A method and system for intelligently identifying potential safety hazards of high slopes of a hydropower station based on inspection images

Abstract

This invention discloses an intelligent identification method and system for safety hazards on high slopes of hydropower stations based on inspection images. Specifically, it relates to the field of intelligent identification technology for safety hazards on high slopes of hydropower stations. The method includes acquiring inspection images continuously collected by an edge computing device during the inspection of high slopes of hydropower stations; identifying slope stability reference areas in each inspection image; extracting abnormal attachment zones and suspected abnormal areas located within the abnormal attachment zones; and outputting a set of suspected anomalies, a set of stable reference areas, and a set of abnormal attachment zones for adjacent frames. The method involves first extracting the slope stability reference areas and abnormal attachment zones from adjacent inspection images on the edge computing device, then mapping the suspected abnormal areas to the corresponding slope attachment positions, and determining real hazards and imaging pseudo-anomalies based on the continuous consistency of slope attachment positions, abnormal attachment zone types, attachment relationships, and risk manifestation conditions in adjacent frames.

Description

Technical Field

[0001] This invention relates to the field of intelligent identification technology for safety hazards on high slopes of hydropower stations, and more specifically, to a method and system for intelligent identification of safety hazards on high slopes of hydropower stations based on inspection images. Background Technology

[0002] In the technology for identifying safety hazards on high slopes of hydropower stations based on inspection images, the mainstream approach in the industry is to address how to use images or videos collected during the inspection process to promptly detect abnormal signs such as cracks, seepage, falling rocks, and loose stones. Typically, abnormal target detection, segmentation recognition, or key frame filtering is performed on single-frame images at the inspection terminal or edge device, and voting, smoothing, or optimal judgment is made by combining the confidence results of multiple adjacent frames. For example, in the scenario of drone inspection of high slopes in the reservoir area of ​​a hydropower station, the drone needs to fly continuously along the slope to complete a large-scale inspection. At the same time, there are hard constraints such as continuous changes in flight attitude, rapid fluctuations in shooting distance and angle, alternation of local shadows and sunlight, wet marks and reflections, vegetation obstruction, and limited computing power and backhaul bandwidth on the edge side, which prevent the full recalculation from the center side. Under this constraint, the mainstream approach will consistently reveal an observable and verifiable defect: the same suspected anomaly is prone to present contradictory identification results in adjacent frames. For example, it may be identified as a crack in one frame and as a shadow boundary in the next frame, or as a seepage source in one frame and disappear due to reflection or occlusion in the next frame. As a result, although the edge device can obtain multiple frames of anomaly evidence, it cannot determine whether these pieces of evidence continuously correspond to the same real location on the slope. Thus, real hidden dangers and imaging false anomalies are still confused under multi-frame conditions. The technical problem this application aims to solve is: how to accurately distinguish between real safety hazards and imaging artifacts on high slopes of hydropower stations, based on the consistency of the attribution of the same slope entity location to anomaly evidence in adjacent frames of inspection images under edge computing conditions. Summary of the Invention

[0003] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an intelligent identification method and system for safety hazards on high slopes of hydropower stations based on inspection images. This method first extracts the slope stability reference area and abnormal attachment zone from adjacent inspection images on an edge computing device, then maps the suspected abnormal area to the corresponding slope attachment position, and determines the real hazards and imaging pseudo-anomalies based on the continuous consistency of slope attachment position, abnormal attachment zone type, attachment relationship and risk manifestation conditions in adjacent frames, thereby solving the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a method for intelligent identification of safety hazards on high slopes of hydropower stations based on inspection images, comprising: S1. Acquire inspection images continuously collected by the edge computing device during the inspection of the high slope of the hydropower station, identify the slope stability reference area in each inspection image, extract the abnormal attachment zone and the suspected abnormal area located in the abnormal attachment zone, and output the suspected abnormal set, stable reference set and abnormal attachment zone set of adjacent frames. S2. Perform reference correspondence identification on the stable reference set in adjacent frames, and map each suspected abnormal region to the corresponding slope attachment position according to the positional relationship between each suspected abnormal region and the corresponding stable reference region and the abnormal attachment zone, and output the attachment position attribution set of the suspected abnormal region. S3. Perform a consistency determination of attachment relationship for the attachment location set of suspected abnormal areas. When the slope attachment location, abnormal attachment zone category and attachment relationship of the same suspected abnormal area are consistent in adjacent frames, output the attachment retention result; otherwise, output the attachment offset result. S4. For the suspected abnormal areas corresponding to the attachment retention results, perform risk performance condition judgment according to the corresponding abnormal attachment zone category. When the abnormal performance continuously meets the risk performance conditions of the corresponding abnormal attachment zone, output the real hidden danger mark; otherwise, output the imaging pseudo-anomaly mark. S5. For suspected abnormal areas marked with real hidden dangers, generate the safety hazard identification results of the high slope of the hydropower station according to the corresponding slope attachment location, abnormal attachment zone category and risk performance condition judgment results, and output them to the edge computing device.

