Image recognition-based automatic alarm method for abnormal behavior of an experimental bench

Abstract

The application discloses an experimental table safety abnormal behavior automatic alarm method based on image recognition and relates to the technical field of automatic alarm, which comprises the following steps: determining the fume hood area corresponding to the experimental table, obtaining a geometric reference map based on the original image frame of the fume hood area; performing cumulative imaging on the difference image corresponding to the geometric reference map to obtain a hinged contact belt map; performing local contrast change on the geometric reference map based on the hinged contact belt map to obtain a texture inversion development image; obtaining a wing surface rolling texture map based on the texture inversion development image; obtaining an adhesion potential field map based on the exhaust structure mask image corresponding to the geometric reference map and the wing surface rolling texture map; and automatically alarming the adhesion migration abnormal event of the sheet-shaped fragments in the fume hood area according to the adhesion potential field map. The application improves the hidden abnormality recognition capability and stability.

Description

Technical Field

[0001] This invention relates to the field of automatic alarm technology, and in particular to an automatic alarm method for abnormal safety behavior of experimental benches based on image recognition. Background Technology

[0002] Fume hoods are essential safety devices in laboratories used to control the diffusion of harmful gases and maintain a local negative pressure exhaust environment. They are widely used in chemical analysis, biological experiments, reagent preparation, material handling, and hazardous sample transfer. Laboratory personnel typically perform operations such as unfolding weighing paper, wrapping aluminum foil, cutting sealing film, sample dispensing, and liquid transfer on the fume hood's work surface. During these operations, thin, flaky residues such as weighing paper fragments, aluminum foil pieces, and sealing film fragments are easily generated. Under the continuous negative pressure airflow inside the fume hood, these flaky fragments can spread along the surface of the fume hood. The work surface of the test bench, the surface of the rear baffle, and the edge of the air guide slits may adhere, lift, roll up, and migrate, gradually moving closer to the rear baffle opening and the air guide slit area. When the flaky fragments enter the vicinity of the exhaust structure, they can easily form local obstructions to the air guide slits and exhaust inlets, thereby changing the original airflow path inside the fume hood, causing local backflow, exhaust disturbance, and a decrease in the ability to capture harmful gases. In severe cases, it may also cause volatile harmful gases, dust, or vapors generated during the experiment to be unable to be discharged in time, thus affecting the safety of the experimental environment and the personal safety of the experimental personnel.

[0003] Existing technologies typically employ manual inspection, fixed-threshold image detection, or ordinary motion detection to assess abnormal conditions inside fume hoods. Manual inspection relies on subjective observation by researchers, making it difficult to promptly detect sheet-like debris slowly migrating along the rear baffle and air guide slits. Fixed-threshold image detection is susceptible to changes in lighting, reflection, and background disturbances, making it difficult to accurately distinguish the true attachment and migration status of sheet-like debris. Ordinary motion detection primarily targets overall moving objects and cannot identify abnormal states such as continuous edge attachment, localized wing roll-up, and slow migration along the exhaust path formed by sheet-like debris under negative pressure airflow. Therefore, existing technologies struggle to accurately determine whether sheet-like debris is continuously approaching the exhaust structure area and cannot visualize the spatial relationship between sheet-like debris and the rear baffle, air guide slits, and exhaust inlet. Consequently, they cannot provide stable and reliable automatic alarms for attachment and migration anomalies that may cause exhaust blockage risks. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies that make it difficult to visualize the spatial relationship between sheet-like fragments and the rear baffle, air guide slit, and exhaust inlet, and to propose an automatic alarm method for abnormal safety behavior of experimental benches based on image recognition.

[0005] To address the problems existing in the prior art, the present invention adopts the following technical solution: An automatic alarm method for abnormal safety behaviors of experimental benches based on image recognition includes: S1. Determine the fume hood area corresponding to the experimental bench, and obtain the geometric reference map based on the original image frame of the fume hood area; S2. Accumulate and image the differential image corresponding to the geometric reference image to obtain the articulated contact zone image; S3. Based on the articulated contact zone map, perform local contrast changes on the geometric reference map to obtain a texture inversion development image; based on the texture inversion development image, obtain the wing surface rollover texture map. S4. Obtain the attraction potential field map based on the exhaust structure mask image and the wing surface rollover texture map corresponding to the geometric reference map; S5. Based on the attraction potential field diagram, automatically alarm for abnormal events of adhesion and migration of sheet-like debris in the fume hood area.

[0006] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention acquires original image frames from multiple perspectives through wireless networking, and obtains a geometric reference map through distortion correction, cross-view geometric registration and image fusion. This ensures that the experimental platform working surface, back baffle and guide slit are in a unified image coordinate system, which can stably express the relative position of sheet-like fragments and exhaust structure, reduce the impact of single-view occlusion, view skew and imaging distortion on the recognition results, and provide a consistent image basis for subsequent alarms.

[0007] 2. This invention obtains the articulated contact zone map by boundary zone cumulative imaging and morphological closing operation, and then extracts the observation strips of the wing surface along the articulated edge to generate texture inversion development image and wing surface rollover texture map. It can continuously image-represent the edge attachment, local lifting and rollover migration state of sheet-like fragments, avoid misjudgment and missed detection caused by relying solely on overall motion detection, and improve the ability and stability of hidden anomaly identification.

[0008] 3. This invention generates an adhesion potential field map by using an exhaust structure mask image, an exhaust distance image, and a wing surface rollover texture map. The adhesion potential field map is then converged to obtain a risk-driven signal, thereby generating alarm control information. This ensures that the alarm object, alarm area, and alarm intensity correspond to the degree of adhesion and migration of the sheet-like debris to the guide slit and the rear baffle opening, thereby improving the accuracy, interpretability, and practicality of the automatic alarm. Attached Figure Description

[0009] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings: Figure 1This is a flowchart illustrating an automatic alarm method for abnormal safety behavior of an experimental platform based on image recognition, provided in an embodiment of the present invention. Detailed Implementation

[0010] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0011] This embodiment provides an automatic alarm method for abnormal safety behavior of an experimental platform based on image recognition. See [link to relevant documentation]. Figure 1 Specifically, including: S1. Determine the fume hood area corresponding to the experimental bench, and obtain the geometric reference map based on the original image frame of the fume hood area; In an embodiment of the present invention, obtaining a geometric reference diagram includes: Identify the fume hood area from the workstation area of ​​the laboratory bench; The work area of ​​the experimental bench refers to the workspace where experimental personnel perform operations such as sampling, weighing, transferring, and observing. The fume hood area refers to the exhaust work space in the experimental bench work area, which includes the fume hood work surface, the back baffle, and the air guide slit. The back baffle is a plate-like structure located behind the fume hood work surface and connected to the exhaust duct, used to guide the airflow inside the fume hood towards the exhaust duct. The air guide slit is a narrow opening structure formed on the back baffle, used to allow air inside the fume hood to enter the exhaust duct through the air guide slit, thus forming an airflow path from the experimental bench work surface towards the exhaust duct.

