Image-based tunnel disease identification system and method and portable device thereof

Abstract

The invention provides an image-based tunnel disease recognition system and method and portable equipment thereof, and the system comprises the portable equipment which is provided with an image collection module, a data storage module, a local feature comparison module and a wireless communication module, the image collection module is used for obtaining image data of a tunnel wall in real time, and the data storage module is used for storing the image data of the tunnel wall; the data storage module is used for storing a local feature template and historical diagnosis data, the local feature comparison module is used for preprocessing acquired image data, extracting features and then comparing the features with the local feature template to obtain a preliminary diagnosis result, and the wireless communication module is used for remotely outputting the preliminary diagnosis result through a wireless protocol. According to the method, image preprocessing and feature extraction are carried out at a device end, the data size of an original image needing to be transmitted is reduced, the dimensionality and the storage size of an image feature vector are effectively reduced by combining dimensionality reduction, quantization and sparse storage technologies, and therefore the data transmission amount is remarkably reduced, and the network bandwidth pressure is relieved.

Description

Technical Field

[0001] The present invention relates to the field of tunnel disease identification, and in particular to an image-based tunnel disease identification system, method and portable device thereof. Background Art

[0002] As an important infrastructure, tunnels are widely used in multiple fields such as transportation, energy, and urban construction. With the increase in the service life of tunnels and the change of the external environment, different types of diseases often appear in the structure and surface of tunnels, such as cracks, water seepage, surface defects, etc. These diseases not only affect the safety of tunnels but also may lead to further damage to their structures. Therefore, it is crucial to detect and effectively repair these diseases in a timely manner to ensure the safety of tunnels and extend their service life.

[0003] Traditional tunnel disease detection methods mainly rely on manual inspections or the use of fixed monitoring devices. Manual inspections generally require staff to enter the tunnel for visual inspections, record the disease conditions, and determine the location and size of the diseases through measuring tools. Although this method is intuitive and easy to implement, it has the disadvantages of low efficiency, long time consumption, and being easily affected by human factors. Fixed monitoring devices such as sensors and surveillance cameras usually need to be installed at key positions in the tunnel, but the installation of such devices is limited and cannot cover the entire tunnel. Moreover, they often lack real-time feedback functions and cannot achieve dynamic monitoring and real-time warning.

[0004] With the development of computer vision and machine learning technologies, image-based tunnel disease detection methods have gradually been applied. By using high-resolution imaging devices (such as high-definition cameras, laser scanners, etc.) to obtain the surface images of tunnels and then analyzing these images through image processing algorithms, disease types such as cracks, water seepage, and surface defects can be identified. Common image processing methods include edge detection, image segmentation, morphological processing, etc. By using these technologies to extract the features of the tunnel surface, diseases can be further identified. However, existing disease detections usually require transmitting high-resolution image data to a remote server, which will result in a large amount of data transmission and storage requirements. Especially in the case of long tunnels or frequent monitoring, the transmission pressure is very high, bringing technical bottlenecks to remote areas and situations with unsatisfactory on-site environments (such as remote tunnels). Therefore, an image-based tunnel disease identification system, method and portable device thereof are proposed to solve the above problems. Summary of the Invention

[0005] The main purpose of the present invention is to provide an image-based tunnel disease identification system, method and portable device thereof, so as to solve the problem that existing disease detections usually require transmitting high-resolution image data to a remote server, which will result in a large amount of data transmission and storage requirements. Especially in the case of long tunnels or frequent monitoring, the transmission pressure is very high.

[0006] To solve the above technical problems, the technical solutions adopted by the present invention are as follows: An image-based tunnel disease recognition system, method and its portable device, including: A portable device, which is provided with an image acquisition module, a data storage module, a local feature comparison module and a wireless communication module. The image acquisition module is used to obtain image data of the tunnel wall in real time. The data storage module is used to store local feature templates and historical diagnosis data. The local feature comparison module preprocesses the acquired image data, extracts features, and then compares them with the local feature templates to obtain a preliminary diagnosis result. The wireless communication module remotely outputs the preliminary diagnosis result through a wireless protocol; A remote server, which receives the preliminary diagnosis results of multiple portable disease recognition devices through a wireless protocol, then generates a comprehensive diagnosis result through analysis, and feedbacks the comprehensive diagnosis result to the portable disease recognition device through a wireless protocol.

