An image recognition algorithm and system suitable for multi-physical field monitoring of cold region engineering

Abstract

The application discloses an image recognition algorithm and system suitable for monitoring of multiple physical fields of cold region projects, and relates to the technical field of cold region project monitoring.The algorithm comprises the following steps: obtaining an original cold region project image of a target region, performing enhancement algorithm processing, and obtaining a target cold region project image; performing edge feature extraction and weighted calculation on the target cold region project image, and performing dynamic template matching by using an SSD algorithm to obtain target pixel displacement; correcting and compensating a device for obtaining the original cold region project image by using a temperature compensation model to form a compensation coefficient, adjusting the target pixel displacement, and obtaining actual physical displacement of the target region; constructing a scene parameter library; performing scene parameter library matching according to the target cold region project image, determining corresponding optimization parameters and a safety threshold according to a matched scene type, and obtaining a final recognition result of the target region based on the corresponding optimization parameters, the safety threshold and the actual physical displacement of the target region.

Description

Technical Field

[0001] This application relates to the field of monitoring technology for cold-region engineering, and more specifically, to an image recognition algorithm and system applicable to multi-physics field monitoring of cold-region engineering. Background Technology

[0002] Infrastructure projects in cold regions (such as the Qinghai-Tibet Railway, Qinghai-Tibet Highway, Sichuan-Tibet Railway, and Xinjiang-Tibet Railway) are constantly affected by factors such as low-temperature freeze-thaw cycles, traffic vibration loads, and snow cover, making them prone to structural cracking, excessive settlement, and sudden drops in stability. In the harsh environment of the uninhabited areas of the Qinghai-Tibet Plateau, characterized by extreme cold, severe oxygen deficiency, and intense radiation, manual operations are unsustainable, and traditional monitoring methods are virtually impossible. Therefore, adopting intelligent monitoring technologies such as image recognition to achieve non-contact automatic identification and all-weather monitoring of roadbed cracks, thermal thaw collapses, and frost heave deformation has become an inevitable choice for overcoming natural limits and ensuring the long-term safety of permafrost projects. However, as the requirements for "accuracy, real-time performance, and scenario adaptability" in cold-region infrastructure construction continue to increase, although the focus has shifted from manual inspection to automated image recognition, the complexity of the special environment in cold regions means that existing technologies still have significant shortcomings in multi-dimensional adaptability and cannot truly meet the stringent requirements of engineering-level monitoring. Summary of the Invention

[0003] The main objective of this application is to provide an image recognition algorithm and system suitable for multi-physics field monitoring in cold region engineering, so as to solve the technical problems in the background art.

[0004] To achieve the above objectives, the first aspect of this application proposes an image recognition algorithm suitable for multi-physics field monitoring in cold-region engineering, comprising: The original cold-region engineering image of the target area is acquired and processed by an enhancement algorithm to obtain the target cold-region engineering image, wherein the target area refers to the selected area in the captured image; Edge features are extracted and weighted for the target cold-region engineering image, and dynamic template matching is performed using the SSD algorithm to obtain the target pixel displacement. The dynamic template means that in cold-region monitoring, the target region in the first frame of the target cold-region engineering image is used as the initial reference template. When the appearance of the target region changes and the matching confidence is lower than a set confidence threshold, the target region matched in the current frame is used as the target reference template for the next moment to form a dynamic matching sample. A temperature compensation model is used to correct and compensate the device that acquires the original cold region engineering image, forming a compensation coefficient. Based on the compensation coefficient, the displacement of the target pixel is adjusted to obtain the actual physical displacement of the target area. A scenario-based parameter library is constructed, which contains optimization parameters and safety thresholds corresponding to different scenario types. The scenario types include at least roadbed scenario type, slope scenario type and glacier scenario type, and each scenario type corresponds to optimization algorithm parameters and safety thresholds. Based on the target cold-region engineering image, the scene-based parameter library is matched, and the corresponding optimization parameters and safety thresholds are determined according to the matched scene type. Based on the corresponding optimization parameters, the safety thresholds, and the actual physical displacement of the target area, the final identification result of the target area is obtained, wherein the final identification result represents the safety assessment status of the target area.

[0005] In some feasible methods, the step of acquiring the original cold-region engineering image of the target area and performing enhancement algorithm processing to obtain the cold-region engineering image includes: The original cold region engineering image of the target area is obtained, and the original cold region engineering image is converted into a grayscale image and normalized to obtain the initial cold region engineering image; The initial cold region engineering image is subjected to contrast-limited adaptive histogram equalization processing, and bilinear interpolation is used to eliminate artifacts to obtain an enhanced intermediate cold region engineering image. Gaussian filtering is applied to the intermediate cold region engineering image to obtain the target cold region engineering image.

[0006] In some feasible methods, the step of performing contrast-limited adaptive histogram equalization on the initial cold-region engineering image and using bilinear interpolation to eliminate artifacts to obtain an enhanced intermediate cold-region engineering image includes: The initial cold region engineering image is subjected to contrast-limited adaptive histogram equalization to obtain the equalized initial cold region engineering image. The equalized initial cold region engineering image is divided into n×n CLAHE sub-blocks, and bilinear interpolation is used to eliminate artifacts between CLAHE sub-blocks to obtain an artifact-free initial cold region engineering image. The initial cold-region engineering image, which has been processed to eliminate artifacts, is then processed using Gamma correction to obtain an enhanced intermediate cold-region engineering image.

[0007] In some feasible methods, the step of performing Gaussian filtering on the intermediate cold-region engineering image to obtain the target cold-region engineering image includes: For the engineering images of the intermediate cold region, the global variance of the enhanced images is calculated to obtain the global variance of the engineering images of the intermediate cold region, wherein the global variance... The calculation formula is: = ; in, H is the grayscale mean, H is the height of the intermediate cold region engineering image, W is the width of the intermediate cold region engineering image, and Gray(i,j) is the grayscale value of the pixel in the i-th row and j-th column of the intermediate cold region engineering image. Based on the global variance of the intermediate cold region engineering image, a filter standard deviation is constructed, wherein the filter standard deviation increases with the increase of the global variance and decreases with the decrease of the global variance; An m×m Gaussian kernel is generated using the filter standard deviation. The Gaussian kernel is then used to perform convolution filtering on the intermediate cold region engineering image, with the boundaries filled using duplicates, to obtain the target cold region engineering image.

[0008] In some feasible methods, the steps of extracting and weighting edge features from the target cold-region engineering image, and using the SSD algorithm for dynamic template matching to obtain the target pixel displacement include: For the current target region in the cold region engineering image of the current frame, the Canny operator is used to extract the edge points of the current target region to form an edge binary map, and the edge points in the edge binary map are assigned a higher weight than the non-edge points to obtain the current target region with enhanced edge matching contribution; The similarity between the initial reference template and the current target region is calculated using the weighted SSD algorithm to obtain the confidence score of the similarity between the initial reference template and the current target region. The SSD calculation formula is as follows: ; Where T is the template image, I is the image window to be matched, Edge(i,j) is the edge binary map, and i and j are pixel coordinates; The confidence level is compared with a preset confidence threshold. When the confidence level is less than the confidence threshold, the current target region is used as a target reference template to form a dynamic matching sample. The target pixel displacement of the current target region is obtained based on the coordinate offset between the target reference template and the initial reference template.

[0009] In some feasible methods, the step of correcting and compensating the device acquiring the original cold-region engineering image using a temperature compensation model to form a compensation coefficient, and adjusting the target pixel displacement based on the compensation coefficient to obtain the actual physical displacement of the target region includes: Using a target of known size with the same material as the target area, a basic ratio calibration is performed to obtain the conversion reference ratio between pixels and actual physical size; The ambient temperature is collected by a temperature sensor attached to the lens of the image acquisition device, and a compensation coefficient is calculated using a temperature compensation model. The displacement of the target pixel is adjusted based on the compensation coefficient to obtain the actual physical displacement of the target area.

[0010] In some feasible approaches, the step of constructing the scenario-based parameter library includes: The monitoring scenarios for cold-region engineering projects are divided into at least three types: roadbed, slope, and glacier. Based on the differences in image texture and features among the different scenario types, the key parameters of the algorithm are optimized. The key parameters of the algorithm include at least the CLAHE sub-block size, Canny threshold, and matching confidence threshold for the corresponding scenario.

[0011] In some feasible methods, the step of performing scene-based parameter library matching based on the target cold-region engineering image, determining the corresponding optimization parameters and safety threshold based on the matched scene type, and obtaining the final identification result of the target region based on the corresponding optimization parameters, the safety threshold, and the actual physical displacement of the target region includes: Based on the characteristics of the target cold-region engineering image, feature matching is performed in the scene-based parameter library to obtain the target scene type; Based on the target scene type, the corresponding optimization algorithm parameters and the safety threshold are retrieved from the scene-specific parameter library to obtain the target optimization algorithm parameters and the target safety threshold for the target scene type. Substitute the actual physical displacement of the target area into the target scene type, and combine it with the corresponding target optimization algorithm parameters for adaptation and correction to obtain the final physical displacement value of the target area after the target scene type adaptation and correction. The final physical displacement value of the target area is compared with the target safety threshold to obtain the safety assessment level; The final physical displacement value of the target area and the safety assessment level constitute the final identification result of the target area.

