Tunnel crack image identification method and tunnel wall uniform image detection system

Abstract

The invention relates to the technical field of tunnel crack detection, and discloses a tunnel crack image identification method and a tunnel wall uniform image detection system. The method comprises the steps that tunnel images are collected, images belonging to cracks are calibrated to form an image set, and a crack target sample library is constructed; performing target detection identification on the collected tunnel image set, optimizing a target identification method according to an identification result, and outputting multiple tunnel crack disease detection images; extracting crack features from the collected tunnel image set; and based on the multiple tunnel crack disease detection images, performing pixel-level segmentation on the tunnel crack disease to complete tunnel crack identification. According to the invention, the traditional layout is broken, the requirements of closer shooting object distance and depth of field are met, and the detection requirement of smaller width of the welding seam is met; according to the algorithm level, a front image preprocessing part of deep learning training is optimized, especially contrast enhancement, multi-level fusion and the like are added, and the recognition accuracy of a tiny target can be further improved.

Description

Technical Field

[0001] The present invention relates to the technical field of tunnel crack detection, and in particular, to a method for identifying tunnel crack images and a tunnel wall uniform image detection system. Background Art

[0002] All kinds of diseases will occur during the construction and operation stages of tunnels, such as cracks, surface peeling, offset and unevenness of joints, water seepage and leakage, etc. Once cracks occur, they will also be affected by the change of air pressure during train operation, causing diseases such as cracks to continuously expand. These tunnel diseases always threaten the safety of subway tunnels. If the diseases are not prevented and treated in time, it will seriously affect the normal operation of the subway and bring huge losses to people's lives and property. As an intuitive manifestation of tunnel diseases, the diseases on the lining surface are one of the most direct and important indicators for monitoring the tunnel state.

[0003] At present, tunnel detection adopts two methods: independent robots or on-vehicle detection. The main technical principles used include: laser three-dimensional scanning technology, two-dimensional linear array camera scanning detection technology, three-dimensional linear array laser scanning technology, etc. In terms of layout, due to the continuous improvement of the detection accuracy index requirements, the existing independent on-vehicle layout cannot well adapt to the influence of depth of field, speed, etc. At the same time, to a certain extent, it is required that the appearance of the vehicle body is simple and cannot be installed in parts such as the front of the vehicle. Therefore, innovative design considers arranging cameras around the vehicle compartment or awning position in terms of layout to reduce the depth of field and improve the high-speed adaptability and defect accuracy. Summary of the Invention

[0004] In view of this, the present invention provides a method for identifying tunnel crack images and a tunnel wall uniform image detection system to solve the above problems.

[0005] To solve the above technical problems, the present invention provides a method for identifying tunnel crack images, including: Collect tunnel images, calibrate the images belonging to cracks to form an image set, and construct a crack target sample library; Based on the crack target sample library, perform target detection and recognition on the collected tunnel image set, and optimize the target recognition method according to the recognition results, and output multiple tunnel crack disease detection images; Perform crack enhancement processing on the collected tunnel image set, and extract crack features; based on the above multiple tunnel crack disease detection images, perform pixel-level segmentation on the tunnel crack diseases to complete tunnel crack recognition.

[0006] As an optional method, when performing target detection and recognition on the tunnel image set, it includes: Convert the tunnel image detection result to the frequency domain, where cracks are represented as high-frequency parts and the background is represented as low-frequency parts; Perform erosion operation on the grayscale image to expand the crack area, making it connected and more obvious; Perform Gaussian low-pass filtering on the original image to remove high-frequency components and retain only the low-frequency background part; Subtract the morphologically enhanced image from the background image after low-pass filtering to generate an image with more prominent high-frequency crack features.

[0007] As an optional method, before extracting crack features, it also includes segmenting the image to extract the morphological features and gradient features of the cracks; the segmentation includes: Perform image threshold segmentation and edge threshold segmentation on the enhanced image; The image threshold segmentation includes: It is best to supplement the processes of the above two threshold segmentations; The edge threshold segmentation includes.

