Image Recognition-Based Automated Tunnel Construction Monitoring System

Abstract

This invention relates to the field of image recognition, specifically to an automated tunnel construction monitoring system based on image data. The system includes an edge computing node that acquires the original tunnel image and constructs an atmospheric scattering physical model using measured data from an optical dust concentration sensor. The edge computing node transforms the original image to the frequency domain, extracts low-frequency brightness features and high-frequency texture features using discrete cosine transform, and adaptively compensates the high-frequency texture features with weights based on the transmittance matrix calculated by the physical model to suppress diffuse light interference. The compensated features are then input into a lightweight convolutional neural network via inverse transform. This scheme introduces dust physical parameters as prior constraints into the frequency domain feature weight calculation, reducing the probability of image edge artifacts and color distortion under conditions of localized strong light and high dust, ensuring the physical authenticity of target edge features in the reconstructed image, and improving the accuracy of feature extraction under harsh conditions.

Description

Technical Field

[0001] This invention relates to the field of image recognition, and more specifically to an automated tunnel construction monitoring system based on image data. Background Technology

[0002] Existing tunnel construction image monitoring solutions primarily rely on general image enhancement algorithms to process images in harsh lighting and dusty environments, which are then input into a neural network for target recognition. Conventional solutions typically employ dehazing algorithms based on dark channel priors or data-driven deep learning image enhancement models. In practice, these solutions directly acquire the raw images captured by tunnel monitoring cameras, statistically analyze the minimum brightness values ​​of local areas at the pixel level to estimate ambient light and transmittance, calculate inverse compensation parameters using the general mathematical expression of atmospheric scattering physics equations, and linearly stretch or nonlinearly map the pixel values ​​of the entire image. After this pixel-level enhancement processing, the solution inputs the restored image into a target detection network to extract features of construction workers and machinery and output the recognition results. These solutions rely solely on the statistical information of the image's two-dimensional pixel matrix for parameter estimation and image restoration throughout the entire processing flow.

[0003] The particle size distribution of dust particles in the enclosed space of a tunnel and the scattering mechanism formed by the position of the light source are physically different from those in an open space. When the existing technology estimates the transmittance based on the statistical information of the two-dimensional pixel matrix of the image itself, it lacks the physical constraint of the dust concentration in the real environment. Under the condition of local strong direct light and high concentration of dust superimposed in the tunnel, the transmittance derived from the pixel statistics deviates from the actual physical scattering law. This causes high-frequency texture details to be incorrectly compensated during inverse transformation reconstruction, resulting in artifacts and color distortion at the edges of construction machinery. Consequently, feature shift occurs when the convolutional neural network extracts the features of the construction target, leading to recognition failure. Summary of the Invention

[0004] The purpose of this invention is to provide an automated tunnel construction monitoring system based on image recognition, which can effectively solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] The image recognition-based automated tunnel construction monitoring system includes edge computing nodes and a cloud server. The edge computing nodes acquire original images of tunnel construction, extract prior information of dark channels and illumination components from the original images, and combine them with measured data from optical dust concentration sensors deployed in the tunnel to construct an atmospheric scattering physical model specific to the tunnel environment.

[0007] The edge computing node converts the original image to the frequency domain, extracts low-frequency brightness features and high-frequency texture features using discrete cosine transform, and performs adaptive weight compensation on the high-frequency texture features using the transmittance matrix calculated by the atmospheric scattering physical model to suppress diffuse light interference caused by dust scattering.

[0008] The edge computing node inputs the compensated frequency domain features into a lightweight convolutional neural network after inverse transformation, extracts the contours of construction machinery and the features of personnel safety helmets, and outputs the recognition results. The measured data of the optical dust concentration sensor directly participates in the weight calculation process of the frequency domain features as a priori parameters in the monitoring system.

[0009] Preferably, the edge computing node performs local minimum filtering on the original image to obtain the prior information of the dark channel, and uses maximum filtering based on Retinex theory to separate the illumination component;

[0010] When constructing the atmospheric scattering physical model, the edge computing node reads the measured data of the optical dust concentration sensor as the dust particle scattering coefficient in the current environment, substitutes the scattering coefficient into the modified atmospheric scattering physical equation to replace the fixed constants of the standard atmospheric model, solves the initial transmittance by combining the prior information of the dark channel, and then uses the illumination component to perform guided filtering optimization on the initial transmittance to generate the transmittance matrix.

[0011] Preferably, before converting the original image to the frequency domain, the edge computing node divides the original image into image blocks of non-overlapping sizes, and independently performs a two-dimensional discrete cosine transform on each image block to convert the spatial domain pixel matrix into a frequency domain coefficient matrix.

[0012] The edge computing node sets the zigzag scan cutoff frequency position in the frequency domain coefficient matrix, defines the coefficient matrix region before the cutoff frequency position as the low-frequency brightness feature, and defines the coefficient matrix region after the cutoff frequency position as the high-frequency texture feature, and retains the sign bit and absolute value information of the high-frequency texture feature for subsequent weight compensation.

[0013] Preferably, when performing the adaptive weight compensation, the edge computing node performs bicubic interpolation on the transmittance matrix to align with the size dimension of the high-frequency texture features, and calculates the reciprocal of the transmittance at each position in the transmittance matrix as the basic compensation factor.

[0014] The edge computing node performs element-wise multiplication of the basic compensation factor with the absolute value information of the high-frequency texture feature, restores the positive and negative polarities according to the sign bit of the high-frequency texture feature, and performs truncation and smoothing processing on abnormal values ​​in the basic compensation factor that exceed the preset safety threshold to generate compensated high-frequency texture features that suppress diffuse light interference.

[0015] Preferably, the edge computing node performs matrix recombination and splicing of the compensated high-frequency texture features and the uncompensated low-frequency brightness features in the frequency domain, and performs inverse discrete cosine transform on the recombined complete frequency domain coefficient matrix to reconstruct the spatial domain enhanced image.