[0012] Secondly, this application provides an image recognition system suitable for multi-physics field monitoring in cold-region engineering projects, applying the aforementioned image recognition algorithm for multi-physics field monitoring in cold-region engineering projects. The system includes: The cold region image combination enhancement and adaptive Gaussian filtering module is used to acquire the original cold region engineering image of the target area, perform enhancement algorithm processing, and obtain the target cold region engineering image, wherein the target area refers to the area selected in the captured image; An improved SSD target matching module is used to extract edge features and perform weighted calculations on the target cold-region engineering image, and to perform dynamic template matching using the SSD algorithm to obtain the target pixel displacement. The dynamic template means that in cold-region monitoring, the target region in the first frame of the target cold-region engineering image is used as the initial reference template. When the appearance of the target region changes and the matching confidence is lower than a set confidence threshold, the target region matched in the current frame is used as the target reference template for the next moment to form a dynamic matching sample. The temperature compensation conversion module is used to correct and compensate the device that acquires the original cold region engineering image through a temperature compensation model, form a compensation coefficient, and adjust the displacement of the target pixel based on the compensation coefficient to obtain the actual physical displacement of the target area. A scenario-based parameter adaptation module is used to build a scenario-based parameter library. The scenario-based parameter library contains optimization parameters and safety thresholds corresponding to different scenario types. The scenario types include at least roadbed scenario type, slope scenario type and glacier scenario type. Each scenario type corresponds to optimization algorithm parameters and safety thresholds. The result module is used to perform scene-based parameter matching based on the target cold region engineering image, and determine the corresponding optimization parameters and safety thresholds based on the matched scene type. Based on the corresponding optimization parameters, safety thresholds and the actual physical displacement of the target area, the final identification result of the target area is obtained, wherein the final identification result represents the safety assessment status of the target area.

[0013] Thirdly, this application provides a computer program that, when executed by a processor, implements the aforementioned steps of the image recognition algorithm applicable to multi-physics monitoring in cold-region engineering.

[0014] The technical solution provided in this application may include the following beneficial effects: This application presents an image recognition algorithm applicable to multi-physics field monitoring in cold-region engineering projects. Addressing the issue of insufficient adaptability to automated monitoring caused by the unique environment of cold regions, an integrated monitoring solution is formed through a full-process algorithm design, effectively meeting the engineering-level requirements for accuracy, real-time performance, and scene adaptability in cold-region engineering safety monitoring. By enhancing the original image of the target area using an algorithmic enhancement process, image quality is optimized for harsh environments such as snow cover and low light, laying the foundation for subsequent accurate identification. Combining edge feature weighting with the dynamic template matching method of the SSD algorithm, it can adapt to changes in the appearance of the target area due to environmental changes during cold-region monitoring, achieving real-time and accurate extraction of target pixel displacement and avoiding the problem of fixed template matching easily failing in cold regions. Relying on a temperature compensation model to correct and compensate for image acquisition equipment and complete the conversion from pixels to displacement, it eliminates equipment deformation errors caused by low temperatures in cold regions, significantly improving the monitoring accuracy of actual physical displacement. Simultaneously, by constructing a scenario-based parameter library covering typical cold-region engineering scenarios such as roadbeds, slopes, and glaciers, automatic matching and parameter adaptation of monitoring scenarios are achieved, solving the pain point of poor scenario adaptability in existing technologies. Finally, by combining scene optimization parameters, safety thresholds, and actual physical displacement to output the safety assessment results of the target area, a complete monitoring system was formed, from image preprocessing and displacement extraction to scene adaptation and safety assessment. This effectively solved the problem of insufficient multi-dimensional adaptability of automated image recognition monitoring for cold region engineering, and realized engineering-level accurate real-time monitoring of cold region engineering. It also provided reliable technical support for early warning and prevention of disaster risks such as structural cracking and excessive settlement in cold region engineering. Attached Figure Description

[0015] The accompanying drawings, which form part of this application, are used to provide a further understanding of the application and to make other features, objects, and advantages of the application more apparent. The illustrative embodiments and descriptions of this application are used to explain the application and do not constitute an undue limitation of the application. In the drawings: Figure 1 A flowchart of an image recognition algorithm for multi-physics field monitoring in cold region engineering provided in this application. Detailed Implementation

[0016] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0017] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this application described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0018] In this application, the terms "upper," "lower," "left," "right," "front," "rear," "top," "bottom," "inner," "outer," "middle," "vertical," "horizontal," "lateral," and "longitudinal" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are primarily for the purpose of better describing this application and its embodiments, and are not intended to limit the indicated device, element, or component to having a specific orientation, or to be constructed and operated in a specific orientation.

[0019] Furthermore, in addition to indicating location or positional relationship, some of the aforementioned terms may also have other meanings. For example, the term "above" may also be used in some cases to indicate a certain dependency or connection relationship. Those skilled in the art can understand the specific meaning of these terms in this application based on the specific circumstances.

[0020] Furthermore, the terms "installation," "setup," "equipped with," "connection," "linked," and "socketing" should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral structure; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, or an internal connection between two devices, components, or parts. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0021] The following explanations of the terms used in this application are provided to aid in understanding the technical solution of this application: CIE standards, CIE (Commission Internationale de l'Éclairage, International Commission on Illumination) is the most authoritative standardization organization in the fields of optics, lighting, and colorimetry. Its standards are international benchmarks in the fields of color measurement, lighting engineering, and display technology.

[0022] CLAHE (Contrast-Limited Adaptive Histogram Equalization) is an image contrast enhancement algorithm that divides an image into multiple sub-blocks and performs histogram equalization independently on each sub-block while limiting the peak value of the histogram to avoid excessive noise enhancement. Finally, the results of the sub-blocks are interpolated and fused to enhance the image contrast.

[0023] Gamma correction is an image processing method that uses a power function to transform the pixel values ​​of an image to achieve non-linear brightness correction, thereby matching the display brightness with human visual perception.

[0024] SSD (Single Shot MultiBox Detector) is a single-stage object detection algorithm that directly predicts the object class and bounding box on feature maps of different scales, balancing detection speed and accuracy.

[0025] The Canny operator is a multi-stage high-precision edge detection algorithm that achieves accurate, single-pixel-width edge extraction through Gaussian filtering, gradient calculation, non-maximum suppression, and dual threshold detection.

[0026] Replicate padding is a boundary padding method that repeatedly copies the pixel values ​​of the outermost edge of an image outward along the boundary. It is used to expand the image boundary and avoid black borders or invalid areas during filtering / convolution.

[0027] like Figure 1 As shown, in a first aspect, this application provides an image recognition algorithm suitable for multi-physics field monitoring in cold region engineering, comprising: S100: Obtain the original cold-region engineering image of the target area, perform enhancement algorithm processing, and obtain the target cold-region engineering image.

[0028] The target area refers to the selected region in the captured image.

[0029] For example, taking the monitoring of roadbed in cold-region highways in Golmud City, Qinghai Province as an application scenario, an industrial camera with a resolution of 1920×1080 was used to capture images of the roadbed area. A 100mm×100mm concrete marker on the roadbed surface was manually selected as the target area. This area has typical cold-region environmental interferences such as snow and ice coverage (30% of the area), low light intensity (200 lux), and snow and ice particle noise.

[0030] At the monitoring site of the roadbed in cold regions, industrial cameras are used to collect images of the entire roadbed in cold regions. In the original images, the area where the concrete markers are located is selected as the target area (either manually or automatically). The image corresponding to this target area is then extracted as the original cold region engineering image (RGB color image, pixel size 200×200). The original image is then subjected to enhancement algorithms, including grayscale conversion and normalization, contrast-limited adaptive histogram equalization and artifact removal, and Gaussian filtering, to obtain the target cold region engineering image with noise reduction and detail enhancement.

[0031] Specifically, obtaining the engineering map of the target cold region may include the following steps: S101, acquire the original cold region engineering image of the target area, convert the original cold region engineering image into a grayscale image and normalize it to obtain the initial cold region engineering image.

[0032] Specifically, grayscale image conversion refers to converting the original cold-region engineering image (RGB three-channel color image) of the target area into a single-channel grayscale image by fusing the red, green, and blue pixel values ​​of the color image according to the visual perception ratio based on the grayscale conversion principle of the CIE standard. This eliminates the interference of color information on subsequent cold-region image recognition. After obtaining the grayscale image, the initial range of its pixel grayscale values ​​is 0 to 255 (0 is pure black, 255 is pure white). Gray-scale normalization processing involves linearly mapping the pixel gray-scale values ​​of the above gray-scale image, normalizing the original gray-scale range of 0 to 255 to the range of 0 to 1, eliminating the gray-scale scale differences caused by different shooting light intensities, and improving the consistency of subsequent image processing. After the above processing, the initial cold-region engineering image is obtained.

[0033] S102, the initial cold region engineering image is subjected to contrast-limited adaptive histogram equalization processing, and bilinear interpolation is used to eliminate artifacts to obtain an enhanced intermediate cold region engineering image.

[0034] Taking the monitoring of roadbed in cold-region highways in Golmud City, Qinghai Province as an application scenario, the initial cold-region engineering image obtained in S101 (pixel size 200×200 after grayscale normalization, grayscale value range 0~1) is processed. Due to the strong reflection of ice and snow in cold regions, the image has problems such as local overexposure (grayscale value close to 1) and obscuring of details of roadbed concrete markers (grayscale value close to 0.1). The following details each sub-step of S102.

[0035] Specifically, obtaining enhanced engineering images of intermediate cold regions may include the following steps: S1021, Perform contrast-limited adaptive histogram equalization on the initial cold region engineering image to obtain the equalized initial cold region engineering image.

[0036] Specifically, the initial cold-region engineering images are affected by the strong reflection of ice and snow in the cold region, and the local gray values ​​are concentrated in the high range of 0.8 to 1, and the gray values ​​of the details in the target area (concrete marker points) are compressed; conventional histogram equalization will further amplify the overexposed areas, so an adaptive processing method of "block equalization + contrast limitation" is adopted.