[0007] In a preferred solution, the storage method of the data storage module is at least one of a dimensionality reduction processing method, a quantization storage method and a sparse storage method; The dimensionality reduction processing method is: through principal component analysis, compress the high-dimensional feature vector into a low-dimensional space, and compress the 512-dimensional feature vector into 128 dimensions; The quantization storage method: through data type conversion, quantize the floating-point representation float32 into a more compact integer representation int8, and quantize each element in the 128-dimensional feature vector into 1 byte; The sparse storage method is: using the characteristics of the sparse matrix, only store non-zero or significant eigenvalue and their position information, sparsify the feature vector, and store the significant values and their index positions.

[0008] The method includes: S1. Image acquisition and preprocessing, collect tunnel wall images in real time, and preprocess the collected images; S2. Local feature extraction and matching, first use a convolutional neural network to extract the global features of the image, then use the SIFT or SURF algorithm to extract key feature points in the tunnel wall image, and select the corresponding SIFT or SURF algorithm to calculate the descriptors of the extracted feature points; S3. Feature matching, accelerate candidate matching through the locality-sensitive hashing algorithm, then perform exact matching on the candidate features through the brute-force matching algorithm, calculate the Euclidean distance between feature points, and obtain a preliminary diagnosis result; S4. Result transmission and server analysis, send the preliminary diagnosis result to the remote server through the wireless communication module, and the remote server performs comprehensive analysis based on the received multi-party data, generates a comprehensive diagnosis result and feedbacks it to the portable device.

[0009] In the preferred solution, the preprocessing in step S1 includes: S11. Denoising, removing image noise through Gaussian filtering, and the formula is: ; where is the Gaussian kernel, is the original image, is the image after filtering; S12. Grayscale conversion, converting the color image into a grayscale image to simplify the calculation and highlight the structural information in the image. The grayscale conversion formula is as follows: ; where is the pixel value of the grayscale image, , , are the pixel values of the red, green, and blue channels in the original image respectively; S13. Histogram equalization, adjusting the contrast of the image to make the brightness distribution of the image more uniform, and further enhancing the details. The formula is as follows: ; where is the gray level 's pixel frequency, is the cumulative distribution function; maps the pixel values of the image to a new range through the CDF, thereby enhancing the contrast of the image: ; where and are the height and width of the image respectively, is the number of gray levels, is the minimum value of the CDF; S14. Image enhancement, highlighting the edges by applying enhanced high-frequency components. The formula is: ; where is the image gradient, is the enhancement factor; makes the image details more prominent by increasing the contrast of the image. The formula is as follows: ; where and are the parameters for adjusting the contrast, means restricting the result within the valid range of the image pixel values.

[0010] In the preferred solution, the convolutional neural network in step S2 is a MobileNet or EfficientNet model, which extracts the global feature vector of the tunnel wall image; The specific method is as follows: Assume the input image is , and the size of this image is , where is the height, is the width, is the number of channels, usually 3, that is, an RGB image; Convolution operation: The input image is subjected to feature extraction through multiple convolutional layers. The convolution operation uses several convolutional kernels to scan the image and generate multiple feature maps. The calculation formula for the convolution operation is: ; Among them: is the input image, is the convolutional kernel (filter), which is a small-sized matrix, is the bias, represents the convolution operation; is the feature map after convolution; Pooling operation: The pooling layer is used to reduce the size of the feature map and reduce the computational amount. The formula for the max pooling operation is: ; Among them, is the output after pooling, represents a sub-matrix in the feature map, and the pooling operation takes the maximum value in this sub-matrix; Fully connected layer: After being processed by multiple convolutional and pooling layers, the obtained high-dimensional feature map is flattened, and then further processed through one or more fully connected layers to output the final feature vector , and the calculation formula for the fully connected layer is: ; Among them: is the weight matrix of the fully connected layer, is the bias, is to flatten the feature map output by the convolutional layer into a one-dimensional vector; Feature dimensionality reduction and quantization. Since the feature vectors extracted by the convolutional neural network usually have a high dimension, dimensionality reduction processing can reduce the storage and computational burden. The specific process is as follows: ; Among them, is the projection matrix of PCA, is the high-dimensional feature vector extracted from the CNN, is the feature vector after dimensionality reduction; Feature quantization. To further reduce the storage and computational burden, the extracted features are quantized. The quantization compresses the feature values from 32-bit floating-point numbers to 8-bit integers (int8). The quantization operation is as follows: ; The storage and computational overhead of the quantized feature vectors are reduced.

