Vertical plastic laying machine for saline-alkali soil sealing area and plastic laying monitoring method based on machine vision

Abstract

The invention discloses a vertical plastic laying machine for saline-alkali soil sealing and a plastic laying monitoring method based on machine vision, and belongs to the technical field of salt isolating membrane laying. The vertical plastic laying machine for saline-alkali soil sealing comprises a chain type furrow opener, a membrane placing mechanism and an industrial camera; the membrane placing mechanism realizes smooth laying of the isolating membrane through a membrane releasing roller, a guide roller, a vertical roller and a guide plate, and an industrial camera monitors in real time. According to the invention, isolation membrane laying and real-time quality management and control can be realized, and the operation efficiency and the treatment effect are improved. The plastic laying monitoring method based on machine vision comprises the steps that an isolating membrane image in the laying process is collected, and a corrected and enhanced image is obtained through self-adaptive illumination elimination processing; extracting illumination invariant features fusing multi-channel color constancy and gradient magnitude; a multi-scale residual error is calculated, and a defect sensitive feature graph is generated through morphological topological weight modulation; and in combination with real-time tension fluctuation, four states of normality, wrinkling, hemming and damage are identified.

Description

Technical Field

[0001] This invention belongs to the field of salt isolation membrane laying technology, specifically a vertical plastic laying machine for sealing saline-alkali land and a plastic laying monitoring method based on machine vision. Background Technology

[0002] In the treatment of saline-alkali land, the core technical means to reduce soil salinity and achieve long-term improvement is to block the infiltration of saline-alkali components by vertically laying a saline-alkali intercepting membrane within the soil strata. The key to this technology lies in the quality of the vertical laying of the membrane, which requires vertical membrane laying equipment.

[0003] For the current mainstream vertical geomembrane laying equipment, you can refer to the patented technologies such as the vertical geomembrane laying equipment with authorization announcement number CN 205776403 U and the vertical geomembrane laying device with authorization announcement number CN 205224120 U.

[0004] As can be seen from the above-mentioned patented technologies, existing vertical film laying equipment generally adopts a design with grooves built into the film roll, which has the following drawbacks: Firstly, the membrane placement and splicing processes are inefficient: Current methods require winding the membrane onto a membrane rod and vertically placing it into the trench to unwind it. After a single roll of membrane is laid, the membrane rod must be removed from the trench, a new roll spliced, and then placed back into the trench to continue the work. This process is cumbersome and labor-intensive. Furthermore, membrane replacement and splicing are limited by the trench space, making it difficult to precisely control the sealing and flatness of the heat-fused splicing, thus affecting the treatment effect.

[0005] Secondly, there is a lack of effective quality monitoring methods: it is impossible to detect and warn of defects such as wrinkles, curling edges, and damage that occur during the laying process in real time, resulting in significant quality risks; existing equipment has limited functions, only completing basic laying, and is not combined with intelligent sensing and control technologies, making it impossible to make dynamic adjustments based on the laying status; in the face of complex outdoor lighting environments, conventional image processing methods are unable to reliably extract effective defect features, making it difficult to implement automated detection. Summary of the Invention

[0006] The present invention aims to at least partially solve one of the technical problems in the related art. To this end, the present invention proposes a vertical plastic spreading machine for sealing off saline-alkali land and a plastic spreading monitoring method based on machine vision.

[0007] This technical solution proposes a vertical plastic spreading machine for sealing saline-alkali land, including a chain-type trencher connected to a traction device. The chain-type trencher includes a frame with a trenching chain connected to it. A protective frame is provided above the trenching chain. The protective frame is further connected to a film-laying mechanism, which includes a film-laying roller. One end of the film-laying roller is fixedly connected to the protective frame, and a film roll is mounted on the film-laying roller, allowing the film roll to rotate relative to the film-laying roller. A guide roller is provided below the film-laying roller, with one end fixedly connected to the protective frame. The guide roller is inclined, and the release film is guided by the guide roller to change from vertical to horizontal unwinding. A vertical roller is also provided to the right of the guide roller, connected to the tail end of the protective frame. The vertical roller is also connected to a guide plate, and the release film is released along the guide plate.

[0008] Furthermore, the guide roller has the same structure as the vertical roller, including a sleeve. The sleeve has slots on both sides for the dustproof film to pass through. The sleeve has a built-in rotating shaft, which is rotatably connected to both ends of the sleeve. One end of the sleeve is connected to a connecting rod, which is connected to the protective frame.

[0009] Furthermore, the tail end of the protective frame is detachably connected to a bracket, and the vertical roller is connected to the bracket.

[0010] Furthermore, the protective frame is also connected to a support arm, and an industrial camera for monitoring the plastic film laying process is connected to the end of the support arm.

[0011] This technical solution also proposes a machine vision-based plastic paving monitoring method, based on the vertical plastic paving machine used for saline-alkali land enclosure and management, including the following steps: S1. Acquire image data during the vertical laying process of the isolation membrane; S2. The original isolation film image is processed by an adaptive illumination elimination method that preserves local contrast through multi-scale illumination estimation to obtain an illumination-corrected and enhanced image. S3. An illumination-invariant feature extraction method that integrates the multi-channel color constancy logarithmic ratio and multi-channel gradient magnitude is used to process the illumination-corrected and enhanced image. By locally normalizing each color channel to reduce the influence of illumination intensity, color constancy features are obtained. At the same time, the comprehensive gradient across channels is calculated to capture illumination-independent structural information, resulting in the multi-channel comprehensive gradient magnitude. The color constancy features and the multi-channel comprehensive gradient magnitude are adaptively fused to generate a single-channel illumination-invariant feature that is robust to illumination changes. S4. Calculate the multi-scale residual between the actual gradient magnitude field and the theoretical gradient magnitude field; use nonlinear morphological topological weights for modulation to adaptively enhance the abnormal region that conforms to the physical defect model in spatial and feature dimensions, and obtain the defect-sensitive feature map. S5. A defect-sensitive feature map is processed using a defect identification method based on the coupling of adaptive decision boundary and laying quality. Real-time tension fluctuations during the laying process are introduced as classification priors. By constructing a coupled confidence scoring function, the feature space distance, the geometric properties of the defect area, and the potential risk coefficient to the seepage prevention function are comprehensively considered to complete the defect category identification. Furthermore, S2 includes: The original isolation film image is converted from the red-green-blue color space to a color space that separates luminance and chrominance, and its luminance component is extracted. This component concentrates the main illumination information of the image and is consistent with human visual perception, serving as the basis for illumination estimation, thus obtaining the original luminance component values. A multi-scale Gaussian-Laplacian pyramid is constructed on the extracted original luminance component values. By weighted fusion of Gaussian smoothing results and Laplacian edge information at different scales, an illumination map representing the low-frequency illumination components in the image is generated. The estimated illumination map is used to correct each color channel of the original isolation film image. Non-uniform illumination is eliminated by dividing by the illumination map, and a non-linear contrast enhancement term based on local standard deviation is introduced to adaptively enhance details in flat and textured regions while maintaining color balance, resulting in an illumination-corrected and enhanced image.