[0016] The lightweight convolutional neural network includes a depthwise separable convolutional layer and a channel attention mechanism layer. The depthwise separable convolutional layer performs spatial feature dimensionality reduction extraction on the spatial enhancement image. The channel attention mechanism layer recalibrates the weights between channels of the extracted spatial feature map. The recalibrated feature map is input into the classifier to output the recognition result corresponding to the outline of the construction machinery and the features of the personnel's safety helmet.

[0017] Preferably, during the weight calculation of the frequency domain features, the measured data of the optical dust concentration sensor is used to construct a mapping lookup table between dust concentration values ​​and frequency domain compensation intensity. When the measured data acquired in real time falls into a certain interval of the mapping lookup table, the corresponding compensation intensity adjustment coefficient is output.

[0018] The edge computing node uses the compensation intensity adjustment coefficient to dynamically shift and adjust the overall numerical distribution of the transmittance matrix, thereby reducing the compensation amplitude of the high-frequency texture features under low dust concentration conditions and increasing the compensation amplitude of the high-frequency texture features under high dust concentration conditions.

[0019] Preferably, when the edge computing node performs guided filtering optimization on the initial transmittance using the illumination component, it constructs a joint guided filtering kernel function that includes spatial distance weights and color similarity weights.

[0020] The edge computing node acquires the timestamp when the optical dust concentration sensor outputs the measured data and the timestamp when the edge computing node acquires the original image, calculates the time difference between the two timestamps, and when the time difference exceeds the synchronization threshold, it uses a linear interpolation algorithm to calculate the interpolation scattering coefficient aligned with the timestamp of the original image based on the measured data of adjacent moments, updates the initial transmittance using the interpolation scattering coefficient, and then performs the guided filtering optimization.

[0021] Preferably, when setting the zigzag scan cutoff frequency position, the edge computing node calculates the variance distribution characteristics of the frequency domain coefficient matrix of the current image block and statistically analyzes the frequency interval boundaries corresponding to the energy concentration region;

[0022] The edge computing node adds a preset frequency redundancy offset to the frequency interval boundary as the dynamic cutoff frequency position, so that the range of the low frequency brightness feature expands or contracts adaptively with the smoothness of the image content. When the original image contains a large area of ​​flat tunnel wall, the dynamic cutoff frequency position is reduced to retain more low frequency information. When it contains dense construction machinery, the dynamic cutoff frequency position is increased to retain more high frequency texture information.

[0023] Preferably, when the edge computing node performs truncation and smoothing processing on abnormal values ​​in the basic compensation factor that exceed the preset safety threshold, it uses a piecewise linear function to map the basic compensation factor, keeping the values ​​within the safety threshold linearly and proportionally mapped, mapping the values ​​greater than the upper limit of the safety threshold to a fixed upper limit constant, and mapping the values ​​less than the lower limit of the safety threshold to a fixed lower limit constant.

[0024] After performing the truncation operation, the edge computing node introduces a morphological closing operation kernel to smooth the local region of the truncated compensation factor matrix, thereby eliminating block effect edge artifacts caused by the discontinuity of high-frequency texture feature compensation between adjacent image blocks due to hard truncation.

[0025] Preferably, the depth-separable convolutional layer includes a spatial convolutional kernel and a pointwise convolutional kernel. The spatial convolutional kernel performs independent convolution operations on the horizontal and vertical textures in the spatial enhancement image using asymmetric rectangular convolutional kernels to extract the elongated outline features of the tunnel construction machinery and the arc-shaped outline features of the personnel's safety helmet.

[0026] The pointwise convolution kernel concatenates and fuses the horizontal and vertical feature maps obtained from independent convolution operations along the channel dimension. The channel attention mechanism layer calculates the global average pooling and global max pooling features of each channel on the concatenated and fused feature map, and outputs the attention weight coefficients of each channel through a shared multilayer perceptron layer and multiplies them by the concatenated and fused feature map.

[0027] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0028] 1. This invention incorporates measured data from an optical dust concentration sensor into the atmospheric scattering physical model construction process of an edge computing node. It uses the measured data to replace the fixed constants in the general scattering equation to solve for the transmittance matrix. In the frequency domain, it uses discrete cosine transform to separate low-frequency brightness features from high-frequency texture features. Adaptive weight compensation is applied to the high-frequency texture features using the transmittance matrix, transforming the dust parameters of the physical environment into a priori processing feature in the frequency domain. This suppresses the interference of dust diffuse light on high-frequency textures, overcomes the problem of transmittance estimation deviating from actual physical laws due to reliance solely on image pixel statistics, reduces the probability of edge artifacts and color distortion under conditions of localized strong light and dust superposition, and ensures the physical authenticity of the edge features of construction targets in the inverse transform reconstructed image. This provides feature-offset-free input data for subsequent lightweight convolutional neural network extraction of construction machinery contours and personnel safety helmet features.

[0029] 2. This invention performs block-independent two-dimensional discrete cosine transform on the original image, dynamically adjusts the cutoff frequency position by combining the frequency interval boundary of the energy concentration region, so that the division of low-frequency brightness features and high-frequency texture features is adaptively adjusted according to the smoothness of the image content; performs bicubic interpolation alignment on the transmittance matrix and calculates the reciprocal as the basic compensation factor; uses a piecewise linear function to truncate and smooth abnormal values ​​that exceed the safety threshold; and combines morphological closing operation to eliminate block effect edge artifacts between adjacent image blocks, maintaining the continuity of the reconstructed image in the spatial dimension after frequency domain compensation and inverse transform, reducing the feature extraction error of the long strip contour of construction machinery and the arc contour of personnel in the depth separable convolutional layer. Attached Figure Description

[0030] Figure 1 This is a flowchart illustrating the overall execution of the image recognition-based automated tunnel construction monitoring system of the present invention.

[0031] Figure 2 This is a flowchart illustrating the construction and dynamic adjustment of the atmospheric scattering physics model of the present invention.