[0011] In the preferred solution, the SIFT algorithm in step S2 is as follows: Extract the key feature points in the tunnel wall image, and perform scale space calculation: ; where is the Gaussian blur function, is the scale factor; Calculate the feature point descriptor: ; where is the gradient direction descriptor of the area around the feature point; The SURF algorithm is as follows: Feature point detection: Detect the feature points in the image through the Hessian matrix. And use the integral image to accelerate the calculation of the Hessian matrix. The calculation formula of the integral image: ; where, is the original image, and the integral image represents the accumulation of all pixel values from the upper left corner of the image to the point ; Calculation of the Hessian matrix: ; where, is the scale factor, and the SURF algorithm uses the determinant of the Hessian matrix to calculate the local extreme points; Feature point descriptor calculation: Construct a fixed-size window around each feature point, calculate the gradient information of the pixels in the window, and calculate the descriptor of each window; Descriptor calculation: The SURF algorithm describes the local features in the image through the Haar wavelet response. Specifically, it is by calculating the integral values of the Haar wavelet responses in the horizontal and vertical directions; Feature point descriptor calculation formula: ; where, is the convolution kernel of the Haar wavelet, is the local area of the image.

[0012] In a preferred embodiment, the locality-sensitive hashing algorithm in step S3 maps high-dimensional features to a low-dimensional space, ensuring that the hash values of similar features are as close as possible. For two feature points and , their hash values are respectively and , and the goal is to make: ; That is, similar features should have the same hash value; Brute-force matching in step S3: ; wherein, and are descriptors of two sets of feature points, is the Euclidean distance between two sets of feature points.

[0013] In a preferred embodiment, when transmitting the preliminary diagnosis result in step S4, Protobuf or CBOR is used to compress the data, and the data is transmitted through the MQTT protocol.

[0014] The device includes: a safety helmet, the safety helmet includes a top cover, a dorsal part is further provided on one side of the top cover, a processing unit and a power supply are provided on the outer side of the dorsal part, two side jaws are provided on the front side of the dorsal part, the two side jaws are symmetrically arranged at two edges of the dorsal part, and a hat band is provided between the two side jaws, and a high-definition camera and a lighting lamp are provided at the front end of one of the side jaws; wherein, the processing unit includes an image processing unit, a wireless communication module, a gyroscope and a data storage module, and a computer program for executing the method according to any one of claims 3-8 is stored in the data storage module.

[0015] In a preferred embodiment, a temperature and humidity sensor and a depth sensor are provided at the front end of the other side jaw; Rotating shaft brackets are provided on both sides of the top cover, rotatable rotating shafts are provided on the rotating shaft brackets, a transparent mask is provided between the two rotating shafts, and projection support plates are further provided inside the lower sides of the side jaws, and HUD projectors are provided on the projection support plates, and the HUD projectors are used to project real-time information on the transparent mask; Heat dissipation holes are further provided on both sides of the dorsal part, a micro fan is provided on the outer side of the dorsal part, air ducts are provided on both sides of the micro fan, the air ducts are communicated with the heat dissipation holes, and the micro fan is used to cool the inside of the safety helmet.

[0016] The present invention provides an image-based tunnel disease identification system, method, and portable device thereof. By performing image preprocessing and feature extraction on the device side, the amount of original image data to be transmitted is reduced. Combining dimensionality reduction, quantization, and sparse storage technologies, the dimensionality and storage size of the image feature vectors are effectively reduced, thereby significantly reducing the data transmission volume and alleviating the network bandwidth pressure. At the same time, by combining with a remote server to comprehensively analyze the diagnostic data from multiple portable devices, a more accurate tunnel disease diagnosis report can be formed to assist tunnel maintenance and repair decisions and provide data support for tunnel operation management. Description of the Drawings

[0017] The present invention will be further described below in conjunction with the drawings and embodiments: Figure 1 is the system flow chart of the present invention; Figure 2 is the structure diagram of the safety helmet of the present invention; Figure 3 is the front sectional view structure diagram of the safety helmet of the present invention; Figure 4 is the front view structure diagram of the safety helmet of the present invention; Figure 5 is the side sectional view structure diagram of the safety helmet of the present invention; In the figure: safety helmet 1; transparent face mask 101; rotating shaft 102; side jaw part 103; temperature and humidity sensor 104; heat dissipation holes 105; air duct 106; micro fan 107; HUD projector 108; rotating shaft bracket 109; dorsal part 110; high-definition camera 111; projection support plate 112; top cover 113; power supply 114; depth sensor 115; lighting lamp 116; processing unit 117. Detailed Embodiments

