CN117315221A - Rock slope crack identification method based on multi-source image data fusion - Google Patents

Abstract

The invention provides a rock slope crack identification method based on multi-source image data fusion, which comprises the following steps: s1: visible light image data and infrared image data are collected; s2: registering images; s3: preprocessing an image; s4: decomposing and fusing the multi-source images; s5: according to the method, high-quality unmanned aerial vehicle image data are acquired from a camera imaging principle model, a visible light image and a thermal infrared image are registered, image pretreatment is respectively carried out on the visible light image and the thermal infrared image of a high-steep rock slope scene, the visible light image and the thermal infrared image after fusion pretreatment are used for identifying the rock slope scene cracks, and a quantitative crack pattern identification method is used for identifying the rock slope scene cracks, so that the definition and accuracy of image acquisition are greatly improved, and compared with other methods, the crack identification effect after multi-source image data fusion in the rock slope scene is more stable and reliable, and the identification accuracy is higher.

Description

A method for identifying cracks in rock slopes based on multi-source image data fusion

Technical Field

The invention relates to the field of rock mass gap research, and in particular to a rock slope crack identification method based on multi-source image data fusion.

Background Art

With the rapid economic development in my country in recent years, various large-scale infrastructure construction is in full swing, and geotechnical engineering problems involving complex geological environments are emerging one after another. With the continuous advancement of geotechnical technology, foundation pit engineering, tunnel engineering, etc. are facing problems such as deeper excavation depth, larger excavation area and more complicated construction process.

At present, there are two main sources of data for automatic information extraction: (1) digital photography and (2) three-dimensional laser scanning. There are two main methods for processing digital images obtained by digital photography. One is to directly perform digital image processing and analysis on digital photos, such as: using image processing algorithms to extract rock mass cracks, and automatically analyzing the extraction results to obtain the parameters of the degree of development of surrounding rock cracks; directly performing statistical feature analysis on the pre-processed images. The other is based on close-range photogrammetry technology. According to the mapping relationship of the reference points in the tunnel face photos taken at different angles, the three-dimensional space point cloud data is calculated, and then the information such as the length of the rock mass crack and the occurrence of the rock formation is analyzed. There are also methods that combine the three, such as the patent announcement number: CN 109187548 discloses a complex rock crack model modeling and identification method. The three-dimensional solid model is modeled by ProE software to facilitate the location of the crack in the rock. The different combinations of the crack inclination angle, rock bridge angle, etc. greatly simplify the complexity of ANASY or other software modeling, and there is no need to perform Boolean operation cutting models in ANSYS software. However, in the above method, the image acquisition clarity of the rock crack is not enough, because when the camera shutter is opened, the light enters the camera and is received by the photosensitive device, that is, the camera collects the target optical signal within the exposure time. The Bayer sensor array in the digital camera contains three matrix patterns of green, red and blue, but the signal pixels collected by it have only one color channel. In actual situations, a pixel should contain three color channel values. Therefore, the Bayer interpolation method is used in the digital camera imaging process to fill in the missing pixel channel values. The interpolation is gamma corrected and lossy compressed to save as a color image. The above digital camera imaging model can be expressed as:

In formula (1), I represents color image, c represents lossy compression, g represents camera response function (CRF), BI represents Bell interpolation, S(t) represents the response of the sensor at time t, and N represents the noise signal. According to formula (1), the accumulated signal received by the sensor during the imaging process of the digital camera is Blurs will be produced in the image. Bayer interpolation and lossy compression processes can cause artifacts in the imaging process and reduce image resolution, affecting the accuracy of crack identification.

Summary of the invention

The above technical problems have not been solved to improve the accuracy of crack identification.

