. The total number of classifiers here is capital M, and lowercase m represents the sequence number of the classifier, that is, the value is 1, 2, ..., M.

[0037] In addition, a weighted fusion layer is added to the network, and the activation function of this layer is: ; in, , h is the weight, and its value is . The distance between the predicted boundary and the true boundary. Stochastic gradient descent is used to optimize the objective function and find its minimum.

[0038] .

[0039] Example 2 Since the calculation of surrounding rock integrity requires the measurement of joint spacing based on the identification of joints, the present invention adopts a binocular camera to capture surrounding rock images in real time. While the deep learning algorithm identifies surrounding rock cracks, it automatically measures the distance between cracks to achieve a new calculation of surrounding rock integrity.

[0040] This embodiment uses stereo vision to calculate depth. The stereo vision implementation consists of a left imager, a right imager, and an optional infrared signal transmitter. The left imager and the right imager can capture the scene and send raw image data to the vision processor, which then calculates the depth value of each pixel in the image by associating points on the left image with the right image and using the shift between the left image and the right image. The infrared signal transmitter can send an invisible static infrared pattern to improve the depth accuracy in low-texture scenes. The basic principle of binocular vision is as follows Figure 2 shown.

[0041] In the figure, P is a point on the object to be measured, O R and O T are the optical centers of the two cameras respectively, the imaging points of point P on the two camera photoreceptors are p and p' respectively, f is the focal length of the camera, B is the center distance of the appearance, Z is the depth information, and if the distance between p and p' is dis, then: ; According to the principle of similar triangles: ; We can get: ; The distance between two points in the image can be calculated by the depth of different points. This method can be used to calculate the radial distance of the tunnel.

[0042] Example 3 Relying on the Pearl River Delta Water Resources Allocation Project, this method is used to extract images of rock mass around tunnels and identify cracks. The specific setting parameters of the CED model are as follows: the parameters of all hidden layers are generated by Gaussian random distribution with a mean of 0 and a variance of 0.01, and the network is trained by the stochastic gradient descent method. The hyperparameters of the training process are the initial learning rate set to 1e-5, the weight decay rate set to 2e-4, and the momentum set to 0.99.

[0043] A total of 463 photos were collected at the project site, of which 240 were used as training data and 29 were used as test data. Some of the pictures are shown in Figure 3 The data labels are manually annotated to classify the pixels of the image group into cracks and non-cracks. Due to the limited number of surrounding rock crack images, during the training process, the dataset was augmented with conventional image data including random rotation, flipping, cropping, Gaussian blurring, etc.

[0044] After the model training is completed, the test image is input into the model, and some calculation results are shown as follows: Figure 4 As can be seen from the figure, the model automatically identifies and divides the crack area from the surrounding rock photos. The crack identification is relatively clear and the positioning is relatively accurate. Since cracks are usually accompanied by water seepage, the seepage rock wall may have a color close to that of the rock mass cracks, such as Figure 4 c, which causes the width of the identified result to be greater than the actual width of the crack. The calculation of the integrity of the surrounding rock mainly considers the joint spacing, which has little to do with the joint width, so the algorithm can meet the actual use requirements.

[0045] like Figure 5 As shown in the figure, after the rock mass cracks are identified based on the CED method, the distance within the image range is calculated by binocular vision ranging. J f = N / L (in, N is the total number of joints, L is the distribution distance within the joint statistics range, J f The joint frequency (also known as joint density) can be calculated. In this figure, the frequency of surrounding rock joints is 1.64m -1 .

[0046] Example 4 This example uses the BSDS500 dataset to conduct a comparative study of different CED network architectures. Specifically, the following variants of the model are compared: HED: The original HED network using the VGG16 backbone. It is used as a baseline model.

[0047] •Res16x-HED: HED with ResNet50 as the backbone. It is an enhanced version of the original HED.

[0048] •Res16x-HED- large: This is Res16x-HED with a large core. It helps to calibrate the performance through the improvement of CED by the present invention.

[0049] •CED-sub-multi: This is the CED network using VGG16 without sub-pixel convolution and multi-scale testing. Therefore, this model just adds a refinement path to HED.

[0050] •CED-multi: This is the CED network using VGG16 without multi-scale testing. This is the basis of the full model, including refinement path and sub-pixel convolutions.

[0051] •CED-sub: This is the CED network using VGG16 without sub-pixel convolution. This model further adds multi-scale testing of CED-sub-multi.

[0052] •CED: This is the full CED model with multi-scale testing, but with VGG16 as the backbone.

[0053] •Res16x-CED: This is an improved version of the CED network of the present invention. This model uses ResNet50 as the backbone network and adds large kernel convolutions to the side outputs of the backbone network.

[0054] 1. Model training. For training, the present invention adopts an improved version of the consensus sampling strategy to prevent mislabeling of the original data. If a pixel is marked as an edge by at least three annotators, the pixel is assigned a positive label. Pixels not marked by any annotator are considered negative. The remaining pixels are ignored during training. The present invention also increases data samples by random rotation, cropping, and multi-scale resizing. For fair comparison, all CED models are initialized from the same HED base model.