[0005] In a preferred embodiment, S1 includes: S1-1. Obtain the slope foreground area in each inspection image, perform structural boundary extraction, and identify the fixed structure edge, stable bare rock texture area and stable support component area based on boundary continuity, texture repetition stability and cross-frame consistency retention, and output the slope stability reference candidate set. S1-2. Based on the candidate set of slope stability references, perform neighborhood relationship aggregation and boundary expansion calculation. Based on the adjacency relationship between each candidate region of slope stability references and the slope surface boundary, drainage trace boundary, support component boundary and repair trace boundary, generate the corresponding abnormal attachment zone and output the abnormal attachment zone set. S1-3. Using the set of abnormal attachment bands as a constraint region, perform in-band texture disturbance detection and appearance continuity destruction identification, and identify regions that meet the conditions of texture abrupt change, boundary breakage, brightness migration abnormality or morphological extension abnormality as suspected abnormal regions, and output the set of suspected abnormal regions located in each abnormal attachment band.

[0006] In a preferred embodiment, the process of outputting the abnormal attachment band set in S1-2 further includes: S1-21. Obtain the slope stability reference candidate set and the slope apparent boundary, drainage trace boundary, support component boundary and repair trace boundary. Construct an adjacency candidate graph with each slope stability reference candidate region and each boundary segment as nodes. The association cost function is constructed with the distance deviation between nodes, direction deviation, texture continuity deviation and occlusion violation. Perform association solution under the boundary single-class ownership constraint, local connectivity constraint and adjacent frame preservation constraint, and output the initial adjacency relationship set and the corresponding association confidence results. S1-22. Based on the initial adjacency set, perform neighborhood relationship aggregation and conflict resolution. Construct a consistency verification set from shared boundary segments, overlapping coverage areas, and multi-class competing adjacency relationships. Perform confidence updates based on boundary category mutual exclusion constraints, region inclusion constraints, and adjacent frame relationship continuation constraints. When the category affiliation of each adjacency relationship remains consistent in two adjacent confidence updates and the conflict relationships in the consistency verification set no longer change, output the confirmed adjacency set.

[0007] In a preferred embodiment, the process of outputting the abnormal attachment band set in S1-2 further includes: S1-23. Perform directional boundary expansion calculations on each slope stability reference candidate area according to the confirmed adjacency relationship set. Construct the zone generation cost function with the expansion width deviation, cross-boundary leakage amount and apparent stability deviation within the zone. Solve the boundary direction expansion range without crossing heterogeneous boundary constraints, covering the corresponding adjacent boundary constraints and expansion termination boundary constraints, and output the candidate abnormal attachment zone and the corresponding zone generation confidence results. S1-24. Perform multi-view consistency verification on the candidate abnormal attachment bands and the confirmed adjacency relationship set in the adjacent frames. Perform joint verification on the band position drift, boundary category jump variable and band apparent fluctuation. Perform confidence correction and band filtering based on the band generation confidence result and consistency verification result, and output the abnormal attachment band set.

[0008] In a preferred embodiment, S2 includes: S2-1. Based on the stable reference set, abnormal attachment band set, and suspected abnormal region set in adjacent frames, construct a cross-frame correspondence map with stable reference regions as reference nodes and suspected abnormal regions as nodes to be assigned. Then, perform reference correspondence solving based on the boundary morphology encoding, texture distribution encoding, and adjacent band category encoding of each stable reference region, and output the stable reference correspondence set of adjacent frames. S2-2. Based on the stable reference set of adjacent frames, perform joint position attribution calculation on the orientation, distance and boundary contact relationship of each suspected anomaly region relative to the corresponding stable reference region, as well as the intra-band position relationship, band edge adjacency relationship and band extension relationship relative to the corresponding anomaly attachment band, and output the candidate set of slope attachment positions of each suspected anomaly region in adjacent frames. S2-3. Perform a consistency check between the candidate set of slope attachment positions of each suspected abnormal region in adjacent frames and the corresponding set of stable references in adjacent frames. Based on the overlap of attachment positions of the same suspected abnormal region in adjacent frames, the consistency of attachment zone categories, and the retention of reference relationships, determine the corresponding slope attachment positions and output the set of attachment positions of suspected abnormal regions.