[0037] The specific steps are as follows: The initial cold region engineering image was divided into 8×8 CLAHE sub-blocks according to the principle of "non-overlapping and equal size" (the entire 200×200 image was divided into 25×25 sub-blocks of size 8×8 pixels). The gray-level distribution of each sub-block was independently counted to avoid the loss of local features caused by the equalization of the entire image. Set a contrast limit coefficient, for example, clipLimit=0.02 (an empirical value adapted to cold-region snow and ice scenes). The core principle is: for the gray-level histogram of each CLAHE sub-block, crop out the gray-level values ​​that exceed the clipLimit threshold, and evenly distribute the cropped gray-level values ​​to other gray-level ranges to avoid image overexposure caused by a single gray-level value having too high a proportion. For each sub-block that has completed the contrast limit, the pixel gray values ​​are remapped based on the corrected gray histogram, so that the gray values ​​in the sub-block are evenly distributed in the range of 0 to 1, thereby improving the local contrast between the target area and the background.

[0038] The final equalized initial cold-region engineering image was obtained, the overexposure problem in the snow and ice areas was suppressed, the gray value of the edge details of the concrete markers was stretched from 0.1-0.2 to 0.2-0.4, and the local contrast was improved by about 30%.

[0039] S1022, the initial cold region engineering image after equalization is divided into n×n CLAHE sub-blocks, and bilinear interpolation is used to eliminate artifacts between CLAHE sub-blocks to obtain an initial cold region engineering image with artifacts eliminated.

[0040] Specifically, after S1021 block equalization, adjacent CLAHE sub-blocks have different grayscale mapping rules, resulting in abrupt changes in grayscale values ​​at the boundaries (artifacts, manifested as "dividing lines" between sub-blocks), which mask the continuous edge features of the cold region target area. Interpolation is needed to achieve a smooth transition of grayscale values.

[0041] The specific steps are as follows: First, following the block division rule of S1021, the initial cold region engineering image after equalization is divided into 8×8 CLAHE sub-blocks with n=8 (completely consistent with the size and number of sub-blocks in S1021), and the artifacts are clearly concentrated at the horizontal / vertical boundary pixel positions of the sub-blocks (such as the 8th column, 16th column, 8th row, 16th row, etc.). Next, for each boundary pixel, the gray value is corrected based on the principle of "weighted average of neighboring pixels". For example, take the four adjacent non-boundary pixels around the boundary pixel (top left, top right, bottom left, bottom right direction), assign weights according to the distance between pixels (the closer the distance, the greater the weight), and obtain the new gray value of the boundary pixel through weighted calculation, which replaces the original abrupt gray value. Then, traverse the entire image to confirm that all sub-block boundary artifacts have been eliminated while preserving edge details in the target area; Ultimately, an initial cold-region engineering image with artifacts eliminated was obtained, with a gray-scale abrupt change elimination rate of over 95% between sub-blocks, and the edge features of the target region were continuous and unbroken.

[0042] S1023, using Gamma correction to eliminate artifacts from the initial cold-region engineering image, an enhanced intermediate cold-region engineering image is obtained.

[0043] Specifically, even after artifact removal, the image still suffers from issues such as high grayscale values ​​in the snow and ice areas (0.7–1) and low grayscale values ​​in the target area (0.2–0.4). Nonlinear grayscale transformation is needed to compress the highlight areas and stretch the shadow details.

[0044] The specific steps are as follows: Gamma correction principle: It adjusts the gray value of each pixel based on a non-linear gray-level mapping relationship. The formula is: ; in, The grayscale value of the pixel in the i-th row and j-th column after correction; The grayscale value of the pixel in the i-th row and j-th column before correction (after artifact removal); This is the Gamma correction factor, set in this embodiment specifically for cold-region ice and snow scenarios. =1.2( When the value is >1, the grayscale value of the highlight area is compressed and the grayscale value of the shadow area is stretched.

[0045] Gray-scale range calibration: After correction, some pixel gray-scale values ​​may exceed the 0-1 range. Therefore, all pixel gray-scale values ​​are re-linearly mapped to the 0.1-0.9 range (to avoid extreme gray-scale values ​​of 0 or 1 affecting subsequent filtering processing).

[0046] The final result is an enhanced engineering image of the intermediate cold region. The grayscale value of the snow and ice area is compressed from 0.7-1 to 0.5-0.7, and the grayscale value of the concrete marker details is stretched from 0.2-0.4 to 0.3-0.6, significantly improving the distinction between the target area and the background.

[0047] S103. Gaussian filtering is applied to the intermediate cold region engineering image to obtain the target cold region engineering image.

[0048] Taking the monitoring of roadbed in cold-region highways in Golmud City, Qinghai Province as the application scenario, Gaussian filtering was performed on the intermediate cold-region engineering image (pixel size 200×200, grayscale value range 0.1~0.9) obtained by S102. The image still has high-frequency noise caused by ice and snow particles in the cold region (manifested as sporadic bright spots and dark spots in the image), which will interfere with subsequent edge feature extraction and displacement calculation.

[0049] Step S103 removes high-frequency noise from ice and snow particles in the intermediate cold-region engineering images while preserving the edge features of the target area (roadbed concrete markers) to the greatest extent possible. This provides a clear and clean image foundation for subsequent edge feature extraction and dynamic template matching, meeting the accuracy requirements of cold-region engineering monitoring.

[0050] Specifically, obtaining the target cold-region engineering image may include the following steps: S1031, For the engineering image of the central cold region, calculate the global variance of the enhanced image to obtain the global variance of the engineering image of the central cold region.

[0051] Wherein, the global variance The calculation formula is: = ; in, H is the grayscale mean, H is the height of the intermediate cold region engineering image (H=200 in this embodiment), and W is the width of the intermediate cold region engineering image (W=200 in this embodiment). is the average gray value of the intermediate cold region engineering image, and its value range is consistent with the image gray value (0.1 to 0.9 in this embodiment). Gray(i,j) is the gray value of the pixel in the i-th row and j-th column of the intermediate cold region engineering image, where i ranges from 1 to 200 and j ranges from 1 to 200.

[0052] Specifically, ice and snow particle noise in engineering images of the central cold region leads to uneven distribution of image pixel gray values. Global variance is the core indicator characterizing the dispersion of image gray values. The larger the global variance, the more noise there is in the image and the more drastic the gray value fluctuations; conversely, the smaller the global variance, the smoother the image and the less noise. The purpose of calculating global variance is to provide a quantitative basis for subsequent adaptive adjustment of the filtering intensity, avoiding the problems of "incomplete noise removal" or "blurred target edges" caused by a fixed filtering intensity.

[0053] The specific steps are as follows: First, calculate the mean grayscale value of the engineering image for the central cold region. Iterate through all pixels in the image, calculate the gray value Gray(i,j) of each pixel, sum the gray values ​​of all pixels, and divide by the total number of pixels H*W to obtain the mean gray value. In other words, the overall average brightness of the image is obtained as a benchmark for judging the degree of deviation of pixel grayscale values.

[0054] Substitute into the global variance formula to calculate For example, the calculation result is the average grayscale value obtained after traversing all pixels of the intermediate cold region engineering image. =0.52, substitute into the formula to calculate the global variance. =0.032 indicates that the image contains moderate levels of ice and snow particle noise, requiring moderate-intensity filtering.

[0055] Finally, the global variance of the engineering images of the intermediate cold region was obtained. =0.032, which serves as the core basis for subsequently constructing the filter standard deviation.

[0056] S1032, Based on the global variance of the intermediate cold region engineering image, construct the filter standard deviation.

[0057] Among them, the standard deviation of the filter increases with the increase of the global variance and decreases with the decrease of the global variance.

[0058] Specifically, the standard deviation of the filter is the core parameter that determines the strength of the Gaussian filter. The larger the standard deviation, the stronger the filter and the better the denoising effect, but it is easy to blur the edges of the target area. The smaller the standard deviation, the weaker the filter and the better it can preserve the edges, but it may not be able to completely remove noise. Considering the noise characteristics of cold-region images, the standard deviation of the filter needs to be linked with the global variance to achieve adaptive filtering, balancing denoising effect and edge preservation.

[0059] The specific steps are as follows: Following the principle that "the standard deviation of the filter increases with the increase of the global variance and decreases with the decrease of the global variance", and taking into account the noise characteristics of cold region engineering images, the range of the standard deviation of the filter is limited, for example, the range is 0.8 to 2.5 (verified by a large number of cold region monitoring experiments, this range can effectively balance the removal of noise in cold regions and the preservation of target edges, avoiding over- or under-filtering).

[0060] Determine the specific value of the filter standard deviation, for example, by combining it with the global variance calculated using S1031. =0.032, corresponding to the adaptive filter standard deviation. In this embodiment, the adaptation relationship between the global variance and the filter standard deviation can be set as follows: when When the noise level is between 0.02 and 0.04 (medium noise), the standard deviation of the filter is taken as 1.5; if <0.02 (low noise), filter standard deviation is 0.8~1.0; if >0.04 (high noise), filter standard deviation is set to 2.0-2.5. The filter standard deviation is selected according to the range of global variance to adapt to the differences in image noise under different snow and ice coverage in cold regions (for example, when there is more snow and ice coverage and more noise, the global variance increases, and the filter standard deviation increases accordingly, thus enhancing the denoising effect).

[0061] The final output can be based on global variance. =0.032, construct the filter standard deviation, for example, filter standard deviation = 1.5, so as to be used for subsequent Gaussian kernel generation.