[0018] Embodiment 1 As Figure 1 shown, an image-based tunnel disease identification system includes a portable device and a remote server. Among them, the portable device is used for real-time acquisition of tunnel wall images, feature extraction and matching, generation of preliminary diagnosis results, and transmission of the results to the remote server through a wireless protocol. The remote server receives the diagnostic results from multiple portable devices, conducts comprehensive analysis, and feeds back a diagnostic report.

[0019] The portable device is provided with an image acquisition module, a data storage module, a local feature comparison module, and a wireless communication module. The image acquisition module is a high-definition camera, which is used to obtain the image data of the tunnel wall in real time. The data storage module is used to store the local feature templates and historical diagnosis data. The local feature comparison module preprocesses the obtained image data, extracts features, and then compares them with the local feature templates to obtain a preliminary diagnosis result. The wireless communication module remotely outputs the preliminary diagnosis result through a wireless protocol.

[0020] The remote server receives the preliminary diagnosis results of multiple portable disease identification devices through a wireless protocol, then generates a comprehensive diagnosis result through analysis, and feedbacks the comprehensive diagnosis result to the portable disease identification device through the wireless protocol to guide subsequent diagnosis or maintenance work.

[0021] In the preferred solution, the storage method of the data storage module is at least one of a dimensionality reduction processing method, a quantization storage method, and a sparse storage method.

[0022] Among them, the dimensionality reduction processing method is: through principal component analysis, the high-dimensional feature vector is compressed into a low-dimensional space, and the 512-dimensional feature vector is compressed into 128 dimensions, which can greatly reduce the storage occupancy, without losing the matching accuracy, reducing the computational complexity during matching, and improving the processing speed.

[0023] The quantization storage method: through data type conversion, the floating-point representation float32 is quantized into a more compact integer representation int8, and each element in the 128-dimensional feature vector is quantized into 1 byte. While compressing storage, it retains good feature recognition, and the computational efficiency is higher.

[0024] The sparse storage method is: using the characteristics of the sparse matrix, only the non-zero or significant eigenvalue and its position information are stored, the feature vector is sparsified, and the significant value and its index position are stored, reducing the computational burden of irrelevant features during matching.

[0025] The dimensionality reduction processing, quantization storage, and sparse storage methods can be effectively used alone or in combination, which can reduce the storage space, improve the computational efficiency, and retain important information at the same time.

[0026] Embodiment 2 Using the method of the image-based tunnel disease identification system described in Embodiment 1, this method includes: S1. Image acquisition and preprocessing, real-time acquisition of tunnel wall images, and preprocessing of the acquired images. Among them, the preprocessing includes: S11. Denoising, removing image noise through Gaussian filtering, removing low-frequency noise, and retaining the detail information of the image. The formula is: ; Among them, is the Gaussian kernel, is the original image, is the filtered image.

[0027] S12. Grayscale conversion: Convert the color image to a grayscale image to simplify calculations and highlight the structural information in the image. The grayscale conversion formula is as follows: ; Among them, is the pixel value of the grayscale image, , , are the pixel values of the red, green, and blue channels in the original image, respectively.

[0028] S13. Histogram equalization: By adjusting the contrast of the image, make the brightness distribution of the image more uniform, and then enhance the details. The formula is as follows: ; Among them, is the gray level of the pixel frequency, is the cumulative distribution function; Map the pixel values of the image to a new range through the CDF to enhance the contrast of the image: ; Among them, and are the height and width of the image, respectively, is the number of gray levels, is the minimum value of the CDF.

[0029] S14. Image enhancement: By applying to enhance the high-frequency components to highlight the edges and further highlight possible cracks and other diseases. The formula is: ; Among them, is the image gradient, is the enhancement factor; By increasing the contrast of the image, make the details of the image more prominent. The formula is as follows: ; Among them, and are the parameters for adjusting the contrast, means to limit the result within the valid range of the image pixel values.

[0030] S2. Local feature extraction and matching. First, use a convolutional neural network to extract the global features of the image, and then use the SIFT or SURF algorithm to extract the key feature points in the tunnel wall image. Select the method corresponding to the SIFT or SURF algorithm to calculate the descriptors for the extracted feature points, and these descriptors are used for the subsequent matching process.