[0032] Figure 3 This is a flowchart of the frequency domain feature extraction and dynamic cutoff frequency division of the present invention;

[0033] Figure 4 This is a flowchart of the high-frequency texture feature adaptive weight compensation and truncation smoothing process of the present invention;

[0034] Figure 5 This is a flowchart of the lightweight convolutional neural network recognition and feature extraction process of the present invention;

[0035] Figure 6 This is a flowchart of the sensor data synchronization and edge computing node interaction control process of the present invention. Detailed Implementation

[0036] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0037] Please refer to Figure 1 This embodiment provides an image recognition-based automated tunnel construction monitoring system, including edge computing nodes and a cloud server. The edge computing nodes are deployed in the field control cabinet in the tunnel construction area and establish data interaction with image acquisition equipment and optical dust concentration sensors deployed in the tunnel through wired communication links. The cloud server is deployed in a remote monitoring center and establishes data interaction with the edge computing nodes through wireless communication links. The edge computing nodes acquire original images of the tunnel construction. The original images are RGB three-channel digital images output by the image acquisition equipment. The pixel value quantization range of each color channel is 0 to 255, and the spatial dimension of the image is M rows and N columns, where M and N are both positive integers. The edge computing nodes perform multi-channel pixel statistical processing on the acquired original images, extracting the dark channel prior information and illumination components of the original images. The dark channel prior information is the statistical result of the minimum pixel value in a local area of ​​the original image, reflecting the ambient light base characteristics of areas without strong direct sunlight in the tunnel construction scene. The illumination components are the illumination intensity distribution information of the corresponding scene in the original image, reflecting the spatial distribution characteristics of artificial light sources and ambient reflected light in the tunnel.

[0038] Edge computing nodes read measured data from optical dust concentration sensors deployed in the same monitoring area within the tunnel. This measured data represents the amount of dust particles per unit volume of air within the tunnel. The edge computing nodes use this measured data as physical constraint parameters for environmental scattering characteristics. Combined with extracted dark channel prior information and illumination components, they construct a tunnel-specific atmospheric scattering physical model. This model characterizes the image pixel value attenuation caused by the absorption and scattering of light by dust particles within the tunnel. The model's input parameters include the original image pixel values, dark channel prior information, illumination components, and measured dust concentration data. The output parameter is a transmittance matrix representing the proportion of light transmitted through the dusty medium.

[0039] Edge computing nodes transform the original image to the frequency domain, using Discrete Cosine Transform (DCT) to convert the spatial pixel matrix into a frequency coefficient matrix. DCT redistributes image energy in the frequency domain by decomposing the image signal into a weighted sum of cosine basis functions at different frequencies. The edge computing nodes then separate low-frequency brightness features and high-frequency texture features from the frequency coefficient matrix. Low-frequency brightness features correspond to large-area brightness variations and smooth regions in the image, including the overall illumination distribution and background features of the tunnel scene. High-frequency texture features correspond to edges, contours, and details with abrupt pixel value changes in the image, including the edge contours of construction machinery and the shape features of personnel's safety helmets.

[0040] Edge computing nodes use a transmittance matrix calculated by an atmospheric scattering physics model to adaptively compensate for high-frequency texture features. Each element in the transmittance matrix corresponds to the proportion of light transmitted at the same spatial location in the original image. The lower the transmittance value, the stronger the dust scattering effect at that location, and the more severe the attenuation of the corresponding high-frequency texture features. Based on the values ​​at each position in the transmittance matrix, the edge computing nodes assign corresponding compensation weights to the frequency domain coefficients of the corresponding spatial locations in the high-frequency texture features. For severely attenuated high-frequency coefficients, the amplitude is increased to suppress the interference of diffuse light generated by dust scattering on high-frequency texture details. During the compensation process, the original values ​​of low-frequency brightness features are retained without adjustment.

[0041] The edge computing node merges the compensated high-frequency texture features with the original low-frequency brightness features in the frequency domain to generate a complete frequency domain coefficient matrix. An inverse discrete cosine transform is then performed on the merged frequency domain coefficient matrix to convert the frequency domain signal back to the spatial domain, generating an enhanced spatial domain image. The edge computing node inputs the enhanced spatial domain image into a lightweight convolutional neural network. This network extracts deep features from the image through multiple convolutional operations, focusing on extracting the contour features of construction machinery and the features of workers' safety helmets. After feature mapping and classification operations, the network outputs recognition results, including the type, spatial location, and contour boundaries of the construction machinery, as well as the spatial location and safety helmet wearing status of the construction workers.

[0042] The measured data from the optical dust concentration sensor serves as a priori parameters, directly participating in the weight calculation process of the frequency domain features. Before performing weight compensation for high-frequency texture features, the edge computing nodes transform the measured dust concentration data into global adjustment parameters for frequency domain compensation. These parameters directly affect the numerical distribution of the transmittance matrix, thereby adjusting the compensation weight amplitude of the high-frequency texture features. This ensures that the frequency domain compensation process is directly constrained by the physical concentration of real dust within the tunnel, rather than relying solely on the pixel statistics of the image itself.

[0043] In this embodiment, in order to clarify the division rules of frequency domain features and the corresponding processing logic, a table corresponding to the division of frequency domain features and compensation processing parameters of the original tunnel construction image is constructed as shown in Table 1.

[0044] Table 1. Correspondence between frequency domain feature division and compensation processing parameters of original images from tunnel construction.

[0045]

[0046] In Table 1, the frequency distribution intervals are based on the zigzag scanning order of the frequency domain coefficient matrix after the two-dimensional discrete cosine transform. The DC component corresponds to the first coefficient of the zigzag scan, representing the average brightness of the entire image. The cutoff frequency position is the dividing point between low-frequency and high-frequency features. Coefficients before the cutoff frequency position are classified as low-frequency brightness features, and coefficients after the cutoff frequency position are classified as high-frequency texture features. The basis for the division of compensation processing methods is as follows: diffuse light generated by dust scattering in the tunnel mainly affects the high-frequency detail information of the image, and its impact on the low-frequency brightness distribution is negligible. Therefore, only high-frequency texture features are weighted for compensation, while the low-frequency brightness features remain unchanged to avoid incorrect adjustment of the overall illumination distribution of the image. Among the core bases for weight adjustment, the value of the corresponding position in the transmittance matrix directly determines the compensation amplitude of a single pixel position, and the measured dust concentration data determines the benchmark value of the global compensation intensity. The two together constitute the constraint conditions for frequency domain weight compensation.