[0062] S1033, an m×m Gaussian kernel is generated using the standard deviation of the filter. The Gaussian kernel is then used to perform convolution filtering on the intermediate cold region engineering image, and the boundaries are filled with duplicates to obtain the target cold region engineering image.

[0063] Specifically, in Gaussian filtering, the Gaussian kernel is essentially an m×m matrix. The values ​​within the matrix follow a Gaussian distribution (large values ​​in the middle and gradually decreasing towards the edges). By convolving the Gaussian kernel with the image, a weighted average of the pixel gray values ​​is achieved, thereby smoothing noise. Boundary padding is used to prevent image boundary pixels from becoming distorted after filtering due to the lack of surrounding pixel references, thus adapting to monitoring scenarios where the target area in cold regions may be close to the boundary.

[0064] The specific steps are as follows: Determine the Gaussian kernel size m. Combining the pixel size of the cold region engineering image (200×200) and the size of the target area (concrete marker point), m=3 can be selected, which generates a 3×3 Gaussian kernel. (This Gaussian kernel size can effectively remove ice and snow particle noise and retain the edge details of the target area to the greatest extent. If a larger Gaussian kernel is selected, the target edge will be blurred, affecting the accuracy of subsequent displacement calculation.)

[0065] With a 3×3 Gaussian kernel generated, a 3×3 Gaussian kernel matrix is ​​generated based on the filtering standard deviation obtained in S1032, for example, a filtering standard deviation of 1.5. The generated 3×3 Gaussian kernel is moved across the intermediate cold region engineering image in a "sliding window" manner, moving one pixel at a time. For the 9 pixels within the window, a weighted average is calculated according to the values ​​of the Gaussian kernel matrix. The calculation result is used as the new grayscale value of the current window's center pixel. This process is repeated throughout the entire image to complete the filtering processing for all pixels.

[0066] Next, duplicate padding is performed at the boundaries. During convolution filtering, when the Gaussian kernel slides to the image boundary, the window may extend beyond the image area, resulting in some locations lacking pixel references. In this case, duplicate padding (i.e., copying the pixel values ​​from the image edges and filling the areas where the window extends beyond the image) is used to ensure that the filtering calculations for boundary pixels have a reference, avoiding distortion of boundary pixel grayscale values ​​and ensuring consistent filtering effects across the entire image. For example, when the Gaussian kernel slides to the leftmost boundary of the image, the portion of the window extending beyond the image on the left is filled with the leftmost edge pixel values; when it slides to the bottom boundary, the portion of the window extending beyond the image on the bottom is filled with the bottom edge pixel values.

[0067] The final output, after the above convolution filtering and boundary filling processing, shows that the removal rate of ice and snow particle noise in the intermediate cold region engineering image is over 90%, the edge features of the target area (concrete marker point) are completely preserved, the gray value distribution is uniform, and there is no obvious noise interference. The resulting target cold region engineering image can be directly used for subsequent edge feature extraction, dynamic template matching and other steps.

[0068] In summary, S103 addresses the interference problem of ice and snow particle noise in cold-region engineering images through a process of "calculating global variance (to determine the degree of noise), constructing adaptive filter standard deviation (to determine the filter strength), generating Gaussian kernel and performing convolution filtering (to remove noise)".

[0069] S200 performs edge feature extraction and weighted calculation on the target cold-region engineering image, and uses the SSD algorithm for dynamic template matching to obtain the target pixel displacement.

[0070] In the context of cold region monitoring, the dynamic template refers to the initial reference template of the target region in the first frame of the target cold region engineering image. When the appearance of the target region changes and the matching confidence is lower than the set confidence threshold, the target region matched in the current frame is used as the target reference template for the next moment, thus forming a dynamic matching sample.

[0071] This embodiment continues the cold-region highway subgrade monitoring scenario in Golmud City, Qinghai Province. It processes the target cold-region engineering image (pixel size 200×200, grayscale range 0.1–0.9, snow and ice particle noise removal rate ≥90%, target area is 100mm×100mm concrete marker points) obtained from S103. The aim is to enhance the target area's recognizability through edge feature extraction, and combine dynamic template matching to capture the pixel displacement of the target area, providing a basis for subsequent actual physical displacement conversion; adapting to the scenario in cold-region subgrade monitoring where the target area may change in appearance due to snow and ice melting or slight subsidence, ensuring the stability and accuracy of the matching.

[0072] In this embodiment, the first frame of the target cold-region engineering image is the first frame image collected after the monitoring is started, and its target area (concrete marker point) is used as the initial reference template; each subsequent frame of the target cold-region engineering image is used as the current frame, and its target area is used as the current target area. It is matched with the initial reference template or the updated target reference template to finally obtain the target pixel displacement of the current target area (i.e., the pixel position offset of the target area in the image).

[0073] Specifically, obtaining enhanced engineering images of intermediate cold regions may include the following steps: S201, for the current target region in the current frame target cold region engineering image, the Canny operator is used to extract the edge points of the current target region to form an edge binary map, and the edge points in the edge binary map are assigned a higher weight than the non-edge points to obtain the current target region with enhanced edge matching contribution.

[0074] Specifically, in the target cold-region engineering image, although the grayscale difference between the target area (concrete marker point) and the background (roadbed soil, residual ice and snow) has been improved, the edge contour still needs to be further enhanced; edge features are the core identification markers of the target area. Extracting the edge and assigning it high weight can improve the accuracy of subsequent template matching, avoid matching deviations caused by background interference, and adapt to complex background scenes in cold regions.

[0075] The specific steps are as follows: Locate the current target region. In the current frame of the cold-region engineering image, locate the current target region corresponding to the initial reference template in the first frame. Its pixel size is still 200×200. The core monitoring area is the concrete marker point (corresponding to the 100×100 pixel range in the center of the image). Ensure that the region range is consistent with the initial reference template to avoid matching region deviation.

[0076] Edge extraction using the Canny operator: The Canny operator is used to extract edges from the current target region. The core principle is to sequentially perform: smoothing and denoising, capturing areas with drastic grayscale changes, and connecting discrete edge points. This filters out complete edges of the target region (such as the outline of concrete markers) and removes subtle grayscale changes in the background (such as the fine texture of residual snow). In this embodiment, considering the grayscale range (0.1–0.9) of the target cold-region engineering image, the low threshold of the Canny operator can be set to 0.15 and the high threshold to 0.3 to ensure the completeness and accuracy of edge extraction. Grayscale changes below the low threshold are considered background interference, while those above the high threshold are considered valid edges. Regions between these thresholds are considered edge points if they are connected to valid edges; otherwise, they are considered background.

[0077] An edge binary map is generated by binarizing the extracted edge points and non-edge points. Specific rules include: setting the pixel value of edge points to 1 and the pixel value of non-edge points to 0; ensuring the edge binary map has the same pixel size as the current target region (200×200), retaining only the edge contour information of the target region, and eliminating interference caused by differences in grayscale values.

[0078] Weighted processing, to enhance the contribution of edge features in subsequent matching, assigns higher weights to edge points in the binary edge image than to non-edge points. Specifically, in the binary edge image, the weight coefficient is set to 2 for positions with a pixel value of 1 (edge ​​point) and 1 for positions with a pixel value of 0 (non-edge point). After weighting, edge points have a higher proportion in subsequent similarity calculations, effectively improving the accuracy of template matching and reducing the influence of background regions on the matching results.

[0079] The final output is the current target region with enhanced edge matching contribution. This current target region retains the original grayscale information and strengthens the edge features through weighting, and can be directly used for similarity calculation in subsequent template matching.

[0080] S202, using the weighted SSD algorithm, calculates the similarity between the initial reference template and the current target region, and obtains the confidence level of the similarity between the initial reference template and the current target region.

[0081] in, The calculation formula is: = ; in, The weighted SSD values ​​of the initial reference template and the current target region are: T is the template image, T(i,j) is the gray value of the pixel in the i-th row and j-th column of the initial reference template, I(i,j) is the image window to be matched, and the gray value of the pixel in the i-th row and j-th column of the current target region is: Edge(i,j) is the edge binary map (the value of the pixel in the i-th row and j-th column of the edge binary map obtained by S201, which takes the value 0 or 1 (0 represents a non-edge point, 1 represents an edge point)), and i,j are the pixel coordinates.

[0082] Specifically, the core of initial reference template matching is to determine the similarity between the current target region and the initial reference template. The higher the similarity, the smaller the positional change of the target region. The core function of the weighted SSD algorithm is to accurately quantify the similarity by calculating the pixel difference between the two and combining it with edge weights, and finally convert it into confidence, providing a basis for judgment for subsequent template updates and displacement calculations.

[0083] The specific steps are as follows: The matching objects are the "initial reference template" and the "current target region with enhanced edge matching contribution". The initial reference template is the target region in the first frame of the target cold region engineering image after S103 processing. Edge extraction and weighting processing have been completed, and it serves as the benchmark for matching all subsequent frames.

[0084] The weighted SSD calculation logic involves calculating the grayscale difference between the initial reference template and the current target region pixel by pixel, squaring the difference (amplifying larger differences and reducing smaller ones), multiplying it by the corresponding weight coefficient (weight 2 for edge points and weight 1 for non-edge points), and finally summing the calculation results for all pixels to obtain the weighted SSD value. The smaller the weighted SSD value, the higher the similarity between the two. The specific calculation can be performed using the weighted SSD calculation formula.

[0085] The weighted SSD values ​​are converted into confidence scores in the range of 0 to 1 (the higher the similarity, the closer the confidence score is to 1). The specific conversion method is as follows: Confidence Score =1 / (1+ In this embodiment, the preset confidence threshold can be 0.7 (verified by cold region monitoring experiments, this threshold can effectively distinguish between normal minor changes and significant changes in appearance in the target area, avoiding incorrect or missed template updates).