[0031] Among them, the convolutional neural network is the MobileNet or EfficientNet model, which extracts the global feature vector of the tunnel wall image. The network converts the input image into a low-dimensional feature vector through multiple convolutional and pooling operations. This feature vector contains the overall information of the tunnel wall, such as shape, texture, etc. The specific method is as follows: Assume the input image is , and the size of this image is , where is the height, is the width, is the number of channels, usually 3, that is, an RGB image.

[0032] Convolution operation: Extract features from the input image through multiple convolutional layers. The convolution operation scans the image with several convolutional kernels to generate multiple feature maps. The calculation formula for the convolution operation is: ; Among them: is the input image, is the convolutional kernel (filter), which is a small-sized matrix, is the bias, represents the convolution operation; is the feature map after convolution.

[0033] Pooling operation: The pooling layer is used to reduce the size of the feature map and reduce the computational amount. The formula for the max pooling operation is: ; Among them, is the output after pooling, represents a sub-matrix in the feature map, and the pooling operation takes the maximum value in this sub-matrix.

[0034] Fully connected layer: After being processed by multiple convolutional and pooling layers, the obtained high-dimensional feature map is flattened, and then further processed through one or more fully connected layers to output the final feature vector , and the calculation formula for the fully connected layer is: ; Among them: is the weight matrix of the fully connected layer, is the bias, Flatten the feature map output by the convolutional layer into a one-dimensional vector.

[0035] Feature dimensionality reduction and quantization. Since the feature vectors extracted by the convolutional neural network usually have a high dimension, dimensionality reduction can reduce the storage and computational burden. The specific process is as follows: ; Among them, is the projection matrix of PCA, is the high-dimensional feature vector extracted from the CNN, is the feature vector after dimensionality reduction.

[0036] Feature quantization. To further reduce the storage and computational burden, the extracted features are quantized. Quantization compresses the feature values from 32-bit floating-point numbers to 8-bit integers (int8). The quantization operation is as follows: ; The storage and computational overhead of the quantized feature vector are reduced.

[0037] Among them, the SIFT algorithm is as follows: Extract the key feature points in the tunnel wall image. Scale space calculation: ; Among them is the Gaussian blur function, is the scale factor.

[0038] Feature point descriptor calculation: ; Among them is the gradient direction descriptor of the area around the feature point.

[0039] The SURF algorithm is as follows: Feature point detection: Detect the feature points in the image through the Hessian matrix. And use the integral image to accelerate the calculation of the Hessian matrix. The calculation formula of the integral image: ; Among them, is the original image, and the integral image represents the accumulation of all pixel values from the upper left corner of the image to the point .

[0040] Calculation of the Hessian matrix: ; Among them, is the scale factor. The SURF algorithm uses the determinant of the Hessian matrix to calculate the local extreme points.

[0041] Feature point descriptor calculation: A window of a fixed size is constructed around each feature point, the gradient information of the pixels within the window is calculated, and the descriptor of each window is calculated.

[0042] Descriptor calculation: The SURF algorithm describes local features in an image through the Haar wavelet response. Specifically, it calculates the integral values of the Haar wavelet responses in the horizontal and vertical directions. Feature point descriptor calculation formula: ; where is the convolution kernel of the Haar wavelet, is the local region of the image.

[0043] S3. Feature matching: Accelerate candidate matching through the locality-sensitive hashing algorithm, and then perform exact matching on the candidate features using the brute-force matching algorithm. Calculate the Euclidean distance between feature points to obtain a preliminary diagnosis result. Among them, the locality-sensitive hashing algorithm maps high-dimensional features to a low-dimensional space, ensuring that the hash values of similar features are as close as possible. For two feature points and , their hash values are respectively and , and its goal is to make: ; That is, similar features should have the same hash value.

[0044] The brute-force matching formula is: ; where and are the descriptors of two sets of feature points, is the Euclidean distance between two sets of feature points.

[0045] S4. Result transmission and server analysis: Send the preliminary diagnosis result to the remote server through the wireless communication module. The remote server conducts comprehensive analysis based on the received multi-party data, generates a comprehensive diagnosis result, and feeds it back to the portable device. Among them, when transmitting the preliminary diagnosis result, Protobuf or CBOR is used to compress the data, and the data is transmitted through the MQTT protocol.

