Method for identifying inclined loose rock mass

Abstract

Provided is a method for identifying inclined loose rock masses. 【Solution means】After collecting the original image, based on the original image, a first detailed feature image and a second detailed feature image are constructed. Cutting processing is performed on any of the images to achieve area-by-area classification, and based on the original block, the first feature block, and the second feature block in the same area, classification is performed on this area based on a three-channel classification model. Correction processing is performed on the classification value of the area to improve the classification accuracy, find the rock area, peel off the background area, and improve the accuracy of identifying loose rock masses. Depending on the roughness situation on the surface of the rock area, the roughness of this rock and the situation of joints and cracks are characterized. The greater the roughness, the more easily this rock mass is eroded, the rock becomes loose, and it is more likely to collapse. Paying attention to this point, based on the surface average roughness of the rock area, it is determined whether this rock is a loose rock mass.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and specifically, to a method for identifying inclined loose rock masses.

Background Art

[0002] The stability of slopes is an important issue in fields such as civil engineering, mining engineering, and transportation infrastructure. The presence of loose rock masses may lead to slope instability, thereby causing geological disasters such as landslides and threatening the safety of people's lives and property. Conventional methods for detecting slope rock masses mainly rely on manual on-site investigations. Such methods require time and labor, and are inefficient, making it difficult to achieve rapid evaluation over a large area. Currently, loose rock masses in rocks are identified by collecting images of the target area. When performing image recognition on the loose rock masses on a rock slope, due to the lush vegetation around the target area, the collected images contain a lot of background information that is not the target. These background information hinder the accurate identification of the loose rock masses, thereby reducing the accuracy of identification.

Summary of the Invention

Problems to be Solved by the Invention

[0003] In view of the above problems in the prior art, the present invention provides a method for identifying inclined loose rock masses that solves the problem in the prior art of low identification accuracy of loose rock masses.

Means for Solving the Problems

[0004] To achieve the above object of the present invention, the present invention adopts the following technical solution. A method for identifying inclined loose rock masses, comprising: S1, a step of collecting an image of a target area to obtain an original image; S2, a step of constructing a first detailed feature image and a second detailed feature image based on color values on the original image; S3. Perform cutting on the original image, the first detailed feature image, and the second detailed feature image respectively to obtain a plurality of original blocks corresponding to the original image, a plurality of first feature blocks corresponding to the first detailed feature image, and a plurality of second feature blocks corresponding to the second detailed feature image. S4. Based on the original block, the first feature block, and the second feature block in the same area, perform classification on this area based on a three-channel classification model to obtain the classification value of this area. S5. Perform correction processing on the classification value of each area to find the rock area. S6. Calculate the surface average roughness of the rock area based on the first contour distribution non-uniformity, the second contour distribution non-uniformity, and the third contour distribution non-uniformity of each pixel point on the rock area. When the surface average roughness is greater than the roughness threshold, set this rock area as a loose rock mass.

[0005] The beneficial effects of the present invention are as follows. After collecting the original image, based on the original image, the first detailed feature image and the second detailed feature image are constructed to prominently show the image detailed features. Cutting processing is performed on both of them to realize area-by-area classification. Based on the original block, the first feature block, and the second feature block in the same area, classification is performed on this area based on a three-channel classification model to obtain the classification value of this area. Further correction processing is performed to improve the classification accuracy, find the rock area, peel off the background area, and improve the accuracy of identifying the loose rock mass. Furthermore, based on the surface average roughness of the rock area, it is determined whether this rock is a loose rock mass. The present invention characterizes the roughness, joint, and crack conditions of this rock according to the roughness situation on the surface of the rock area. The greater the roughness, the easier this rock mass is eroded, the rock becomes looser, and it is more likely to collapse.

[0006] Furthermore, the above S2 includes S21. In the original image, a sub-step centered on each pixel point. S22. A sub-step of calculating the first color feature value and the second color feature value respectively based on the color values near the center. S23. A sub-step of forming a first detailed feature image with new color values of pixel points centered around the first color feature value; S24. A sub-step of forming a second detailed feature image with new color values of pixel points centered around the second color feature value.