The present invention proposes a rock slope crack identification method based on multi-source image data fusion, including: S1: visible light image data and infrared image data acquisition: by using a multi-source camera mounted on an unmanned aerial vehicle, the unmanned aerial vehicle performs terrain-like tilt photogrammetry to obtain image data and thus obtain initial terrain information of the study area, and performs spatial plane fitting on the terrain unit, creates a flight plane parallel to it, constructs multi-angle structural surfaces on the flight plane, and the camera is perpendicular to the structural surface, and simultaneously obtains visible light image and thermal infrared image data;

S2: Image registration: Use scale-invariant feature transformation method to align the content information of visible light image and thermal infrared image;

S3: Image preprocessing: Improved Laplacian sharpening and multi-scale Retinex enhancement of color restoration are performed on visible light images and thermal infrared images respectively. The improved Laplacian sharpening is to perform second-order differential on image pixels to improve edge details, which can be expressed as:

Where I(x, y) represents the input visible light image with pixel coordinates (x, y), γ and Γ are parameters for adjusting the sharpness and edge brightness respectively;

Combining formulas (2) and (3), the Laplace operator is obtained by second-order differential calculation, which is expressed as

represents the Laplacian operator, and Λ represents the weight parameter. The image I _LS (x, y) after Laplacian sharpening can be expressed as

The multi-scale Retinex enhancement process of color restoration is as follows: the multi-scale retinal MSR is obtained by taking the sum of different weighted scales of the single-scale retinal SSR, and then the color restoration function C(x, y) is added to the multi-scale retinal MSR, as shown in formula (9). This function adjusts the percentages of the three color channels of the image:

represents the i-th reflection component of the image I(x, y), G(x, y) represents the covering support function to obtain the reflection component, N represents different scales, w _n represents the weight and satisfies M represents the total number of pixels in the input image, represents the enhanced image of the output;

S4: Decomposition and fusion of multi-source images: Extract the three-channel (R, G, B) images of the enhanced images I _LS and R _MSRCR in S2 respectively, and perform decomposition and fusion strategies on them, retaining the low-rank part and significant part of each single channel, abandoning the sparse noise part of the single channel, and then using the average strategy to fuse the low-rank part of each channel, and using the summation strategy to fuse the significant part of each channel;

S5: Crack identification: The fused image obtained in S4 is obtained by applying the Frangi vesselness filter and extracting linear targets from it.

Preferably, in S4, MSRCR and Laplacian sharpening are used to process the input visible light image respectively, and the processed images I _LS and R _MSRCR are obtained. Then, I _LS and R _MSRCR are decomposed and their three-channel (R, G, B) images are extracted respectively; the three-channel images of I _LS and R _MSRCR are subjected to decomposition and fusion strategies, and the sparse noise part of a single channel is discarded in the decomposed result, and the low-rank part and the significant part in the I _LS and R _MSRCR images are retained. The process can be expressed as

Where i represents one of the [R, G, B] channels, and represents the single-channel low-rank part of R _MSRCR and I _LS , and Represents the single channel salient part of R _MSRCR and I _LS , on this basis, the average strategy is used to fuse the low rank part of each channel, and the summation strategy is used to fuse the salient part of each channel, which is defined as

represents the low-rank partial fusion result, Indicates the significant part of the fusion result. The present invention realizes single-channel image fusion in the following way

I ₁ ⁱ represents the single-channel reconstructed image, and finally fused and The final decomposition and fusion output result I ₁ of the visible light image is achieved.

Preferably, in S4, Laplacian sharpening is used to preprocess the thermal infrared image V, and then the preprocessed thermal infrared image V _LS and the enhanced visible light image I ₁ are decomposed and fused according to formulas (25) to (31).

in, and represents the single-channel low-rank part of I ₁ and V _LS , and represents the single channel significant portion of _I1 and _VLS , Represents the low-rank fusion result of a single channel, represents the significant fusion result of a single channel, Represents a single-channel reconstructed image, and finally fused and The final decomposition and fusion output result F of the visible light image and thermal infrared image is achieved.

Preferably, the image to be processed F is first converted into a grayscale image and filtered using a Gaussian filter to reduce noise. A binary image is obtained based on the filtered image. In order to extract the skeleton of rock cracks, morphological closure and directional closure strategies are used to generate a skeletonized binary image. Finally, the distribution of cracks is displayed in the multi-source fusion image data.