[0012] Furthermore, S3 includes: The intensity value of each pixel in the illumination-corrected and enhanced image is compared with the local mean value representing the local illumination level to obtain color constancy features. The horizontal and vertical gradients of each color channel in the illumination-corrected and enhanced image are calculated separately, and the gradient energies of all channels are combined to obtain a multi-channel integrated gradient magnitude map, which can effectively capture the edge information of the image, as well as structural information related to defects and relatively insensitive to changes in overall illumination. The color constancy features of each channel are summed to obtain preliminary brightness-invariant features. At the same time, a spatially adaptive fusion weight is calculated based on the multi-channel integrated gradient magnitude map. This weight is larger in areas with strong gradients and smaller in areas with weak gradients. The weighted gradient magnitude features are added to the preliminary brightness-invariant features to obtain a single-channel illumination-invariant feature map.

[0013] Furthermore, in S4, by introducing the theoretical tension field during the laying process as prior knowledge, the ideal gradient distribution of the membrane surface under defect-free conditions is simulated to obtain the theoretical gradient amplitude field; the actual gradient amplitude is calculated based on the single-channel illumination-invariant feature map; the absolute value of the residual between the actual gradient amplitude field and the theoretical gradient amplitude field is calculated at multiple spatial scales, and the residual is nonlinearly transformed and weighted by local fluctuations to obtain a multi-scale resonance energy map. The calculation at different scales can capture defect features of different sizes. Small scales correspond to subtle anomalies such as damage or scratches, while large scales correspond to macroscopic deformations such as wrinkles or curling; using the multi-channel integrated gradient amplitude map and the local statistical information of the illumination-corrected and enhanced images, a morphological topological weight map that helps identify the topological structure of potential defect regions is constructed. The multi-scale resonance energy and the morphological topological weight map are fused, and the original illumination-invariant feature map is enhanced to obtain a defect-sensitive feature map.

[0014] Furthermore, S5 includes: For each defect category, a prototype vector in the enhanced feature space is defined. This prototype vector is learned from sample features during the training phase. During the inference phase, the Euclidean distance from the feature points to each type of prototype is calculated, and this distance is dynamically adjusted according to the real-time tension fluctuation rate to form an adaptive decision boundary. For each candidate abnormal region in the image, the feature distance similarity, the matching degree between the region's geometric attributes and the category prior, and the risk coefficient of the defect type to the seepage prevention function are fused to calculate the comprehensive confidence score of the region belonging to each category. Finally, the category with the highest confidence score is determined, and the defect identification result of the membrane laying is obtained.

[0015] Furthermore, the categories include normal, wrinkled, rolled edges, and damaged.

[0016] The above technical solution has the following advantages: This invention introduces industrial cameras and image recognition technology to construct a full-process monitoring system from image acquisition, preprocessing, feature enhancement to defect identification. It proposes an adaptive illumination elimination method that combines multi-scale illumination estimation and local contrast preservation, effectively overcoming the interference of uneven outdoor illumination. An illumination-invariant feature integrating the logarithmic ratio of color constancy and the multi-channel gradient amplitude is designed to enhance robustness to illumination changes. Defect feature enhancement is achieved by combining membrane tension priors with multimodal feature resonance, significantly improving the distinguishability of defect features. Finally, an adaptive decision-making method coupling real-time tension fluctuations and defect risk coefficients is used for identification, realizing a deep fusion of engineering physical information and visual perception, and enabling adaptive adjustment of the laying process through closed-loop control.

[0017] The vertical film spreading machine of the present invention, through the optimized structural design of the film placement mechanism, can achieve continuous and smooth laying of the release film, eliminating the need to frequently remove the film rod from the groove for film replacement and splicing, greatly simplifying the operation process and improving film placement efficiency; at the same time, the design of the guide roller, vertical roller and guide plate ensures the stability of the release film laying posture, effectively reducing the occurrence of problems such as film offset and wrinkles. Attached Figure Description

[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below.

[0019] Figure 1 This is a three-dimensional structural view of the vertical plastic spreading machine used for sealing saline-alkali land in Example 1.

[0020] Figure 2 yes Figure 1 Main view of the vertical plastic spreading machine.

[0021] Figure 3 yes Figure 1 Enlarged schematic diagram of the structure of region A in the middle.

[0022] Figure 4 This is a structural diagram of the film feeding roller.

[0023] Figure 5 It is a top-down view of the structure of the vertical roller and the separator.

[0024] Figure 6 This is a three-dimensional structural view of the vertical plastic spreading machine used for sealing saline-alkali land in Example 2.

[0025] Figure 7 This is a flowchart of the steps in the method of Example 3.

[0026] Figure 8 This is a flowchart of the single-channel illumination-invariant feature map obtained in Example 3.

[0027] Explanation of reference numerals in the attached figures: 1. Traction equipment; 2. Frame; 3. Grooving chain; 4. Protective frame; 5. Film roll; 51. Separating membrane; 6. Guide roller; 7. Vertical roller; 71. Sleeve; 711. Groove; 72. Shaft; 8. Film feeding roller; 9. Bracket; 10. Guide plate; 11. Connector; 12. Support arm; 13. Industrial camera. Detailed Implementation

[0028] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are merely illustrative of the technical solution of the present invention and are therefore intended to limit the scope of protection of the present invention.

[0029] Example 1 like Figure 1 - Figure 5 As shown in the figure, this embodiment proposes a vertical plastic paving machine for sealing off saline-alkali land, including a chain-type ditch opener. The chain-type ditch opener is connected to a traction device 1 for driving. The traction device 1 can be an agricultural tractor or an engineering traction machine, providing stable power output for the overall operation and ensuring smooth operation under complex geological conditions of saline-alkali land.

[0030] The chain-type trencher includes a frame 2, on which a trenching chain 3 is connected, and a protective frame 4 is installed above the trenching chain 3. The frame 2 is welded from high-strength alloy steel and has an overall frame structure, possessing good structural rigidity and deformation resistance, and can withstand soil resistance and vibration impact during trenching operations. The trenching chain 3 is rotatably connected to the frame 2 through a sprocket assembly. The cutting teeth of the trenching chain 3 are forged from wear-resistant alloy material, and the spacing of the cutting teeth is reasonably set to ensure high vertical accuracy and smooth sidewalls of the excavated trench, meeting the technical requirements for laying the isolation membrane 51.

[0031] The protective frame 4 is fixedly installed on the upper part of the frame 2, directly above the trenching chain 3. Its main function is to block the mud and stones splashed during the trenching process, so as to avoid damage to other parts of the equipment and prevent safety hazards to the operators.