[0055] 2. Results. The results are summarized in Table 1. First, the proposed improvement approach improves the ODS score of the baseline HED by 1.3% (CED-sub-multi vs. HED). Second, further adding sub-pixel convolutions brings a small improvement of 0.1% under the standard evaluation criteria (CED-sub-multi vs. CED-multi). Sub-pixel convolutions are critical when using strict criteria. Third, multi-scale tests increase by 0.7-0.9% (CED-sub-multi vs. CED-sub or CED-multivs. CED). In addition, Res16x-HED improves HED by 0.5%, demonstrating the effectiveness of the better backbone network. Res16x-HED-large adds additional convolution filters with large kernels in the side output, further improving the performance of Res16x-HED by 0.8%. Res16x-CED performs 1.7% higher than Res16x-HED-large, which is a strong performance improvement brought by the CED architecture. Finally, the present invention notes that the improved model of the present invention (Res16x-CED) is the recognition accuracy record for a single model on BSDS500.

[0056] Table 1 Comparison of accuracy of each model

Claims

1. A rock mass fracture identification method based on a boundary detection neural network model, characterized in that: The following steps are included Step 1: Obtain the image of the rock mass crack to be detected; Step 2: preprocessing the rock mass crack image to be detected; Step 3: Build a fine boundary detection (CED) model and pre-train the model using open-source wall and pavement crack datasets. When building a CED model, you first need to build the output edge map of the overall edge detection algorithm - HED. HED is an edge detection algorithm based on a convolutional neural network. By calculating the confidence of each layer of the network for the edge of the image, a set of complete edge images is obtained. In the network, the image is continuously downsampled through continuous pooling operations, that is, the edge features of smaller images are continuously collected, and then the boundaries of the small images are synthesized upward through bilinear interpolation, that is, by fitting and averaging the resolution of the image, the edge map of the large image is finally portrayed; Step 4: After completing the CED model training, input the test image into the CED model and display some calculation results.

2. The method according to claim 1, characterized in that In step 1, a binocular camera is used to capture the surrounding rock image in real time. While the deep learning algorithm identifies the surrounding rock cracks, the distance between the cracks is automatically measured to achieve a complete new calculation of the surrounding rock. The radial distance calculation formula of the tunnel is as follows: ； Among them, P is a point on the object to be measured, O R and O T are the optical centers of the two cameras, and the imaging points of point P on the two camera photoreceptors are p and p' respectively. f is the focal length of the camera, B is the center distance of the camera, Z is the depth information, X R From point P to O R The vertical distance, X T From point P to O T The vertical distance between p and p' is dis, then: .

3. The method according to claim 1, characterized in that In step 3, the training data set is recorded as , where the sample Represents the original input image, and the binary edge map corresponding to the image is recorded as ; The set of all network hyperparameters of the model is recorded as W , j represents the serial number of each pixel. Assuming that the image has 10 pixels, the value of j is 1, 2, ..., 10; Assuming that the number of sub-network layers is M , each layer is associated with a classifier, and the weights of each classifier are expressed as , where classifier represents a model that classifies the input pixels.

4. The method according to claim 1, characterized in that: In step 4, the objective function of the model is recorded as: ； in, represents the loss function of the output image, which is calculated for all pixels of the training image. m represents the sequence number of the classifier. , , and and denote the set of labels belonging to the edge and not belonging to the edge, respectively. , To balance the weight represents the sigmoid activation function, represents the processing result of the j-th pixel and the m-th classifier, Indicates that the above processing results are calculated using the sigmoid activation function. ) also represents the above calculation result; In addition, a weighted fusion layer is added to the network, and the activation function of this layer is: ； in, ; w is the weight matrix of multiple classifiers in the algorithm, h is the "fusion weight" that divides the image into multiple levels, Dist(Y,Y fuse ) represents the distance between the predicted boundary and the true boundary. The stochastic gradient descent method is used to optimize the objective function and find its minimum value. Y is the true value, Y fuse is the predicted value is the activation function, M is the total number of classifiers, m is the sequence number of the classifier, and h m is the "fusion weight" of the m-th classifier, represents the "boundary" position of the image predicted by the mth classifier, h is the weight, and its value is .

5. The method according to claim 1, characterized in that In step 4, the CED model consists of two components: a forward propagation path and a reverse refinement path. The forward propagation path generates a high-dimensional low-resolution feature map with rich semantic information, and the reverse refinement path fuses the feature map with the intermediate features along the forward propagation path.

6. The method according to claim 5, characterized in that Each backward refinement module fuses the top-down feature map of the backward path with the feature map of the current layer in the forward path, and further merges the map by a small factor 2x before passing it down the path; this module has two core components, namely fusion and upsampling.

7. The method according to claim 6, characterized in that The fusion process is as follows: the dimensionality of the two feature maps is reduced by additional convolutions, and the two low-dimensional feature maps with equal channels are concatenated. Specifically, the number of channels of the input forward path feature map is k h After convolution and ReLU operations, the channel is reduced to, which is much smaller than k h , perform the same operation on the feature map of the previous refinement module to obtain k u generate ; Connect the above feature maps into a The new feature map of the channel is reduced to a 3 × 3 convolutional layer with The feature map of the channel.

8. The method according to claim 6, characterized in that The upsampling process is as follows: sub-pixel convolution is used to upsample the fused feature map; assuming the network has i input channels and o The desired output channels, the kernel size of the convolutional layer is recorded as ( o , i , r , c ),in r and c represent the nuclear width and nuclear height, respectively; The resolution of the output feature map is the resolution of the input feature map k times, the kernel size of the convolutional layer is ( o × k 2 , i , r , c ), thus generating o × k 2 The feature maps of feature channels have the same resolution; sub-pixel convolution is then used to assemble the output feature maps into o feature channels but high resolution k times the feature map.

Images