[0007] The beneficial effects of the above further invention are as follows. The present invention calculates the first color feature value and the second color feature value respectively based on the neighboring color values of each center, constructs the first detailed feature image by setting the new color value centered around the first color feature value, constructs the second detailed feature image by setting the new color value centered around the second color feature value, characterizes the color distribution situation at the center by the first color feature value and the second color feature value, constructs a feature image again with these two distribution situations, and prominently shows the imaging features of the rock on the image.

[0008] JPEG0007693178000002.jpg42170

[0009] The beneficial effects of the above further invention are as follows. The present invention takes the color values at four positions, namely (x,y + 1), (x,y - 1), (x + 1,y), and (x - 1,y), which are the neighboring ranges at the center, and based on the difference between the color values at positions (x,y + 1) and (x,y - 1), and the difference between the color values at positions (x + 1,y) and (x - 1,y), reflects the color distribution situation in four directions at this center.

[0010] JPEG0007693178000003.jpg44170

[0011] The beneficial effects of the above further invention are as follows. The present invention takes the color values at four positions, namely (x + 1,y + 1), (x - 1,y - 1), (x - 1,y + 1), and (x + 1,y - 1), which are the neighboring ranges at the center, and based on the difference between the color values at positions (x + 1,y + 1) and (x - 1,y - 1), and the difference between the color values at positions (x - 1,y + 1) and (x + 1,y - 1), reflects the color distribution situation in another four directions at this center.

[0012] Furthermore, the three-channel classification model in the S4 includes a first channel processing unit, a second channel processing unit, a third channel processing unit, a first multiplier M1, a second multiplier M2, a first Maxpool layer, a first Avgpool layer, a second Maxpool layer, a second Avgpool layer, a first adder A1, a second adder A2, a first Concat layer, and a fully connected layer. The input end of the first channel processing unit is used to input a first feature block, the input end of the second channel processing unit is used to input an original block, and the input end of the third channel processing unit is used to input a second feature block. The first input end of the first multiplier M1 is connected to the output end of the first channel processing unit, its second input end is connected to the output end of the second channel processing unit, and its output end is connected to the input ends of the first Maxpool layer and the first Avgpool layer respectively. The first input end of the second multiplier M2 is connected to the output end of the second channel processing unit, its second input end is connected to the output end of the third channel processing unit, and its output end is connected to the input ends of the second Maxpool layer and the second Avgpool layer respectively. The first input end of the first adder A1 is connected to the output end of the first Maxpool layer, its second input end is connected to the output end of the first Avgpool layer. The first input end of the second adder A2 is connected to the output end of the second Maxpool layer, its second input end is connected to the output end of the second Avgpool layer. The input end of the first Concat layer is connected to the output ends of the first adder A1 and the second adder A2 respectively, and its output end is connected to the input end of the fully connected layer. The output end of the fully connected layer is the output end of the three-channel classification model.

[0013] The beneficial effects of the further invention are as follows. The present invention employs three channel processing units to process the first feature block, the original block, and the second feature block respectively, and adopts the output features of the first channel processing unit to emphasize the output features of the second channel processing unit, and adopts the output features of the third channel processing unit to emphasize the output features of the second channel processing unit, so as to fuse the distribution features and the original image features, significantly show the feature situation of this area, make the features prominent, and further adopt the Maxpool layer and the Avgpool layer to extract features, so as to extract prominent features and global features, combine the prominent features and the global features for classification, and improve the classification accuracy.

[0014] Furthermore, the structures of the first channel processing unit, the second channel processing unit, and the third channel processing unit are the same, and all include a first Conv block, a second Conv block, a third Conv block, a fourth Conv block, a fifth Conv block, a sixth Conv block, a second Concat layer, a third Concat layer, and a fourth Concat layer. The input end of the first Conv block is the input end of the first channel processing unit, the second channel processing unit, or the third channel processing unit, and its output end is connected to the input ends of the second Conv block and the third Conv block respectively. The input end of the second Concat layer is connected to the output ends of the second Conv block and the third Conv block respectively, and its output end is connected to the input end of the fourth Conv block. The input end of the third Concat layer is connected to the output ends of the second Conv block and the third Conv block respectively, and its output end is connected to the input end of the fifth Conv block. The input end of the fourth Concat layer is connected to the output ends of the fourth Conv block and the fifth Conv block respectively, and its output end is connected to the input end of the sixth Conv block. The output end of the sixth Conv block is the output end of the first channel processing unit, the second channel processing unit, or the third channel processing unit. The size of the convolution kernel of the second Conv block and the fourth Conv block is 3*3. The size of the convolution kernel of the third Conv block and the fifth Conv block is 5*5.