[0012] Specifically, the workstation area of ​​the experimental bench is photographed to obtain a workstation observation image including the experimental bench work surface, the fume hood work surface, the back baffle, and the guide slit. The work surface boundary, the vertical boundary of the back baffle, the opening boundary of the guide slit, and the side wall boundary of the fume hood are extracted from the workstation observation image. The range of the fume hood work surface is determined according to the intersection of the work surface boundary and the vertical boundary of the back baffle. The exhaust side boundary is determined according to the position of the guide slit opening boundary on the back baffle. The left and right boundaries are determined according to the side wall boundary of the fume hood. The area of ​​the experimental bench work surface that is continuously connected to the fume hood work surface, the area where the back baffle is located, and the area where the guide slit is located are collectively determined as the fume hood area.

[0013] The raw image frames of the fume hood area are acquired through wireless networking. Specifically, camera nodes are arranged at corresponding positions on the front, sides, and rear baffle of the fume hood area, so that the shooting range of each camera node covers a part of the fume hood area. Each camera node is connected to the same wireless mesh network through a wireless communication link. The wireless mesh network transmits the image data and acquisition time information collected by each camera node. The image data with consistent acquisition time information are used as the original image frames of the fume hood area at the same time. The original image frames include at least one image frame covering the junction of the experimental table working surface and the fume hood working surface, at least one image frame covering the rear baffle, and at least one image frame covering the air guide slit.

[0014] Distortion correction is performed on the original image frames to obtain corrected image frames; Specifically, the pixel grayscale distribution in each original image frame is continuously scanned. Edge pixels are extracted at locations where pixel grayscale changes continuously. These edge pixels are then connected according to their spatial positions to form the edges of the experimental work surface, fume hood sidewalls, rear baffles, and air guide seams. Straight lines are fitted to each of these edges to obtain the fitted straight lines. The fitted straight lines are compared with the actual edge positions in the original image frame to calculate the positional offset of the actual edge positions relative to the fitted straight lines. Based on the positional offset, a pixel coordinate correction relationship is established for each pixel position in the original image frame. Each pixel in the original image frame is then remapped to a new image coordinate position according to the pixel coordinate correction relationship. The grayscale values ​​of surrounding adjacent pixels are extracted from the empty areas formed by the remapping. Grayscale interpolation is then performed to fill the empty areas, resulting in a corrected image frame with continuous edge positions and straight edges restored to a flat state.

[0015] Extract structural features of the fume hood from the corrected image frame; Wireless networking refers to the establishment of data transmission relationships between multiple camera nodes through wireless communication; the original image frame refers to image data directly acquired by the camera node without correction processing; distortion correction processing refers to the image processing process of correcting edge curvature, proportional offset and position deviation caused by lens imaging in the original image frame; the structural features of the fume hood refer to the image content in the corrected image frame that can characterize the fixed structural position of the fume hood, including the edge of the back baffle, the edge of the guide slit and the edge of the working surface. Specifically, gradient calculations are performed on the pixel grayscale values ​​in the corrected image frame to obtain the direction and intensity of grayscale changes at each pixel position. Pixel regions with continuous grayscale change intensity and consistent grayscale change direction are identified as candidate edge regions. Connectivity analysis is performed on the pixels in the candidate edge regions, and interconnected edge pixels are combined to form edge contours. The direction of the edge contours is statistically analyzed to obtain the edge direction of the experimental platform working surface, the edge direction of the fume hood side wall, the edge direction of the rear baffle, and the edge direction of the guide slit. Edge contours with the same edge direction and continuous spatial position are connected to form structure lines. The experimental platform working surface region, the fume hood rear baffle region, and the guide slit region are determined based on the regions enclosed by the structure lines. The structure lines and corresponding regions are used as the structural features of the fume hood.

[0016] Based on the structural features of the fume hood, cross-view geometric registration processing is performed on the corrected image frame to obtain the registered image frame. Cross-view geometric registration refers to the process of converting corrected image frames from different shooting angles into the same image coordinate relationship according to the positional relationship of the fixed structure of the fume hood; registered image frames refer to image frames that have a unified spatial correspondence after cross-view geometric registration. Specifically, structural lines and structural regions belonging to the same fume hood entity structure are selected from each calibration image frame. The entity structure includes the edge of the experimental table working surface, the edge of the fume hood side wall, the edge of the rear baffle opening, and the edge of the guide slit opening. The endpoints, intersections, and corners of the structural lines corresponding to the same entity structure in different calibration image frames are used as registration points. The image coordinates of one calibration image frame are used as reference coordinates. Based on the positional relationship between the registration points in each calibration image frame and the corresponding registration points in the reference coordinates, the planar projection transformation relationship from each calibration image frame to the reference coordinates is calculated. The pixels in each calibration image frame are remapped to the reference coordinates according to the planar projection transformation relationship to obtain a registered image frame with the same spatial correspondence with the reference coordinates.

[0017] Image fusion processing is performed on the registered image frames to obtain a geometric reference map.

[0018] Image fusion processing refers to the process of combining the image content corresponding to the same fume hood area from multiple registered image frames into a single image; geometric reference map refers to the image obtained by fusing registered image frames and used to represent the fixed spatial relationship of the fume hood area.

[0019] Specifically, each registered image frame is placed in the same reference coordinate system. Pixels representing the same spatial location in each registered image frame are aligned. When multiple pixel values ​​exist at the same spatial location, the coordinate offset distance between the corresponding pixels and the spatial location is obtained. The corresponding distance weight is calculated based on each coordinate offset distance. The distance weight is multiplied by the corresponding pixel value and summed. The sum is then divided by the sum of all distance weights to obtain the fused pixel value at that spatial location. When only one pixel value exists at the same spatial location, that pixel value is retained. When no pixel value exists at the same spatial location, interpolation is performed based on the pixel values ​​of the surrounding adjacent spatial locations. The pixels after fusion, retention, and interpolation are combined to form a geometric reference map covering the experimental platform working surface, the fume hood back panel, and the guide slit.

[0020] S2. Accumulate and image the differential image corresponding to the geometric reference image to obtain the articulated contact zone image; In an embodiment of the present invention, a hinged contact zone diagram is obtained, including: Identify flaky debris in the fume hood area; Flake-like fragments refer to thin, sheet-like solid fragments located on the working surface of a fume hood, the working surface of an experimental table, the surface of a back panel, or near a flow guide slit. Flake-like fragments have a structural form where the thickness is less than the length and width, and can undergo edge rolling, surface adhesion, and position migration under the action of airflow. Flake-like fragments include weighing paper corner pieces, aluminum foil corner pieces, and sealing film fragments. Specifically, the working surface of the fume hood, the working surface of the experimental table, the surface of the rear baffle, and the periphery of the guide slit are defined as the fragment search range in the geometric reference map. The grayscale of the pixels within the fragment search range is smoothed to reduce imaging noise. The grayscale gradient of the smoothed image is calculated to obtain the grayscale change position between the bearing surface and the solid edge. The interconnected grayscale change positions are connected to form closed or nearly closed candidate contours. For each candidate contour, the contour length, contour width, circumscribed rectangle, edge bending point, internal grayscale continuous area, and contact position with the bearing surface are extracted. The candidate contour corresponding to the candidate contour with a sheet-like extended shape of circumscribed rectangle, a bent or warped shape of edge, and internal grayscale continuous and located within the fragment search range is determined as a sheet-like fragment area.