[0046] Example 3 Further illustrate in combination with Example 1 and 2. As Figures 2-5The structure shown is a portable device for tunnel disease identification based on images, including a safety helmet 1. The safety helmet 1 includes a top cover 113, and a dorsal part 110 is also provided on one side of the top cover 113. The two are of an integrally formed structure. A processing unit 117 and a power supply 114 are arranged on the outer side of the dorsal part 110. The power supply 114 is used for power supply. Two side jaws 103 are arranged on the front side of the dorsal part 110 facing forward. The two side jaws 103 are symmetrically arranged at the two edges of the dorsal part 110, facilitating the wear by the detection personnel. And a cap strap is arranged between the two side jaws 103. A high-definition camera 111 and a lighting lamp 116 are arranged at the front end of one of the side jaws 103. The high-definition camera 111 serves as the image acquisition module in Embodiment 1 to collect image data of the tunnel wall, and the lighting lamp 116 can provide lighting conditions for its shooting.

[0047] Among them, the processing unit 117 includes an image processing unit, a wireless communication module, a gyroscope and a data storage module. A computer program for executing the method described in any one of claims 3-8 is stored in the data storage module.

[0048] It should be noted that: The image processing unit includes a processor, a memory and an acceleration unit. Among them, the processor can be an embedded processor for image processing tasks. The memory is a RAM for storing image data and processing results. The acceleration unit is a GPU or an NPU neural network processing unit for accelerating convolutional neural network CNN and other computationally intensive tasks. The power supply 114 consists of a rechargeable lithium battery and a battery management system.

[0049] In a preferred solution, a temperature and humidity sensor 104 and a depth sensor 115 are arranged at the front end of the other side jaw 103. The temperature and humidity sensor 104 is used to detect the environmental conditions inside the tunnel and provide environmental condition information. The depth sensor 115 can be a lidar for measuring distances and obtaining depth images.

[0050] Rotating shaft brackets 109 are arranged on both sides of the top cover 113. Rotatable rotating shafts 102 are arranged on the rotating shaft brackets 109. A transparent face shield 101 is arranged between the two rotating shafts 102. The transparent face shield 101 can be flipped through the two rotating shafts 102 to achieve the effect of opening and closing. During use, it can provide protection for the face of the detection personnel.

[0051] Projection support plates 112 are also arranged inside the lower sides of the side jaws 103. An HUD projector 108 is arranged on the projection support plates 112. The HUD projector 108 is used to project real-time information on the transparent face shield 101, so that the preliminary diagnosis result and the comprehensive diagnosis result can be projected onto the transparent face shield 101 for the detection personnel to observe.

[0052] On both sides of the dorsal part 110, there are also heat dissipation holes 105. On the outer side of the dorsal part 110, there is also a micro fan 107. On both sides of the micro fan 107, there are air ducts 106. The air ducts 106 are communicated with the heat dissipation holes 105. The micro fan 107 is used to cool the inside of the safety helmet.

[0053] It should be noted that the above electrical components are connected by electrical connection means. In addition, these are common technical means in the art, so no detailed description will be given here.

[0054] The above embodiments of the present invention solve the problems of huge transmission pressure and storage requirements generated when high-resolution image data is transmitted to a remote server in the existing tunnel disease detection method. Especially when the tunnel is long or needs to be monitored frequently, the traditional method may lead to slow system response or overloaded storage. The present invention effectively reduces the storage space requirement of image data and significantly reduces the data transmission volume by introducing dimensionality reduction processing, quantization storage, and sparse storage technologies. Through the implementation of local feature comparison and preliminary diagnosis, the dependence on the remote server is greatly reduced, and the diagnostic efficiency and real-time performance are improved. In addition, combined with the comprehensive analysis of wireless communication and the remote server, it can provide more accurate and timely support for tunnel disease monitoring, thus providing a strong guarantee for the maintenance and safety management of the tunnel.

[0055] The above embodiments are only the preferred technical solutions of the present invention and should not be regarded as limitations of the present invention. The protection scope of the present invention should be the technical solutions recorded in the claims, including the equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, the equivalent replacement improvements within this scope are also within the protection scope of the present invention.