[0015] The beneficial effects of the above further invention are as follows. The present invention constructs two side paths in the channel processing unit, uses different convolution kernels, and shares the output features of the second Conv block and the third Conv block to extract image features with different scales.

[0016] Furthermore, the S5 includes: S51, a sub-step of setting each area as a reference area; JPEG0007693178000004.jpg18170S53, a sub-step of correcting the classification value of the reference area based on the classification value of the peripheral area of the reference area when the distance is greater than the distance threshold; When the classified value after correction or the classified value not involved in the correction is greater than the classification threshold, mark the corresponding area as a candidate area, and include a sub-step of forming a rock area from a plurality of candidate areas.

[0017] The beneficial effects of the further invention are as follows. The present invention calculates the distance between the classified values of the reference area and the surrounding area. When the distance is greater than the distance threshold, it indicates that the difference in the classified values between the reference area and the surrounding area is large. Therefore, the classified value of the surrounding area is adopted to correct the classified value of the reference area, and based on the classified value after correction and the classified value not involved in the correction, a rock area is selected.

[0018] JPEG0007693178000005.jpg19170

[0019] The beneficial effects of the further invention are as follows. The present invention corrects the classified value of the reference area based on the average value of the classified values of a plurality of surrounding areas. The greater the distance, the greater the degree of correction to the classified value of the reference area, thereby accurately extracting the rock area and removing the background area.

[0020] Furthermore, the said S6 is S61, a sub-step of calculating the first contour distribution non-uniformity, the second contour distribution non-uniformity, and the third contour distribution non-uniformity of each pixel point on each area of the rock area; S62, a sub-step of calculating the rock roughness of each area of the rock area based on the first contour distribution non-uniformity, the second contour distribution non-uniformity, and the third contour distribution non-uniformity of each pixel point on each area of the rock area; S63, a sub-step of calculating the surface average roughness based on the rock roughness of each area of the rock area; S64, a sub-step of setting this rock area as a loose rock mass when the surface average roughness is greater than the roughness threshold.

[0021] JPEG0007693178000006.jpg55170

[0022] The beneficial effects of the above further invention are as follows. The present invention first calculates the contour distribution non-uniformity of each pixel point on each area of the rock region, characterizes the roughness degree of each pixel point according to the contour distribution non-uniformity of the pixel point, further calculates the rock roughness of each area of the rock region, takes the average value of the rock roughness of each area on the rock region to obtain the surface average roughness of the rock region. When the surface average roughness of the rock region is large, it indicates that the weathering or erosion of the rock is serious, and this rock is more likely to collapse. When the surface average roughness of the rock region is small, the rock surface is smooth, the rock surface has been polished for a long time, and the rock becomes harder.

[0023] JPEG0007693178000007.jpg39170

[0024] The beneficial effects of the above further invention are as follows. The present invention adds the three contour distribution non-uniformities of each pixel point on the rock region of the original image, the first detail feature image, and the second detail feature image to reflect the contour distribution situation of this pixel point, and adds all pixel points to reflect the rock roughness of each area.

Brief Description of the Drawings

[0025]

Figure 1

Figure 2

Figure 3

Embodiments for Carrying Out the Invention

[0026] As shown in FIG. 1, it is a method for identifying slope loose rock mass, S1, an image acquisition step of collecting an image of a target area to obtain an original image, S2, a feature image construction step of constructing a first detail feature image and a second detail feature image based on the color values on the original image, S3. For the original image, the first detailed feature image, and the second detailed feature image, perform cutting respectively to obtain a plurality of original blocks corresponding to the original image, a plurality of first feature blocks corresponding to the first detailed feature image, and a plurality of second feature blocks corresponding to the second detailed feature image, which is an image cutting step; S4. Based on the original blocks, the first feature blocks, and the second feature blocks in the same area, perform classification on this area based on a 3-channel classification model to obtain the classification value of this area, which is an area classification step; S5. Perform correction processing on the classification value of each area to find the rock area, which is a classification value correction step; S6. Calculate the surface average roughness of the rock area based on the first contour distribution non-uniformity, the second contour distribution non-uniformity, and the third contour distribution non-uniformity of each pixel point on the rock area. When the surface average roughness is greater than the roughness threshold, identify this rock area as a loose rock mass, which is a loose rock mass identification step.