[0008] As an optional method, extracting crack features includes: Perform multi-level fusion filtering on the segmented image to remove non-crack features similar to cracks in the image. The multi-level fusion filtering includes: Establish connected regions to identify sets of pixel points with the same pixel value and connected to each other in the image; According to the characteristics of the connected regions, filter out small region noises with an area and / or number of pixel points less than a preset value, and retain the long-strip and large-range crack feature regions; Mark the positions of the connected regions suspected of being cracks in the threshold segmentation map, match the positions of the connected regions with the feature map generated by the edge threshold algorithm, retain the regions with matching sizes, and remove the noise edges with mismatched positions or large size differences; Extract the regions suspected of being cracks after multi-level filtering processing from the original image to complete the target extraction.

[0009] On the other hand, the present invention also provides a tunnel wall uniform image detection system, which uses the above-mentioned method for identifying tunnel crack images to identify cracks, including: An image acquisition device, a synchronous control device, and a subordinate industrial control computer; The image acquisition device is composed of a light source module formed by a plurality of line array cameras and a preset strong light source; The synchronous control device includes a main control machine and a frequency division device. The main control machine is used to set frequency division parameters and send acquisition instructions to each subordinate industrial control computer for controlling cameras; the frequency division device is used for transcoding and controlling the image acquisition device to acquire images; The subordinate industrial control computer includes an image acquisition card and a large-capacity storage device.

[0010] As an alternative, the image acquisition device at least includes 8 linear array cameras and a light source module formed by encapsulating a corresponding number of preset strong light sources; both have dust and waterproof settings, and only a data interface and an integrated power supply and control interface are reserved for external connection for each interface.

[0011] As an alternative, after receiving the axle encoder signal of the train, the frequency conversion device converts it into a camera acquisition signal and a laser stroboscopic signal with a preset frequency to control the image acquisition device to perform acquisition.

[0012] The beneficial effects of the present invention are as follows: The present invention breaks the traditional layout, adapts to closer shooting object distances and depth of field requirements, and at the same time meets the detection requirements for smaller weld widths; at the algorithm level, the pre-image preprocessing part of deep learning training is optimized, especially the addition of contrast enhancement, multi-level fusion, etc., which can further improve the recognition accuracy of tiny targets. Description of the Drawings

[0013] Figure 1 It is a flow chart of the method for identifying tunnel crack images provided by the embodiment of the present invention; Figure 2 It is a schematic diagram of the tunnel crack image recognition process provided by the embodiment of the present invention; Figure 3 It is a schematic diagram of the structure of the tunnel wall uniform image detection system provided by the embodiment of the present invention; Figure 4 It is a flow chart of the target recognition of the tunnel image set provided by the embodiment of the present invention. Detailed Embodiments

[0014] In order to enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below in conjunction with the specific embodiments.

[0015] The light in the subway tunnel is relatively dim, the environment is complex, and there is a large amount of interference noise. In the actual process of image acquisition using a line scan camera with a high line frequency and resolution, it is inevitable to add interference noise and subtle deformation. The width of the crack itself is relatively small, accounting for a low pixel width in the field of view. All these will make it difficult to detect tunnel diseases such as cracks.

[0016] In this embodiment, deep learning and basic image processing are fused at the algorithm level to finally achieve the target detection of the tunnel surface, and at the same time, the crack diseases are segmented with pixel-level accuracy.

[0017] Please refer to Figure 1 And Figure 2 , this embodiment provides a method for identifying tunnel crack images, including: Collect tunnel images, calibrate the images belonging to cracks to form an image set, and construct a crack target sample library; Based on the crack target sample library, perform target detection and recognition on the collected tunnel image set, and optimize the target recognition method according to the recognition results, and output multiple tunnel crack disease detection images; Perform crack enhancement processing on the collected tunnel image set to extract crack features; based on the above-mentioned multiple tunnel crack disease detection images, perform pixel-level segmentation on the tunnel crack disease to complete tunnel crack recognition. Specifically, the image threshold segmentation in this embodiment adopts the following method: First, set a sliding window of a certain size, place the sliding window in the middle of the entire image, that is, the area with the best image acquisition quality, and use the mean and standard deviation of this area as the reference values for the gray-scale transformation of the entire image; Secondly, sweep the sliding window across the entire image, calculate the mean and standard deviation of each sliding window, and then compare them with the reference values to divide into different brightness regions. Design different binary algorithms for different brightness regions, and finally complete the binaryzation of the entire image.