The rock slope crack identification method based on multi-source image data fusion proposed in the present invention has the following beneficial effects: high-quality UAV image data is obtained based on the camera imaging principle model, visible light images and thermal infrared images are registered, image preprocessing is performed on visible light images and thermal infrared images of steep rock slope scenes, the preprocessed visible light images and thermal infrared images are fused, and a quantitative crack pattern recognition method is used to identify cracks in rock slope scenes, which greatly improves the clarity and accuracy of image acquisition. In rock slope scenes, the crack identification effect after multi-source image data fusion is more stable and reliable than other methods and has a higher recognition accuracy.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for describing the embodiments are briefly introduced below.

Figure 1 is a block diagram of the overall technical method;

Figure 2 shows the DJI M300 RTK UAV equipped with the Zenmuse H20T multi-source camera;

Figure 3 shows the image data acquisition results of the zoom camera (left) and the thermal infrared camera (right);

FIG4 is a schematic diagram of the route planning of the UAV multi-angle close-up photography terrain unit;

Figure 5 is a schematic diagram of the route planning of the UAV for multi-angle close-up photography of the structure surface;

Figure 6 is a supplement to the blind spot route;

Figure 7 shows the 24-hour thermal infrared image data monitoring results of a high and steep rock slope;

FIG8 is a schematic diagram of the registration of a visible light image and a thermal infrared image;

Figure 9 shows the decomposition and fusion of visible light images;

Figure 10 shows the decomposition and fusion of multi-source images;

Figure 11 shows the visible light image used in the experiment and its corresponding thermal infrared image;

FIG12 is a diagram showing multi-source decomposition and fusion of visible light images and thermal infrared images;

FIG13 is a diagram showing the results of a global fusion experiment;

Figure 14 shows the global crack segmentation result;

FIG15 is an input visible light image;

Figure 16 shows the local crack segmentation result;

FIG17 is a schematic diagram of skeleton segmentation;

DETAILED DESCRIPTION

The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.

Example

This embodiment carried out data collection work near the dam site of Xulong Hydropower Station in Deqin County, Yunnan Province, and the time for collecting high-quality visible light images and thermal infrared images was 7:30 in the morning.

As shown in Figure 1, the present invention proposes a rock slope crack identification method based on multi-source image data fusion, including: S1: visible light image data and infrared image data acquisition; S2: image registration; S3: image preprocessing; S4: decomposition and fusion of multi-source images; S5: crack identification.

Specifically, in S1, a multi-source camera is carried by a UAV, and the UAV is used for terrain-like tilt photogrammetry to obtain image data and thus obtain the initial terrain information of the study area, and the terrain unit is fitted into a spatial plane to create a flight plane parallel to it. On the flight plane, a multi-angle structural surface of the component is constructed, and the camera is perpendicular to the structural surface, and visible light image and thermal infrared image data are obtained at the same time; in this embodiment, a DJI M300 RTK UAV is selected to carry a Zenmuse H20T multi-source camera (as shown in FIG. 2 ), and the zoom camera image data acquisition parameters and the thermal infrared camera image data acquisition parameters are set according to the above-mentioned digital camera imaging principle as shown in Table 1, and the typical image data collected under this parameter is shown in FIG. 3 .

Table 1 Zoom camera image data acquisition parameters and thermal infrared camera image data acquisition parameters

Table 2 Related working parameters of UAV operation

For the collection of visible light images and UAV thermal infrared image data, the Digital Elevation Model (DEM) is imported into the UAV ground controller to plan the flight route, so that the UAV can adjust the flight altitude according to the terrain changes in the study area to perform terrain-like oblique photogrammetry and obtain clearer image data.

Then import the image data into the drone, establish a terrain-simulating model of the study area, and obtain the initial terrain information of the study area. Multi-angle approach route planning for terrain units. According to the terrain feature information of the study area, it is divided into multiple terrain units, and a spatial plane fitting is performed for each terrain unit, as shown in Figure 4. Outside the specific distance of the fitting plane (between 5 and 100m), a close flight plane parallel to it is created. Set a multi-angle approach route in the flight plane, and the camera angle is automatically set to be perpendicular to the fitting plane. The image overlap rate is set to 80% or more in the heading and 60% or more in the lateral direction. Multi-angle approach route planning for structural surfaces. According to the development grouping characteristics of the structural surfaces determined by the field investigation, multi-angle approach routes for structural surfaces are planned on the flight plane (as shown in Figure 5). Set a corresponding approach flight for each group of structural surfaces, and the camera angle is perpendicular to the structural surface. The image overlap rate setting is consistent with the steps of multi-angle approach route planning for terrain units. Supplementary route planning for key areas. On the basis of the above routes, multi-angle routes are set for hidden structural surfaces and other areas of interest that are prone to photographic blind spots, as well as research areas, to conduct panoramic close-up photography to eliminate photographic blind spots.