[0086] For example, assuming the weighted SSD value of the initial reference template and the current target region is 0.43, substituting it into the confidence formula, we get the confidence = 1 / (1+0.43)≈0.70. This confidence is equal to the preset threshold, which is considered a qualified match, and the template is not updated for the time being.

[0087] The final output shows the confidence level of the similarity between the initial reference template and the current target region (for example, the example value in this embodiment is 0.70), which is used for subsequent comparison with a preset confidence threshold.

[0088] S203, compare the confidence level with the preset confidence level threshold. When the confidence level is less than the confidence level threshold, use the current target region as the target reference template to form a dynamic matching sample, and obtain the target pixel displacement of the current target region based on the coordinate offset between the target reference template and the initial reference template.

[0089] Specifically, in cold region monitoring, the target area (concrete marker point) may undergo significant changes in appearance due to factors such as snow melting, slight roadbed settlement, and changes in ambient light, resulting in a decrease in similarity with the initial reference template (confidence level below the threshold). If the initial reference template is continued to be used for matching in this case, it will lead to matching deviation. Therefore, it is necessary to update the reference template to form dynamic matching samples to ensure the accuracy of subsequent matching. At the same time, the pixel displacement of the target area is obtained by calculating the coordinate offset, which provides a basis for subsequent temperature compensation and physical displacement conversion.

[0090] The specific steps are as follows: First, a confidence level comparison is performed: the confidence level obtained in S202 is compared with a preset confidence threshold (such as 0.7 in the previous embodiment), and two cases are considered: If the confidence level is ≥0.7: it means that the current target area is highly similar to the initial reference template, the appearance of the target area has not changed significantly, there is no need to update the reference template, and the coordinate offset is calculated directly based on the initial reference template; if the confidence level is <0.7: it means that the current target area is not highly similar to the initial reference template, the appearance of the target area has changed significantly (such as changes in snow and ice coverage, or slight damage to the marker points), and the reference template needs to be updated.

[0091] Next, when the confidence level is less than 0.7, the current target region of the current frame (after edge extraction and weighting processing) is used as the new target reference template, replacing the original initial reference template, for matching the target region in the next frame. Simultaneously, this new target reference template is associated with and stored along with the corresponding current frame information (acquisition time, confidence level, pixel displacement) to form a dynamic matching sample, facilitating subsequent tracking and matching accuracy calibration. For example, assuming the confidence level of the current target region in a certain frame is 0.65 (less than 0.7), the current target region of that frame is used as the new target reference template to form a dynamic matching sample.

[0092] Then, the coordinate system is set. To accurately calculate the coordinate offset, the coordinate system of the target cold region engineering image is set. For example, the upper left corner of the image is the origin (0,0), the horizontal direction to the right is the positive x-axis, and the vertical direction downward is the positive y-axis. The pixel coordinates are in units of pixels, and each pixel corresponds to a coordinate point (x,j), where x is the horizontal pixel number and y is the vertical pixel number, corresponding to the row and column of the image (i corresponds to the y-axis and j corresponds to the x-axis).

[0093] Using the center pixel coordinates of the reference template (initial reference template or updated target reference template) as a reference, calculate the offset between the center pixel coordinates of the current target region and the center pixel coordinates of the reference template. This offset is the target pixel displacement. Specifically, the calculation method can be: Target pixel displacement (Δx, Δy) = (Current target region center x coordinate - Reference template center x coordinate, Current target region center y coordinate - Reference template center y coordinate).

[0094] For example, the center pixel coordinates of the initial reference template are (100, 100) (corresponding to the center position of a 200×200 pixel image); if the center pixel coordinates of the current target area are (102, 101), then the target pixel displacement Δx = 2 pixels and Δy = 1 pixel, that is, the current target area is offset horizontally to the right by 2 pixels and vertically downward by 1 pixel relative to the initial reference template.

[0095] If the reference template has been updated (forming dynamic matching samples), the target pixel displacement in subsequent frames will be calculated based on the updated target reference template to ensure the continuity and accuracy of displacement calculation. If the target area does not change significantly (confidence ≥ 0.7), the initial reference template will always be used as the reference to avoid displacement calculation deviation caused by frequent template updates.

[0096] The final output shows the target pixel displacement of the current target region (in the example, Δx=2 pixels, Δy=1 pixel); if the confidence level is <0.7, a dynamic matching sample (updated target reference template and associated information) is generated for template matching in the next frame.

[0097] The S300 uses a temperature compensation model to correct and compensate the equipment that acquires the original cold-region engineering images, forming a compensation coefficient. Based on the compensation coefficient, the displacement of the target pixel is adjusted to obtain the actual physical displacement of the target area.

[0098] Processing is performed based on the target pixel displacement obtained from S200 (in example, Δx=2 pixels, Δy=1 pixel). The image acquisition device used is a 1920×1080 resolution industrial camera (fixed mount) mentioned in S100. The core objective is to address the pixel displacement deviation caused by thermal expansion and contraction of the image acquisition device (lens, body) due to the low temperature environment in cold regions. By correcting the deviation through temperature compensation and combining the conversion benchmark between pixels and physical dimensions, the pixel displacement is converted into actual physical displacement. This provides accurate quantitative data for the risk assessment of disasters such as roadbed settlement and displacement in cold regions, ensuring that the monitoring results meet engineering-grade accuracy requirements. It is understandable that the ambient temperature at the monitoring site in cold regions fluctuates greatly (e.g., the temperature range during the monitoring period in this embodiment is -25℃ to 5℃). The lens and body material of the image acquisition equipment will undergo slight deformation due to temperature changes, resulting in slight stretching or shift in the acquired image, which in turn causes a deviation in the target pixel displacement obtained by S200. By calculating the compensation coefficient through the temperature compensation model, this deviation can be corrected. Then, combined with the conversion ratio between pixels and actual physical size, the pixel displacement is converted into actual physical displacement at the millimeter (mm) level, ensuring the accuracy of the monitoring data.

[0099] Specifically, obtaining the actual physical displacement of the target region may include the following steps: S301 uses a target of known size with the same material as the target area to perform basic ratio calibration, and obtains the conversion reference ratio between pixels and actual physical size.

[0100] Specifically, the target pixel displacement obtained by S200 is only the pixel offset in the image and cannot directly reflect the actual physical displacement of the target area (such as the number of millimeters of roadbed settlement). It is necessary to establish a conversion relationship of "how many actual physical dimensions correspond to 1 pixel" through basic scale calibration, so as to provide a benchmark for the subsequent conversion of pixel displacement to physical displacement. At the same time, it is necessary to ensure that the target material is consistent with the target area to avoid light reflection and imaging deviation caused by material differences, and to adapt to cold region monitoring scenarios.

[0101] The specific steps are as follows: Choose a concrete target with the same material as the target area (concrete marker point) to ensure that the surface texture and reflective properties of the target are consistent with the target area. This will avoid imaging deviations caused by material differences and affect calibration accuracy. The target can be square, with a known actual physical size of 100mm × 100mm (consistent with the actual size of the concrete marker point in the target area). The target surface should have a clearly marked border for easy positioning and size measurement in the image.

[0102] The concrete target is fixedly placed next to the target area (roadbed concrete marker point), ensuring that the target and the target area are on the same plane and at the same shooting angle, and that the distance between the target and the image acquisition device is consistent with the distance between the target area and the image acquisition device (e.g., both are 5m). Alternatively, compensation calculations can be performed when the distances are known. For example, if the distance between the target and the image acquisition device is 4m and the distance between the target area and the image acquisition device is 5m, then the distance between the target and the image acquisition device can be increased by 1m during the calculation. The cold region engineering image containing the target is acquired using the industrial camera used in S100. The acquisition parameters (resolution, exposure time, focal length) are consistent with the parameters of the original cold region engineering image acquired in S100 to avoid calibration deviations caused by parameter differences.

[0103] For the acquired image containing the target, the same image enhancement processing steps as S101 to S103 (grayscale conversion, normalization, CLAHE enhancement, and Gaussian filtering, etc.) are used to obtain a clear target enhanced image. In the target enhanced image, the target's border is located, the pixel size of the target's border in the image is measured, and the coordinates of the target's upper left, upper right, lower left, and lower right corners are read through the image coordinate system (consistent with S203, with the upper left corner as the origin (0,0)). The pixel length of the target in the x-axis and y-axis directions is calculated.

[0104] Based on the known actual physical size of the target and the measured pixel size, a conversion ratio between pixels and the actual physical size is calculated. In this embodiment, the measured pixel size of the target in the image is 100×100 pixels (corresponding to the actual physical size of the target, 100mm×100mm). Therefore, the conversion ratio is calculated as follows: Conversion ratio = Actual physical size of the target ÷ Pixel size of the target image. For example, 100mm (actual width of the target) ÷ 100 pixels (width of the target image) = 1mm / pixel, that is, 1 pixel corresponds to an actual physical size of 1 millimeter.

[0105] The final output yields the conversion benchmark ratio between pixels and actual physical size. In this embodiment, 1 pixel corresponds to 1 mm, which serves as the core benchmark for subsequent pixel displacement to be converted into actual physical displacement.

[0106] S302 collects the ambient temperature through a temperature sensor attached to the lens of the image acquisition device, calculates the compensation coefficient using a temperature compensation model, and adjusts the target pixel displacement based on the compensation coefficient to obtain the actual physical displacement of the target area.