[0021] Edge extraction is performed on the geometric reference image to obtain the boundary zone image of the sheet-like fragments; The boundary zone image of a sheet-like fragment refers to the continuous strip-shaped image region formed around the outer contour of the sheet-like fragment in the geometric reference map. The pixels in the boundary zone image are used to represent the location regions where the gray-level changes between the sheet-like fragment and the surrounding background are continuous. Specifically, the edge processing range is defined as the region of the sheet-like fragment and the adjacent bearing surface outside the region. The grayscale of the pixels within the edge processing range is calculated to obtain the grayscale change of each pixel position along the horizontal and vertical directions. The grayscale changes in the two directions are combined into a grayscale change intensity map. The position of continuous grayscale change is traced along the outer contour of the sheet-like fragment region in the grayscale change intensity map to obtain the outer contour line of the sheet-like fragment. The neighborhood of the outer contour line of the sheet-like fragment is expanded so that the adjacent pixels inside the outer contour line belonging to the sheet-like fragment and the adjacent pixels outside the outer contour line belonging to the bearing surface together form a continuous strip region. The continuous strip region is written into an image with the same size as the geometric reference map, and the pixels outside the continuous strip region are set to blank to obtain the boundary zone image of the sheet-like fragment.

[0022] Generate a difference image based on the geometric reference map; A difference image is an image formed by calculating the grayscale difference between the geometric reference image at the current time and the geometric reference image at the adjacent time according to the same pixel coordinates. The pixel grayscale values ​​in the difference image are used to represent the degree of change of the same spatial location at different times. Specifically, the geometric reference image at the current moment and the geometric reference image at the adjacent moment are obtained under the same camera viewpoint and the same image coordinate relationship. The geometric reference image at the current moment and the geometric reference image at the adjacent moment are paired according to the same pixel coordinates. The difference of the pixel gray value at each pair of the same pixel coordinates is calculated and the absolute value is taken. The obtained absolute difference is written into the corresponding pixel coordinate position to form a difference image with the same size as the geometric reference image.

[0023] Based on the boundary zone image, cumulative imaging processing is performed on the difference image to obtain the boundary zone cumulative image; Boundary band cumulative image refers to an image formed by continuously superimposing the pixel gray values ​​in the difference images corresponding to multiple time points within the pixel range defined by the boundary band image of the sheet fragment. The high gray-level regions in the boundary band cumulative image are used to represent the location regions where the edge of the sheet fragment continuously changes position. Specifically, the boundary zone image and the difference image are placed under the same image coordinate relationship. The cumulative range in the difference image is determined according to the continuous band-shaped region marked in the boundary zone image. Within this cumulative range, the pixel gray value of the difference image at the current time is extracted. The pixel gray value of the difference image at the current time is added to the historical cumulative gray value already formed at the same pixel coordinate. The unmarked positions in the boundary zone image are left blank. The gray value after addition is written into an image with the same size as the geometric reference image to obtain the boundary zone cumulative image.

[0024] Morphological closing operations are performed on the cumulative boundary zone image to obtain the articulated contact zone image.

[0025] Morphological closing operation refers to an image processing operation that dilates the bright areas in the cumulative boundary zone image and then erodes them. The dilation process is used to expand the range of the bright areas corresponding to the edges of the sheet-like fragments, and the erosion process is used to restore the boundary morphology of the bright areas, thereby connecting the bright areas that are close to each other in the cumulative boundary zone image. The articulated contact zone image refers to the strip-shaped image used to represent the contact relationship between the edges of the sheet-like fragments and the working surface of the experimental table, the working surface of the fume hood, the surface of the back baffle, or the edge of the guide slit. The strip-shaped area in the articulated contact zone image corresponds to the edge part of the sheet-like fragment that does not completely detach from the bearing surface as the sheet-like fragment is rolled up as a whole. This edge part forms the contact fulcrum when the sheet-like fragment attaches, lifts, and migrates.

[0026] Specifically, pixels with gray values ​​in the boundary band cumulative image are identified as pixels to be processed. Pixels adjacent to the edges and corners of each pixel to be processed are identified as neighboring pixels. The gray value of the pixel to be processed is copied to the position of the neighboring pixels that does not have a gray value. This expands the neighborhood into a connected gray region, where closely adjacent but unconnected banded gray regions in the boundary band cumulative image are then processed. Each pixel at the edge of the connected gray region is checked to see if it can find the corresponding gray support position of the original pixel to be processed in its adjacent edge and corner neighborhood. Pixels that cannot find a gray support position are removed from the connected gray region, causing the outer edge of the connected gray region to shrink back to the banded range corresponding to the sheet-like fragment boundary band. The remaining connected gray region is written into a blank image with the same size as the geometric reference image, and pixels outside the connected gray region are left blank, resulting in the articulated contact band map.

[0027] S3. Based on the articulated contact zone map, perform local contrast changes on the geometric reference map to obtain a texture inversion development image; based on the texture inversion development image, obtain the wing surface rollover texture map. In an embodiment of the present invention, obtaining a texture inversion developed image includes: Based on the articulated contact zone diagram, determine the articulated edge region of the sheet-like fragment in the geometric reference diagram; The articulated edge region refers to the image area located at the edge of the sheet fragment and in contact with the working surface of the experimental table, the working surface of the fume hood, the surface of the back baffle, or the edge of the guide slit. The articulated edge region corresponds to the edge part of the sheet fragment that can still maintain contact under the action of air flow. This edge part is used to limit the sheet fragment from detaching from the bearing surface as a whole and to form a contact support area when the sheet fragment undergoes local rolling and position migration. Specifically, the articulated contact zone image and the geometric reference image are mapped to each other using the same image width, image height, and pixel coordinate origin. A continuous strip-shaped pixel region with grayscale values ​​is extracted from the articulated contact zone image. The pixel coordinates of each continuous strip-shaped pixel region are mapped to the same pixel coordinate position in the geometric reference image. The corresponding edge pixels of the sheet-like fragment, the fragment surface pixels located inside the edge of the sheet-like fragment, and the bearing surface pixels located outside the edge of the sheet-like fragment are extracted from the geometric reference image with the mapped pixel coordinate position as the center. The continuous image region formed by the edge pixels of the sheet-like fragment, the fragment surface pixels, and the bearing surface pixels is determined as the articulated edge region of the sheet-like fragment. The bearing surface is the working surface of the experimental table, the working surface of the fume hood, the surface of the back baffle, or the surface where the edge of the guide slit is located.

