Road crack detection method based on multi-scale Retinex fused with wavelet transform
By fusion of wavelet transform and multi-scale Retinex algorithm, the problem of manual detection in traditional methods is solved, and the problem of halo phenomenon in the Retinex algorithm is efficient, accurate enhancement and clarity improvement of road crack images is achieved, and it is suitable for automatic detection.
Patent Information
- Application Number
- CN202210528070.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-05-16
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2042-05-16
AI Technical Summary
Traditional road crack detection methods rely on artificial vision, are time-consuming and labor-intensive and difficult to objectively evaluate the degree of road deterioration. In addition, the Retinex algorithm is prone to 'halo' under insufficient or uneven light conditions, affecting image recognition.
The multi-scale Retinex method based on fusion wavelet transformation is used to divide the road crack images into low-frequency and high-frequency components. The crack profile information is extracted through the top hat transformation, and the multi-scale Retinex enhances the high-frequency components, combined with image entropy calculation, avoiding halo The phenomenon occurs and the reconstruction obtains a clear image of road cracks.
It realizes efficient and accurate enhancement of road crack images, highlights the detailed information of the image texture, avoids halo phenomenon, improves the clarity and contrast of the image, enhances the crack feature information, and is suitable for automatic detection.
Smart Images

Figure CN114881979B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of image processing, and in particular relates to a road crack detection method based on multi-scale Retinex fused with wavelet transform. Background Art
[0002] Cracks are the main road disease and a potential threat to road safety. Detecting and repairing cracks is the focus of intelligent transportation system research. Timely and accurate inspection of pavement cracks can help transportation agencies reduce road maintenance costs and extend the service life of the road. However, traditional crack detection mainly relies on manual visual inspection, which is time-consuming and laborious, and it is difficult to objectively evaluate the degree of road deterioration. Therefore, it is necessary to develop automatic, accurate and efficient pavement crack detection methods. With the advancement of computer vision and civil engineering image processing, image-based automatic crack detection technology has gradually replaced traditional manual visual inspection.
[0003] Road pavement images are often obtained under adverse weather and lighting conditions. Images taken under insufficient or uneven lighting conditions will have poor contrast, low brightness, blurred local details, sudden brightness changes, and are often accompanied by significant noise. This makes it difficult for machine vision to extract and analyze information from these images. Therefore, crack image enhancement becomes an important criterion for subsequent identification. Many researchers are also committed to crack image enhancement, which requires highlighting crack feature information while increasing brightness and contrast.
[0004] Retinex is a theory of human color vision proposed by Edwin Land. Unlike traditional linear and nonlinear methods that can only enhance a certain type of image features, Retinex can strike a balance between dynamic range compression, edge enhancement, and color constancy, so it can adaptively enhance various types of images. Retinex is most commonly used to enhance medical images such as blood vessels and retina. Considering the similarity between the texture features of road crack images and medical images, the literature Wang Q., Zhang N., Jiang K., Ma C., Zhou Z., Dai C., Tunnel Lining Crack Recognition Based on Improved Multiscale Retinex and Sobel Edge Detection, (2021) Mathematical Problems in Engineering, 2021, art.no.9969464, Cited 0 times., DOI:10.1155 / 2021 / 9969464. and the literature
[14] Bin Qian, Z. Tang and W. Xu,"Pavement crack detection based on improved tensor The application of Retinex image enhancement algorithm to crack image enhancement is disclosed. The research and development of such algorithms has greatly improved the effect of crack image enhancement processing. However, because the Gaussian operator of the Retinex algorithm cannot estimate the illumination well in the strong light-shadow transition area and the area with large brightness changes, the output image will have a "halo" phenomenon, which will affect the recognition of subsequent images.
[0005] Therefore, it is urgent to design a road crack image processing method that can make the road crack image clearer and avoid the "halo" phenomenon. Summary of the invention
[0006] The present invention provides a road crack detection method based on multi-scale Retinex fused with wavelet transform, through which a clearer road crack picture can be obtained without the appearance of "halo" phenomenon.