[0047] In this embodiment, by incorporating measured data from an optical dust concentration sensor into the construction process of the atmospheric scattering physical model, the scattering characteristic parameters of the model are made to fit the real physical environment inside the tunnel, rather than using fixed constants of a general atmospheric model. Low-frequency and high-frequency features are separated by frequency domain transformation, and adaptive weight compensation is performed only on high-frequency texture features that are severely affected by dust scattering, avoiding incorrect adjustments to the overall brightness of the image. Measured dust concentration data are used as prior parameters to directly participate in the weight calculation process of frequency domain features, making the compensation process constrained by real physical environment parameters. This overcomes the problem of transmittance estimation deviating from actual physical laws due to relying solely on image pixel statistics, ensuring the authenticity and integrity of construction target features in the enhanced image.

[0048] In an alternative embodiment, refer to Figure 2The edge computing node performs local minimum filtering on the original image to obtain dark channel prior information. The edge computing node performs pixel-by-pixel minimum value calculation on the pixel values ​​of the RGB three color channels of the original image to generate a single-channel minimum pixel value map. Then, it performs local minimum filtering on the minimum pixel value map. The filtering window is a rectangular window of a preset size, and it slides through the entire minimum pixel value map with a step size of 1. During the traversal, the minimum value of all pixel values ​​within the current window is taken as the output value at the current center position. The final generated single-channel image is the dark channel prior information of the original image. The calculation process of the dark channel prior information is quantified and described by the following formula:

[0049]

[0050] in, For the prior information of the dark channel in coordinates Pixel value at that location, The c-th color channel of the original image is located at coordinates Pixel value at that location, For coordinates centered at the local filtering window, c is the color channel index of the original image, corresponding to the R, G, and B color channels respectively.

[0051] Edge computing nodes use Retinex-based maximum filtering to separate the illumination component of the original image. The edge computing node performs pixel-by-pixel maximum calculation on the pixel values ​​of the RGB three color channels of the original image to generate a single-channel maximum pixel value map. Then, local maximum filtering is performed on the maximum pixel value map. The size of the filtering window is consistent with the filtering window size in the dark channel prior extraction process. The filtering window slides through the entire maximum pixel value map with a step size of 1. During the traversal, the maximum value of all pixel values ​​within the current window is taken as the output value at the current center position. The final generated single-channel image is the illumination component of the original image. The calculation process of the illumination component is quantitatively described by the following formula:

[0052]

[0053] in, For the illumination component in coordinates The pixel value at the given location, and the physical meaning of the other parameters are consistent with the corresponding parameters in formula (1).

[0054] When constructing the atmospheric scattering physics model, edge computing nodes read measured data from optical dust concentration sensors as the dust particle scattering coefficient in the current environment. This scattering coefficient is then substituted into the modified atmospheric scattering physics equation, replacing the fixed constants in the standard atmospheric model. The modified atmospheric scattering physics equation is quantitatively described by the following formula:

[0055]

[0056] in, The c-th color channel of the original image is located at coordinates The observed pixel value at that location, The pixel value at the corresponding location represents an ideal, clear image free from dust scattering. Let be the global ambient light intensity of the c-th color channel. coordinates The transmittance at that point. The relationship between transmittance and dust scattering coefficient is quantitatively described by the following formula:

[0057]

[0058] in, This represents the dust particle scattering coefficient corresponding to the measured data from the optical dust concentration sensor. This represents the distance between the target in the scene and the image acquisition device.

[0059] The initial transmittance is calculated by combining the edge computing nodes with prior information from the dark channel. The calculation process of the initial transmittance is quantitatively described by the following formula:

[0060]

[0061] in, coordinates Initial transmittance at the location, This is the preset defogging intensity adjustment factor; the physical meanings of the other parameters are consistent with the corresponding parameters in the aforementioned formula.

[0062] Edge computing nodes utilize the illumination component to perform guided filtering optimization on the initial transmittance, generating a transmittance matrix. The core of guided filtering is constructing a joint guided filtering kernel function that includes spatial distance and color similarity weights. Using the illumination component as the guiding image, the initial transmittance is linearly and smoothly optimized. The optimization process is quantified by the following formula:

[0063]

[0064] in, This is the optimized transmittance value at pixel i, derived from the guide filter output. To guide the value at pixel i in the image, in this embodiment, the guiding image is the separated illumination component. and For filtering window The linear coefficients within the joint guided filter kernel function. The weight calculation process of the joint guided filter kernel function is quantitatively described by the following formula:

[0065]

[0066] in, To guide the image in the filter window The mean within, To guide the image in the filter window within variance, This represents the total number of pixels within the filtering window. These are the preset regularization parameters. The weights of the joint guided filter kernel function between pixels i and j include both spatial distance weights and color similarity weights. The spatial distance weights are determined by the relative positions of the pixels within the filter window, while the color similarity weights are determined by the differences in pixel values ​​in the guided image.

[0067] The edge computing node acquires the timestamp of the measured data output by the optical dust concentration sensor, and the timestamp of the original image acquired by the edge computing node. The time difference between the two timestamps is calculated. When the time difference exceeds a synchronization threshold, a linear interpolation algorithm is used to calculate an interpolation scattering coefficient aligned with the original image timestamp based on the measured data from adjacent time points. The initial transmittance is then updated using the interpolation scattering coefficient, followed by guided filtering optimization. The linear interpolation calculation process is quantified by the following formula:

[0068]

[0069] in, These are the interpolated scattering coefficients aligned with the timestamps of the original image. This is the timestamp of the original image acquisition. The timestamp of the most recent measured dust concentration data prior to the original image acquisition timestamp. This is the timestamp of the most recent measured dust concentration data after the original image acquisition timestamp. for The dust particle scattering coefficient at time t. for The dust particle scattering coefficient at that time.