[0028] Based on the articulated edge region, strip regions are extracted from the geometric reference map to obtain the observation strip image of the wing surface; The wing surface observation strip image refers to a continuous strip image area extending from the articulated edge region toward the non-contact side of the sheet fragment. The wing surface observation strip image corresponds to the thin sheet surface area of ​​the sheet fragment that can undergo changes such as lifting, bending, rolling and adhering relative to the articulated edge region. This strip image area covers the air contact area near the edge of the sheet fragment as well as the surface area of ​​the sheet fragment. In an embodiment of the present invention, based on the articulated edge region, strip region extraction is performed on the geometric reference map to obtain an observation strip image of the wing surface, including: Boundary point sampling is performed on the hinged edge region to obtain a set of hinged edge points; Boundary point sampling refers to the process of extracting continuously distributed edge pixels point by point along the edge direction of the hinged edge region. The pixels obtained by boundary point sampling are used to represent the position distribution of the contact between the edge of the sheet fragment and the working surface of the experimental table, the working surface of the fume hood, the surface of the back baffle, or the edge of the guide slit. The hinged edge point set refers to the set of edge pixels obtained by boundary point sampling and arranged in the edge extension order. The edge pixels in the hinged edge point set correspond to the edge positions in the sheet fragment that form contact support. Specifically, the hinged edge region is placed under the same pixel coordinate relationship as the geometric reference map. Continuous edge lines connecting the edge of the sheet-like fragment to the bearing surface are extracted in the hinged edge region. Adjacent connected edge pixels are traced pixel by pixel along the extension direction of the continuous edge lines. The horizontal and vertical coordinates of each traced edge pixel are recorded as a hinged edge point. All hinged edge points are arranged in the tracing order along the continuous edge lines to obtain a set of hinged edge points.

[0029] Determine the normal direction information of the set of hinged edge points; Specifically, for each hinged edge point in the set of hinged edge points, the preceding and following hinged edge points adjacent to it in the arrangement order are selected. The edge extension direction at the hinged edge point is determined according to the coordinate change from the preceding to the following hinged edge point. The direction perpendicular to the edge extension direction and facing the interior of the sheet-like fragment body is determined as the normal direction of the hinged edge point. The normal directions corresponding to each hinged edge point are recorded according to the arrangement order of the set of hinged edge points to obtain the normal direction information of the set of hinged edge points.

[0030] Based on the set of hinged edge points and normal direction information, a strip region mask image is generated in the geometric reference map; Specifically, each hinge edge point in the set of hinge edge points is used as the starting pixel for strip generation. The normal direction information corresponding to the hinge edge point is used as the pixel extension direction from the hinge edge to the interior of the sheet-like fragment body. In the geometric reference map, adjacent pixels are visited sequentially from the starting pixel along the pixel extension direction. During the visit, the pixel positions still located on the surface of the sheet-like fragment body are recorded. The surface of the sheet-like fragment body is determined by the pixel regions that are continuously distributed within the sheet-like fragment area and have grayscale differences from the surrounding bearing surfaces. When the boundary position of the surface of the sheet-like fragment body is visited, the pixel extension corresponding to the hinge edge point is stopped. The pixel positions corresponding to all hinge edge points are merged into strip candidate regions. Adjacent connected pixels in the strip candidate regions are connected and sorted to connect the extended pixels generated by adjacent hinge edge points into continuous strip regions. The pixel positions of the continuous strip regions are written into a blank image with the same size as the geometric reference map. The pixel positions not written into the continuous strip regions are kept blank, resulting in a strip region mask image.

[0031] Based on the strip region mask image, strip regions are extracted from the geometric reference image to obtain the wing surface observation strip image.

[0032] Normal direction information refers to the data in the vertical direction relative to the extension direction of the set of hinge edge points. Normal direction information is used to represent the extension direction from the hinge edge region toward the interior of the sheet fragment body. Strip region mask image refers to the strip pixel identification image formed in the geometric reference map based on the set of hinge edge points and the corresponding normal direction information. The strip region in the strip region mask image corresponds to the thin sheet surface region in the sheet fragment that can undergo changes such as lifting, rolling and attaching relative to the hinge edge region.

[0033] Specifically, the strip region mask image and the geometric reference image are mapped to each other using the same image width, image height, and pixel coordinate origin. The pixel positions of the continuous strip region that have been written are extracted from the strip region mask image. The corresponding pixels are located in the geometric reference image according to the pixel positions of the continuous strip region. The grayscale value or color value of the corresponding pixel is copied. The copied grayscale value or color value is written into the same pixel coordinate position in a blank image with the same size as the geometric reference image. For local missing pixels in the continuous strip region caused by edge bending, the pixel values ​​of its adjacent strip region are interpolated to fill in the missing pixels. The pixel positions outside the continuous strip region are left blank, forming a wing surface observation strip image that only contains the wing surface region of the sheet-like fragment extending from the hinge edge into the body.

[0034] Based on the observed strip image of the wing surface, local contrast changes are performed on the geometric reference image to obtain a texture inversion and development image.

[0035] Local contrast variation refers to the change in pixel grayscale difference within the corresponding area of ​​the strip image observed on the wing surface. Local contrast variation is used to represent the relationship between the light and dark changes formed between the surface of the sheet-like fragment and the surrounding air, the bearing surface, and the illumination. Texture inversion developed image refers to the image formed by local contrast variation within the corresponding area of ​​the strip image observed on the wing surface. The area of ​​light and dark variation in the texture inversion developed image is used to represent the area of ​​light and dark reversal formed on the surface of the sheet-like fragment due to changes in the direction of warping, changes in the surface adhesion state, or changes in the direction of surface reflection.

[0036] Specifically, the wing surface observation strip image and the geometric reference image are mapped using the same image width, image height, and pixel coordinate origin. The written wing surface strip pixel positions are extracted from the wing surface observation strip image. For each wing surface strip pixel position, a local pixel region consisting of a center pixel and surrounding adjacent pixels is established in the geometric reference image. The grayscale values ​​of all pixels within the local pixel region are obtained. The difference between the grayscale value of the center pixel and the grayscale values ​​of each adjacent pixel within the local pixel region is calculated, and the absolute value is taken. All the obtained absolute differences are summed and divided by the number of adjacent pixels to obtain the local contrast change value of the corresponding wing surface strip pixel position. The contrast change value is written to the same pixel coordinate position in a blank image with the same size as the geometric reference image. The local contrast change value corresponding to the pixel position of the adjacent wing strip is subjected to continuous gray-scale transition processing, so that the local contrast change value is continuously distributed on the surface of the sheet fragment. The gray-scale is enhanced for the position where the local contrast change value is higher than the adjacent area, and the gray-scale is weakened for the position where the local contrast change value is lower than the adjacent area. This makes the area of ​​light and dark change on the surface of the sheet fragment, caused by changes in the direction of the warp, changes in the attachment state, and changes in surface reflection, form a continuous bright and dark flip area in the image. The pixel position outside the observation strip area of ​​the wing surface is kept blank, resulting in a texture inversion development image.