[0032] The protective frame 4 is also connected to a membrane placement mechanism to achieve orderly unwinding, posture adjustment, and precise laying of the isolation membrane 51. The membrane placement mechanism includes a membrane placement roller 8, one end of which is fixedly connected to the protective frame 4. A membrane roll 5 is fitted onto the membrane placement roller 8, and the membrane roll 5 can rotate relative to the membrane placement roller 8. Specifically, one end of the membrane placement roller 8 is firmly connected to the inner wall of the protective frame 4 by welding or bolting, while the other end is a free end, facilitating the quick placement and replacement of the membrane roll 5. The membrane roll 5 is formed by winding a special isolation membrane 51 for saline-alkali interception and seepage control. The membrane roll 5 is movably fitted onto the membrane placement roller 8, with a clearance fit between the membrane roll 5 and the membrane placement roller 8. The surface of the membrane placement roller 8 is smoothed to ensure that the membrane roll 5 can rotate flexibly relative to the membrane placement roller 8, allowing the isolation membrane 51 to be unwound smoothly during the laying process and avoiding problems such as jamming, tearing, and damage.

[0033] A guide roller 6 is provided below the unwinding roller 8. One end of the guide roller 6 is fixedly connected to the protective frame 4. The connection method can be welding or bolt fastening to ensure the connection strength. The guide roller 6 is set at an angle, and the isolation film 51 is guided by the guide roller 6 to change from vertical unwinding to horizontal unwinding.

[0034] A vertical roller 7 is also provided on the right side of the guide roller 6. The vertical roller 7 is connected to the tail end of the protective frame 4. The vertical roller 7 is also connected to the guide plate 10. The isolation film 51 is released along the guide plate 10. After the isolation film 51 is adjusted to a vertical position by the vertical roller 7, it is released smoothly downward along the inner surface of the guide plate 10 and falls accurately into the groove, realizing the vertical and flat laying of the isolation film 51.

[0035] The guide roller 6 has the same structure as the vertical roller 7, including a sleeve 71. The sleeve 71 has slots 711 on both sides for the dustproof membrane to pass through. The sleeve 71 has a built-in rotating shaft 72, which is rotatably connected to both ends of the sleeve 71. Specifically, the sleeve 71 is made of stainless steel and has corrosion resistance and wear resistance. The sleeve 71 has symmetrical slots 711 on both sides. The width of the slots 711 is slightly larger than the thickness of the isolation membrane 51, allowing the isolation membrane 51 to pass through and acting as a limit to prevent the membrane from shifting left or right during movement. The rotating shaft 72 is built into the axial position of the sleeve 71. The two ends of the rotating shaft 72 are rotatably connected to the sleeve 71 through rolling bearings, so that the sleeve 71 can rotate synchronously with the movement of the isolation membrane 51, minimizing frictional damage between the membrane and the sleeve 71.

[0036] One end of the sleeve 71 is connected to a connecting rod, which is connected to the protective frame 4. The connecting rod is a rigid metal rod, and the other end of the connecting rod is fixedly connected to the protective frame 4 to achieve stable installation of the guide roller 6.

[0037] The tail end of the protective frame 4 is detachably connected to a bracket 9, on which a vertical roller 7 is connected. The bracket 9 and the tail end of the protective frame 4 are detachably connected by bolts. This connection method facilitates adjustment of the installation position and angle of the vertical roller 7 according to operational needs, and also facilitates the disassembly and maintenance of the vertical roller 7. The bracket 9 and the support frame of the protective frame 4 are also equipped with a connector 11, which positions the bracket 9. The connector 11 can be a conventional component such as a wire rope or a tie rod.

[0038] The vertical film spreading machine in this embodiment achieves continuous and smooth laying of the release film 51 by optimizing the structural design of the film placement mechanism. It eliminates the need to frequently remove the film rod from the groove for film replacement and splicing, greatly simplifying the operation process and improving the film placement efficiency. At the same time, the design of the guide roller 6, vertical roller 7 and guide plate ensures the stability of the laying posture of the release film 51, effectively reducing the occurrence of problems such as film offset and wrinkles.

[0039] Example 2 like Figure 6As shown, this embodiment proposes a vertical plastic paving machine for saline-alkali land enclosure and treatment with real-time monitoring function. Its structure is based on the technical solution of Embodiment 1 and improved. The vertical plastic paving machine also adds a monitoring mechanism to solve the drawback of the lack of real-time monitoring of the laying quality of existing equipment. The specific technical solution is as follows: In order to monitor the laying quality of the isolation film 51 in real time, a support arm 12 is also connected to the protective frame 4. The end of the support arm 12 is connected to an industrial camera 13 for monitoring the plastic film laying process. The industrial camera 13 is electrically connected to the controller and the display screen. The controller can be integrated into the cab of the traction equipment 1.

[0040] The industrial camera 13 is a high-definition industrial camera with high resolution and high frame rate. It can clearly capture the real-time status of the isolation membrane 51 during the laying process, and promptly detect quality problems such as wrinkles, curling, damage, and displacement of the membrane. The captured images are fed back to the cab display of the traction equipment 1 in real time through the data transmission module, so that the operator can observe and take adjustment measures in time to ensure the laying quality.

[0041] During operation, the industrial camera 13 transmits the captured images to the display screen in real time via the data transmission module. Operators can directly observe the laying status of the isolation membrane 51. If any quality problems are found, adjustment measures can be taken in a timely manner to achieve dynamic control of the laying quality.

[0042] The vertical plastic film laying machine in this embodiment, while inheriting the advantages of high-efficiency film laying, effectively makes up for the shortcomings of existing equipment in quality control by adding a real-time monitoring mechanism. It can promptly detect and deal with potential quality problems during the laying process, ensure the long-term stability of seepage interception and isolation in saline-alkali land, and provide more reliable equipment support for the management of saline-alkali land.

[0043] Example 3 This embodiment provides a machine vision-based method for monitoring plastic paving, such as... Figure 7 As shown, it includes the following steps: S1. Vertical plastic spreading machine and methods for monitoring spreading quality. To achieve real-time identification and early warning of defects in the installation of the isolation membrane, an industrial camera and specialized image recognition technology are introduced to construct a full-process monitoring system. The specific solution is as follows: Industrial cameras capture the surface condition of the membrane roll in real time during the initial release phase, focusing on detecting defects in the membrane roll itself (such as pre-existing wrinkles and edge damage) and instant wrinkles during the unwinding process.

[0044] S2, Image data acquisition of the isolation membrane Systematically collect image data of the isolation membrane during the vertical laying process to ensure that the collected data can comprehensively and accurately reflect various defects that may occur in actual operations.

[0045] The data acquisition system mainly consists of a high-resolution industrial camera deployed at the exit of the membrane roll. This camera is synchronized with the laying machine's walking encoder and triggers shooting at fixed time or displacement intervals to ensure that the images can continuously cover the surface state of the separator membrane from unwinding to initial entry into the groove.

[0046] The data collection process needs to cover different lighting conditions (such as early morning, noon, and evening), different laying stages (such as the initial section, continuous laying section, and film splicing section), and working scenarios under saline-alkali land conditions with different soil characteristics, in order to construct a diverse and representative image dataset.

[0047] The collected images of the separator membrane are labeled with corresponding categories based on the actual condition of the separator membrane in the images. For example, they are divided into four categories: "normal" (the membrane surface is flat and without abnormalities), "wrinkles" (the membrane surface has wavy or linear raised textures), "edge curling" (the membrane edge is warped or folded), and "damage" (the membrane surface has tears, holes or obvious scratches).