[0107] Specifically, in cold and low-temperature environments, the lens (glass material) and body (metal material) of the image acquisition device will shrink and deform due to the decrease in temperature, resulting in a slight change in the lens focal length and a slight shift in the image. This will cause a deviation in the target pixel displacement obtained by S200 (for example, lens shrinkage at low temperatures may cause the pixel displacement to be smaller). By collecting the real-time ambient temperature by a temperature sensor and substituting it into the temperature compensation model to calculate the compensation coefficient, this deviation can be corrected to ensure the accuracy of the pixel displacement. Combined with the conversion reference ratio obtained by S301, the actual physical displacement of the target area can be obtained.

[0108] The specific steps are as follows: A high-precision temperature sensor (e.g., measurement range -40℃ to 85℃, accuracy ±0.1℃) is attached to the lens surface of the image acquisition device. This temperature sensor works synchronously with the image acquisition device, collecting and monitoring the ambient temperature (i.e., the lens surface temperature, consistent with the operating temperature of the image acquisition device) in real time, and using the temperature data for subsequent compensation coefficient calculations. For example, when a frame of image is acquired, the ambient temperature collected by the temperature sensor is -10℃.

[0109] The temperature compensation model and compensation coefficient calculation are based on the correlation between ambient temperature and image acquisition device deformation. A linear correction formula is used to calculate the compensation coefficient used to correct pixel displacement deviation. The value of the compensation coefficient is related to the ambient temperature; the lower the temperature, the greater the deformation of the image acquisition device, and the closer the compensation coefficient is to 1.0 (smaller deviation) or slightly greater than 1.0 (adjusted according to the actual deformation direction). In this embodiment, the temperature compensation model uses the linear formula: K = 1 + ×( -T), where K is a compensation coefficient used to correct the deviation of the target pixel displacement. This is a temperature correction factor (empirical value). The standard reference temperature is 25°C, and T is the real-time ambient temperature collected by the temperature sensor.

[0110] Next, the target pixel displacement (Δx, Δy) obtained in S200 is multiplied by the compensation coefficient K to obtain the corrected target pixel displacement. The core purpose of the correction is to offset the pixel displacement deviation caused by device deformation and ensure the accuracy of the pixel displacement. Then, combined with the conversion reference ratio obtained in S301 (1 pixel corresponds to 1 mm), the corrected target pixel displacement is multiplied by the conversion reference ratio to obtain the actual physical displacement of the target area. The direction of the physical displacement is consistent with the direction of the pixel displacement.

[0111] The final output is the actual physical displacement of the target area. This actual physical displacement data accurately reflects the actual physical changes in the target area (such as roadbed settlement and horizontal displacement), and can be directly used for subsequent scenario-based parameter library matching and safety assessment.

[0112] S400 constructs a scenario-based parameter library, which contains optimization parameters and safety thresholds corresponding to different scenario types.

[0113] The scene types include at least roadbed scene type, slope scene type and glacier scene type, and each scene type has corresponding optimization algorithm parameters and safety thresholds.

[0114] Taking the Qinghai-Tibet Plateau cold region engineering full-area monitoring as the application scenario, the goal is to solve the problems of insufficient generalization of algorithm parameters, low monitoring accuracy, and lack of quantitative basis for risk warning caused by differences in image texture and target features in different monitoring scenarios in cold regions. By dividing scenarios, optimizing parameters in a targeted manner, and setting safety thresholds, a scenario-based parameter library that can be directly called is constructed to achieve "automatic parameter adaptation + accurate risk warning" for monitoring processes under different cold region scenarios.

[0115] Specifically, the step of constructing the scenario-based parameter library includes: S401 divides the monitoring scenarios for cold-region engineering into at least three types: roadbed, slope, and glacier. Based on the differences in image texture and features among these different scenario types, the key parameters of the algorithm are optimized.

[0116] The key parameters of the algorithm include at least the CLAHE sub-block size, Canny threshold, and matching confidence threshold for the corresponding scene.

[0117] It should be noted that, based on the criteria of "image texture features + monitoring target attributes," the monitoring scenarios for cold-region engineering are divided into three types, with the characteristics of each type clearly defined: Roadbed scene type: The monitoring object is the roadbed of highways / railways in cold regions. The image texture shows that the background (soil / asphalt pavement) is uniform with a small amount of scattered ice and snow, the core target (concrete marker point) has clear edges, and the gray value is concentrated in the range of 0.2 to 0.8. Slope scene type: The monitored object is a cold region rock and soil slope. The image texture is characterized by a rough and uneven background (rock / frozen soil / thick layer of ice and snow), and the core target (slope crack) has thin and discontinuous edges with a large gray value range (0.1~0.9). Glacier scene type: The monitored object is a glacier in a cold region. The image texture is mainly composed of crystalline textures in the background (large area of ​​ice and snow), while the core target (glacier crevasses) has wide but blurry edges and high gray values ​​(0.6-0.95).

[0118] Furthermore, for the roadbed scene type, the CLAHE sub-block size can be optimized to 8×8 pixels. This size is suitable for the uniform image texture of the roadbed scene, which can avoid the loss of edge details of concrete markers due to excessively large sub-blocks, and also avoid image artifacts caused by excessively small sub-blocks. The Canny threshold can be optimized to a low threshold of 0.15 and a high threshold of 0.3. This threshold combination is suitable for the clear edge features of markers in the roadbed scene, which can effectively filter out the weak background interference caused by ice and snow particles and extract only the effective contours of the markers. The matching confidence threshold can be optimized to 0.7. This threshold is suitable for the stable appearance of concrete markers in the roadbed scene, which can avoid the erroneous update of the matching template caused by slight ice and snow coverage and ensure the stability of the matching results.

[0119] For slope scenes, the CLAHE sub-block size can be optimized to 6×6 pixels. This smaller sub-block size can accurately capture local grayscale changes under the rough texture of the slope scene, enhance the contrast between small cracks and the background, and prevent the rough texture from obscuring the crack target. The Canny threshold can be optimized to a low threshold of 0.1 and a high threshold of 0.25. This lower threshold combination improves the sensitivity of edge extraction and can completely capture the edges of fine and discontinuous cracks in the slope scene. At the same time, by combining high and low thresholds, the situation where rock and snow textures are misjudged as valid edges is reduced. The matching confidence threshold can be optimized to 0.65. This slightly lower threshold is suitable for the characteristics of slope scenes where cracks are easy to expand and the frequency of target appearance changes is higher than that of roadbeds. It can update the matching template in time and ensure the real-time matching of the template and the crack.

[0120] For glacier scenes, the CLAHE sub-block size can be optimized to 10×10 pixels. This larger sub-block size can smooth out interference from large areas of ice surface crystal texture in glacier scenes, focus on the grayscale difference between glacier cracks and the ice surface background, and avoid texture misjudgment caused by over-enhancement. The Canny threshold can be optimized to a low threshold of 0.08 and a high threshold of 0.2. This further reduced threshold combination can offset the edge weakening problem caused by strong light reflection in glacier scenes and completely extract the blurred edges of glacier cracks. The matching confidence threshold can be optimized to 0.6. This lowest threshold is adapted to the features of glacier scenes where glaciers are constantly moving, cracks are constantly expanding, and the target appearance is updated most frequently, ensuring that the matching template is updated in a timely manner with the slight changes in the glacier's shape and avoiding matching bias.

[0121] The scenario-based parameter library is stored in the form of a mapping table of scenario type-algorithm module-parameter value-safety threshold (algorithm module-parameter value-safety threshold is the parameter set). The monitoring system can automatically call the corresponding parameter group according to the current identification scenario without manual debugging.

[0122] The S500 performs scene-based parameter matching based on the target cold-region engineering image, and determines the corresponding optimization parameters and safety thresholds according to the matched scene type. Based on the corresponding optimization parameters, safety thresholds and the actual physical displacement of the target area, the final recognition result of the target area is obtained.

[0123] The final identification result represents the security assessment status of the target area.

[0124] This embodiment continues the engineering monitoring scenario of the Qinghai-Tibet Plateau cold region. It retrieves the parameter library content by matching scene features, corrects the actual physical displacement, completes the safety assessment, and outputs the final identification result representing the safety status of the target area.

[0125] Specifically, obtaining the final identification result of the target region may include the following steps: S501. Based on the characteristics of the target cold-region engineering image, feature matching is performed in the scene parameter library to obtain the target scene type.

[0126] Specifically, scene feature matching compares the core features of the current target cold-region engineering image with the feature templates of three types of scenes—roadbed, slope, and glacier—pre-stored in the scene-based parameter library. The scene type with the highest similarity is selected as the target scene type. The matching process does not require complex algorithms and is based solely on the core feature dimension of the image.

[0127] It should be noted that the feature extraction method for the target cold region engineering image can be a conventional feature extraction method, and this application does not limit the feature extraction method.

[0128] S502, based on the target scene type, retrieve the corresponding optimization algorithm parameters and safety threshold from the scene-based parameter library to obtain the target optimization algorithm parameters and target safety threshold for the target scene type.

[0129] Specifically, the scenario-based parameter library can adopt a "scenario type-parameter set" binding storage format. After matching the target scenario type, the corresponding parameter set can be retrieved through a unique scenario identifier (such as roadbed SC001, slope SC002, glacier SC003).

[0130] For example, the scenario-based parameter library is shown in Table 1, which provides corresponding safety threshold parameters for different monitoring scenarios. The various thresholds in Table 1 are the safety thresholds, which are used as the basis for judgment.