Thermal infrared imaging technology was used to collect thermal infrared image data of high and steep rock slope scenes. Based on the visible light image acquisition strategy, 24-hour thermal infrared image data monitoring of the same area of the high and steep rock slope was carried out (as shown in Figure 7), and data was collected every 1.5 hours.

For S2: Image Registration

After completing the multi-angle close-to-route mission, the visible light image and its corresponding thermal infrared image data are obtained. Due to the parallax between the color camera and the thermal infrared camera, the image data collected by the two cameras are not completely corresponding. The patent of this invention uses the scale-invariant feature transform (SIFT) method to align the content information of the visible light image and the thermal infrared image, as shown in Figure 8.

For S3: image preprocessing, mainly for visible light images and thermal infrared images, Laplacian sharpening and Retinex enhancement processing are first implemented respectively. Laplacian sharpening enhances edge detail features, and Retinex enhancement improves image contrast and clarity. Since the cracks on the surface of the rock structure happen to be the edge part, and some cracks are not obvious enough in the image, it is necessary to enhance the edge detail features. The present invention improves the traditional Laplace sharpening strategy to enhance the edges of images taken on steep rock slopes. The edge details are improved by performing second-order differentials on image pixels, which can be expressed as:

Where I(x, y) represents the pixel coordinates of the input visible light image (x, y), and γ and Γ are parameters for adjusting the sharpness and edge brightness, respectively. Combining formulas (2) and (3), the Laplace operator is calculated by second-order differential, which is expressed as:

represents the Laplacian operator, and Λ represents the weight parameter. The image I _LS (x, y) after Laplacian sharpening can be expressed as:

The essence of the Retinex principle is to treat the image as a multiplication of the illumination component and the reflection component from the perspective of the multiplicative model, which can be expressed as:

I(x,y)＝L(x,y)·R(x,y) (6)

L(x, y) represents the illumination component at pixel (x, y), and R(x, y) represents the reflection component at pixel (x, y). Among them, the reflection component refers to the high-frequency information in the image and is independent of the illumination component. The Retinex strategy aims to estimate the image brightness and obtain the reflection image by removing the brightness component of the original image. However, it cannot achieve a balance in terms of dynamic range, edge enhancement, and color constancy. The enhancement of rock slope images needs to fully consider the above three conditions. Therefore, the patent of the present invention adopts a multi-scale Retinex algorithm (MSRCR) for color restoration in the image preprocessing process, which is a variant algorithm of Retinex. MSRCR shows good performance in enhancing images taken in different visual scenes (such as underwater scenes, low-light scenes, and foggy scenes). The steps of MSRCR are as follows: first, the multi-scale retina (MSR) is calculated by taking the sum of different weighted scales of the single-scale retina (SSR), which is a balance between the dynamic range and color fidelity of the image. Therefore, MSR can handle both tonal contrast and dynamic range, and avoid the occurrence of halo phenomenon. Although MSR has made progress in improving the performance of SSR, it introduces color distortion. Subsequently, a color restoration function C(x, y) was added to MSR to solve the color distortion problem, as shown in formula (9), which adjusts the percentage of the three color channels of the image.

represents the i-th reflection component of the image I(x, y), G(x, y) represents the covering support function to obtain the reflection component, N represents different scales, w _n represents the weight and satisfies M represents the total number of pixels in the input image, Represents the output enhanced image.