[0048] S3. Image data preprocessing and illumination effect elimination of the isolation membrane During the image acquisition process of the isolation membrane, uneven image brightness is caused by changes in ambient light, machine shadows, and membrane surface reflection. Conventional techniques such as global histogram equalization or Retinex method may over-enhance local areas or fail to adapt to dynamic lighting changes.

[0049] This invention employs an adaptive illumination elimination method that combines multi-scale illumination estimation with local contrast preservation to process the original isolation film image. This method aims to eliminate the effects of uneven illumination while retaining details and color information. The specific steps are as follows: 1) Color space conversion and luminance component extraction The original image of the isolation film is converted from the red-green-blue color space to a color space that separates luminance and chrominance, and its luminance component is extracted. This component contains the main illumination information of the image and is consistent with human visual perception, forming the basis for illumination estimation. It is represented as follows: , In the formula, Indicates the pixel position The original luminance component value is obtained by fusing the intensity values ​​of the red, green and blue channels according to a specific weight. The specific weight is set based on the human visual sensitivity to the brightness of different colors and is preset as the standard color space conversion coefficient. This indicates the row coordinate index in the image, i.e., the position in the vertical direction; This indicates the column coordinate index in the image, i.e., the position in the horizontal direction; This indicates the location of the original isolation membrane image. The pixel intensity value of the red channel; This indicates the location of the original isolation membrane image. The pixel intensity value of the green channel; This indicates the location of the original isolation membrane image. The pixel intensity value of the blue channel.

[0050] 2) Illumination estimation based on multi-scale Gaussian Laplace pyramid A multi-scale Gaussian Laplacian pyramid is constructed based on the extracted raw luminance component values. By weighted fusion of Gaussian smoothing results and Laplacian edge information at different scales, an illumination map representing the low-frequency illumination components in the image is generated, as follows: , In the formula, Indicates the location The estimated illumination intensity value at that location represents the low-frequency illumination component, and is obtained by weighted fusion of multi-scale Gaussian smoothing component and Laplacian edge component. This represents the total number of layers in the multi-scale pyramid, i.e., the total number of scales used. It is a preset hyperparameter, with an example value of 4. Indicates the index of the current scale. ,in This usually represents the coarsest scale (maximum smoothing). Indicates the first The weighting coefficients of the Gaussian smoothing components corresponding to each scale are used to control the contribution of low-frequency information in the final illumination map at that scale, and are set in the following way: ; The standard deviation is expressed as A two-dimensional Gaussian filter kernel is used for the first... The brightness image is smoothed at each scale to extract low-frequency illumination information corresponding to that scale. The standard deviation of the two-dimensional Gaussian filter kernel is used. The value increases with the scale, and the setting method is expressed as follows: ; Represents the two-dimensional convolution operator; This represents the global balance weighting coefficient, used to adjust the influence of high-frequency edge information on the final illumination map estimation. An example value is 0.1. Indicates the first The weighting coefficients of the Laplacian components corresponding to each scale are used to control the contribution of edge information to the illumination map at that scale, and are set in the following way: .

[0051] 3) Adaptive correction and local contrast enhancement based on illumination map Each color channel of the original isolation film image is corrected using the estimated illumination map, by dividing by the illumination map to eliminate uneven illumination, and a non-linear contrast enhancement term based on local standard deviation is introduced to adaptively enhance details in flat and textured regions while maintaining color balance, expressed as: , In the formula, Indicates illumination correction and enhancement of images In the Each color channel, position The intensity values ​​at that location, after being processed by illumination normalization and local contrast enhancement, are more robust to changes in illumination and have clearer details. Indicates the color channel index. These represent the red, green, and blue channels, respectively. Represents the original image of the isolation membrane. In the Each channel, location The intensity value at that location; This represents a very small positive integer, used to prevent the denominator from being zero and to ensure numerical stability. Examples of its values ​​are shown below. ; Indicates the first The contrast enhancement intensity coefficient for each channel controls the extent of detail enhancement for that channel. Considering that the human visual system is most sensitive to green, followed by red, and least sensitive to blue, the optimal setting is chosen to maintain a natural color balance while enhancing detail. , , . This represents the hyperbolic tangent activation function, used to constrain the enhancement term to... Within the specified range, prevent image distortion caused by over-enhancement; Indicates the first Each channel is located in The local standard deviation within a certain neighborhood window is used to quantify the local contrast or texture richness at that location, and is used to correct the image in the middle. The above is obtained by calculating the standard deviation of pixel intensity within this neighborhood; This represents the intermediate corrected image, which is the initial illumination corrected image. It serves as an intermediate feature between illumination correction and enhancement. In position Strength value . This represents the standardization parameter, used to scale the range of local standard deviations and control the starting position of the saturation region of the hyperbolic tangent activation function. An example value is 10.

[0052] In practical implementation, a neighborhood of a certain size is used... A rectangular window centered on the hyperparameter is used, the size of which is sufficient to calculate meaningful local statistics, but not so large as to obscure local details. For example... A window of pixels.

[0053] S4. Illumination-invariant feature extraction from isolation membrane image data Images that have been illuminated and enhanced may still retain non-uniform illumination or color distortion introduced by the correction. Features obtained by directly using color intensity values ​​or simple transformations are still sensitive to changes in illumination. Conventional methods using color histograms will lose key spatial structure information, while texture descriptors such as local binary modes are not adaptable to changes in overall illumination and color.

[0054] This invention employs an illumination-invariant feature extraction method that integrates the logarithmic ratio of multi-channel color constancy with the gradient magnitude of multi-channel gradients. It weakens the influence of illumination intensity by locally normalizing each color channel, while simultaneously calculating the comprehensive gradient across channels to capture illumination-independent structural information. The two methods are then adaptively fused to generate a single-channel illumination-invariant feature map robust to illumination changes. Figure 8 As shown, the specific steps are as follows: 1) Calculate the log-ratio characteristic of multi-channel color constancy For each color channel of the illumination-corrected and enhanced image, its color constancy feature is calculated. Specifically, by simulating the local adaptation ability of human vision, the intensity value of each pixel is compared with the local mean value representing the local illumination level, thereby suppressing the influence of local illumination changes, as expressed as: , In the formula, Indicates the first Each color channel, position The color constancy eigenvalues ​​at a given location are obtained by taking the natural logarithm of the local normalized intensity values, which can effectively suppress local light intensity variations. Indicates the first The contribution weight coefficients of each channel can be used to adjust the importance of different color channels in the feature. To simplify and maintain a balanced contribution of each channel, it is preferable to set the contribution weight coefficients of the three channels to the same value, for example... If you wish to emphasize the characteristics of a particular channel, you can adjust accordingly. This represents a logarithmic function, with the default base being the natural constant. Indicates the first Each channel is located in The local mean value within a certain neighborhood is used as an estimate of the local ambient light level at that location, in illumination correction and image enhancement. The above is obtained by averaging the pixel intensity within the neighborhood.

[0055] In practical implementation, a neighborhood of a certain size is used... A rectangular window centered on the hyperparameter is used, the size of which is sufficient to calculate meaningful local statistics, but not so large as to obscure local details. For example... A window of pixels.