[0037] In an embodiment of the present invention, obtaining the wing surface rollover texture map includes: Based on the observed stripe image of the wing surface, extract the current stripe region image from the geometric reference image at the current moment; Specifically, the geometric reference image at the current moment and the wing surface observation strip image are mapped to each other using the same image width, image height, and pixel coordinate origin. The pixel positions of the continuous strip regions that have been written are extracted from the wing surface observation strip image. The corresponding pixels are located in the geometric reference image at the current moment according to the pixel positions of the continuous strip regions. The grayscale value or color value of the corresponding pixel is copied. The copied grayscale value or color value is written into the same pixel coordinate position in a blank image with the same size as the geometric reference image at the current moment. The pixel positions outside the continuous strip regions are left blank, thus obtaining the current strip region image.

[0038] Based on the observed strip images of the wing surface, images of adjacent strip regions are extracted from the geometric reference images at adjacent time points; The current strip region image refers to the image region extracted from the geometric reference image at the current moment according to the pixel positions corresponding to the strip images observed on the wing surface. The pixels in the current strip region image are used to represent the surface state, edge morphology, brightness distribution, and local upturn state of the sheet debris wing surface region at the current moment. The adjacent strip region image refers to the image region extracted from the geometric reference image at an adjacent moment according to the same strip positions as the current strip region image. The pixels in the adjacent strip region image are used to represent the surface state, edge morphology, brightness distribution, and local upturn state of the sheet debris wing surface region at the adjacent moment. Specifically, the geometric reference image and the wing surface observation strip image at adjacent time points are correlated using the same image width, image height, and pixel coordinate origin. The pixel positions of the continuous strip regions that have been written are extracted from the wing surface observation strip image. The corresponding pixels are located in the geometric reference image at adjacent time points according to the pixel positions of the continuous strip regions. The grayscale value or color value of the corresponding pixel is copied. The copied grayscale value or color value is written into the same pixel coordinate position in a blank image with the same size as the geometric reference image at adjacent time points. The pixel positions outside the continuous strip regions are left blank, thus obtaining the adjacent strip region image.

[0039] Pixel displacement is calculated between the current strip region image and the adjacent strip region images to obtain the strip region pixel displacement information; Specifically, the current strip region image and adjacent strip region images are mapped to the same image width, image height, and pixel coordinate origin. Strip pixel positions with grayscale values ​​are extracted from the current strip region image. A current pixel block, consisting of the center pixel, its adjacent edge pixels, and corner pixels, is extracted from each strip pixel position. In adjacent strip region images, multiple candidate pixel blocks are extracted along the horizontal and vertical adjacent pixel directions, centered on the same pixel coordinate position. Each candidate pixel block consists of a candidate center pixel, its adjacent edge pixels, and corner pixels. The difference between the grayscale values ​​of the corresponding positions in the current pixel block and each candidate pixel block is calculated and its absolute value is taken. The absolute differences of each corresponding position are added together to obtain the pixel block difference value. The candidate pixel block with the smallest pixel block difference value is determined as the corresponding pixel block of the current pixel block in the adjacent strip region image. Based on the horizontal and vertical coordinate differences between the center pixel coordinates of the current pixel block and the center pixel coordinates of the corresponding pixel block, the strip region pixel displacement information of that strip pixel position is obtained.

[0040] Generate a motion field image of the strip region based on the pixel displacement information of the strip region; Strip region pixel displacement information refers to the positional change data of the pixel position in the current strip region image relative to the corresponding pixel position in the adjacent strip region image. Strip region pixel displacement information is used to represent the local displacement changes formed when the sheet debris wing surface region undergoes bending, lifting, rolling, and attachment changes between adjacent time points; strip region motion field image refers to the image formed based on strip region pixel displacement information. Different image regions in the strip region motion field image correspond to the motion direction, motion range, and motion continuity of different positions of the sheet debris wing surface region. Specifically, the pixel displacement information of the strip region is arranged with the current strip region image using the same pixel coordinate relationship. In a blank image with the same size as the current strip region image, a corresponding motion recording position is established for each strip pixel position. The horizontal and vertical coordinate differences corresponding to the strip pixel position are converted into displacement amplitude. The displacement amplitude is written into the blank image as the gray value of the motion recording position. For pixel positions between adjacent strip pixel positions that have not been written with gray values, the gray value that is closest to them in space is used to fill them. The gray values ​​corresponding to the displacement amplitude are processed to make the gray value changes between adjacent strip pixel positions continuous, thus obtaining a strip region motion field image that represents the local displacement distribution within the observed strip on the wing surface.

[0041] The motion field image of the strip region is developed to obtain a deformation motion developed image.

[0042] Deformation motion developed image refers to the image formed after developing the motion field image of the strip area. The bright and dark change areas in the deformation motion developed image are used to represent the local bending, local rolling, local adhesion and local lifting changes of the wing surface area of ​​the sheet-like debris under the action of airflow. Specifically, the pixel positions in the strip region motion field image that record displacement amplitude are used as development processing positions. For each development processing position, the corresponding gray value and the gray values ​​of adjacent pixel positions are extracted. Based on the change relationship between the gray values ​​of the development processing position and the gray values ​​of adjacent pixel positions, local motion development gray values ​​are formed. The local motion development gray values ​​are written into the same pixel coordinate positions in a blank image with the same size as the strip region motion field image. The missing pixels between adjacent development processing positions are filled by interpolation using the local motion development gray values ​​of the adjacent development processing positions. The gray value distribution after interpolation is processed to be continuous, so that the local motion changes caused by lifting, attaching, rolling and position migration in the wing area of ​​the sheet-like fragment are presented as a continuous gray area, thus obtaining the deformation motion development image.

[0043] The texture inversion development image and the deformation motion development image are superimposed and fused to obtain the wing surface rollover texture map.

[0044] It should be noted that the wing surface rollover texture map refers to an image used to represent the state of the wing surface area of ​​sheet-like fragments as they are lifted, bent, rolled, attached, and moved in position under the action of airflow. The continuous bright and dark change areas in the wing surface rollover texture map correspond to the position areas where the relationship between the surface of the sheet-like fragment and the air, the bearing surface, and the light changes. The local grayscale change direction in the wing surface rollover texture map corresponds to the rolling direction and attachment direction of the wing surface area of ​​the sheet-like fragment. The continuous band-shaped change area in the wing surface rollover texture map corresponds to the wing surface change area formed by the sheet-like fragment extending from the hinged edge area to the non-contact side.