[0070] Edge computing nodes construct a lookup table mapping dust concentration values ​​to frequency domain compensation intensities. When real-time measured data falls into a certain interval of the lookup table, the corresponding compensation intensity adjustment coefficient is output. The edge computing nodes use this adjustment coefficient to dynamically shift and adjust the overall numerical distribution of the transmittance matrix, thereby reducing the compensation amplitude for high-frequency texture features under low dust concentration conditions and increasing the compensation amplitude for high-frequency texture features under high dust concentration conditions.

[0071] In this embodiment, in order to clarify the adaptation rules of compensation parameters under different dust concentration conditions, a mapping table of scattering coefficient and compensation intensity adjustment coefficient corresponding to different dust concentration ranges is constructed as shown in Table 2.

[0072] Table 2. Mapping table of scattering coefficient and compensation intensity adjustment coefficient for different dust concentration ranges.

[0073]

[0074] In Table 2, β1, β2, β3, and β4 are preset dust concentration range boundaries, determined based on statistical data of dust concentration distribution in the tunnel construction environment. k1, k2, k3, k4, and k5 are compensation intensity adjustment coefficients for the corresponding ranges, with values ​​increasing with increasing dust concentration. In the transmittance matrix adjustment method, positive offset involves superimposing a fixed positive offset on all elements of the transmittance matrix, increasing the overall transmittance value and decreasing the corresponding compensation amplitude; negative offset involves superimposing a fixed negative offset on all elements of the transmittance matrix, decreasing the overall transmittance value and increasing the corresponding compensation amplitude. The high-frequency feature compensation amplitude is based on the compensation amplitude in the medium concentration range and gradually increases with increasing dust concentration to ensure sufficient compensation for severely attenuated high-frequency texture features under high dust concentration conditions and to avoid image noise amplification due to overcompensation under low dust concentration conditions.

[0075] In this embodiment, local minimum filtering and maximum filtering based on Retinex theory are used to extract prior information of the dark channel and illumination components, respectively, providing basic image-dimensional data for the construction of the atmospheric scattering physics model. By substituting the measured data of the optical dust concentration sensor as the scattering coefficient into the atmospheric scattering physics equation, the fixed constant term in the general atmospheric model is replaced, making the scattering characteristics of the model fit the real physical environment inside the tunnel. The initial transmittance is optimized by joint steering filtering based on illumination components, eliminating block effects and noise in the transmittance matrix and improving the accuracy of transmittance estimation. The problem of asynchronous dust concentration measured data and image acquisition time is solved by timestamp synchronization and linear interpolation processing, ensuring the time alignment of scattering coefficient and image data. By constructing a mapping lookup table between dust concentration and compensation intensity, the frequency domain compensation amplitude is dynamically adjusted with dust concentration, making the compensation process adaptable to different tunnel construction conditions.

[0076] In another alternative embodiment, refer to Figure 3Before transforming the original image to the frequency domain, the edge computing node divides the original image into non-overlapping image blocks of a preset square size. During the division process, if the number of rows or columns of pixels in the original image is not divisible by the image block size, the edges of the original image are zero-padded to ensure that the number of rows and columns of pixels in the padded image is an integer multiple of the image block size, thus ensuring that all image blocks are of the same size and do not overlap. The edge computing node independently performs a two-dimensional discrete cosine transform on each image block, converting the spatial domain pixel matrix into a frequency domain coefficient matrix. Before the transform, the pixel values ​​of each image block are centered by subtracting 128 from each pixel value, adjusting the range of pixel values ​​to -128 to 127, improving the numerical stability of the transform process. The calculation process of the two-dimensional discrete cosine transform is quantitatively described by the following formula:

[0077]

[0078] in, For size Image patch in spatial coordinates Pixel value at that location, The transformed frequency domain coefficient matrix in frequency coordinates The coefficient values ​​at each position, where u is the horizontal frequency index and v is the vertical frequency index. and The weighting coefficients for the transformation have the following values: when u=0, When u>0, , The rules for taking values ​​and Consistent.

[0079] Edge computing nodes set the zigzag scan cutoff frequency position in the frequency domain coefficient matrix. A zigzag scan is a one-dimensional sequential arrangement of the two-dimensional frequency domain coefficient matrix from low to high frequency. The scan path starts from the DC component (0,0) and alternates along the diagonal, converting the two-dimensional matrix into a one-dimensional coefficient sequence. The coefficients at the beginning of the sequence correspond to low-frequency components, and the coefficients at the end correspond to high-frequency components. The edge computing node defines the coefficient matrix region before the cutoff frequency position as low-frequency brightness features and the coefficient matrix region after the cutoff frequency position as high-frequency texture features. The sign bit and absolute value information of the high-frequency texture features are retained for subsequent weight compensation. The sign bit represents the positive or negative attribute of the frequency domain coefficients, and the absolute value information represents the magnitude of the frequency domain coefficients.

[0080] When setting the cutoff frequency position for zigzag scanning, the edge computing node calculates the variance distribution characteristics of the frequency domain coefficient matrix of the current image patch and statistically analyzes the frequency interval boundaries corresponding to the energy concentration areas. The variance distribution characteristics are calculated using the one-dimensional coefficient sequence after zigzag scanning. The magnitude of the variance value characterizes the complexity of the image patch content; the lower the variance value, the smoother the image patch content and the more concentrated the energy is in low-frequency components; the higher the variance value, the more complex the image patch content and the wider the frequency range of energy distribution. The rule for defining the energy concentration area is: perform energy accumulation calculation on the one-dimensional coefficient sequence after zigzag scanning. When the proportion of accumulated energy to the total energy of the current image patch reaches a preset energy threshold, the corresponding sequence index is the frequency interval boundary. The edge computing node adds a preset frequency redundancy offset to the frequency interval boundary as the dynamic cutoff frequency position, allowing the range of low-frequency brightness features to adaptively expand or contract with the smoothness of the image content. When the original image contains a large area of ​​flat tunnel walls, the dynamic cutoff frequency position is lowered to retain more low-frequency information; when it contains dense construction machinery, the dynamic cutoff frequency position is raised to retain more high-frequency texture information. The calculation process for the dynamic cutoff frequency position is quantitatively described by the following formula:

[0081]

[0082] in, For the dynamic cutoff frequency position, This represents the frequency interval boundary corresponding to the energy concentration region of the frequency domain coefficient matrix of the current image block. This is the preset frequency redundancy offset.