[0027] In this embodiment, the step S3 of performing the same method of cutting on the original image, the first detailed feature image, and the second detailed feature image is specifically as follows. Perform cutting on the original image, the first detailed feature image, and the second detailed feature image, that is, cut the original image into a plurality of areas, each area is an original block, cut the first detailed feature image into a plurality of areas, each area is a first feature block, cut the second detailed feature image into a plurality of areas, and each area is a second feature block.

[0028] Since the image is two-dimensional, perform K - 1 horizontal cuts and K - 1 vertical cuts on the original image, the first detailed feature image, and the second detailed feature image to obtain K * K original blocks, K * K first feature blocks, and K * K second feature blocks, where K is a positive integer greater than or equal to 3.

[0029] The above S2 is S21. In the original image, a sub-step centered on each pixel point; S22. A sub-step of calculating a first color feature value and a second color feature value respectively based on the neighborhood color values near the center; S23. A sub-step of constructing a first detailed feature image by using the first color feature value as the new color value of the pixel point at the center; S24. A sub-step of constructing a second detailed feature image by using the second color feature value as the new color value of the pixel point at the center, including.

[0030] The present invention calculates a first color feature value and a second color feature value respectively based on the neighborhood color values near each center, constructs a first detailed feature image by using the first color feature value as the new color value of the center, constructs a second detailed feature image by using the second color feature value as the new color value of the center, characterizes the color distribution situation at the center by the first color feature value and the second color feature value, constructs a feature image again with these two distribution situations, and significantly shows the imaging characteristics of the rock on the image.

[0031] JPEG0007693178000008.jpg42170

[0032] The present invention takes the color values at four positions, namely (x, y + 1), (x, y - 1), (x + 1, y), and (x - 1, y), which are the neighborhood ranges near the center, and reflects the color distribution situations in four directions at this center based on the difference between the color values at positions (x, y + 1) and (x, y - 1), and the difference between the color values at positions (x + 1, y) and (x - 1, y).

[0033] JPEG0007693178000009.jpg42170

[0034] The present invention takes the color values at four positions, namely (x + 1, y + 1), (x - 1, y - 1), (x - 1, y + 1), and (x + 1, y - 1), which are the neighborhood ranges near the center, and reflects the color distribution situations in another four directions at this center based on the difference between the color values at positions (x + 1, y + 1) and (x - 1, y - 1), and the difference between the color values at positions (x - 1, y + 1) and (x + 1, y - 1).

[0035] As shown in FIG. 2, the three-channel classification model in S4 includes a first channel processing unit, a second channel processing unit, a third channel processing unit, a first multiplier M1, a second multiplier M2, a first Maxpool layer, a first Avgpool layer, a second Maxpool layer, a second Avgpool layer, a first adder A1, a second adder A2, a first Concat layer, and a fully connected layer. The input end of the first channel processing unit is used to input a first feature block, the input end of the second channel processing unit is used to input an original block, and the input end of the third channel processing unit is used to input a second feature block. The first input end of the first multiplier M1 is connected to the output end of the first channel processing unit, its second input end is connected to the output end of the second channel processing unit, and its output end is connected to the input ends of the first Maxpool layer and the first Avgpool layer respectively. The first input end of the second multiplier M2 is connected to the output end of the second channel processing unit, its second input end is connected to the output end of the third channel processing unit, and its output end is connected to the input ends of the second Maxpool layer and the second Avgpool layer respectively. The first input end of the first adder A1 is connected to the output end of the first Maxpool layer, and its second input end is connected to the output end of the first Avgpool layer. The first input end of the second adder A2 is connected to the output end of the second Maxpool layer, and its second input end is connected to the output end of the second Avgpool layer. The input end of the first Concat layer is connected to the output ends of the first adder A1 and the second adder A2 respectively, and its output end is connected to the input end of the fully connected layer. The output end of the fully connected layer is the output end of the three-channel classification model.