[0045] Specifically, the texture inversion developed image and the deformation motion developed image are mapped to each other using the same image width, image height, and pixel coordinate origin. The texture inversion developed gray value and the deformation motion developed gray value at the same pixel coordinate position are added together and averaged to obtain the fused gray value at that pixel coordinate position. The fused gray value is written into the same pixel coordinate position in a blank image with the same size as the geometric reference image. The developed gray value is retained for pixel positions with a single developed gray value, while pixel positions without a developed gray value are left blank. The continuous gray area after fusion is used as the rolling texture expression of the sheet-like fragment wing surface region to obtain the wing surface rolling texture map.

[0046] S4. Obtain the attraction potential field map based on the exhaust structure mask image and the wing surface rollover texture map corresponding to the geometric reference map; In an embodiment of the present invention, obtaining the attraction potential field diagram includes: Extract the exhaust structure region from the geometric reference image to obtain the exhaust structure mask image; An exhaust structure mask image is an image used to represent the spatial location of the exhaust structure inside a fume hood. The marked areas in the exhaust structure mask image correspond to the opening of the back baffle, the guide slit, and the location of the air flow inlet connected to the exhaust channel. The unmarked areas in the exhaust structure mask image correspond to the working surface of the experimental table, the working surface of the fume hood, and the surface of the non-exhaust area. The marked areas in the exhaust structure mask image are used to represent the location area where air inside the fume hood is continuously drawn into the exhaust channel. In an embodiment of the present invention, the exhaust structure region is extracted from the geometric reference map to obtain an exhaust structure mask image, including: Determine candidate regions for exhaust structures in the fume hood from the geometric reference diagram; Specifically, the image range of the rear baffle of the fume hood is located in the geometric reference map. The image area with a narrow opening shape within the image range of the rear baffle is taken as the candidate area of ​​the guide slit. The image area within the image range of the rear baffle that is connected to the candidate area of ​​the guide slit and faces the exhaust channel is taken as the candidate area of ​​the rear baffle opening. The candidate areas of the guide slit and the candidate areas of the rear baffle opening are jointly determined as the candidate areas of the exhaust structure in the fume hood.

[0047] Edge extraction is performed on the candidate regions to obtain the edge image; The candidate region of the exhaust structure in the fume hood refers to the image area in the geometric reference diagram that may correspond to the back baffle opening, the guide slit, and the exhaust inlet position. This image area corresponds to the solid opening of the air entering the exhaust channel inside the fume hood and its surrounding structure. The edge image refers to the image formed by the gray-scale change positions corresponding to the boundaries of the solid opening and the boundaries of the surrounding structure in the candidate region. The linear regions in the edge image correspond to the boundary positions between the exhaust structure and the non-exhaust surface. Specifically, within the candidate region, the grayscale value of each pixel is calculated along both the horizontal and vertical directions to obtain the difference between adjacent pixels. The horizontal and vertical differences are combined to form the grayscale change intensity. The grayscale change intensity is mapped to the pixel position within the candidate region to obtain the grayscale change distribution map of the candidate region. The pixel positions where the grayscale change is continuously distributed and extends around the boundaries of the guide slot candidate region and the rear baffle opening candidate region are written into a blank image with the same size as the geometric reference map. The remaining pixel positions are left blank to obtain the edge image.

[0048] Line segments are extracted from the edge image to obtain a line segment image; Specifically, pixels with grayscale values ​​in the edge image are taken as edge pixels. The edge pixels are divided into multiple connected edge regions according to their edge-adjacent and corner-adjacent relationships. In each connected edge region, any edge pixel at one end is selected as the tracking starting point. The edge pixels that are adjacent at one edge or corner are tracked point by point to the edge pixel at the other end, resulting in an edge pixel sequence arranged in spatial extension order. The horizontal and vertical coordinates of each edge pixel in the edge pixel sequence are fitted with a straight line to obtain the corresponding fitted straight line. The starting point and ending point of the line segment are determined on the fitted straight line based on the starting and ending point coordinates of the edge pixel sequence. The pixel positions between the starting point and ending point of the line segment are written into a blank image with the same size as the geometric reference image. The pixel positions that do not fall between the starting point and ending point of the line segment are left blank. The line segments corresponding to all connected edge regions are written into the same blank image to obtain the line segment image.

[0049] Generate region contour images based on line segment images; Specifically, extract line segments corresponding to the outer perimeter of the rear baffle opening, guide slit, or exhaust inlet from the line segment image. Determine the starting and ending points of each line segment. Connect the endpoints of line segments that are spatially adjacent and belong to the same outer perimeter of the opening. For the endpoints of adjacent line segments that are not directly connected but belong to the same outer perimeter, fill in connecting pixels between the two endpoints in a straight line connection manner, so that multiple line segments and connecting pixels together form a closed boundary. Merge the pixel positions that have repetition or overlap in the closed boundary. Fill in the pixel positions that have breaks in the closed boundary according to the endpoint connection relationship. Write the closed boundary after merging and filling into a blank image with the same size as the geometric reference image. Keep the pixel positions outside the closed boundary blank to obtain the region contour image.

[0050] The region contour image is filled to obtain the exhaust structure mask image.

[0051] A region contour image refers to a closed or nearly closed image formed by connecting line segments in a line segment image. The region contour image corresponds to the outer perimeter boundary of the ventilation structure entity opening in the geometric reference diagram. Region filling refers to the process of writing the internal image position enclosed by the outer perimeter boundary in the region contour image into the ventilation structure region.

[0052] Specifically, the closed boundaries in the region contour image are used as filling boundaries. In a blank image with the same size as the geometric reference image, each row of pixels is processed in a line-by-line scanning manner. The boundary pixel positions where the closed boundary intersects with the pixels in the same row are determined. The pixel positions between two adjacent boundary pixel positions are written as the exhaust structure region. The pixel positions on the closed boundary are written as the exhaust structure boundary. The pixel positions not inside the closed boundary are kept blank. The exhaust structure regions after line-by-line scanning are connected and sorted. The interconnected exhaust structure regions are retained as continuous filling regions. The continuous filling regions formed by the same back baffle opening, the same guide slit, or the same exhaust inlet are used as the exhaust structure mask image.

[0053] The exhaust structure mask image is processed by distance transformation to obtain the exhaust distance image; The exhaust distance map refers to the spatial distance distribution image formed based on the exhaust structure mask image. Each pixel position in the exhaust distance map corresponds to its spatial distance relationship with the exhaust structure area. Positions closer to the rear baffle opening and the guide slit in the exhaust distance map correspond to smaller spatial distances, while positions farther away from the rear baffle opening and the guide slit correspond to larger spatial distances. The continuous change in grayscale in the exhaust distance map is used to represent the spatial distribution state in which the influence of airflow gradually weakens from the exhaust structure outward. Specifically, the exhaust structure mask image and the geometric reference image are mapped to each other using the same image width, image height, and pixel coordinate origin. The pixel positions of the exhaust structure regions that have been written are extracted from the exhaust structure mask image. The pixel coordinate distance between each non-exhaust structure region pixel position and each exhaust structure region pixel position is calculated. The smallest pixel coordinate distance is selected as the exhaust distance value corresponding to the non-exhaust structure region pixel position. The exhaust distance value of the exhaust structure region pixel position is recorded as zero. The exhaust distance value corresponding to each pixel position is written to the same pixel coordinate position in a blank image with the same size as the geometric reference image. The exhaust distance value is then subjected to grayscale mapping to obtain the exhaust distance image.