[0056] It should be noted that human visual perception of brightness depends not only on absolute light intensity, but also on the local context, that is, the average brightness of the surrounding area. This method simulates the adaptation process of the visual system to the local background by calculating the ratio of the intensity of each pixel to the average intensity of its local neighborhood, thereby suppressing the influence of uniform illumination changes and preserving relative contrast information.

[0057] 2) Calculate the multi-channel integrated gradient magnitude The horizontal and vertical gradients are calculated for each color channel of the illumination-corrected and enhanced image, and the gradient energies of all channels are combined to obtain a multi-channel integrated gradient magnitude map. This map effectively captures edge information and structural information related to defects that is relatively insensitive to changes in overall illumination, and is represented as follows: , In the formula, Indicates the location The multi-channel integrated gradient magnitude at the location can enhance edges that are significant in multiple channels, and is insensitive to changes in illumination. It is a multi-channel integrated gradient magnitude map The value of the element in the x-th row and y-th column; Indicates the first Each color channel is located in position The horizontal gradient component at a certain point is used for illumination correction and image enhancement. exist The approximate value of the spatial first derivative in the (horizontal) direction is obtained by using the Sobel gradient operator. Indicates the first Each color channel is located in position The vertical gradient component at a certain point is used for illumination correction and image enhancement. exist The approximate value of the spatial first derivative in the (vertical) direction is obtained by using the Sobel gradient operator.

[0058] 3) Adaptive feature fusion to generate illumination-invariant feature maps The color constancy features of each channel are summed to obtain the preliminary brightness-invariant features. Simultaneously, a spatially adaptive fusion weight is calculated based on the multi-channel integrated gradient magnitude map. This weight is larger in regions of strong gradients and smaller in regions of weak gradients. The weighted gradient magnitude features are then added to the preliminary brightness-invariant features to obtain the single-channel illumination-invariant feature map, represented as follows: , In the formula, Indicates the location Spatial adaptive fusion weights at the location, with a value range of This is used to dynamically adjust the contribution of gradient features to the final feature. It is obtained by mapping the local average gradient magnitude through a Sigmoid-like function. The value is larger in texture edge regions and smaller in flat regions. The calculation method is expressed as follows: ; Represents the natural exponential function; Represents adaptive fusion weights in the control space The hyperparameter that increases with the rate of increase of the input value; the larger the value, the steeper the weight function curve. An example value is 5.0. Indicates the multi-channel integrated gradient magnitude at a level of The local mean value within a certain neighborhood centered at a certain size is used to smooth noise and obtain an overall gradient level representation of the region; it is a multi-channel integrated gradient magnitude map. In It is the arithmetic mean of a certain size neighborhood window centered on the center, obtained through average pooling operation; Indicates the location The single-channel illumination-invariant eigenvalues ​​at a given location, combined with the local illumination invariance provided by color constancy and the structural information provided by the integrated gradient, exhibit robustness to both global and local illumination changes. It is an illumination-invariant feature map The value of the element in the x-th row and y-th column.

[0059] In practical implementation, a neighborhood of a certain size is used... A rectangular window centered on the hyperparameter is used, the size of which is sufficient to calculate meaningful local statistics, but not so large as to obscure local details. For example... A window of pixels.

[0060] It should be noted that, The term represents the position of the red, green, and blue channels. The color constancy eigenvalues ​​at the location are summed to obtain preliminary fusion characteristics, which can effectively suppress the influence of local illumination changes.

[0061] S5, Enhanced Image Defect Features of Separator Film Different types of laying defects may not be significantly distinguishable in the feature space. Conventional edge enhancement or frequency domain filtering methods often homogenize the entire image and cannot selectively enhance the physical form of specific defects such as periodic ripples of wrinkles, high curvature edges of curled edges, and random sharp fracture patterns of damage.

[0062] This invention employs a defect feature enhancement method combining prior membrane tension with multimodal eigenre resonance. By introducing the theoretical tension field during the laying process as prior knowledge, it simulates the ideal gradient distribution on the membrane surface in a defect-free state, calculates the multi-scale residual between the actual gradient amplitude field and the theoretical gradient amplitude field, and modulates it using nonlinear morphological topological weights. This adaptively enhances anomalous regions that conform to the physical defect model in both spatial and feature dimensions. The specific steps are as follows: 1) Construction of the ideal tension field and characteristic gradient field based on laying parameters Under ideal laying conditions, the release film should be in a uniform tension state, and its surface should appear as a flat area in the image. Combining the traveling speed of the paving machine, the speed of the unwinding roller, and the theoretical elastic modulus of the film, a theoretical gradient amplitude field characterizing the flatness of the film surface under ideal conditions is constructed. The value of this theoretical gradient amplitude field approaches zero in flat areas and has a preset theoretical gradient value at known mechanical clamps or seam positions. At the same time, the actual gradient amplitude field is calculated using an illumination-invariant feature map. By comparing the theoretical gradient amplitude field with the actual gradient amplitude field, abnormal gradient areas caused by defects can be highlighted. a) Theoretical gradient magnitude field definition Represents the theoretical gradient magnitude field. Indicates the location The theoretical gradient magnitude at a certain point is used to simulate the smoothness distribution of a defect-free ideal film surface, and its calculation method is expressed as follows: , In the formula, This represents the baseline gradient coefficient, reflecting the background gradient level caused by slight natural sagging of the membrane surface. It is negatively correlated with the laying speed and the elastic modulus of the membrane material. An example value is shown below. ; Indicates the column coordinate index of a line in the image; This represents the horizontal decay coefficient, which controls the rate at which the gradient decays from the sides of the image toward the center. An example value is one-quarter of the image width. This indicates the total number of known fixed feature points in the image, such as the location of mechanical clamps or pre-set seams; Indices representing fixed feature points ; Indicates the first The theoretical gradient peak at a fixed feature point can be obtained statistically from defect-free calibration images. That is, at the known fixed feature point position, the average gradient amplitude of multiple ideal laying images near that position can be calculated. Alternatively, it can be estimated by mechanical simulation based on the physical shape of the fixture and the pressure on the membrane. Indicates the first The row coordinates of a fixed feature point; Indicates the first The column coordinates of a fixed feature point; Indicates the first The spatial radius of the influence of a peak can be statistically determined from a defect-free calibration image by observing the spatial distance from the center of the gradient peak attenuation to the background level, or it can be calculated based on the physical dimensions of the fixture according to the image resolution.

[0063] It should be noted that the theoretical gradient magnitude field This is used to model the theoretical distribution of the gradient (i.e., brightness variation) of the film surface in an image under ideal (defect-free) layup conditions. The simulated gradient is a gentle brightness gradient that gradually increases from the image center to the edges due to slight natural sagging of the membrane surface or viewing angle. It is a baseline gradient, strongest at the edges and weakest at the center line. This simulation addresses the local gradient effects caused by known fixed feature points (such as mechanical clamps and seams) on the membrane surface. These points compress or lift the membrane surface, creating localized, attenuated gradient peaks around them. The theoretical gradient amplitude field is then used to... It can be compared with the actual observed gradient field. Areas where the two differ significantly are abnormal gradient regions that may be caused by unexpected defects (such as wrinkles or damage).