[0036] The present invention employs three channel processing units to process a first feature block, an original block, and a second feature block respectively, emphasizes the output features of a second channel processing unit by adopting the output features of a first channel processing unit, and emphasizes the output features of the second channel processing unit by adopting the output features of a third channel processing unit, so as to fuse the distribution features and the original image features, prominently show the feature situation of this area, make the features prominent, and further adopt a Maxpool layer and an Avgpool layer to extract features, thereby extracting prominent features and global features, combining the prominent features and the global features for classification, and improving the classification accuracy.

[0037] As shown in FIG. 3, the structures of the first channel processing unit, the second channel processing unit, and the third channel processing unit are the same, and each includes a first Conv block, a second Conv block, a third Conv block, a fourth Conv block, a fifth Conv block, a sixth Conv block, a second Concat layer, a third Concat layer, and a fourth Concat layer. The input end of the first Conv block is the input end of the first channel processing unit, the second channel processing unit, or the third channel processing unit, and its output end is connected to the input ends of the second Conv block and the third Conv block respectively. The input ends of the second Concat layer are connected to the output ends of the second Conv block and the third Conv block respectively, and its output end is connected to the input end of the fourth Conv block. The input ends of the third Concat layer are connected to the output ends of the second Conv block and the third Conv block respectively, and its output end is connected to the input end of the fifth Conv block. The input ends of the fourth Concat layer are connected to the output ends of the fourth Conv block and the fifth Conv block respectively, and its output end is connected to the input end of the sixth Conv block. The output end of the sixth Conv block is the output end of the first channel processing unit, the second channel processing unit, or the third channel processing unit. The size of the convolutional kernels of the second Conv block and the fourth Conv block is 3*3, and the size of the convolutional kernels of the third Conv block and the fifth Conv block is 5*5.

[0038] In this embodiment, the size of the convolutional kernels of the first Conv block and the sixth Conv block is 1*1.

[0039] The present invention constructs two side paths in the channel processing unit, uses different convolutional kernels, and shares the output features of the second Conv block and the output features of the third Conv block, thereby extracting image features with different scales.

[0040] In this embodiment, each Conv block includes a convolutional layer, a ReLU layer, and a BN layer.

[0041] S5 includes the following sub-steps.

[0042] S51. Set each area as a reference area.

[0043] In the present invention, since the image is cut horizontally and vertically in step S3, there are multiple peripheral areas for each area.

[0044] S53. When the distance is greater than the distance threshold, correct the classification value of the reference area based on the classification value of the peripheral area of the reference area.

[0045] S54. When the corrected classification value or the classification value not involved in the correction is greater than the classification threshold, mark the corresponding area as a candidate area, and a rock area is formed by a plurality of candidate areas.

[0046] In the present invention, the peripheral area is an area in contact with the reference area.

[0047] In this embodiment, the distance threshold is a threshold set for each distance and is used to find a reference area where the difference between the surrounding area and the classification value is large.

[0048] The present invention calculates the distance between the classification values of the reference area and the surrounding area. When the distance is greater than the distance threshold, it indicates that the difference between the classification values of the reference area and the surrounding area is large. Therefore, the classification value of the surrounding area is adopted to correct the classification value of the reference area, and based on the corrected classification value and the classification value not involved in the correction, a rock area is selected.

[0049] In the present invention, since the image is cut at S3, there are a plurality of areas on the original image, a plurality of areas on the first detailed feature image, and a plurality of areas on the second detailed feature image. Since the cutting method is the same, after classifying the same area on the three images by adopting a three-channel classification model, the classification value of each area on the original image, the classification value of each area on the first detailed feature image, and the classification value of each area on the second detailed feature image are obtained, and the classification values are the same in the same area of the three images.

[0050] In this embodiment, for the same area, when the classification value of this area has been corrected, in step S54, the corrected classification value is adopted to compare with the classification threshold, and the classification value before correction is no longer adopted for comparison with the classification threshold.

[0051] In this embodiment, the classification threshold is a threshold set for each classification value and is used to find an area where the classification value is large.