[0018] The core of edge threshold segmentation adopts the following method: The core idea is to use edge threshold segmentation to handle application scenarios such as tunnel cracks that are extremely small and difficult to segment and are easily hidden in large-area stains. For cracks, the edge is an extremely important structural feature. The edge appears as those pixels with abrupt gray-scale changes in the digital image. There is always one side of the background gray-scale of the edge that is either much brighter than the gray-scale of the line pixels or much darker than the gray-scale of the line pixels. Therefore, the edge can be judged by the change of the image gradient. At the same time, since the development direction of cracks is usually irregular.

[0019] When performing target detection and recognition on the tunnel image set, considering that in the tunnel wall crack detection results, the proportion of pixel points occupied by cracks is relatively low and the width is small, usually only 4-6 pixel points, but the cracks have obvious continuous strip features. When the result is converted to the frequency domain, the cracks appear as the darker high-frequency part in the image, and the background appears as the low-frequency part. The specific method of image enhancement is mainly to perform erosion operation on the gray-scale image, expand the area of the cracks morphologically, so that the cracks are connected and become more obvious. Then perform Gaussian low-pass filtering operation with limited frequency on the original image, filter out the high-frequency components in the original image, and only retain the low-frequency background part. Finally, subtract the two result images, and thus a high-frequency image with more obvious high-frequency features such as cracks can be obtained. The specific framework is as Figure 4 shown.

[0020] In this embodiment, before extracting the crack features, it also includes segmenting the image to extract the morphological features and gradient features of the cracks; the segmentation includes: Perform image threshold segmentation and edge threshold segmentation on the enhanced image; The image threshold segmentation includes: First, set a sliding window of a certain size, place the sliding window in the middle of the entire image, that is, the area with the best image acquisition quality, and use the mean and standard deviation of this area as the reference values for the gray-scale transformation of the entire image; Second, sweep the sliding window across the entire image, calculate the mean and standard deviation of each sliding window, and then compare them with the reference values to divide them into different brightness regions. Design different binary algorithms for different brightness regions, and finally complete the binarization of the entire image; The edge threshold segmentation judges the edge through the change of the image gradient. The core idea of the edge threshold segmentation is to use the edge threshold segmentation to process application scenarios such as tunnel cracks that are extremely small and difficult to segment and are easily hidden in large-area stains.

[0021] For cracks, the edge is an extremely important structural feature. The edge appears as those pixels with sudden gray-scale changes in the digital image. The background gray scale of the edge always has one side, either much brighter than the gray scale of the line pixels or much darker than the gray scale of the line pixels. Therefore, the edge can be judged by the change of the image gradient. Extracting crack features includes: Perform multi-level fusion filtering on the segmented image to remove non-crack features similar to cracks in the image. The multi-level fusion filtering includes: Use connected component filtering or the method of constructing a custom template to filter out noise. On this basis, by analyzing the morphological features of cracks, a multi-level fusion filtering algorithm for crack feature maps based on connected components and edge thresholds is designed to gradually filter out non-crack components in the image, and finally extract the suspected crack regions from the original image.

[0022] Specifically, first perform connected component filtering: Establish connected components, that is, a set of pixel points in the image that have the same pixel value and are connected to each other. Most of the noise points in the image are small-area connected components, and the number of pixel points contained in the region contour is relatively small. The crack features usually appear as long strips and large-scale connected component regions. Therefore, find the contours of each connected component in the binary image, and then filter out the vast majority of scattered point noises based on the characteristics of the connected components.

[0023]

[0024] Where N is the number of white pixel points in each connected component. Represents each connected component, and the range of n is the maximum number of connected components in the image. Thresh is the threshold required for determination. The connected components with N less than the threshold tresh are determined as scattered point noises and need to be filtered out. Taking cracks as the main, usually a minimum width of two pixels can ensure detectability. Thus, calculate the number of pixels and deduce the length to judge whether it is a real crack.

[0025] Subsequently, fusion filtering is performed: First, mark the connected region positions of each suspected crack in the threshold segmentation map, and then compare them with the corresponding positions in the feature map based on the edge threshold algorithm. If the size at this position matches in the edge threshold image, the edge is the crack itself and should be retained; if the size difference is too large or the position does not match, the edge is a noise edge and should be filtered out.