[0083] In this embodiment, in order to clarify the adaptive partitioning rules of frequency domain features corresponding to different image content, a dynamic cutoff frequency position and frequency domain feature partitioning parameter table corresponding to different image content types is constructed as shown in Table 3.

[0084] Table 3. Dynamic cutoff frequency position and frequency domain feature division parameters corresponding to different image content types.

[0085]

[0086] In Table 3, the frequency domain coefficient variance feature represents the variance value of the frequency domain coefficient matrix of the current image block. The classification of the variance value is determined based on the statistical distribution characteristics of common image content in tunnel construction scenarios. The boundary of the energy concentration frequency interval is the frequency index value when the accumulated energy reaches a preset energy threshold under the corresponding image content type. This value increases with the increase of image content complexity. The frequency redundancy offset is a preset fixed offset value, used to reserve a certain frequency redundancy based on the boundary of the energy concentration interval, avoiding the misclassification of effective high-frequency information into high-frequency texture features. The dynamic cutoff frequency position increases with the increase of image content complexity, and the corresponding proportion of low-frequency features decreases with the increase of image content complexity, ensuring that more low-frequency brightness information is retained in smooth areas and more high-frequency detail information is retained in complex texture areas, achieving adaptive division of low-frequency and high-frequency features.

[0087] In this embodiment, by dividing the original image into non-overlapping image blocks and performing independent two-dimensional discrete cosine transforms, the computational complexity of frequency domain transformation is reduced, while the division of frequency domain features can adapt to the content features of different image blocks. By converting the two-dimensional frequency domain coefficient matrix into a one-dimensional sequence through zigzag scanning, the ordered arrangement of low-frequency and high-frequency components is achieved, providing a unified benchmark for feature division. By calculating the variance distribution and energy concentration interval of the frequency domain coefficient matrix, the cutoff frequency position is dynamically adjusted, so that the division range of low-frequency brightness features and high-frequency texture features adaptively expands or contracts with the smoothness of the image content. This avoids the problem of excessive high-frequency noise retention in smooth areas or loss of effective details in complex texture areas caused by a fixed cutoff frequency, ensuring the integrity of effective features in the subsequent frequency domain compensation process.

[0088] In yet another alternative embodiment, refer to Figures 4 to 6 During adaptive weight compensation, the edge computing nodes perform bicubic interpolation on the transmittance matrix to align with the size dimension of high-frequency texture features. The spatial size of the transmittance matrix is ​​consistent with the spatial size of the original image, while the high-frequency texture features are frequency domain coefficients obtained from a block-based two-dimensional discrete cosine transform. Therefore, the transmittance matrix needs to be decomposed into transmittance sub-matrices with the same number and size as the image blocks, according to the original image block partitioning method. Bicubic interpolation is performed on each transmittance sub-matrix to ensure that the size of the interpolated transmittance sub-matrix is ​​consistent with the size of the frequency domain coefficient matrix of the corresponding image block, ensuring that each element of the transmittance matrix corresponds one-to-one with the frequency domain coefficients of the high-frequency texture features in space. The bicubic interpolation process calculates the weighted average of the 16 neighboring pixels around the target pixel to obtain the interpolated pixel value, ensuring a smooth numerical transition in the interpolated weight matrix.

[0089] The reciprocal of the transmittance at each location in the transmittance matrix calculated by the edge computing node is used as the basic compensation factor. The calculation process of the basic compensation factor is quantitatively described by the following formula:

[0090]

[0091] in, coordinates The basic compensation factor at the location, This represents the value of the transmittance matrix at the corresponding coordinates. This is a preset minimum constant used to avoid calculation errors caused by a denominator of zero.

[0092] Edge computing nodes perform element-wise multiplication of the basic compensation factor with the absolute value information of the high-frequency texture features, and restore the positive and negative polarities based on the sign bit of the high-frequency texture features, so that the compensated frequency domain coefficients retain their original positive and negative attributes, only adjusting the magnitude of the coefficients. Edge computing nodes perform truncation and smoothing processing on abnormal values ​​in the basic compensation factor that exceed a preset safety threshold, generating compensated high-frequency texture features that suppress diffuse light interference. The truncation and smoothing processing uses a piecewise linear function to map the basic compensation factor, and the calculation process of the piecewise linear function is quantitatively described by the following formula:

[0093]

[0094] in, The compensation factor after truncation processing. The lower limit of the preset safety threshold, This is the preset upper limit of the safety threshold.

[0095] After performing the truncation operation, the edge computing node introduces a morphological closing kernel to smooth the local region of the truncated compensation factor matrix, eliminating block artifacts caused by discontinuities in high-frequency texture feature compensation between adjacent image blocks due to hard truncation. The morphological closing operation first performs dilation and then erosion to fill small holes in the matrix and smooth numerical abrupt changes in adjacent regions, while maintaining the overall size of the matrix unchanged during the operation.