[0064] It should also be noted that components or preset positions that will inevitably come into contact with the film surface physically and may change its local flatness will produce specific textures or edges in the image, even in an ideal defect-free state. Therefore, they need to be modeled into the theoretical field to avoid misjudging them as defects.

[0065] b) Actual gradient magnitude field definition Represents the actual gradient magnitude field. Indicates the location The actual gradient magnitude at that point is derived from the illumination-invariant feature map. The calculated actual gradient magnitude is expressed as follows: , In the formula, Represents the gradient operator; This represents the L2 norm, also known as the Euclidean norm.

[0066] In practical implementation, illumination-invariant feature maps partial derivatives and It is obtained by approximate calculation using discrete difference methods such as the Sobel operator.

[0067] 2) Multi-scale residual resonance energy calculation The absolute value of the residual between the actual gradient magnitude field and the theoretical gradient magnitude field is calculated at multiple spatial scales. This residual is then subjected to a nonlinear transformation and local fluctuation weighting to obtain a multi-scale resonance energy map. Calculations at different scales can capture defect characteristics of different sizes. Small scales correspond to subtle anomalies such as damage or scratches, while large scales correspond to macroscopic deformations such as wrinkles or curling edges, as shown below: , In the formula, Indicates the first Scale, location The absolute value of the gradient residual at a given location represents the magnitude of the deviation between the observed gradient and the ideal gradient at that location under a specific smoothing scale. It is calculated as follows: ; Indicates scale index. ; This represents the total number of levels in the multi-scale pyramid; it is a preset hyperparameter, and examples of its values ​​are shown below. ; Indicates the first Scale, location The resonant energy at that location characterizes the scale of that position. The following represents the confidence level of the defect; the higher the value, the higher the probability of a defect. Indicates the first The energy threshold at each scale is used to filter out small noise residuals and can be set as a multiple of the average value of the theoretical gradient magnitude field at that scale (e.g., 0.1-0.5 times). Indicates the first The scaling parameter for each scale is used to adjust the sensitivity of the nonlinear transformation and can be set as the standard deviation estimate of the residual distribution at that scale. Indicates the use of smoothing the actual gradient magnitude field The A Gaussian kernel of each scale, Standard deviation It increases with increasing scale, for example Small-scale analysis focuses on subtle anomalies, while large-scale analysis focuses on macroscopic deformations. Represents the use of smoothing theoretical gradient magnitude fields The A Gaussian kernel of each scale, Standard deviation Slightly larger than the same scale ,For example This is to make the theoretical field smoother and avoid oversensitivity to minute theoretical features; Indicates that in The local standard deviation of the absolute value of the gradient residual within a certain neighborhood window centered on the center is used to measure the degree of fluctuation of the residual in that region. This represents the standardization parameter, used to scale the numerical range of local standard deviations. Examples of possible values ​​are shown below. ; This parameter represents the intensity of the contribution of local volatility. It is used to amplify the energy of regions where the residuals are not only large in absolute value but also fluctuate violently in their local neighborhood. Examples of values ​​are provided. .

[0068] In practical implementation, the Gaussian kernel and Standard deviation with scale index Increase it as you go; for example, you can set the Gaussian kernel. The standard deviation is Gaussian kernel The standard deviation is The neighborhood window size is a preset hyperparameter, for example... Pixel.

[0069] In practical implementation, a neighborhood of a certain size is used... A rectangular window centered on the hyperparameter is used, the size of which is sufficient to calculate meaningful local statistics, but not so large as to obscure local details. For example... A window of pixels.

[0070] 3) Nonlinear morphological topological weight modulation and feature fusion enhancement By utilizing multi-channel integrated gradient magnitude maps and local statistical information from illumination-corrected and enhanced images, a morphological topological weight map is constructed to aid in identifying the topological structure of potential defect regions. Multi-scale resonant energy is fused with the morphological topological weight map, and the original illumination-invariant feature map is enhanced to obtain a defect-sensitive feature map, represented as follows: , In the formula, Indicates the location The morphological topological weight at a location indicates that a larger value suggests that the location is more likely to belong to a defective topological structure. The calculation method is expressed as follows: ,and It is a morphological topological weight graph The value of the element in the x-th row and y-th column; This represents the global enhancement intensity coefficient, which controls the overall magnitude of defect feature enhancement. An example value is shown below. ; Indicates the first The fusion weights of the resonance energies at each scale are used to weight and sum the resonance energies at different scales to form a comprehensive energy map. Their values ​​can be pre-set according to the significance of different defect types at different scales; for example, small defects are more dependent on finer scales, so a weighted value is preferred. ; This indicates taking the smaller of the two values; This represents the upper limit pruning threshold for topological weights, preventing excessively large weights at individual points from causing feature distortion. An example value is... . This indicates a morphological dilation operation. This indicates a morphological erosion operation. Morphological dilation and erosion operations use small-sized structural elements, such as... Rectangular structural elements; This represents the multi-channel integrated gradient magnitude map. In position The value after performing morphological dilation operation at that location; This represents the multi-channel integrated gradient magnitude map. In position The value after performing morphological etching operation at the location; This indicates the positive part operation, which sets a negative number to zero. Indicates illumination correction and enhancement of images In The local standard deviation within a neighborhood window of a certain size centered on the target; Indicates illumination correction and enhancement of images In The local mean within a certain neighborhood window centered on the locant; Indicates the location The intensity value of the defect-sensitive feature map at the location, and It is a defect-sensitive feature map The value of the element in the x-th row and y-th column; This represents the element-wise multiplication operator.

[0071] In practical implementation, a neighborhood of a certain size is used... A rectangular window centered on the hyperparameter is used, the size of which is sufficient to calculate meaningful local statistics, but not so large as to obscure local details. For example... A window of pixels.

[0072] S6. Constructing a defect identification system for the isolation membrane laying. Conventional multi-class classifiers may ignore the differences in spatial distribution, morphological continuity, and degree of harm to seepage prevention performance of different defects, and are sensitive to class imbalance in training data.