[0026]

[0027] In the formula represents the suspected crack region in the threshold image, represents the suspected crack region in the crack feature map based on the edge threshold algorithm. The maximum range of n is the number of suspected crack regions in the threshold image. t 3 represents the minimum area of the suspected crack region. Regions smaller than this threshold are not considered suspected crack regions. r 3 is a calculated value. The calculation result of r 3 is used as the judgment condition for the value of When > t 3 this calculation is performed, and then the value of 3 is determined according to the size relationship between r and 0.

[0028] In this way, the core content of this embodiment in the entire algorithm framework is to improve the overall contrast of the image and locally enhance the brightness of the crack region. Generally, there are two ways to improve the image quality (enhancement or restoration). Using image enhancement technology can directly use the algorithm to improve the overall and local contrast of pixels without considering the reasons for the decline in image quality. The specific approach is to first improve the overall image quality (image enhancement) using a regional processing method. The original image is successively sharpened and filtered to obtain a preprocessed image, and then after the crack data is recognized by the deep learning network localization model, local point-by-point pixel processing is performed. The method used is the histogram equalization operation in the histogram correction technology. This operation can have a certain enhancement effect on images in complex changing scenes and can effectively suppress noise.

[0029] Compared with the traditional method, this embodiment takes into account the overall image processing to improve the accuracy of crack detection rather than simply intercepting a partial crack area for experiments, aiming to meet the needs of engineering practice. This algorithm optimizes the images collected by the line array camera, reducing the influence of illumination and the structural characteristics of the acquisition device. The algorithm has strong adaptability to the subway tunnel with a complex environment, can exclude most of the noise interference, and has good filtering ability for scatter noise, pipelines, large stains, and reticular noise. However, it cannot filter out the scratches well, which are extremely similar to cracks in morphology.

[0030] On the other hand, please refer to Figure 3 , the vehicle-mounted tunnel surface image acquisition system mainly uses an industrial high-resolution line array camera as the core of the image acquisition system. Considering that the light in the tunnel is dim and unevenly distributed, a laser light source is used for supplementary lighting. The laser light source and the line array camera need to be aligned within the same center line. For convenient use and to ensure stability, they are encapsulated into an integrated image acquisition device. This embodiment also provides a tunnel wall uniform image detection system, which uses the above-mentioned method for identifying tunnel crack images to identify cracks, including: Image acquisition device: It is composed of a module encapsulated by a group (8 pieces) of line array cameras and a special strong light source. This module ensures the strict matching of the camera and the laser, and has a certain dust and waterproof function. At the same time, each interface is unified and only the data interface and the integrated power supply and control interface are reserved externally; Synchronous control device: It is jointly composed of a main control machine and a frequency conversion device. The main industrial control machine sets the frequency conversion parameters and sends acquisition instructions to the subordinate industrial control machines that control the cameras. The frequency conversion device receives the wheel axle encoder signal, and then converts it into a camera acquisition signal and a laser stroboscopic signal with a preset frequency to control the image acquisition device to perform acquisition; The subordinate industrial control machine includes an image acquisition card and a large-capacity storage device. Since the data volume of line scanning imaging is large. Therefore, higher requirements are put forward for the disk capacity, etc.

[0031] In an optional scenario, since the width requirement for tunnel crack detection is 0.5 m, the more pixel points the crack occupies, the easier it is to detect. Therefore, the acquisition accuracy is set to about 0.2 mm, so that the crack width has 2 - 4 pixels, which is convenient for algorithm design. Based on this, the system uses 8 line-scanning 2D cameras to complete the full-range surface scan of the tunnel around the carriage awning for one week. Among them, there are 2 cameras on the roof, and 3 cameras on the left and right sides of the vehicle body respectively. Among them, the height from the roof to the ground covered by the three cameras on one side is about 3.34 m.

[0032] In this way, this embodiment breaks the traditional layout, adapts to the requirements of a closer shooting object distance and depth of field, and at the same time meets the detection requirements for a smaller weld width; at the algorithm level, the pre-image preprocessing part of deep learning training is optimized, especially the addition of contrast enhancement, multi-level fusion, etc., which can further improve the recognition accuracy of tiny targets.