[0052] JPEG0007693178000011.jpg19170

[0053] The present invention corrects the classification value of the reference area based on the average value of the classification values of a plurality of surrounding areas. The greater the distance, the greater the degree of correction to the classification value of the reference area, thereby accurately extracting the rock area and removing the background area.

[0054] S6 is S61, a sub-step of calculating the first contour distribution non-uniformity, the second contour distribution non-uniformity, and the third contour distribution non-uniformity of each pixel point on each area of the rock region; S62, a sub-step of calculating the rock roughness of each area of the rock region based on the first contour distribution non-uniformity, the second contour distribution non-uniformity, and the third contour distribution non-uniformity of each pixel point on each area of the rock region; S63, a sub-step of calculating the surface average roughness based on the rock roughness of each area of the rock region; S64, including a sub-step of setting this rock region as a loose rock mass when the surface average roughness is greater than the roughness threshold.

[0055] In this embodiment, the roughness threshold is a threshold set for the surface average roughness.

[0056] JPEG0007693178000012.jpg55170

[0057] The present invention first calculates the contour distribution non-uniformity of each pixel point on each area of the rock region, characterizes the degree of roughness at each pixel point by the contour distribution non-uniformity of the pixel point, further calculates the rock roughness of each area of the rock region, takes the average value of the rock roughness of each area on the rock region to obtain the surface average roughness of the rock region. When the surface average roughness of the rock region is large, it indicates that the weathering or erosion of the rock is serious, and this rock is more likely to collapse. When the surface average roughness of the rock region is small, the rock surface is smooth, the rock surface has been polished for a long time, and the rock becomes harder.

[0058] JPEG0007693178000013.jpg39170

[0059] The present invention adds the three contour distribution non-uniformities of each pixel point on the rock regions of the original image, the first detailed feature image, and the second detailed feature image to reflect the contour distribution situation at this pixel point, and adds all the pixel points to reflect the rock roughness of each area.

[0060] In this embodiment, all threshold values are specifically set based on experiments or experience.

[0061] After collecting the original image, the present invention constructs a first detailed feature image and a second detailed feature image based on the original image, prominently shows the image detailed features, performs cutting processing on both, realizes area-based classification, and based on the original block, the first feature block, and the second feature block in the same area, classifies this area based on a three-channel classification model to obtain the classification value of this area, and further performs correction processing to improve the classification accuracy, find the rock area, peel off the background area, and improve the accuracy of identifying the loose rock mass. Furthermore, based on the surface average roughness of the rock area, it is determined whether this rock is a loose rock mass. The present invention characterizes the roughness and the joint and crack conditions of this rock according to the roughness situation on the surface of the rock area. The greater the roughness, the easier it is for this rock mass to be eroded, the rock becomes loose, and it is more likely to collapse.