S4: Image decomposition and fusion

The present invention realizes image decomposition and fusion by improving the low-rank latent representation algorithm. The algorithm can be divided into three steps: (1) decomposing the input image into a low-rank part, a significant part and a sparse noise part; (2) using two different fusion strategies to fuse the low-rank part and the significant part of the image; (3) reconstructing the image. The global structural information and brightness information of the input image are mainly in the low-rank part, while the local structural information and significant features are in the significant part. In addition to the sparse noise part, the low-rank part and the significant part are both key information that needs to be retained in the image. The mathematical expression of LatLRR can be defined as:

X＝XZ+LX+E (11)

Where X, Z, L, and E represent the observed data matrix, low-rank coefficients, significant coefficients, and sparse noise, respectively. XZ and LX represent the low-rank part and the significant part, respectively. Formula (11) is optimized and solved in the following way:

||·|| _* and ||·|| ₁ represent the nuclear norm and L ₁ norm respectively, and λ represents the balance parameter (its value is greater than 0). The low-rank part I _Lrr and the significant part I _S of the decomposed image are expressed as:

I _Lrr ＝XZ (13)

_IS = LX (14)

After decomposing the image and removing the sparse noise part, the average and strategy are used to fuse the low-rank part and the salient part respectively, which is expressed as

F _Lrr (x,y)＝w ₁ I _1Lrr (x,y)+w ₂ I _2Lrr (x,y) (15)

F _S (x,y)＝s ₁ I _1S (x,y)+s ₂ I _2S (x,y) (16)

_w1 and _w2 are weight parameters of the low-rank part, _I1Lrr and _I2Lrr are the low-rank parts of the decomposition, and _F1Lrr represents the low-rank fusion result. _s1 and _s2 are weight parameters of the significant part in the average and strategy process, _I1S and _I2S are the significant parts of the decomposition, and _Fs represents the significant fusion result. After obtaining the low-rank fusion result and the significant fusion result, the final reconstructed image F can be fused using _F1Lrr and _Fs , which is defined as

F(x,y)＝F _Lrr (x,y)+F _S (x,y) (17)

Specifically: After the visible light image is preprocessed, the patent of the present invention performs visible light image enhancement according to the improved low-rank latent representation algorithm (as shown in Figure 9). Specifically, MSRCR and Laplacian sharpening are used to process the input visible light image respectively, and the processed images I _LS and R _MSRCR are obtained. Then I _LS and R _MSRCR are decomposed and their three-channel (R, G, B) images are extracted respectively. The three-channel images of I _LS and R _MSRCR are subjected to a decomposition and fusion strategy. In the decomposed result, the patent of the present invention abandons the sparse noise part of a single channel and retains the low-rank part and the significant part in the I _LS and R _MSRCR images. The process can be expressed as

Where i represents one of the [R, G, B] channels, and represents the single-channel low-rank part of R _MSRCR and I _LS , and Represents the single channel salient part of R _MSRCR and I _LS . On this basis, the average strategy is used to fuse the low-rank part of each channel, and the summation strategy is used to fuse the salient part of each channel, which is defined as

represents the low-rank partial fusion result, Indicates the significant part of the fusion result. The present invention realizes single-channel image fusion in the following way

Represents a single-channel reconstructed image, and finally fused and The final decomposition and fusion output result I ₁ of the visible light image is achieved.

Decomposition and fusion of infrared light: For the fusion of visible light image and thermal infrared image data (as shown in FIG10 ), repeat the image decomposition step in S4. The input images are the enhanced visible light image I ₁ and the thermal infrared image V. First, Laplacian sharpening is used to preprocess the thermal infrared image V LS, and then the preprocessed thermal infrared image V _LS and the enhanced visible light image I ₁ are decomposed and fused according to formulas (25) to (31).

in, and represents the single-channel low-rank part of I ₁ and V _LS , and represents the single channel significant portion of _I1 and _VLS , Represents the low-rank fusion result of a single channel, represents the significant fusion result of a single channel, Represents a single-channel reconstructed image, and finally fused and The final decomposition and fusion output result F of the visible light image and thermal infrared image is achieved.

Finally, the cracks are identified by the fused output structure. The new version of FracPaQ is released on the Zenodo platform. The image is analyzed by using the Frangi vesselness filter and linear targets are extracted from it. The specific process is as follows: First, the processed image F is converted into a grayscale image and filtered using a Gaussian filter to reduce noise. A binary image is obtained based on the filtered image. In order to extract the skeleton of the rock cracks, morphological closure and directional closure strategies are used to generate a skeletonized binary image. Finally, the distribution of cracks is displayed in the multi-source fused image data.