[0054] The articulated contact zone image and the wing surface rollover texture image are fused to obtain the folded wing coupling layer image; The folded wing coupling layer image refers to the image used to represent the correspondence between the contact state of the edge of the sheet-like fragment and the folding state of the wing surface. The continuous strip-shaped area in the folded wing coupling layer image corresponds to the location area where the edge of the sheet-like fragment remains in contact and the wing surface area of ​​the sheet-like fragment undergoes changes such as lifting, bending, folding, and attaching. The local brightness and darkness variation area in the folded wing coupling layer image is used to represent the local folding state and local migration state of the wing surface area of ​​the sheet-like fragment under the action of airflow. Specifically, the articulated contact zone image and the wing surface roll-up texture image are mapped to each other using the same image width, image height, and pixel coordinate origin. The articulated contact zone pixel positions with grayscale values ​​are extracted from the articulated contact zone image, and the wing surface roll-up pixel positions with grayscale values ​​are extracted from the wing surface roll-up texture image. For pixel positions where both articulated contact zone and wing surface roll-up grayscale values ​​exist at the same pixel coordinate location, the grayscale values ​​are added together to obtain a grayscale sum. This sum is divided by two to obtain a fused grayscale value, which is then written to the same pixel coordinate location in a blank image with the same size as the geometric reference image. For pixel positions where only articulated contact zone grayscale values ​​exist, these values ​​are used as the fused grayscale values ​​and written to the same pixel coordinate location in the blank image. For pixel positions where only wing surface roll-up grayscale values ​​exist, these values ​​are used as the fused grayscale values ​​and written to the same pixel coordinate location in the blank image. Pixel positions without grayscale values ​​are left blank, resulting in the folded wing coupling layer image.

[0055] The exhaust distance image and the folding wing coupling layer image are fused to obtain the attraction potential field map.

[0056] The adhesion potential field map refers to the spatial distribution image used to represent the changes in the approach and attachment of sheet-like fragments to the exhaust structure area. The gray-scale distribution in the adhesion potential field map corresponds to the spatial approach relationship between the wing area of ​​the sheet-like fragment and the exhaust structure area. The continuous gray-scale enhancement area in the adhesion potential field map corresponds to the location area where the sheet-like fragments attach and migrate along the surface of the rear baffle, the edge of the guide slot, or the exhaust inlet area under the action of airflow.

[0057] Specifically, the exhaust distance image and the wing coupling layer image are mapped to each other using the same image width, image height, and pixel coordinate origin. The wing coupling pixel positions with gray values ​​are extracted from the wing coupling layer image, and the exhaust distance gray values ​​at the same pixel coordinate positions are extracted from the exhaust distance image. The corresponding exhaust proximity gray values ​​are generated based on the exhaust distance gray values, which are inversely related to the exhaust distance gray values. The gray value at the wing coupling pixel position is multiplied by the corresponding exhaust proximity gray value to obtain the attraction gray value at that pixel position. The attraction gray value is written into the same pixel coordinate position in a blank image with the same size as the geometric reference image. Pixel positions in the wing coupling layer image that do not have gray values ​​are left blank, resulting in an attraction potential field map that represents the approximate distribution of the sheet-like fragment wing region towards the rear baffle opening and the guide slot.

[0058] S5. Based on the attraction potential field diagram, automatically alarm for abnormal events of adhesion and migration of sheet-like debris in the fume hood area.

[0059] In embodiments of the present invention, automatic alarms are provided for abnormal attachment migration events, including: Image convergence processing is performed on the attraction potential field diagram to obtain the risk-driven signal; Specifically, the attraction potential field map and the geometric reference map are mapped to the same image width, image height, and pixel coordinate origin. Pixel positions with gray values ​​are extracted from the attraction potential field map. The gray value corresponding to each pixel position with a gray value is taken as the attraction gray value. All attraction gray values ​​are summed to obtain the attraction gray sum value. The attraction gray sum value is divided by the number of pixels with gray values ​​to obtain the attraction average gray value. The attraction gray sum value and the attraction average gray value are written together into the risk driving signal, so that the risk driving signal simultaneously includes the overall gray scale of the sheet-like fragment attachment migration area and the gray level of the unit area.

[0060] Based on risk-driven signals, alarm control information is generated. Image convergence processing refers to the process of centrally calculating the pixel grayscale distribution in the adhesion potential field map. The data obtained from image convergence processing is used to represent the overall degree of adhesion and migration of sheet-like debris to the exhaust structure area. Risk driving signal refers to the control input data formed by image convergence processing of the adhesion potential field map. Risk driving signal is used to characterize the degree to which abnormal events of sheet-like debris adhesion and migration drive the alarm output. Specifically, the following steps are taken: the sum of the gray values ​​of the gravitation and the average gray value of the gravitation are obtained from the risk-driven signal; the sum of the gray values ​​of the gravitation are converted into the alarm output duration; the average gray value of the gravitation is converted into the alarm output intensity; the alarm output duration and the alarm output intensity are written into the alarm control information; and the alarm control information is written to indicate that the alarm object is an abnormal event of sheet-like fragment attachment and migration, and the alarm area is the experimental station position corresponding to the pixel area with gray value in the gravitation potential field map.

[0061] Based on alarm control information, an automatic alarm is triggered for abnormal events related to the attachment and migration of sheet-like fragments.

[0062] Alarm control information refers to alarm execution data generated based on risk-driven signals. Alarm control information includes alarm output mode, alarm output intensity, and alarm output location. Abnormal events of attachment and migration of sheet-like fragments refer to the abnormal state in which sheet-like fragments attach, roll up, and tend to move towards the exhaust structure under the action of airflow along the working surface of the experimental table, the working surface of the fume hood, the surface of the rear baffle, or the edge of the guide slit.

[0063] Specifically, the alarm output duration, alarm output intensity, alarm object, and alarm area are extracted from the alarm control information. The actual workstation position corresponding to the alarm area is located in the experimental platform workstation area. The workstation position corresponding to the alarm area is converted into area coordinates in the display terminal. The experimental platform position corresponding to the area coordinates is highlighted on the display terminal interface. The alarm object is written into the alarm text content in the display terminal. The alarm output duration is converted into the continuous output time of the audible and visual alarm device. The alarm output intensity is converted into the brightness level of the alarm light and the volume level of the alarm sound. The continuous output time, brightness level, and volume level are written into the control interface corresponding to the audible and visual alarm device. The alarm light in the experimental platform workstation area is driven to continuously output a flashing light signal, and the alarm speaker is driven to continuously output an alarm sound signal. During the alarm duration, the current attraction potential field map is continuously received. The current attraction potential field map is re-processed with image convergence and the risk driving signal is updated. The updated risk driving signal is rewritten into the alarm control information. The brightness of the alarm light, the volume of the alarm sound, and the alarm duration are updated synchronously, so that the experimental personnel continuously receive dynamic alarm prompts corresponding to the adhesion and migration state of the sheet-like fragments.