[0007] A road crack detection method based on multi-scale Retinex fused with wavelet transform, comprising:
[0008] An original road crack image is obtained, and the road crack image is divided into a first low-frequency image component and a first high-frequency image component by using a wavelet transform method. The crack overview image information in the first low-frequency image component is extracted by a top-hat transform method. The crack image information in the first high-frequency image component is enhanced by a multi-scale Retinex method to obtain enhanced crack image information. The entropy of the crack overview image information and the entropy of the enhanced crack image information are respectively calculated by an image entropy method. The proportion of the entropy of the crack overview image information in the total entropy is multiplied by the first low-frequency image component to obtain a second low-frequency image component. The proportion of the entropy of the crack image information in the total entropy is multiplied by the first high-frequency image component to obtain a second high-frequency image component. The second low-frequency image component and the second high-frequency image component construct a road crack image component. The road crack image component is reconstructed by wavelet transform to obtain a road crack enhanced image.
[0009] The top hat transform method is used to extract the crack overview image information in the first low-frequency image component. bh for:
[0010]
[0011] Where I is the original road crack image, D d is a disk-shaped structure operator of scale d, L d , θ is a linear structure operator with scale d and direction θ, d is the scale and θ is the angle.
[0012] set up d max is 13 pixels, θ∈[0,180°), set θ sep is the step size selected for the angle θ, that is, every θ sep Select a value of θ. For the disk-type operator, the pixels larger than the scale d in the original road crack image are retained as crack pixels. For the line-type operator, every θ sep A value of θ is selected, and pixels along the direction of the value of θ and with a scale greater than d are retained as crack pixels, and the remaining pixels are removed as noise points.
[0013] The multi-scale Retinex method is used to enhance the crack image information in the first high-frequency image component to obtain enhanced crack image information I MSR for:
[0014]
[0015] Among them, I is the original road crack image, G k is the Gaussian function of the kth domain scale, and n is the domain size type.
[0016] The pixel point at the kth domain scale is the (x, y) Gaussian function Gk (x, y) is:
[0017]
[0018] Among them, σ k is the Gaussian kernel of the kth domain scale.
[0019] The road crack image component I' is:
[0020]
[0021] Among them, I bh is the crack profile image information, I MSR To enhance the crack image information, H1 is the entropy of the crack overview image information, H2 is the entropy of the enhanced crack image information, and H1+H2 is the total entropy.
[0022] The entropy H1 of the crack profile image information is:
[0023]
[0024] Among them, M1 and N1 are the width and height of the crack overview image, respectively, and p(x1,y1) is the pixel value of the pixel point (x1,y1).
[0025] The entropy H2 of the enhanced crack image information is:
[0026]
[0027] Among them, M2 and N2 are the width and height of the enhanced crack image, respectively, and p(x2, y2) is the pixel value of the pixel point (x2, y2).
[0028] Compared with the prior art, the present invention has the following beneficial effects:
[0029] Firstly, the original image is divided into high-frequency component image and low-frequency component image through wavelet transform. The high-frequency component is processed by multi-scale Retinex algorithm to enhance the crack image, highlighting the image texture detail information and making up for the loss of useful information due to too many decomposition layers of wavelet transform. The low-frequency component image is enhanced by bowler-hat transform, and the image components are obtained based on the image entropy calculation method. Then, the image components are reconstructed by wavelet transform reconstruction technology to avoid the occurrence of halo phenomenon and obtain a road crack enhanced image with higher definition. BRIEF DESCRIPTION OF THE DRAWINGS
[0030] Figure 1 A block diagram of a road crack detection method based on multi-scale Retinex fused with wavelet transform provided for a specific implementation method;
[0031] Figure 2 The original image, the image processed by the single-scale Retinex algorithm, the image processed by the multi-scale Retinex algorithm, and the image processed by the multi-scale Retinex algorithm based on fusion wavelet transform provided for the specific implementation method;
[0032] Figure 3 The brightness, contrast and information entropy curves of the original image, the image processed by the single-scale Retinex algorithm, the image processed by the multi-scale Retinex algorithm and the image processed by the multi-scale Retinex algorithm based on the fusion wavelet transform provided in the specific implementation method are shown in FIG. Figure 3 (a) is the brightness contrast curve. Figure 3 (b) is the contrast curve. Figure 3 (c) is the information entropy curve;
[0033] Figure 4 The original image provided in the specific implementation manner, the crack image obtained by processing using the multi-scale Retinex algorithm and the multi-scale Retinex algorithm based on fusion wavelet transform, wherein: Figure 4 (a) is the original image, Figure 4 (b) is the crack image obtained by using the multi-scale Retinex algorithm. Figure 4 (c) is the crack image obtained by processing the multi-scale Retinex algorithm based on fusion wavelet transform. DETAILED DESCRIPTION
[0034] The present invention provides a road crack detection method based on multi-scale Retinex fused with wavelet transform, such as Figure 1 As shown, the specific steps are as follows:
[0035] (1) An original road crack image is provided, and a wavelet transform method is used to divide the original road crack image into a low-frequency image component, i.e., a low-frequency overview, and multiple high-frequency image components, i.e., multiple high-frequency details.