[0096] Edge computing nodes perform matrix recombination and stitching of compensated high-frequency texture features and uncompensated low-frequency brightness features in the frequency domain. Following the reverse order of zigzag scanning, the low-frequency brightness features and compensated high-frequency texture features are recombined into a complete two-dimensional frequency domain coefficient matrix. An inverse discrete cosine transform (ICT) is then performed on the recombined complete frequency domain coefficient matrix to reconstruct the spatially enhanced image. The ICT is the inverse operation of the two-dimensional discrete cosine transform, converting the frequency domain coefficient matrix back into a spatial domain pixel matrix. Each image block generates a corresponding spatial image block after the inverse transform. All spatial image blocks are stitched together according to the original partitioning order, and after removing zero-filled areas at the edges, a complete spatially enhanced image is generated.

[0097] The lightweight convolutional neural network includes depthwise separable convolutional layers and channel attention mechanism layers. The depthwise separable convolutional layers perform spatial feature dimensionality reduction extraction on the spatial enhancement image, and the channel attention mechanism layers recalibrate the weights between channels of the extracted spatial feature map. The recalibrated feature map is then input into the classifier to output the recognition result corresponding to the outline of the construction machinery and the features of the personnel's safety helmet. The depthwise separable convolutional layer contains spatial convolution kernels and pointwise convolution kernels. The spatial convolution kernels perform independent convolution operations on the horizontal and vertical textures in the spatial domain enhanced image using asymmetric rectangular convolution kernels, respectively, to extract the elongated contour features of tunnel construction machinery and the arc-shaped contour features of personnel safety helmets. A 1×k convolution kernel is used to extract horizontal texture features in the horizontal direction, and a k×1 convolution kernel is used to extract vertical texture features in the vertical direction, where k is the size of the convolution kernel. The pointwise convolution kernels concatenate and fuse the horizontal and vertical feature maps obtained from the independent convolution operations along the channel dimension. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function to normalize and nonlinearly transform the feature maps output by the convolution.

[0098] The channel attention mechanism layer calculates the global average pooling and global max pooling features of each channel on the concatenated and fused feature map. It then shares the attention weight coefficients of each channel from the multilayer perceptron layer and multiplies them by the concatenated and fused feature map to achieve channel weight recalibration. The calculation process of the channel attention weights is quantified by the following formula:

[0099]

[0100] in, Given the channel attention weight matrix corresponding to the input feature map F, The activation function is sigmoid, and the MLP is a shared multilayer perceptron. This is a global average pooling operation performed on the input feature map F. This is the global max pooling operation performed on the input feature map F.

[0101] In this embodiment, to clarify the hierarchical structure and operational logic of the lightweight convolutional neural network, a table showing the correspondence between the operational parameters and feature output dimensions of each layer of the lightweight convolutional neural network is constructed as shown in Table 4.

[0102] Table 4. Correspondence between computational parameters and feature output dimensions of each layer in a lightweight convolutional neural network.

[0103]

[0104] In Table 4, H and W represent the height and width of the input spatial domain enhancement image, respectively, in pixels; The target number of categories is preset, including different categories of construction machinery and categories of personnel's helmet wearing status. In the depthwise separable convolutional layer, the convolutional results of the two branches are concatenated along the channel dimension and then fused and transformed using a 1×1 pointwise convolutional kernel. The input and output feature dimensions of the channel attention mechanism layer are kept consistent, and only the feature weights of each channel are recalibrated to increase the weights of channels containing effective target features and suppress the weights of channels containing background and noise. A convolutional layer with a stride of 2 is used to downsample the feature map, reducing the spatial dimension of the feature map and increasing the receptive field; a convolutional layer with a stride of 1 is used to maintain the spatial dimension of the feature map and extract deep detail features. The fully connected classification layer maps the one-dimensional feature vector after global average pooling to the category space and outputs the predicted probability of each category. Finally, the output layer decodes to obtain the recognition results of the construction machinery category, contour position, and personnel's helmet wearing status.

[0105] In this embodiment, bicubic interpolation is used to align the transmissivity matrix with the dimensions of high-frequency texture features, ensuring that the spatial positions of the compensation weights and frequency domain coefficients correspond. The reciprocal of the transmissivity is used as the basic compensation factor, enabling adaptive adjustment to increase the compensation magnitude for higher-frequency coefficients with more severe attenuation. Piecewise linear functions are used to truncate and smooth the basic compensation factor, avoiding image noise amplification caused by overcompensation. Morphological closing operations eliminate block artifacts caused by truncation, ensuring the continuity of the frequency domain features after compensation. Frequency domain matrix reconstruction and inverse discrete cosine transform reconstruct a spatially enhanced image, providing high-quality input data for subsequent feature extraction. A depthwise separable convolutional layer with asymmetric kernels is used to adapt to the extraction of elongated outlines of construction machinery and curved outlines of personnel safety helmets in tunnel construction scenarios, reducing the computational complexity of the network and adapting to the computational constraints of edge computing nodes. A channel attention mechanism layer is used to recalibrate the weights between channels of the feature map, increasing the weights of effective target features, suppressing background noise interference, and ensuring the stability of target recognition results.

Claims

1. An automated tunnel construction monitoring system based on image recognition, characterized in that, It includes edge computing nodes and cloud servers. The edge computing nodes acquire original images of tunnel construction, extract the dark channel prior information and illumination components of the original images, and combine them with the measured data of optical dust concentration sensors deployed in the tunnel to construct an atmospheric scattering physical model specific to the tunnel environment. The edge computing node converts the original image to the frequency domain, extracts low-frequency brightness features and high-frequency texture features using discrete cosine transform, and performs adaptive weight compensation on the high-frequency texture features using the transmittance matrix calculated by the atmospheric scattering physical model to suppress diffuse light interference caused by dust scattering. The edge computing node inputs the compensated frequency domain features into a lightweight convolutional neural network after inverse transformation, extracts the outline of construction machinery and the features of personnel safety helmets, and outputs the recognition results. The measured data of the optical dust concentration sensor is used as a priori parameters to directly participate in the weight calculation process of the frequency domain features. The edge computing node performs local minimum filtering on the original image to obtain the prior information of the dark channel, and uses maximum filtering based on Retinex theory to separate the illumination component; When constructing the atmospheric scattering physical model, the edge computing node reads the measured data of the optical dust concentration sensor as the dust particle scattering coefficient in the current environment, substitutes the scattering coefficient into the modified atmospheric scattering physical equation to replace the fixed constants of the standard atmospheric model, solves the initial transmittance by combining the prior information of the dark channel, and then uses the illumination component to perform guided filtering optimization on the initial transmittance to generate the transmittance matrix. When performing the adaptive weight compensation, the edge computing node performs bicubic interpolation on the transmittance matrix to align with the size dimension of the high-frequency texture features, and calculates the reciprocal of the transmittance at each position in the transmittance matrix as the basic compensation factor. The edge computing node performs element-wise multiplication of the basic compensation factor with the absolute value information of the high-frequency texture feature, restores the positive and negative polarities according to the sign bit of the high-frequency texture feature, and performs truncation and smoothing processing on abnormal values ​​in the basic compensation factor that exceed the preset safety threshold to generate compensated high-frequency texture features that suppress diffuse light interference.