[0009] In a preferred embodiment, S3 includes: S3-1. Based on the attachment location attribution set of suspected abnormal areas, construct attachment determination units according to the slope attachment location, abnormal attachment band category and attachment relationship of the same suspected abnormal area in adjacent frames, and perform position correspondence verification, band category correspondence verification and relationship type correspondence verification on each attachment determination unit, and output the attachment correspondence result set. S3-2. Based on the attachment corresponding result set, perform relationship conflict resolution, and form a conflict judgment group by changing the slope attachment position, switching of abnormal attachment zone categories, and transferring attachment relationships. Then, perform conflict merging and main conflict identification based on the position deviation direction, zone category maintenance status, and relationship transfer path in each conflict judgment group, and output the attachment consistency judgment result corresponding to each suspected abnormal area. S3-3. Assign the results of each suspected abnormal region according to the attachment consistency judgment results. When the slope attachment position, abnormal attachment band category and attachment relationship of the same suspected abnormal region are consistent in adjacent frames, it is determined to be the attachment retention result; otherwise, it is determined to be the attachment offset result.

[0010] In a preferred embodiment, S4 includes: S4-1. Obtain the suspected abnormal regions, abnormal attachment band categories, and adjacent frame inspection images corresponding to the attachment retention results. Extract the directional extension, boundary fitting, apparent continuous change, and intra-band occupancy change of each suspected abnormal region in adjacent frames according to the abnormal attachment band category, and construct the corresponding risk performance feature set.

[0011] In a preferred embodiment, S4 further includes: S4-2. Based on the risk performance feature set, perform in-category condition verification on each suspected abnormal region according to the risk performance constraints corresponding to each abnormal attachment zone category, and form a risk performance judgment group by directional extension continuity, boundary fitting continuity, apparent continuous change consistency and in-band occupancy retention, and output the risk performance results corresponding to each suspected abnormal region. S4-3. Based on the risk performance results corresponding to each suspected abnormal area, perform the labeling and attribution. When the risk performance judgment group of the same suspected abnormal area remains valid in adjacent frames, output the real hidden danger label; otherwise, output the imaging pseudo-anomaly label.

[0012] In a preferred embodiment, S5 includes: S5-1. Take the suspected abnormal areas with real hazard markers, the corresponding slope attachment locations, abnormal attachment zone categories and risk performance condition judgment results, and perform hazard merging calculation according to the area merging relationship of the same slope attachment location, the category merging relationship of the same abnormal attachment zone category and the result merging relationship of the same risk performance condition judgment results, and output the high slope hazard identification unit set. S5-2. Based on the high slope hazard identification unit set, generate the hydropower station high slope safety hazard identification results according to the slope attachment location, abnormal attachment zone category and risk performance condition judgment results corresponding to each high slope hazard identification unit, and output the hydropower station high slope safety hazard identification results to the edge computing device.

[0013] A smart identification system for safety hazards on high slopes of hydropower stations based on inspection images, the system comprising an anomaly extraction module, an attachment mapping module, a judgment module, a discrimination module, and a result generation module: The anomaly extraction module is used to acquire inspection images continuously collected by edge computing devices during the inspection of high slopes in hydropower stations. It identifies stable reference areas on the slope in each inspection image, extracts abnormal attachment zones and suspected abnormal areas located within the abnormal attachment zones, and outputs a set of suspected anomalies, a set of stable reference zones, and a set of abnormal attachment zones in adjacent frames. The attachment mapping module is used to perform reference correspondence identification on the stable reference set in adjacent frames, and based on the positional relationship between each suspected abnormal region and the corresponding stable reference region and the abnormal attachment zone, it maps each suspected abnormal region to the corresponding slope attachment position and outputs the attachment position attribution set of the suspected abnormal region. The determination module is used to perform a consistency determination of the attachment relationship of the attachment position set of suspected abnormal areas. When the slope attachment position, abnormal attachment band category and attachment relationship of the same suspected abnormal area are consistent in adjacent frames, the attachment retention result is output; otherwise, the attachment offset result is output. The discrimination module is used to determine the risk performance conditions of suspected abnormal areas corresponding to the adhesion retention results based on their corresponding abnormal adhesion band categories. When the abnormal performance continuously meets the risk performance conditions of the corresponding abnormal adhesion band, a real hazard mark is output; otherwise, an imaging pseudo-anomaly mark is output. The results generation module is used to generate safety hazard identification results for high slopes of hydropower stations for suspected abnormal areas with real hazard markers, based on the corresponding slope attachment location, abnormal attachment zone category and risk performance conditions, and output the results to the edge computing device.