[0036] (2) Using bowler-hat transform to extract crack overview image information in low-frequency image components. Since the bowler-hat transform algorithm has a short average response time and high segmentation accuracy, this application uses bowler-hat transform to extract rough crack overview information. It uses disk-shaped structural elements of different diameter scales and linear structural elements of different lengths and directions to perform opening operations on the low-frequency image components after wavelet decomposition, and superimposes the maximum difference between the calculation results of the two groups of structural elements at different scales as the extracted crack overview image information. Using the top hat transform method to extract the crack overview image information in the first low-frequency image component bh for:
[0037]
[0038] Where I is the original road crack image, D d is a disk-shaped structural element with a scale of d, L d , θ is a linear structure element with a scale of d and a direction of θ, where d is the scale and θ is the angle. d max is 13 pixels, θ∈[0,180°), set θ sep is the step size selected for the angle θ, that is, every θ sep Select a value of θ. For the disk-shaped structural element, the pixels larger than the scale d in the original road crack image are retained as crack pixels. For the linear structural element, every θ sep A value of θ is selected, and pixels along the direction of the value of θ and with a scale greater than d are retained as crack pixels, and the remaining pixels are removed as noise points. The maximum difference between the calculation results of the disk-shaped structure element and the linear structure element at different scales is superimposed as the extracted crack overview image information.
[0039] (3) Multi-scale Retinex is used to enhance the detail information of the crack image in the high-frequency image component to obtain enhanced crack image information. In order to improve the image enhancement effect, multi-scale Retinex expands the single-scale Retinex Gaussian function G(x, y) with a fixed scale σ into k Gaussian functions G with different scales. k (x, y). Generally, the value of k is 3, which means that three different neighborhood scales of small, medium and large Gaussian functions G1, G2, G3 are used, and the neighborhood scale is described by Gaussian kernels σ1, σ2, σ3. Multiscale Retinex uses the image I(x, y) and the scale σ k The Gaussian function G k The kth incident image component L is estimated by phase convolution k (x,y). k The components are expressed as:
[0040] L k (x,y)=I(x,y)*G k (x,y)
[0041]
[0042] According to the L k The component calculation can obtain the expression of R(x,y) in the logarithmic domain, which is expressed as:
[0043]
[0044]
[0045] where wi is the weight coefficient, which is 1 / 3. The image is restored from the logarithmic domain to the spatial domain to calculate I MSR , we can get the enhanced crack image information after multi-scale Retinex algorithm enhancement, expressed as:
[0046]
[0047] Among them, I is the original road crack image, G k is the Gaussian function of the kth domain scale, n is the domain size type, n=3.
[0048] (3) Image entropy is an important indicator to measure the amount of useful information contained in an image. In order to ensure the accuracy of the image after wavelet reconstruction, the image entropy is used to measure the energy of the crack information in the image enhanced by the bowler-hat transform and multi-scale Retinex algorithms. The calculation formula of image entropy is expressed as:
[0049]
[0050] Where p(x,y) is the pixel value of the pixel point (x,y), M and N are the width and height of the image respectively. bh is the image after bowler-hat transformation, and its entropy is H1, I MSR is the image enhanced by multi-scale Retinex, and its entropy is H2. The image components that need to be reconstructed are expressed as:
[0051]
[0052]
[0053]
[0054] Among them, I bh is the crack profile image information, I MSR To enhance the crack image information, H1 is the entropy of the crack overview image information, H2 is the entropy of the enhanced crack image information, H1+H2 is the total entropy, M1 and N1 are the width and height of the crack overview image, and p(x1,y1) is the pixel value of the pixel point (x1,y1). M2 and N2 are the width and height of the enhanced crack image, and p(x2,y2) is the pixel value of the pixel point (x2,y2). The image component I' can be reconstructed by inverse wavelet transform to obtain the enhanced crack image.