Claims

1. An image-based tunnel disease identification system, characterized in that: include: A portable device, wherein an image acquisition module, a data storage module, a local feature comparison module and a wireless communication module are provided, wherein the image acquisition module is used to acquire image data of the tunnel wall in real time, the data storage module is used to store local feature templates and historical diagnostic data, the local feature comparison module pre-processes the acquired image data and extracts features, and then compares the features with the local feature template to obtain a preliminary diagnosis result, and the wireless communication module remotely outputs the preliminary diagnosis result through a wireless protocol; The remote server receives preliminary diagnosis results of multiple portable disease identification devices through a wireless protocol, generates a comprehensive diagnosis result through analysis, and feeds back the comprehensive diagnosis result to the portable disease identification device through a wireless protocol.

2. The method of the image-based tunnel disease identification system according to claim 1 is characterized by: The storage method of the data storage module is at least one of a dimensionality reduction processing method, a quantitative storage method and a sparse storage method; The dimensionality reduction method is: through principal component analysis, the high-dimensional feature vector is compressed into a low-dimensional space, and the 512-dimensional feature vector is compressed into 128 dimensions; Quantization storage method: Through data type conversion, the floating point representation float32 is quantized to a more compact integer representation int8, and each element in the 128-dimensional feature vector is quantized to 1 byte; The sparse storage method is: using the characteristics of sparse matrices, only storing non-zero or significant eigenvalues ​​and their position information, performing sparse processing on the eigenvectors, and storing significant values ​​and their index positions.

3. A method for using the image-based tunnel disease identification system according to claim 1 or 2, characterized in that: The method includes: S1, image acquisition and preprocessing, real-time acquisition of tunnel wall images, and preprocessing of the acquired images; S2, local feature extraction and matching, first use the convolutional neural network to extract the global features of the image, then use the SIFT or SURF algorithm to extract the key feature points in the tunnel wall image, and choose the corresponding SIFT or SURF algorithm to calculate the descriptor of the extracted feature points; S3, feature matching, accelerates candidate matching through the local sensitive hashing algorithm, and then uses the brute force matching algorithm to accurately match the candidate features and calculate the Euclidean distance between feature points to obtain preliminary diagnosis results; S4, result transmission and server analysis, the preliminary diagnosis result is sent to the remote server through the wireless communication module, and the remote server performs comprehensive analysis based on the received multi-party data, generates a comprehensive diagnosis result and feeds it back to the portable device.

4. The method of the image-based tunnel disease identification system according to claim 3 is characterized by: The preprocessing in step S1 includes: S11, denoising, remove image noise through Gaussian filtering, the formula is: ； in, is the Gaussian kernel, is the original image, is the filtered image; S12, grayscale conversion, converting color images into grayscale images, simplifying calculations, and highlighting the structural information in the image. The grayscale conversion formula is as follows: ； in, is the pixel value of the grayscale image, , , are the pixel values ​​of the red, green, and blue channels in the original image respectively; S13, histogram equalization, by adjusting the contrast of the image, makes the brightness distribution of the image more uniform, thereby enhancing the details. The formula is as follows: ； in, Is grayscale The pixel frequency, is the cumulative distribution function; The image's pixel values ​​are mapped to a new range using CDF, thereby enhancing the image's contrast: ； in, and are the height and width of the image, is the number of gray levels, is the minimum value of CDF; S14, image enhancement, by applying enhanced high-frequency components to highlight the edges, the formula is: ； in, is the image gradient, is the enhancing factor; By increasing the contrast of the image, the image details are made more prominent. The formula is as follows: ； in, and is the parameter for adjusting contrast. Indicates that the result is restricted to the valid range of image pixel values.

5. The method of the image-based tunnel disease identification system according to claim 3 is characterized by: The convolutional neural network in step S2 is a MobileNet or EfficientNet model, which extracts the global feature vector of the tunnel wall image; The specific method is: Assume that the input image is , the image size is ,in, is the height, is the width, is the number of channels, usually 3, i.e. RGB image; Convolution operation: The input image is extracted through multiple convolution layers. The convolution operation uses several convolution kernels to scan the image and generate multiple feature maps. The calculation formula of the convolution operation is: ； in: is the input image, is the convolution kernel (filter), which is a small-sized matrix. is the bias, Represents the convolution operation; It is the feature map after convolution; Pooling operation: The pooling layer is used to reduce the size of the feature map and reduce the amount of calculation. The maximum pooling operation formula is: ； in, is the output after pooling, Represents a submatrix in the feature map, and the pooling operation takes the maximum value in the submatrix; Fully connected layer: After being processed by multiple layers of convolution and pooling layers, the high-dimensional feature map is flattened and then further processed by one or more layers of fully connected layers to output the final feature vector , the calculation formula of the fully connected layer is: ； in: is the weight matrix of the fully connected layer, is the bias, It is to flatten the feature map output by the convolutional layer into a one-dimensional vector; Feature dimensionality reduction and quantization. Since the feature vectors extracted by convolutional neural networks are usually high in dimensionality, dimensionality reduction can reduce the storage and computational burden. The specific process is as follows: ； in, is the projection matrix of PCA, is the high-dimensional feature vector extracted from CNN, is the feature vector after dimensionality reduction; Feature quantization, in order to further reduce the storage and computational burden, the extracted features are quantized. Quantization compresses the feature values ​​from 32-bit floating point numbers to 8-bit integers (int8). The quantization operation is as follows: ； The storage and computational overhead of the quantized feature vectors are reduced.