[0014] The technical effects and advantages of this invention are as follows: 1. This solution first identifies the stable reference area and abnormal attachment zone of the slope and maps the suspected anomaly to the attachment position of the slope. Then, it performs consistency judgment on the attachment position, attachment zone type and attachment relationship in adjacent frames. Under the condition of edge computing, it can determine whether the abnormal evidence of multiple frames continuously corresponds to the same slope entity position, thereby relatively improving the problem of easy confusion between real hidden dangers and imaging false anomalies. 2. By limiting the anomaly detection range to the anomalous attachment zone generated by the slope surface boundary, drainage trace boundary, support component boundary, and repair trace boundary, and combining the texture disturbance and appearance continuity disruption within the zone to identify suspected anomaly areas, the interference of shadows, reflections, and non-target textures in the full-frame image on the detection results can be reduced, thereby relatively improving the targeting of suspected anomaly extraction. 3. Establish stable reference correspondence between adjacent frames by using boundary morphology encoding, texture distribution encoding and adjacent band category encoding of stable reference areas, and solve the slope attachment position of suspected abnormal areas on this basis. This can maintain cross-frame attribution benchmark under the conditions of view fluctuation and shooting distance change, thereby relatively improving the continuity and consistency of attachment position attribution. 4. Conflict merging and main conflict identification are performed on changes in slope attachment location, switching of abnormal attachment zone categories, and transfer of attachment relationships. This can distinguish between local location disturbances and actual attachment relationship migration, thereby relatively suppressing misjudgments caused by single location shifts or single category changes. 5. By extracting the directional extension, boundary adhesion, apparent continuous change, and intra-band occupancy change according to the abnormal attachment band category, and performing intra-category condition verification based on the corresponding risk performance constraints, the suspected anomalies of attachment retention can be further distinguished into real hidden dangers and imaging pseudo-anomalies, thereby relatively improving the reliability of the risk judgment results on the edge side. Attached Figure Description

[0015] Figure 1 This is a flowchart outlining the method steps of the present invention; Figure 2 This is a schematic diagram of the system module structure of the present invention. Detailed Implementation