[0055] Since the Retinex algorithm is prone to produce halos during processing, in order to avoid image distortion caused by the halo phenomenon and obtain better crack enhancement effects, wavelet transform can be introduced for image processing. Wavelet transform decomposes the image into sub-bands of different frequency bands. According to the scale space correlation theory, noise cannot be transmitted in the scale space. As the number of image decomposition layers increases, it can be considered that the sub-band image does not contain noise, but the cost is the loss of useful feature components in the image. The multi-scale Retinex algorithm can highlight the edge, texture and other details of the image, and can make up for the deficiency of wavelet transform in losing useful information due to too many decomposition layers. Therefore, this paper adopts a multi-scale Retinex algorithm fused with wavelet transform for crack image enhancement.
[0056] Performance Analysis:
[0057] In order to fully illustrate the advantages of this algorithm, Figure 2 As shown, Figure 2 (a) is the original image, Figure 2 (b) is a picture processed by the single-scale Retinex algorithm. Figure 2 (c) in the figure is the image processed by the multi-scale Retinex algorithm and Figure 2 (d) is the image processed by the algorithm of this application. The experimental results show that compared with the traditional algorithm, the improved algorithm proposed in this application can better highlight the details of the cracks, improve the overall visual effect of the image to a certain extent, and reduce the "halo artifact" phenomenon to a certain extent. The processed image is clearer and the image quality has been greatly improved, which is conducive to the subsequent feature recognition of cracks.
[0058] At the same time, in order to quantitatively compare the crack image enhancement performance of each algorithm, experimental tests were carried out on five image data sets including concrete road cracks, asphalt road cracks, rubber runway cracks, bridge cracks, and tunnel cracks. The comparisons were made from the brightness mean (reflecting the overall light and dark effect of the image), contrast (reflecting the richness of the image grayscale levels), information entropy (reflecting the amount of image information), and time (the time consumed by the image processing process).
[0059] Brightness mean algorithm:
[0060]
[0061] Contrast algorithm:
[0062]
[0063] Where g(x,y) is the gradient value of the pixel (x,y), and M and N are the width and height of the image respectively.
[0064] like Figure 3 (a)- Figure 3 As shown in (c), through the comparison of brightness mean, contrast and information entropy, it can be seen that the image enhancement method proposed in this application can perform good preprocessing on the collected road surface damage images. The enhancement algorithm effectively improves the contrast distinction between the image "highlight" area (non-crack area) and the "dark" area (crack line), highlights the crack feature information in the road image, and achieves the purpose of facilitating crack line extraction; at the same time, the introduction of the bowler-hat transform algorithm makes the average response time shorter and improves the segmentation accuracy.
[0065] The image processing time is shown in Table 1. By comparison, the algorithm in this paper greatly reduces the computational complexity, thereby reducing the processing time and making it easier to apply to real-time image acquisition and processing.
[0066] Table 1: Image processing time comparison table
[0067]
[0068] Compared with the method provided in this application, the performance of the traditional method is significantly worse, not only reducing the amount of information, but also consuming more time. The enhancement method provided in this application effectively improves the contrast distinction between the non-crack area and the crack line of the road damage image, highlights the crack feature information in the image, and achieves the purpose of facilitating crack line extraction. In a comprehensive comparison, the algorithm proposed in this paper still has obvious advantages in crack image enhancement performance.
[0069] The traditional multi-scale Retinex algorithm and the multi-scale Retinex fusion wavelet transform algorithm provided by this application are used to enhance the crack dataset image and perform crack feature recognition to verify the crack image enhancement performance of the algorithm and its effect on improving the crack recognition accuracy. The crack detection results are shown in Figure 2. Figure 4 (a)-(c) show that the crack recognition performance based on the above two algorithms is shown in Table 2.
[0070] Table 2: Crack detection system identification performance
[0071]
[0072] Depend on Figure 4 From the results of (a)-(c) and Table 2, it can be seen that the comprehensive performance of the method provided by the present application is higher than that of the traditional multi-scale Retinex algorithm, indicating that the image enhancement processing method provided by the present application can perform a good preprocessing on the collected road surface damage images, and can achieve a satisfactory effect of image enhancement processing. It also reduces or eliminates the shortcomings of its own algorithm to a certain extent, improves the clarity of the image, and is conducive to the subsequent crack feature extraction and identification.