[0131] Roadbed 5 3 5 0 20 slope 8 5 3 -2 30 glacier 15 10 1 -5 50 S503, substitute the actual physical displacement of the target area into the target scene type, and combine it with the corresponding target optimization algorithm parameters for adaptation and correction, to obtain the final physical displacement value of the target area after target scene type adaptation and correction.

[0132] Specifically, the target optimization algorithm parameters are the parameters corresponding to the actual physical displacement of the target area under the corresponding target scene type. Based on the target optimization algorithm parameters, the displacement value is made to fit the current target scene type to obtain the real motion characteristics of the target area, that is, the final physical displacement value of the target area after target scene type adaptation and correction.

[0133] S504 compares the final physical displacement value of the target area with the target safety threshold to obtain the safety assessment level.

[0134] Specifically, after calculating the final physical displacement value of the target area in the aforementioned steps, the final physical displacement value of the target area is compared with the target safety threshold in the scenario-based parameter library, and the safety assessment level is determined based on the comparison result.

[0135] For example, safety level: final physical displacement value < warning threshold; Warning level: Warning threshold ≤ final physical displacement value < danger threshold; Hazard level: Final physical displacement value ≥ hazard threshold.

[0136] It should be noted that the target scene type, target optimization algorithm parameters, and target safety threshold are directly called from the mapping table in the scene-based parameter library.

[0137] S505, the final physical displacement value of the target area and the safety assessment level constitute the final identification result of the target area.

[0138] Specifically, the final identification result can be presented in a structured manner, including two main elements: "final physical displacement (including direction) + safety assessment level," which intuitively represents the safety assessment status of the target area. For example, the final identification result = target scene type + final physical displacement value (including displacement direction) + safety assessment level. For instance, for a roadbed scene: "Target scene type: roadbed scene; final physical displacement value of the target area: 2.10mm (vertical settlement); safety assessment level: safe."

[0139] The following uses "vibration monitoring of roadbed in cold regions" as a specific example to illustrate the actual application process of this application, including the experimental environment, operation steps, and verification of implementation results.

[0140] Experimental environment (1) Hardware configuration: Computer: Intel Core i7-10700K processor, 16GB RAM, NVIDIA GeForce RTX 3060 graphics card; Optical equipment: Industrial camera (1920×1080 resolution, 30fps), installed 5m beside the roadbed, with the lens pointing towards the concrete marker on the roadbed; Sensor: DS18B20 temperature sensor (measuring range -55°C to 125°C, accuracy ±0.5°C), attached next to the camera lens to collect ambient temperature data.

[0141] (2) Monitoring objects: The roadbed of the cold-region highway (located in Golmud City, Qinghai Province, with a minimum winter temperature of -25°C) has 100mm×100mm white concrete markers set on its surface (as monitoring targets). The displacement parameters of vehicles passing through the roadbed need to be obtained through image recognition.

[0142] (3) Image data: The roadbed image sequence (300 frames in total) was collected during the vehicle traffic period. Typical cold-region interference exists: low light (light intensity of 200 lux in the evening), snow and ice reflection (30% snow cover around the marker point), and snow and ice noise (snowflake particles cover 10% of the image area).

[0143] Operating steps (1) Image preprocessing: 1.1 Input the original image of the first frame (grayscale range [0,255], signal-to-noise ratio 18dB); 1.2 Enhanced execution of composition: The grayscale is normalized to [0,1] to obtain the grayscale image G1; Applying an 8×8 sub-block CLAHE (ClipLimit=0.02) yields the enhanced image G2, which improves local contrast by a factor of 2. Performing a Gamma=1.2 correction maps the grayscale range to [0.1, 0.9], resulting in image G3, where the brightness of the snow and ice reflection area is reduced by 30%. 1.3 Perform adaptive Gaussian filtering: Calculate the global variance of G3 =0.08, dynamically set =0.8 + 1.7 × (0.08 / 0.1) = 2.16; By applying a 3×3 Gaussian kernel to G3 filter, the denoised image G4 is obtained, and the noise percentage is reduced from 10% to below 1%, with the signal-to-noise ratio improved to 42dB.

[0144] (2) Target matching and template updating: 2.1 Manually select the concrete marker points in the first frame as the initial template T (200px*200px); 2.2 Preprocessing the second frame image Perform edge extraction: Applying the Canny operator (threshold [0.1, 0.3]), we obtain the binary edge map E; Calculate the weighted SSD: assign double the weight to edge points in E and single the weight to non-edge points, and calculate by traversing the image window. ; Find the smallest The corresponding window is used as the matching target, and the output pixel displacement Δpx = 2.5px; 2.3 At frame 50, due to increased snow cover, the matching confidence drops to 0.75 (<0.8), and the current matching window is automatically updated to the new template. Continue matching.

[0145] (3) Temperature compensation and displacement calculation: 3.1 Calibration Base Scale: Given that the actual width of the marker point is 100mm, corresponding to a pixel width of 200px, we obtain... =100 / 200=0.5mm / px; 3.2 Read the current temperature =-10°C, calculate the compensation coefficient =1+0.02×(20-(-10)) / 10=1.06; 3.3 Calculate the actual displacement: D = 0.5 × 1.06 × 2.5 = 1.325 mm.

[0146] (4) Construction of a scenario-based parameter library (5) Scene adaptation verification: 5.1 Switching to the slope scene, the algorithm automatically calls the scene-specific parameters: CLAHE sub-block 10×10, Canny threshold [0.08, 0.25], template update confidence 0.75; 5.2 Repeat the above steps for the slope image sequence (300 frames). The matching success rate stabilized from 95% in the roadbed scene to 93%, verifying the effectiveness of scene adaptation.

[0147] Implementation Result Verification (1) Accuracy verification: The roadbed displacement (1.35mm) was measured synchronously by total station. The error between the result (1.325mm) and the result of this algorithm was 0.025mm, with an error rate of 1.85%, which meets the engineering accuracy requirements; (2) Robustness verification: Compared with the traditional SSD algorithm, the number of mismatches in the 300-frame sequence of this algorithm was reduced from 90 to 12, and the tracking interruption rate was reduced from 40% to 3%; (3) Efficiency verification: The processing time for a single frame image is 0.3 seconds, which meets the real-time monitoring requirement of 30fps. The scene switching adaptation time is 8 minutes, which is 360 times more efficient than manual debugging (2 days).

[0148] Secondly, this application provides an image recognition system suitable for multi-physics field monitoring in cold-region engineering projects, applying the aforementioned image recognition algorithm for multi-physics field monitoring in cold-region engineering projects. The system includes: The cold region image combination enhancement and adaptive Gaussian filtering module is used to acquire the original cold region engineering image of the target area, perform enhancement algorithm processing, and obtain the target cold region engineering image, wherein the target area refers to the area selected in the captured image; An improved SSD target matching module is used to extract edge features and perform weighted calculations on the target cold-region engineering image, and to perform dynamic template matching using the SSD algorithm to obtain the target pixel displacement. The dynamic template means that in cold-region monitoring, the target region in the first frame of the target cold-region engineering image is used as the initial reference template. When the appearance of the target region changes and the matching confidence is lower than a set confidence threshold, the target region matched in the current frame is used as the target reference template for the next moment to form a dynamic matching sample. The temperature compensation conversion module is used to correct and compensate the device that acquires the original cold region engineering image through a temperature compensation model, form a compensation coefficient, and adjust the displacement of the target pixel based on the compensation coefficient to obtain the actual physical displacement of the target area. A scenario-based parameter adaptation module is used to build a scenario-based parameter library. The scenario-based parameter library contains optimization parameters and safety thresholds corresponding to different scenario types. The scenario types include at least roadbed scenario type, slope scenario type and glacier scenario type. Each scenario type corresponds to optimization algorithm parameters and safety thresholds. The result module is used to perform scene-based parameter matching based on the target cold region engineering image, and determine the corresponding optimization parameters and safety thresholds based on the matched scene type. Based on the corresponding optimization parameters, safety thresholds and the actual physical displacement of the target area, the final identification result of the target area is obtained, wherein the final identification result represents the safety assessment status of the target area.

[0149] In summary, the image recognition algorithm and system provided in this application, suitable for multi-physics field monitoring in cold region engineering, have the following beneficial effects: (1) The accuracy of image recognition in cold regions has been significantly improved. The combined enhancement module and adaptive filtering work together to improve the signal-to-noise ratio of cold-region images to over 40dB and enhance the sharpness of target edges by 3 times. The temperature compensation module reduces the displacement conversion error in the -30°C environment from 15% to less than 3%, meeting engineering-grade accuracy requirements (allowable error <5%).

[0150] (2) The robustness of target matching is greatly enhanced. The edge weighting mechanism and dynamic template update logic have increased the target matching success rate in cold regions from 70% to over 95%, and reduced the false matching rate to below 5%. In particular, for target appearance changes caused by snow and ice cover, the tracking interruption rate has been reduced from 40% to 5%, solving the tracking instability problem of existing algorithms in complex cold environments.

[0151] (3) Scene adaptation efficiency is significantly improved The scenario-based parameter library enables automatic adaptation to scenarios such as roadbeds, slopes, and glaciers, eliminating the need for manual debugging of algorithm parameters. The adaptation cycle is shortened from 2-3 days to within 10 minutes, reducing the application threshold of the algorithm in different cold-region engineering scenarios and covering more than 3 times the scope of existing technologies.

[0152] (4) The algorithm is highly versatile and easy to integrate and implement. The algorithm adopts a modular design, does not depend on specific hardware or development platform, and can be integrated into various cold-region engineering monitoring systems (such as camera-based displacement monitoring equipment and UAV inspection systems), reducing hardware costs by 50%; and has moderate computational complexity (single frame processing time < 0.5 seconds), making it suitable for real-time monitoring needs at engineering sites.