[0033] The above are only the preferred embodiments of the present invention. It should be noted that the above preferred embodiments should not be regarded as limiting the present invention, and the protection scope of the present invention should be subject to the scope defined by the claims. For those of ordinary skill in the art, without departing from the spirit and scope of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as within the protection scope of the present invention.

Claims

1. A method for identifying a tunnel crack image, characterized in that: include: Collect tunnel images, calibrate images belonging to cracks to form an image set, and build a crack target sample library; Based on the crack target sample library, target detection and recognition are performed on the collected tunnel image set, and according to the recognition result, the target recognition method is optimized to output a variety of tunnel crack disease detection images; Perform crack enhancement processing on the collected tunnel image set to extract crack features; Based on the above-mentioned multiple tunnel crack disease detection images, the tunnel crack diseases are segmented at the pixel level to complete the tunnel crack identification.

2. The method for recognizing a tunnel crack image according to claim 1, characterized in that: When performing target detection and recognition on the tunnel image set, it includes: The tunnel image detection results are converted into the frequency domain, where cracks appear as high-frequency parts and the background appears as low-frequency parts; Perform an erosion operation on the grayscale image to expand the crack area, making it connected and more obvious; Perform a Gaussian low-pass filter on the original image to remove high-frequency components and retain only the low-frequency background part; The morphological enhancement image is subtracted from the background image after low-pass filtering to generate an image with more prominent high-frequency features of cracks.

3. The method for recognizing a tunnel crack image according to claim 1, characterized in that: Before extracting crack features, the image is segmented to extract the morphological features and gradient features of the cracks; The segmentation includes: Perform image threshold segmentation and edge threshold segmentation on the enhanced image; The image threshold segmentation includes: setting a sliding window of a preset size, placing the sliding window in the middle of the entire image, and using the mean and standard deviation of the area as reference values ​​for grayscale transformation of the entire image; The sliding window is swept across the entire image, the mean and standard deviation of each sliding window are calculated, and then compared with the reference value to divide it into different brightness areas; Different binarization algorithms are designed for areas with different brightness, and finally the entire image is binarized; The edge threshold segmentation determines the edge through the change of image gradient.

4. The method for recognizing a tunnel crack image according to claim 3, characterized in that: The extracting of crack features comprises: The segmented image is subjected to multi-level fusion filtering to remove non-crack features similar to cracks in the image, wherein the multi-level fusion filtering includes: Establish a connected domain to identify the set of pixels with the same pixel value and connected to each other in the image; According to the characteristics of the connected domain, small area noise with an area and / or pixel count smaller than a preset value is filtered out, and long strips and large-scale crack feature areas are retained; Mark the connected domain position of the suspected crack in the threshold segmentation map, match the connected domain position with the feature map generated by the edge threshold algorithm, retain the area with matching size, and remove the noise edge with mismatched position or too large size difference; The suspected crack area is extracted from the original image after multi-level filtering to complete the target extraction.

5. A tunnel wall uniform image detection system, which uses the tunnel crack image recognition method as described in any one of claims 1 to 4 to perform crack recognition, characterized in that: include: Image acquisition device, synchronization control device, slave industrial computer; The image acquisition device includes a plurality of linear array cameras and a light source module encapsulated by a preset strong light source; The synchronous control device includes a main control machine and a sub-frequency conversion device, wherein the main control machine is used to set sub-frequency conversion parameters and send acquisition instructions to each slave industrial computer used to control the camera; the sub-frequency conversion device is used to perform transcoding and control the image acquisition device to collect; The slave industrial computer includes an image acquisition card and a large-capacity storage device.

6. A tunnel wall uniform image detection system according to claim 5, characterized in that: The image acquisition device includes at least 8 linear array cameras and a light source module encapsulated by a corresponding number of preset strong light sources; they are all dustproof and waterproof, and each interface is unified to the outside and only retains a data interface and an integrated power supply and control interface.

7. A tunnel wall uniform image detection system according to claim 5, characterized in that: After receiving the axle encoder signal of the train, the sub-frequency conversion device converts it into a camera acquisition signal and a laser stroboscopic signal of a preset frequency to control the image acquisition device to perform acquisition.

Images