Through the above method, after the data collection was completed, 13 pairs of visible light images and their corresponding thermal infrared images (as shown in Figure 11) were selected to carry out fusion experimental research, global fusion experimental research and local fusion experimental research.

Figure 12 shows (a) the input visible light image; (b) the input thermal infrared image; (c) the visible light image after decomposition and fusion; (d) the thermal infrared image after Laplacian sharpening; (e) the fusion result of the decomposition and fusion of the visible light image and the input thermal infrared image; (f) the decomposition and fusion result of multi-source images.

FracPaQ is used to identify cracks in the above results, and the crack identification results are shown in Table 3. According to the crack identification results, it can be proved that the crack identification effect of the present invention after multi-source image data fusion in the rock slope scene is better.

Table 3 Crack recognition results of fusion test (black bold indicates the best result)

The global and local crack recognition experiments were carried out using 13 pairs of visible light images and corresponding thermal infrared images in the fusion experiment. Figure 13 shows the results of the global fusion experiment, where (a) is the input visible light image; (b) is the input thermal infrared image; and (c) is the fusion result of the present invention.

According to the fusion results in Figure 13, FracPaQ was used to identify cracks in the above results.

Figure 14 is a diagram of the global crack recognition effect, where (a) the input visible light image; (b) the input thermal infrared image; (c) the input visible light image recognition result; (d) the input thermal infrared image recognition result; (e) the recognition result after fusion of the patent of the present invention.

Table 4 shows the results of global crack recognition. To verify the generalization of the present invention, local crack recognition experiments were carried out on the rock slope area extraction in 13 pairs of images.

Table 4 Global crack recognition results (black bold indicates the best result)

FIG15 is a fusion result of a local crack recognition experiment, where (a) the input visible light image; (b) the input thermal infrared image; (c) the fusion result of the present invention;

Figure 16 is a diagram of the local crack recognition effect, where the local crack segmentation results are: (a) input visible light image; (b) input thermal infrared image; (c) input visible light image recognition result; (d) input thermal infrared image recognition result; (e) recognition result after fusion of the patent of the present invention.

Table 5 shows the coordinate correspondence between the global image and the local image, and Table 6 shows the number of local crack identification results. According to the crack identification results, it can be proved that the crack identification effect of the present invention after multi-source image data fusion in the rock slope scene is more stable and reliable than other methods, and the recognition accuracy is higher. Based on the crack identification results, the morphological features are extracted to obtain the crack skeleton, as shown in Figure 17.

Table 5 Correspondence between global image and local image coordinates

Global resolution Local resolution Upper left corner coordinates Upper right corner coordinates Lower left corner coordinates Lower right corner coordinates 1# 960×768 304×301 (656,374) (960,374) (656,675) (960,675) 2# 960×768 310×418 (650,350) (960,350) (650,768) (960,768) 3# 960×768 480×392 (480,376) (960,376) (480,768) (960,768) 4# 960×768 551×407 (252,361) (803,361) (252,768) (803,768) 5# 960×768 544×521 (20,142) (564,142) (20,663) (564,663) 6# 960×768 356×462 (604,6) (960,6) (604,468) (960,468) 7# 960×768 470×474 (346,1) (816,1) (346,475) (816,475) 8# 960×768 484×698 (0,16) (484,16) (0,714) (484,714) 9# 960×768 406×499 (554,269) (960,269) (554,768) (960,768) 10# 960×768 614×412 (346,2) (960,2) (346,414) (960,414) 11# 960×768 589×298 (371,286) (960,286) (371,584) (960,584) 12# 960×768 315×245 (356,194) (671,194) (356,439) (671,439) 13# 960×768 500×522 (460,1) (960,1) (460,523) (960,523)

Table 6 Local crack identification results (black bold indicates the best result)

Claims (4)