[0073] This invention employs a defect identification method based on the coupling of adaptive decision boundary and laying quality. It not only utilizes enhanced features but also introduces real-time tension fluctuations during the laying process as a classification prior. By constructing a coupled confidence scoring function, it comprehensively considers the feature space distance, the geometric attributes of the defect area, and the potential risk coefficient to the seepage prevention function, ultimately achieving accurate identification of defect categories. The specific steps are as follows: 1) Calculation of the coupled physical prior category prototype and dynamic decision boundary For each defect category, a prototype vector is defined in the enhanced feature space. This prototype vector is learned from sample features during the training phase. During the inference phase, the Euclidean distance from feature points to each prototype is calculated, and this distance is dynamically adjusted according to the real-time tension fluctuation rate to form an adaptive decision boundary, expressed as: , In the formula, Indicate category The prototype vector is the average center of the feature vectors of all training samples of this class, and is calculated as follows: ; Indicates the defect category index. ; Indicates position Features to categories The prototype's adjusted Euclidean distance is the distance modulated by physical priors, when the normalized tension fluctuation rate... When the value is increased, the adjustment distance for defect categories that are sensitive to tension fluctuations will decrease, thereby widening the decision boundary of the category and making the system more sensitive to the detection of such defects when the operating conditions are unstable. Indicates the position Global average pooling is performed on the local feature blocks centered on the model to convert the spatial feature map into feature vectors. The size of the local feature blocks is a preset hyperparameter, for example... Pixel; Indicate category The intrinsic boundary relaxation factor reflects the natural dispersion of this type of feature in space. Specifically, it is calculated by taking the average distance from the feature vector of all samples of each class to its prototype during the training phase, and then normalizing it as the value of the feature vector. Intrinsic boundary relaxation factor ; Indicates that the training set belongs to category The total number of samples; Indicates the sample index; Indicates belonging to a category The set of training sample indices; Indicates the first Defect-sensitive feature maps of each training sample; Indicate category The sensitivity coefficient to tension fluctuations is used to quantitatively represent the sensitivity of different defect categories to tension fluctuations. The normalized tension fluctuation rate incorporates real-time physical process states into the classification decision of image features, enabling the system to adaptively adjust the classifier's sensitivity according to operating conditions, thereby improving the defect detection rate under abnormal operating conditions. The calculation method is expressed as follows: ,in, This represents the real-time tension fluctuation value. The rated tension is a preset engineering parameter that can be set according to the membrane material characteristics, laying speed, and designed laying depth. For example, for a specific membrane material and a laying speed of 0.5 m / s, the rated tension can be set. .

[0074] In one implementation, the real-time tension fluctuation value Based on the traveling speed of the paving machine and actual unwinding roll speed The simulation was conducted, specifically using the travel speed of the paving machine based on the ideal speed-rotation speed matching equation. Obtain the theoretical unwinding roll speed Then the real-time fluctuation value can be simulated as ,in The simulated tension coefficient can be calibrated based on empirical data or determined through regression analysis of historical data.

[0075] In one implementation, the normalized tension volatility Settings can be based on expert knowledge or historical data, for example: "damage" is the most sensitive ( "Folds" is the next best word. "Curled edges" are generally ( “Normal” is not sensitive ( ) 2) Defect category determination based on a combination of geometric risk and confidence level For each candidate anomalous region in the image, the similarity of feature distance, the matching degree between the region's geometric attributes and the prior class, and the risk coefficient of this type of defect to the anti-seepage function are fused to calculate its comprehensive confidence score for each category. The category with the highest confidence score for that region is then determined, denoted as: , , In the formula, Indicates candidate anomaly regions Category The comprehensive confidence score integrates low-level features, mid-level geometric attributes, and high-level risk priors. The higher the score, the higher the confidence level for that category. The candidate anomaly region is a connected region extracted from the enhanced feature map through thresholding; Indicates candidate anomaly regions Features to categories at all coordinate locations within the range The average value of the adjusted Euclidean distance of the prototype; This represents the geometric feature weight, used to balance the contributions of shape features and curvature features. Examples of possible values ​​are shown below. ; This indicates the Sigmoid activation function, which has a threshold effect, significantly reducing its contribution when the matching degree is below a certain level. Indicates the measurement of candidate anomaly regions Geometric properties and categories A function of prior knowledge matching degree; The mean is Standard deviation is Gaussian distribution in The probability density value at that location; Indicates candidate anomaly regions The average curvature of the boundary is used to first extract candidate anomaly regions. The contour is defined, and then the curvature of each point on the contour is calculated using an image curvature estimation algorithm (such as the rate of change of the angle between adjacent contour point vectors). Finally, the absolute values ​​of the curvature of all points are averaged to obtain candidate anomaly regions. Mean curvature of the boundary ; Indicate category The mean of typical boundary curvature is obtained by calculating the average boundary curvature of the true defect area for each category of samples, and then calculating the mean of the curvature of all samples in that category. Indicate category The typical boundary curvature standard deviation is obtained by calculating the average boundary curvature of the true defect area for each category of samples, and then calculating the standard deviation of the curvature of all samples in that category. Indicate category The basic risk coefficient is an hyperparameter, set by expert knowledge based on the degree of harm that defects cause to the seepage prevention function; This represents the risk sharpening index, used to amplify the scores of high-risk categories. An example value is shown below. ; Indicates candidate anomaly regions The final classification Indicates taking such that Category index that reaches the maximum value .

[0076] In one implementation, the function Candidate anomaly regions can be calculated Geometric characteristics (such as aspect ratio, density, extensibility, etc.) and category The similarity between preset typical geometric feature templates (such as cosine similarity or reciprocal Euclidean distance) is obtained. For example, wrinkles may appear as thin stripes (high aspect ratio), while damage may appear as irregular patches (low density).

[0077] In one implementation, categories Basic risk coefficient Based on expert knowledge or engineering experience, defects are ranked according to their potential harm to the seepage prevention function and assigned numerical values. For example, the basic risk coefficient for the normal category is set to 1.0, the basic risk coefficient for the rolled edge category is set to 1.5, the basic risk coefficient for the wrinkled category is set to 2.0, and the basic risk coefficient for the damaged category is set to 3.0. The higher the value, the higher the risk of the defect and the more it is amplified in the scoring.

[0078] In practical implementation, candidate anomaly regions This refers to the defect-sensitive feature map The image segmentation technique extracts connected regions that may contain defects; specifically, it identifies defect-sensitive feature maps. Thresholding segmentation (such as adaptive or fixed thresholding) is applied, followed by connectivity component analysis of the binary image. Each connected component is then identified as a candidate anomaly region. .

[0079] In the specific implementation, when the final category is determined Not "normal" and its corresponding overall confidence score When the preset alarm threshold is exceeded, the system triggers the corresponding defect alarm.

[0080] Example 4 In this embodiment, a control system can also be deployed on the vertical plastic spreading machine to achieve closed-loop adaptive adjustment of the vertical plastic spreading machine's spreading strategy based on the identification results of the separation film spreading defect. Specifically, when the system identifies a defect and triggers an alarm, it not only provides audible and visual warnings and records defect information, but also feeds back key parameters such as the defect type, location, confidence level, and real-time calculated normalized tension fluctuation rate to the control system of the plastic spreading machine. Based on a pre-set adjustment strategy library, this system automatically generates and executes corresponding correction commands. For example, if a "wrinkle" defect is detected, the control system may determine that the film tension is insufficient, and then slightly increase the traveling speed of the paving machine to increase the film tension and smooth out the wrinkles; if a "curling" or "damage" defect is detected, the system will immediately issue the highest level alarm and instruct the paving machine to stop moving and wait for the operator to inspect and handle it on-site to prevent the defect from expanding further.

[0081] In addition, if the tension fluctuation rate continuously monitored by the system remains at a high level for a long period of time, it can provide an early warning of abnormal operating conditions, even if no specific defects are detected, prompting operators to check the mechanical components or the installation status of the membrane roll.