2. The image recognition-based automated tunnel construction monitoring system according to claim 1, characterized in that, Before converting the original image to the frequency domain, the edge computing node divides the original image into image blocks of non-overlapping sizes, and independently performs a two-dimensional discrete cosine transform on each image block to convert the spatial domain pixel matrix into a frequency domain coefficient matrix. The edge computing node sets the zigzag scan cutoff frequency position in the frequency domain coefficient matrix, defines the coefficient matrix region before the cutoff frequency position as the low-frequency brightness feature, and defines the coefficient matrix region after the cutoff frequency position as the high-frequency texture feature, and retains the sign bit and absolute value information of the high-frequency texture feature for subsequent weight compensation.

3. The image recognition-based automated tunnel construction monitoring system according to claim 1, characterized in that, The edge computing node performs matrix recombination and splicing of the compensated high-frequency texture features and the uncompensated low-frequency brightness features in the frequency domain, and performs inverse discrete cosine transform on the recombined complete frequency domain coefficient matrix to reconstruct the spatial domain enhanced image. The lightweight convolutional neural network includes a depthwise separable convolutional layer and a channel attention mechanism layer. The depthwise separable convolutional layer performs spatial feature dimensionality reduction extraction on the spatial enhancement image. The channel attention mechanism layer recalibrates the weights between channels of the extracted spatial feature map. The recalibrated feature map is input into the classifier to output the recognition result corresponding to the outline of the construction machinery and the features of the personnel's safety helmet.

4. The image recognition-based automated tunnel construction monitoring system according to claim 1, characterized in that, In the process of calculating the weight of the frequency domain features, the measured data of the optical dust concentration sensor are used to construct a mapping lookup table between dust concentration values ​​and frequency domain compensation intensity. When the measured data acquired in real time falls into a certain interval of the mapping lookup table, the corresponding compensation intensity adjustment coefficient is output. The edge computing node uses the compensation intensity adjustment coefficient to dynamically shift and adjust the overall numerical distribution of the transmittance matrix, thereby reducing the compensation amplitude for the high-frequency texture features under low dust concentration conditions.

5. The image recognition-based automated tunnel construction monitoring system according to claim 1, characterized in that, When the edge computing node performs guided filtering optimization on the initial transmittance using the illumination component, it constructs a joint guided filtering kernel function that includes spatial distance weights and color similarity weights. The edge computing node acquires the timestamp when the optical dust concentration sensor outputs the measured data and the timestamp when the edge computing node acquires the original image, calculates the time difference between the two timestamps, and when the time difference exceeds the synchronization threshold, it uses a linear interpolation algorithm to calculate the interpolation scattering coefficient aligned with the timestamp of the original image based on the measured data of adjacent moments, updates the initial transmittance using the interpolation scattering coefficient, and then performs the guided filtering optimization.

6. The image recognition-based automated tunnel construction monitoring system according to claim 2, characterized in that, When setting the zigzag scan cutoff frequency position, the edge computing node calculates the variance distribution characteristics of the frequency domain coefficient matrix of the current image block and statistically analyzes the frequency interval boundaries corresponding to the energy concentration region. The edge computing node adds a preset frequency redundancy offset to the frequency interval boundary as the dynamic cutoff frequency position, so that the range of the low frequency brightness feature expands or contracts adaptively with the smoothness of the image content. When the original image contains a large area of ​​flat tunnel wall, the dynamic cutoff frequency position is reduced to retain more low frequency information. When it contains dense construction machinery, the dynamic cutoff frequency position is increased to retain more high frequency texture information.

7. The image recognition-based automated tunnel construction monitoring system according to claim 1, characterized in that, When the edge computing node performs truncation and smoothing processing on abnormal values ​​in the basic compensation factor that exceed the preset safety threshold, it uses a piecewise linear function to map the basic compensation factor, keeping the values ​​within the safety threshold linearly and proportionally mapped, mapping the values ​​greater than the upper limit of the safety threshold to a fixed upper limit constant, and mapping the values ​​less than the lower limit of the safety threshold to a fixed lower limit constant. After performing the truncation operation, the edge computing node introduces a morphological closing operation kernel to perform local region smoothing on the truncated compensation factor matrix.

8. The image recognition-based automated tunnel construction monitoring system according to claim 3, characterized in that, The depth-separable convolutional layer includes a spatial convolution kernel and a pointwise convolution kernel. The spatial convolution kernel performs independent convolution operations on the horizontal and vertical textures in the spatial enhancement image using asymmetric rectangular convolution kernels to extract the elongated outline features of the tunnel construction machinery and the arc-shaped outline features of the personnel's safety helmet. The pointwise convolution kernel concatenates and fuses the horizontal and vertical feature maps obtained from independent convolution operations along the channel dimension. The channel attention mechanism layer calculates the global average pooling and global max pooling features of each channel on the concatenated and fused feature map, and outputs the attention weight coefficients of each channel through a shared multilayer perceptron layer and multiplies them by the concatenated and fused feature map.

Images