[0064] It should be noted that the automatic alarm for abnormal adhesion and migration of sheet-like debris in the fume hood area is because, under the continuous traction of negative pressure airflow, weighing paper corner pieces, aluminum foil corner pieces, or sealing film fragments do not drift into the exhaust area as a whole. Instead, they form a state of edge adhesion and wing-surface rolling migration, where one edge of the sheet-like debris continuously adheres to the working surface of the experimental table, the surface of the rear baffle, or the edge of the guide slit, while the other side periodically lifts and rolls up under the shearing action of the airflow, gradually migrating along the exhaust path towards the rear baffle opening and the guide slit area. This process causes the sheet-like debris to form local obstruction near the exhaust inlet, thereby changing the airflow path inside the fume hood, causing exhaust disturbance, local backflow, and a decrease in the ability to capture harmful gases. Therefore, it is necessary to automatically alarm for abnormal adhesion and migration of sheet-like debris in the fume hood area based on the spatial distribution of the continuous approach and adhesion of sheet-like debris to the exhaust structure area in the adhesion potential field diagram, so that the experimental personnel can promptly remove the sheet-like debris and restore the smooth exhaust state of the fume hood.

[0065] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An automatic alarm method for abnormal safety behavior of an experimental platform based on image recognition, characterized in that, Includes the following steps: S1. Determine the fume hood area corresponding to the experimental bench, and obtain the geometric reference map based on the original image frame of the fume hood area; S2. Accumulate and image the differential image corresponding to the geometric reference image to obtain the articulated contact zone image; S3. Based on the articulated contact zone map, perform local contrast changes on the geometric reference map to obtain a texture inversion development image; based on the texture inversion development image, obtain the wing surface rollover texture map. S4. Obtain the attraction potential field map based on the exhaust structure mask image and the wing surface rollover texture map corresponding to the geometric reference map; S5. Based on the attraction potential field diagram, automatically alarm for abnormal events of adhesion and migration of sheet-like debris in the fume hood area.

2. The automatic alarm method for abnormal safety behavior of an experimental platform based on image recognition according to claim 1, characterized in that, Obtain the geometric reference diagram, including: Identify the fume hood area from the workstation area of ​​the laboratory bench; The raw image frames of the fume hood area are acquired through wireless networking. Distortion correction is performed on the original image frames to obtain corrected image frames; Extract structural features of the fume hood from the corrected image frame; Based on the structural features of the fume hood, cross-view geometric registration processing is performed on the corrected image frame to obtain the registered image frame. Image fusion processing is performed on the registered image frames to obtain a geometric reference map.

3. The automatic alarm method for abnormal safety behavior of an experimental platform based on image recognition according to claim 1, characterized in that, The articulated contact zone diagram is obtained, including: Identify flaky debris in the fume hood area; Edge extraction is performed on the geometric reference image to obtain the boundary zone image of the sheet-like fragments; Generate a difference image based on the geometric reference map; Based on the boundary zone image, cumulative imaging processing is performed on the difference image to obtain the boundary zone cumulative image; Morphological closing operations are performed on the cumulative boundary zone image to obtain the articulated contact zone image.

4. The automatic alarm method for abnormal safety behavior of an experimental platform based on image recognition according to claim 1, characterized in that, Obtain the texture-inverted developed image, including: Based on the articulated contact zone diagram, determine the articulated edge region of the sheet-like fragment in the geometric reference diagram; Based on the articulated edge region, strip regions are extracted from the geometric reference map to obtain the observation strip image of the wing surface; Based on the observed strip image of the wing surface, local contrast changes are performed on the geometric reference image to obtain a texture inversion and development image.

5. The automatic alarm method for abnormal safety behavior of an experimental platform based on image recognition according to claim 4, characterized in that, Based on the articulated edge region, strip regions are extracted from the geometric reference map to obtain the observed strip image of the wing surface, including: Boundary point sampling is performed on the hinged edge region to obtain a set of hinged edge points; Determine the normal direction information of the set of hinged edge points; Based on the set of hinged edge points and normal direction information, a strip region mask image is generated in the geometric reference map; Based on the strip region mask image, strip regions are extracted from the geometric reference image to obtain the wing surface observation strip image.

6. The automatic alarm method for abnormal safety behavior of an experimental platform based on image recognition according to claim 1, characterized in that, The obtained wing surface rollover texture map includes: Based on the observed stripe image of the wing surface, extract the current stripe region image from the geometric reference image at the current moment; Based on the observed strip images of the wing surface, images of adjacent strip regions are extracted from the geometric reference images at adjacent time points; Pixel displacement is calculated between the current strip region image and the adjacent strip region images to obtain the strip region pixel displacement information; Generate a motion field image of the strip region based on the pixel displacement information of the strip region; The motion field image of the strip region is developed to obtain a deformation motion developed image; The texture inversion development image and the deformation motion development image are superimposed and fused to obtain the wing surface rollover texture map.

7. The automatic alarm method for abnormal safety behavior of an experimental platform based on image recognition according to claim 1, characterized in that, The attraction potential field diagram is obtained, including: Extract the exhaust structure region from the geometric reference image to obtain the exhaust structure mask image; The exhaust structure mask image is processed by distance transformation to obtain the exhaust distance image; The articulated contact zone image and the wing surface rollover texture image are fused to obtain the folded wing coupling layer image; The exhaust distance image and the folding wing coupling layer image are fused to obtain the attraction potential field map.

8. The automatic alarm method for abnormal safety behavior of an experimental platform based on image recognition according to claim 7, characterized in that, The exhaust structure region is extracted from the geometric reference image to obtain the exhaust structure mask image, including: Determine candidate regions for exhaust structures in the fume hood from the geometric reference diagram; Edge extraction is performed on the candidate regions to obtain the edge image; Line segments are extracted from the edge image to obtain a line segment image; Generate region contour images based on line segment images; The region contour image is filled to obtain the exhaust structure mask image.

9. The automatic alarm method for abnormal safety behavior of an experimental platform based on image recognition according to claim 1, characterized in that, Automatic alarms are triggered for abnormal attachment migration events, including: Image convergence processing is performed on the attraction potential field diagram to obtain the risk-driven signal; Based on risk-driven signals, alarm control information is generated. Based on alarm control information, an automatic alarm is triggered for abnormal events related to the attachment and migration of sheet-like fragments.

Images