[0082] Through a closed-loop process of "real-time monitoring - intelligent identification - dynamic adjustment", the original passive quality inspection is transformed into proactive process quality control. This enables the vertical plastic laying machine to adaptively optimize its operating parameters based on real-time feedback on the laying quality, thereby reducing defects at the source and ensuring the final quality and long-term seepage prevention stability of the vertical laying of the isolation membrane.

[0083] The above description is merely a preferred embodiment of the present invention and does not limit the implementation and protection scope of the present invention. Those skilled in the art should realize that any equivalent substitutions and obvious changes made based on the description and illustrations of the present invention should be included within the protection scope of the present invention.

Claims

1. A vertical plastic paving machine for sealing saline-alkali land, comprising a chain-type trencher, the chain-type trencher being driven and connected to a traction device; the chain-type trencher includes a frame, a trenching chain connected to the frame, and a protective frame provided above the trenching chain, characterized in that, The protective frame is also connected to a film placement mechanism, which includes a film feeding roller. One end of the film feeding roller is fixedly connected to the protective frame, and a film roll is sleeved on the film feeding roller. The film roll can rotate relative to the film feeding roller. Below the film feeding roller, there is a guide roller. One end of the guide roller is fixedly connected to the protective frame. The guide roller is inclined, and the isolation film is guided by the guide roller to change from vertical unwinding to horizontal unwinding. On the right side of the guide roller, there is also a vertical roller, which is connected to the tail end of the protective frame. The vertical roller is also connected to a guide plate, and the isolation film is released along the guide plate.

2. The vertical plastic paving machine for sealing saline-alkali land according to claim 1, characterized in that, The guide roller has the same structure as the vertical roller, including a sleeve with slots on both sides for the dustproof film to pass through. The sleeve has a built-in rotating shaft, which is rotatably connected to both ends of the sleeve. One end of the sleeve is connected to a connecting rod, which is connected to the protective frame.

3. The vertical plastic paving machine for sealing saline-alkali land according to claim 2, characterized in that, The tail end of the protective frame is detachably connected to a bracket, and the vertical roller is connected to the bracket.

4. The vertical plastic paving machine for sealing saline-alkali land according to claim 3, characterized in that, The protective frame is also connected to a support arm, and an industrial camera for monitoring the plastic film laying process is connected to the end of the support arm.

5. A machine vision-based method for monitoring plastic paving, characterized in that, Based on the vertical plastic paving machine for saline-alkali land enclosure and management according to any one of claims 1-4, the specific method is as follows: S1. Acquire image data during the vertical laying process of the isolation membrane; S2. The original isolation film image is processed by an adaptive illumination elimination method that preserves local contrast through multi-scale illumination estimation to obtain an illumination-corrected and enhanced image. S3. An illumination-invariant feature extraction method that integrates the multi-channel color constancy logarithmic ratio and multi-channel gradient magnitude is used to process the illumination-corrected and enhanced image. By locally normalizing each color channel to reduce the influence of illumination intensity, color constancy features are obtained. At the same time, the comprehensive gradient across channels is calculated to capture illumination-independent structural information, resulting in the multi-channel comprehensive gradient magnitude. The color constancy features and the multi-channel comprehensive gradient magnitude are adaptively fused to generate single-channel illumination-invariant features. S4. Calculate the multi-scale residual between the actual gradient magnitude field and the theoretical gradient magnitude field; use nonlinear morphological topological weights for modulation to adaptively enhance the abnormal region that conforms to the physical defect model in spatial and feature dimensions, and obtain the defect-sensitive feature map. S5. A defect-sensitive feature map is processed using a defect identification method based on the coupling of adaptive decision boundary and laying quality. Real-time tension fluctuations during the laying process are introduced as classification priors. By constructing a coupled confidence scoring function, the defect category identification is completed by comprehensively considering the feature space distance, the geometric attributes of the defect area, and the potential risk coefficient to the seepage prevention function.

6. The machine vision-based plastic laying monitoring method according to claim 5, characterized in that, S2 include: The original isolation film image is converted from the red-green-blue color space to a color space with separated luminance and chrominance, and its luminance component is extracted to obtain the original luminance component values. A multi-scale Gaussian Laplacian pyramid is constructed on the extracted original luminance component values. By weighted fusion of Gaussian smoothing results and Laplacian edge information at different scales, an illumination map representing the low-frequency illumination components in the image is generated. The estimated illumination map is used to correct each color channel of the original isolation film image. Non-uniform illumination is eliminated by dividing by the illumination map, and a non-linear contrast enhancement term based on local standard deviation is introduced to obtain the illumination-corrected and enhanced image.

7. The machine vision-based plastic laying monitoring method according to claim 6, characterized in that, S3 include: The color constancy feature is obtained by comparing the intensity value of each pixel in the illumination-corrected and enhanced image with the local mean value that represents the local illumination level. For each color channel of the illumination-corrected and enhanced image, the gradients in the horizontal and vertical directions are calculated separately, and the gradient energies of all channels are combined to obtain a multi-channel integrated gradient magnitude map. The color constancy features of each channel are summed to obtain preliminary brightness-invariant features. At the same time, based on the multi-channel integrated gradient magnitude map, spatially adaptive fusion weights are calculated, and the weighted gradient magnitude features are added to the preliminary brightness-invariant features to obtain a single-channel illumination-invariant feature map.

8. The machine vision-based plastic laying monitoring method according to claim 7, characterized in that, In S4, the theoretical tension field during the laying process is introduced as prior knowledge to simulate the ideal gradient distribution of the membrane surface under defect-free conditions, thus obtaining the theoretical gradient amplitude field. The actual gradient amplitude is calculated based on the single-channel illumination-invariant feature map. The absolute value of the residual between the actual gradient amplitude field and the theoretical gradient amplitude field is calculated at multiple spatial scales, and the residual is nonlinearly transformed and weighted by local fluctuations to obtain a multi-scale resonance energy map. Using the multi-channel integrated gradient amplitude map and the local statistical information of the illumination-corrected and enhanced images, a morphological topological weight map is constructed to help identify the topological structure of potential defect regions. The multi-scale resonance energy map is fused with the morphological topological weight map, and the original illumination-invariant feature map is enhanced to obtain a defect-sensitive feature map.

9. The machine vision-based plastic laying monitoring method according to claim 8, characterized in that, S5 include: For each defect category, a prototype vector in the enhanced feature space is defined. This prototype vector is learned from sample features during the training phase. During the inference phase, the Euclidean distance from the feature points to each type of prototype is calculated, and this distance is dynamically adjusted according to the real-time tension fluctuation rate to form an adaptive decision boundary. For each candidate abnormal region in the image, the feature distance similarity, the matching degree between the region's geometric attributes and the category prior, and the risk coefficient of the defect type to the seepage prevention function are fused to calculate the comprehensive confidence score of the region belonging to each category. Finally, the category with the highest confidence score is determined, and the defect identification result of the membrane laying is obtained.

10. The machine vision-based plastic laying monitoring method according to claim 9, characterized in that, The categories include normal, wrinkled, rolled edges, and damaged.

Images