1. A rock slope crack identification method based on multi-source image data fusion, characterized by comprising: S1: Visible light image data and infrared image data acquisition: The UAV is equipped with a multi-source camera, and the UAV is used for terrain-like tilt photogrammetry to obtain image data and then obtain the initial terrain information of the study area. The terrain unit is fitted with a spatial plane, and a flight plane parallel to it is created. On the flight plane, multi-angle structural surfaces are constructed, and the camera is perpendicular to the structural surface, and visible light image and thermal infrared image data are obtained at the same time; S2: Image registration: Use scale-invariant feature transformation method to align the content information of visible light image and thermal infrared image; S3: Image preprocessing: Improved Laplacian sharpening and multi-scale Retinex enhancement of color restoration are performed on visible light images and thermal infrared images respectively. The improved Laplacian sharpening is to perform second-order differential on image pixels to improve edge details, which can be expressed as: Where I(x,y) represents the input visible light image with pixel coordinates (x,y), γ and Γ are parameters for adjusting the sharpness and edge brightness respectively; Combining formulas (2) and (3), the Laplace operator is obtained by second-order differential calculation, which is expressed as represents the Laplacian operator, and Λ represents the weight parameter. The image I _LS (x, y) after Laplacian sharpening can be expressed as The multi-scale Retinex enhancement process of color restoration is as follows: the multi-scale retinal MSR is obtained by taking the sum of different weighted scales of the single-scale retinal SSR, and then the color restoration function C(x, y) is added to the multi-scale retinal MSR, as shown in formula (9). This function adjusts the percentages of the three color channels of the image: represents the i-th reflection component of the image I(x, y), G(x, y) represents the covering support function to obtain the reflection component, N represents different scales, w _n represents the weight and satisfies M represents the total number of pixels in the input image, represents the enhanced image of the output; S4: Image decomposition and fusion: Extract the three-channel (R, G, B) images of the enhanced images I _LS and R _MSRCR in S2 respectively, and perform decomposition and fusion strategies on them, retain the low-rank part and significant part of each single channel, abandon the sparse noise part of the single channel, and then use the average strategy to fuse the low-rank part of each channel, and use the summation strategy to fuse the significant part of each channel; S5: Crack identification: The fused image obtained in S4 is obtained by applying the Frangi vesselness filter and extracting linear targets from it. 2. The rock slope crack identification method based on multi-source image data fusion according to claim 1 is characterized in that, in S4, MSRCR and Laplacian sharpening are used to process the input visible light image respectively, and the processed images I _LS and R _MSRCR are obtained. Then, I _LS and R _MSRCR are decomposed and their three-channel (R, G, B) images are extracted respectively; the three-channel images of I _LS and R _MSRCR are subjected to decomposition and fusion strategies, and the sparse noise part of a single channel is abandoned in the decomposed result, and the low-rank part and the significant part in the I _LS and R _MSRCR images are retained. The process can be expressed as Where i represents one of the [R, G, B] channels, and represents the single-channel low-rank part of R _MSRCR and I _LS , and Represents the single channel salient part of R _MSRCR and I _LS , on this basis, the average strategy is used to fuse the low rank part of each channel, and the summation strategy is used to fuse the salient part of each channel, which is defined as F ₁ ^i_Lrr represents the fusion result of the low-rank part, and F ₁ ^i_S represents the fusion result of the significant part. The present invention realizes single-channel image fusion in the following way Represents a single-channel reconstructed image, and finally fused and The final decomposition and fusion output result I ₁ of the visible light image is achieved. 3. The rock slope crack identification method based on multi-source image data fusion according to claim 1 is characterized in that, in S4, Laplacian sharpening is used to pre-process the thermal infrared image V, and then the pre-processed thermal infrared image V _LS and the enhanced visible light image I ₁ are decomposed and fused according to formulas (25) to (31); in, and represents the single-channel low-rank part of I ₁ and V _LS , and represents the single channel significant portion of _I1 and _VLS , Represents the low-rank fusion result of a single channel, represents the significant fusion result of a single channel, Represents a single-channel reconstructed image, and finally fused and The final decomposition and fusion output result F of the visible light image and thermal infrared image is achieved. 4. The rock slope crack identification method based on multi-source image data fusion according to claim 3 is characterized in that the image to be processed F is first converted into a grayscale image and filtered using a Gaussian filter to reduce noise, and a binary image is obtained based on the filtered image. In order to extract the skeleton of the rock cracks, morphological closure and directional closure strategies are used to generate a skeletonized binary image. Finally, the distribution of the cracks is displayed in the multi-source fusion image data.