Claims

1. A method for identifying slope loose rock mass, comprising the steps of: S1, collecting images of a target area to obtain an original image; S2. Constructing a first detail feature image and a second detail feature image based on color values ​​in the original image; S3: performing cutting on the original image, the first detailed feature image, and the second detailed feature image, respectively, to obtain a plurality of original blocks corresponding to the original image, a plurality of first feature blocks corresponding to the first detailed feature image, and a plurality of second feature blocks corresponding to the second detailed feature image; S4, according to the original block, the first feature block and the second feature block of the same area, a classification is performed on the area based on a three-channel classification model to obtain a classification value of the area; S5: performing a correction process on the classification value of each area to find a rock area; S6. Calculating an average surface roughness of the rock region based on the first contour distribution unevenness in the rock region of the original image of each pixel point on the rock region, the second contour distribution unevenness in the rock region of the first detailed feature image, and the third contour distribution unevenness in the rock region of the second detailed feature image, and when the average surface roughness is greater than a roughness threshold, determining the rock region as a loose rock body; The S2 is S21, a sub-step of centering each pixel point in the original image; S22, a sub-step of calculating a first color feature value and a second color feature value based on the center neighborhood color value, respectively; S23, constructing a first detailed feature image by using the first color feature value as a new color value of the pixel point at the center; S24, the second color feature value is set as a new color value of the pixel point at the center to construct a second detailed feature image; The three-channel classification model in S4 includes a first channel processing unit, a second channel processing unit, a third channel processing unit, a first multiplier M1, a second multiplier M2, a first Maxpool layer, a first Avgpool layer, a second Maxpool layer, a second Avgpool layer, a first adder A1, a second adder A2, a first Concat layer, and a fully connected layer; an input end of the first channel processing unit is used for inputting a first feature block, an input end of the second channel processing unit is used for inputting an original block, and an input end of the third channel processing unit is used for inputting a second feature block; A first input end of the first multiplier M1 is connected to an output end of a first channel processing unit, a second input end of the first multiplier M1 is connected to an output end of a second channel processing unit, an output end of the first multiplier M1 is respectively connected to an input end of a first Maxpool layer and an input end of a first Avgpool layer, a first input end of the second multiplier M2 is connected to an output end of a second channel processing unit, a second input end of the second multiplier M2 is connected to an output end of a third channel processing unit, an output end of the second multiplier M2 is respectively connected to an input end of a second Maxpool layer and an input end of a second Avgpool layer, A method for identifying slope loose rock bodies, characterized in that: a first input end of an adder A1 is connected to an output end of a first Maxpool layer, a second input end of the first adder A1 is connected to an output end of a first Avgpool layer, a first input end of the second adder A2 is connected to an output end of a second Maxpool layer, a second input end of the second adder A2 is connected to an output end of a second Avgpool layer, input ends of the first Concat layer are respectively connected to the output end of the first adder A1 and the output end of the second adder A2, an output end of the first Concat layer is connected to an input end of a fully connected layer, and an output end of the fully connected layer is an output end of a three-channel classification model.

2. The first channel processing unit, the second channel processing unit, and the third channel processing unit have the same structure, and each includes a first Conv block, a second Conv block, a third Conv block, a fourth Conv block, a fifth Conv block, a sixth Conv block, a second Concat layer, a third Concat layer, and a fourth Concat layer; The input end of the first Conv block is an input end of a first channel processing unit, a second channel processing unit or a third channel processing unit, the output end of the first Conv block is connected to the input end of the second Conv block and the input end of the third Conv block, respectively, the input end of the second Concat layer is connected to the output end of the second Conv block and the output end of the third Conv block, respectively, the output end of the second Concat layer is connected to the input end of a fourth Conv block, and the input end of the third Concat layer is connected to the input end of the fourth Conv block. the output end of the third Concat layer is connected to the output end of the second Conv block and the output end of the third Conv block, the output end of the third Concat layer is connected to the input end of the fifth Conv block, the input ends of the fourth Concat layer are connected to the output end of the fourth Conv block and the output end of the fifth Conv block, the output end of the fourth Concat layer is connected to the input end of the sixth Conv block, and the output end of the sixth Conv block is the output end of the first channel processing unit, the second channel processing unit or the third channel processing unit; The size of the convolution kernel of the second Conv block and the fourth Conv block is 3*3; The method for identifying slope loose rock masses according to claim 1, wherein the size of the convolution kernel of the third Conv block and the fifth Conv block is 5*5.

3. The above-mentioned S5 is S51, a sub-step of setting each area as a reference area; The method for identifying loose rock bodies on slopes according to claim 1, further comprising the sub-step of: S54, when the classification value after the correction or the classification value not involved in the correction is greater than the classification threshold, marking the corresponding area as a candidate area, and constituting a rock region with a plurality of candidate areas.

4.

5. The above-mentioned S6 is S61, a sub-step of calculating a first contour distribution unevenness in the rock region of the original image, a second contour distribution unevenness in the rock region of the first detailed feature image, and a third contour distribution unevenness in the rock region of the second detailed feature image for each pixel point on each area of ​​the rock region; S62, a sub-step of calculating rock roughness of each area of ​​the rock region based on a first contour distribution unevenness in the rock region of the original image of each pixel point on each area of ​​the rock region, a second contour distribution unevenness in the rock region of the first detailed feature image, and a third contour distribution unevenness in the rock region of the second detailed feature image; S63, a sub-step of calculating an average surface roughness based on the rock roughness of each area of ​​the rock region; 2. The method for identifying loose rock mass on a slope according to claim 1, further comprising the sub-step of: S64, when the surface average roughness is greater than the roughness threshold, determining the rock area as a loose rock mass.

6.

7.

Images