[0153] Thirdly, this application provides a computer program that, when executed by a processor, implements the aforementioned steps of the image recognition algorithm applicable to multi-physics monitoring in cold-region engineering.

[0154] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0155] Obviously, those skilled in the art should understand that the various units or steps of this application described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device, or fabricating them separately as individual integrated circuit modules, or fabricating multiple modules or steps into a single integrated circuit module. Thus, this application is not limited to any particular combination of hardware and software.

[0156] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. An image recognition algorithm suitable for multi-physics field monitoring in cold region engineering, characterized in that, include: The original cold-region engineering image of the target area is acquired and processed by an enhancement algorithm to obtain the target cold-region engineering image, wherein the target area refers to the selected area in the captured image; Edge features are extracted and weighted for the target cold-region engineering image, and dynamic template matching is performed using the SSD algorithm to obtain the target pixel displacement. The dynamic template means that in cold-region monitoring, the target region in the first frame of the target cold-region engineering image is used as the initial reference template. When the appearance of the target region changes and the matching confidence is lower than a set confidence threshold, the target region matched in the current frame is used as the target reference template for the next moment to form a dynamic matching sample. A temperature compensation model is used to correct and compensate the device that acquires the original cold region engineering image, forming a compensation coefficient. Based on the compensation coefficient, the displacement of the target pixel is adjusted to obtain the actual physical displacement of the target area. A scenario-based parameter library is constructed, which contains optimization parameters and safety thresholds corresponding to different scenario types. The scenario types include at least roadbed scenario type, slope scenario type and glacier scenario type, and each scenario type corresponds to optimization algorithm parameters and safety thresholds. Based on the target cold-region engineering image, the scene-based parameter library is matched, and the corresponding optimization parameters and safety thresholds are determined according to the matched scene type. Based on the corresponding optimization parameters, the safety thresholds, and the actual physical displacement of the target area, the final identification result of the target area is obtained, wherein the final identification result represents the safety assessment status of the target area.

2. The image recognition algorithm for multi-physics field monitoring in cold region engineering as described in claim 1, characterized in that, The steps of acquiring the original cold-region engineering image of the target area and performing enhancement algorithm processing to obtain the target cold-region engineering image include: The original cold region engineering image of the target area is obtained, and the original cold region engineering image is converted into a grayscale image and normalized to obtain the initial cold region engineering image; The initial cold region engineering image is subjected to contrast-limited adaptive histogram equalization processing, and bilinear interpolation is used to eliminate artifacts to obtain an enhanced intermediate cold region engineering image. Gaussian filtering is applied to the intermediate cold region engineering image to obtain the target cold region engineering image.

3. The image recognition algorithm for multi-physics field monitoring in cold region engineering as described in claim 2, characterized in that, The steps of performing contrast-limited adaptive histogram equalization on the initial cold-region engineering image and using bilinear interpolation to eliminate artifacts to obtain an enhanced intermediate cold-region engineering image include: The initial cold region engineering image is subjected to contrast-limited adaptive histogram equalization to obtain the equalized initial cold region engineering image. The equalized initial cold region engineering image is divided into n×n CLAHE sub-blocks, and bilinear interpolation is used to eliminate artifacts between CLAHE sub-blocks to obtain an artifact-free initial cold region engineering image. The initial cold-region engineering image, which has been processed to eliminate artifacts, is then processed using Gamma correction to obtain an enhanced intermediate cold-region engineering image.

4. The image recognition algorithm for multi-physics field monitoring in cold region engineering as described in claim 2, characterized in that, The step of performing Gaussian filtering on the intermediate cold region engineering image to obtain the target cold region engineering image includes: For the engineering images of the intermediate cold region, the global variance of the enhanced images is calculated to obtain the global variance of the engineering images of the intermediate cold region, wherein the global variance... The calculation formula is: = ； in, H is the grayscale mean, H is the height of the intermediate cold region engineering image, W is the width of the intermediate cold region engineering image, and Gray(i,j) is the grayscale value of the pixel in the i-th row and j-th column of the intermediate cold region engineering image. Based on the global variance of the intermediate cold region engineering image, a filter standard deviation is constructed, wherein the filter standard deviation increases with the increase of the global variance and decreases with the decrease of the global variance; An m×m Gaussian kernel is generated using the filter standard deviation. The Gaussian kernel is then used to perform convolution filtering on the intermediate cold region engineering image, with the boundaries filled using duplicates, to obtain the target cold region engineering image.

5. The image recognition algorithm for multi-physics field monitoring in cold region engineering as described in claim 1, characterized in that, The steps of extracting and weighting edge features from the target cold-region engineering image, and using the SSD algorithm for dynamic template matching to obtain the target pixel displacement include: For the current target region in the cold region engineering image of the current frame, the Canny operator is used to extract the edge points of the current target region to form an edge binary map, and the edge points in the edge binary map are assigned a higher weight than the non-edge points to obtain the current target region with enhanced edge matching contribution; The similarity between the initial reference template and the current target region is calculated using the weighted SSD algorithm to obtain the confidence score of the similarity between the initial reference template and the current target region. The SSD calculation formula is as follows: ； Where T is the template image, I is the image window to be matched, Edge(i,j) is the edge binary map, and i and j are pixel coordinates; The confidence level is compared with a preset confidence threshold. When the confidence level is less than the confidence threshold, the current target region is used as a target reference template to form a dynamic matching sample. The target pixel displacement of the current target region is obtained based on the coordinate offset between the target reference template and the initial reference template.

6. The image recognition algorithm for multi-physics field monitoring in cold region engineering as described in claim 1, characterized in that, The step of correcting and compensating the device acquiring the original cold-region engineering image using a temperature compensation model to form a compensation coefficient, and adjusting the target pixel displacement based on the compensation coefficient to obtain the actual physical displacement of the target region includes: Using a target of known size with the same material as the target area, a basic ratio calibration is performed to obtain the conversion reference ratio between pixels and actual physical size; The ambient temperature is collected by a temperature sensor attached to the lens of the image acquisition device, and a compensation coefficient is calculated using a temperature compensation model. The displacement of the target pixel is adjusted based on the compensation coefficient to obtain the actual physical displacement of the target area.

7. The image recognition algorithm for multi-physics field monitoring in cold region engineering as described in claim 1, characterized in that, The steps for constructing the scenario-based parameter library include: The monitoring scenarios for cold-region engineering projects are divided into at least three types: roadbed, slope, and glacier. Based on the differences in image texture and features among the different scenario types, the key parameters of the algorithm are optimized. The key parameters of the algorithm include at least the CLAHE sub-block size, Canny threshold, and matching confidence threshold for the corresponding scenario.

8. The image recognition algorithm for multi-physics field monitoring in cold region engineering as described in claim 1, characterized in that, The step of matching the scene-based parameter library according to the target cold-region engineering image, determining the corresponding optimization parameters and safety threshold according to the matched scene type, and obtaining the final recognition result of the target region based on the corresponding optimization parameters, the safety threshold, and the actual physical displacement of the target region includes: Based on the characteristics of the target cold-region engineering image, feature matching is performed in the scene-based parameter library to obtain the target scene type; Based on the target scene type, the corresponding optimization algorithm parameters and the safety threshold are retrieved from the scene-specific parameter library to obtain the target optimization algorithm parameters and the target safety threshold for the target scene type. Substitute the actual physical displacement of the target area into the target scene type, and combine it with the corresponding target optimization algorithm parameters for adaptation and correction to obtain the final physical displacement value of the target area after the target scene type adaptation and correction. The final physical displacement value of the target area is compared with the target safety threshold to obtain the safety assessment level; The final physical displacement value of the target area and the safety assessment level constitute the final identification result of the target area.

9. An image recognition system suitable for multi-physics field monitoring in cold region engineering, characterized in that, The image recognition algorithm applicable to multi-physics field monitoring in cold region engineering, as described in any one of claims 1-8, comprises: The cold region image combination enhancement and adaptive Gaussian filtering module is used to acquire the original cold region engineering image of the target area, perform enhancement algorithm processing, and obtain the target cold region engineering image, wherein the target area refers to the area selected in the captured image; An improved SSD target matching module is used to extract edge features and perform weighted calculations on the target cold-region engineering image, and to perform dynamic template matching using the SSD algorithm to obtain the target pixel displacement. The dynamic template means that in cold-region monitoring, the target region in the first frame of the target cold-region engineering image is used as the initial reference template. When the appearance of the target region changes and the matching confidence is lower than a set confidence threshold, the target region matched in the current frame is used as the target reference template for the next moment to form a dynamic matching sample. The temperature compensation conversion module is used to correct and compensate the device that acquires the original cold region engineering image through a temperature compensation model, form a compensation coefficient, and adjust the displacement of the target pixel based on the compensation coefficient to obtain the actual physical displacement of the target area. A scenario-based parameter adaptation module is used to build a scenario-based parameter library. The scenario-based parameter library contains optimization parameters and safety thresholds corresponding to different scenario types. The scenario types include at least roadbed scenario type, slope scenario type and glacier scenario type. Each scenario type corresponds to optimization algorithm parameters and safety thresholds. The result module is used to perform scene-based parameter matching based on the target cold region engineering image, and determine the corresponding optimization parameters and safety thresholds based on the matched scene type. Based on the corresponding optimization parameters, safety thresholds and the actual physical displacement of the target area, the final identification result of the target area is obtained, wherein the final identification result represents the safety assessment status of the target area.

10. A computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the image recognition algorithm for multi-physics field monitoring of cold-region engineering as described in any one of claims 1 to 8.

Images