[0073] The road crack detection method based on multi-scale Retinex fused with wavelet transform provided by this application first introduces the Retinex algorithm for crack enhancement. Based on the advantage that this method can better handle the problem of uneven image illumination, it well solves the inherent defects of the original road damage image illumination in image enhancement. At the same time, the wavelet transform is introduced on the basis of the original multi-scale Retinex to eliminate the halo artifact effect produced by the traditional algorithm, thereby reducing the image distortion. This makes up for the shortcomings of the original algorithm.
[0074] This application applies the above method to the preprocessing step of crack detection. The experimental results verify the effectiveness and feasibility of the algorithm given in this article. Based on the traditional multi-scale Retinex enhancement algorithm, the recognition accuracy of the crack system is only 75.1%, while based on the improved enhancement algorithm proposed in this application, the accuracy is improved to 95.8%, and the performance far exceeds the traditional method. In addition, the calculation time for each picture is relatively short. The experimental results fully demonstrate that the road damage image is more conducive to the extraction of crack lines after being enhanced by the new algorithm. Finally, it should be pointed out that improving the time complexity of the improved algorithm proposed in this study has obvious significance for the real-time requirements of the algorithm application in this article. In summary, this application provides an efficient solution for the preprocessing link of road damage images for automatic road crack recognition. At the same time, this method can also be applied to other target detection and recognition systems such as wall cracks and steel plate cracks, and has a wide range of application value.
Claims
1. A road crack detection method based on multi-scale Retinex fused with wavelet transform, characterized in that: include: An original road crack image is obtained, and the road crack image is divided into a first low-frequency image component and a first high-frequency image component by using a wavelet transform method, and the crack overview image information in the first low-frequency image component is extracted by using a top-hat transform method, and the crack image information in the first high-frequency image component is enhanced by using a multi-scale Retinex method to obtain enhanced crack image information, and the entropy of the crack overview image information and the entropy of the enhanced crack image information are respectively calculated by using an image entropy method, and the proportion of the entropy of the crack overview image information in the total entropy is multiplied by the first low-frequency image component to obtain a second low-frequency image component, and the proportion of the entropy of the crack image information in the total entropy is multiplied by the first high-frequency image component to obtain a second high-frequency image component, and the second low-frequency image component and the second high-frequency image component construct a road crack image component, and the road crack image component is reconstructed by wavelet transform to obtain a road crack enhanced image; The top hat transform method is used to extract the crack overview image information in the first low-frequency image component. bh for: Among them, I is the original road crack image, D d is a disk-shaped structural element with a scale of d, L d, θ is a linear structure element with a scale of d and a direction of θ, where d is the scale and θ is the angle; set up d max is 13 pixels, θ∈[0,180°), set θ sep is the step size selected for the angle θ, that is, every θ sep Select a value of θ. For the disk-shaped structural element, the pixels larger than the scale d in the original road crack image are retained as crack pixels. For the linear structural element, every θ sep A value of θ is selected, and pixels along the direction of the value of θ and with a scale greater than d are retained as crack pixels, and the remaining pixels are removed as noise points. The maximum difference between the calculation results of the disk-shaped structure element and the linear structure element at different scales is superimposed as the extracted crack overview image information; The multi-scale Retinex method is used to enhance the crack image information in the first high-frequency image component to obtain enhanced crack image information I MSR for: Among them, I is the original road crack image, G k is the Gaussian function of the kth domain scale, and n is the domain size type; The entropy H1 of the crack profile image information is: Among them, M1 and N1 are the width and height of the crack overview image, respectively, and p(x1,y1) is the pixel value of the pixel point (x1,y1); The entropy H2 of the enhanced crack image information is: Among them, M2 and N2 are the width and height of the enhanced crack image, respectively, and p(x2, y2) is the pixel value of the pixel point (x2, y2).
2. The road crack detection method based on multi-scale Retinex fused with wavelet transform according to claim 1 is characterized in that: The pixel point at the kth domain scale is the (x, y) Gaussian function G k (x, y) is: Among them, σ k is the Gaussian kernel of the kth domain scale.
3. The road crack detection method based on multi-scale Retinex fused with wavelet transform according to claim 1 is characterized in that: The road crack image component I' is: Among them, I bh is the crack profile image information, I MSR To enhance the crack image information, H1 is the entropy of the crack overview image information, H2 is the entropy of the enhanced crack image information, and H1+H2 is the total entropy.