6. The method of the image-based tunnel disease identification system according to claim 5 is characterized by: The SIFT algorithm in step S2 is: Extract key feature points from the tunnel wall image and calculate in scale space: ； in is the Gaussian blur function, is the scale factor; Feature point descriptor calculation: ； in It is the gradient direction descriptor of the area around the feature point; The SURF algorithm is: Feature point detection: detect feature points in the image through the Hessian matrix. And use the integral map to speed up the calculation of the Hessian matrix. The calculation formula of the integral map is: ； in, is the original image, integral image Represents the distance from the upper left corner of the image to point The accumulation of all pixel values, Calculation of the Hessian matrix: ； in, is the scale factor. The SURF algorithm uses the determinant of the Hessian matrix to calculate local extreme points. Feature point descriptor calculation: construct a fixed-size window around each feature point, calculate the gradient information of the pixels in the window, and calculate the descriptor of each window; Descriptor calculation: The SURF algorithm describes the local features in the image through Haar wavelet response, specifically by calculating the integral value of the Haar wavelet response in the horizontal and vertical directions; Feature point descriptor calculation formula: ； in, is the convolution kernel of the Haar wavelet, is a local area of ​​the image.

7. The method of the image-based tunnel disease identification system according to claim 3 is characterized by: The local sensitive hashing algorithm in step S3 maps high-dimensional features to low-dimensional space, ensuring that the hash values ​​of similar features are as close as possible. and , whose hash values ​​are and , whose goal is to make: ； That is, similar features should have the same hash value; Brute force matching in step S3: ； in, and are the descriptors of two sets of feature points, is the Euclidean distance between two sets of feature points.

8. The method of the image-based tunnel disease identification system according to claim 3 is characterized by: When transmitting the preliminary diagnosis result in step S4, Protobuf or CBOR is used to compress the data, and the data is transmitted through the MQTT protocol.

9. A portable device for identifying tunnel defects based on images, characterized by: The safety helmet (1) comprises a top cover (113), a dorsal portion (110) is further provided on one side of the top cover (113), a processing unit (117) and a power supply (114) are provided on the outer side of the dorsal portion (110), two side jaws (103) are provided on the front side of the dorsal portion (110), the two side jaws (103) are symmetrically arranged at two edges of the dorsal portion (110), a cap strap is provided between the two side jaws (103), and a high-definition camera (111) and an illumination lamp (116) are provided at the front end of one of the side jaws (103); The processing unit (117) comprises an image processing unit, a wireless communication module, a gyroscope and a data storage module, and the data storage module stores a computer program for executing the method according to any one of claims 3 to 8.

10. The portable device for identifying tunnel defects based on images according to claim 9, characterized in that: A temperature and humidity sensor (104) and a depth sensor (115) are provided at the front end of the other side jaw (103); A rotating shaft bracket (109) is provided on both sides of the top cover (113), a rotatable rotating shaft (102) is provided on the rotating shaft bracket (109), a transparent mask (101) is provided between the two rotating shafts (102), and projection support plates (112) are provided inside the lower sides of the side jaw (103), a HUD projector (108) is provided on the projection support plate (112), and the HUD projector (108) is used to project real-time information on the transparent mask (101); Heat dissipation holes (105) are also provided on both sides of the back side portion (110), a micro fan (107) is also provided on the outer side of the back side portion (110), air ducts (106) are provided on both sides of the micro fan (107), the air ducts (106) are connected to the heat dissipation holes (105), and the micro fan (107) is used to cool the inside of the helmet.

Images