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

[0017] Refer to the instruction manual appendix Figure 1-2 The present invention provides an intelligent identification method for safety hazards on high slopes of hydropower stations based on inspection images, comprising: S1. Acquire inspection images continuously collected by the edge computing device during the inspection of the high slope of the hydropower station, identify the slope stability reference area in each inspection image, extract the abnormal attachment zone and the suspected abnormal area located in the abnormal attachment zone, and output the suspected abnormal set, stable reference set and abnormal attachment zone set of adjacent frames. This specific implementation method is used to continuously solve the slope stability reference area, abnormal attachment zone, and suspected abnormal area in the inspection image, so that the subsequent slope attachment location attribution, attachment relationship consistency determination, and risk performance condition determination are all based on inputs with entity attribution significance. The basic idea is as follows: First, the slope foreground is separated from the continuous inspection image and the structural boundary is extracted to form a slope stability reference candidate set that can be maintained across frames; then, the adjacency relationship between the slope stability reference candidate set and the slope appearance boundary, drainage trace boundary, support component boundary, and repair trace boundary is used to perform correlation solving, conflict resolution, directional boundary expansion, and multi-view consistency verification to form an abnormal attachment zone set; finally, texture disturbance and appearance continuity failure areas are identified within the abnormal attachment zone, and a suspected abnormal area set is output; Here, the apparent boundary of a slope refers to the boundary where the apparent properties of the slope foreground area change continuously and form a closed or semi-closed separation; the drainage trace boundary refers to the boundary formed by water erosion, wet trace deposition, or infiltration drag marks; the repair trace boundary refers to the material interface boundary formed by shotcreting, patching, sealing, or reinforcement treatment; the abnormal attachment zone refers to the strip-shaped area surrounding the aforementioned boundaries and adjacent to the slope stability reference area, used to define the range of abnormal appearance retention; the implementation process includes the following steps: First, the edge computing device reads the inspection images arranged in the acquisition sequence and performs slope foreground region extraction on each image to obtain the effective image range required for subsequent structural boundary extraction. Inputs include the original inspection images, UAV pose records, and the image time sequence index for the current batch. Processing actions include: removing non-slope areas based on the high-brightness, low-texture features of the sky region, the specular reflection features of the water surface region, the non-closed edge features of the equipment shadow region, and the periodic swaying texture features of the vegetation occlusion region, retaining continuous, contiguous slope foreground regions. Subsequently, structural boundaries are extracted from the slope foreground regions, with the structural boundaries being gray... The boundary is determined by the degree transition result, texture direction change result, and contour closure result. Regions where all three are true are denoted as strong boundaries, and regions where only two are true are denoted as candidate boundaries. Then, boundary continuity, texture repetition stability, and cross-frame consistency retention calculations are performed on strong boundaries and candidate boundaries. Boundary continuity is the ratio of the length of consecutive pixels along the main extension direction of the same boundary segment to the total length of the segment. Texture repetition stability is the repetition matching ratio of texture descriptors on both sides of the boundary in adjacent sampling windows. Cross-frame consistency retention is the joint difference between the current frame boundary segment and the corresponding segment in the adjacent frame in terms of position offset, direction change, and texture distribution change. When a region satisfies the conditions of boundary continuity, texture repetition stability, and cross-frame consistency retention, it is identified as the edge of a fixed structure, a stable bare rock texture area, or a stable support component area, and written into the slope stability reference candidate set. The output is the slope stability reference candidate set, and each record in the candidate set includes at least the candidate region identifier, boundary coordinates, texture code, region category, and inter-frame retention record, and is written into the stability reference candidate record table for subsequent adjacency solution calls. When a local region cannot form a complete boundary in a frame due to cloud shadows, fog, or vegetation occlusion, the stable reference record confirmed in the previous frame is retained and an occlusion status mark is written. If boundary continuity and texture repetition stability results cannot be formed in two consecutive frames, the region is not written into the slope stability reference candidate set. After forming the slope stability reference candidate set, the edge computing device performs adjacency candidate construction based on the spatial proximity and directional association between each candidate region and various boundaries to establish the relational framework required for the generation of abnormal attachment zones. The inputs are the slope stability reference candidate set and the slope appearance boundaries, drainage trace boundaries, support component boundaries, and repair trace boundaries extracted from the current inspection images. The processing actions include: constructing an adjacency candidate graph using each slope stability reference candidate region and each boundary segment as nodes. The connecting edges in the graph are only allowed if the projected distance from the candidate region boundary to the boundary segment falls within the statistical interval of the effective adjacency distance of the current batch, and if the main extension direction of the candidate region is related to the boundary segment. The connection is established when the direction satisfies the adjacent direction constraint and the connection path does not cross the identified occluded area. Then, the association cost function is calculated for each connection edge. The association cost function consists of the distance deviation between nodes, the direction deviation, the texture continuity deviation, and the occlusion violation. The distance deviation is the difference between the nearest projected distance from the candidate region boundary to the boundary segment and the average adjacent distance of the same type of candidate region. The direction deviation is the difference between the angle between the main extension direction of the candidate region and the tangential direction of the boundary segment. The texture continuity deviation is the difference between the texture outside the candidate region and the adjacent texture of the boundary segment in the coding space. The occlusion violation is the proportion of the length of the connection path crossing the occluded area to the total length of the connection path. Subsequently, association solving is performed under boundary single-class attribution constraints, local connectivity constraints, and adjacent frame retention constraints. Boundary single-class attribution constraints are used to restrict the same boundary segment to only one boundary category in the same round of solving. Local connectivity constraints are used to restrict the adjacency results formed by the same stable reference candidate region to remain spatially connected. Adjacency frame retention constraints are used to prioritize retaining adjacency relationships that have been confirmed in the previous frame and still satisfy the association conditions in the current frame. The output is the initial adjacency relationship set and the corresponding association confidence results. The association confidence results are jointly determined by the relative ranking position of the association cost among all candidate connections in the same candidate region and the inter-frame retention status, and are written into the adjacency relationship record table for subsequent conflict resolution calls. When a candidate region does not have a boundary segment that satisfies the above connection conditions, a non-adjacency mark is written and the subsequent expansion processing of that candidate region is skipped. Subsequently, the edge computing device performs neighborhood relationship aggregation and conflict resolution on the initial adjacency relationship set to consolidate spatially overlapping, category-competitive, and inter-frame discontinuous adjacency relationships into a confirmed adjacency relationship set. The input is the initial adjacency relationship set and its associated confidence results. The processing actions include: first, identifying shared boundary segments, cross-coverage regions, and multi-category competing adjacency relationships, and combining the aforementioned conflicts involved in the same stable reference candidate region into a consistency verification set; then, performing a group-by-group confidence update based on boundary category mutual exclusion constraints, region inclusion constraints, and adjacent frame relationship continuation constraints. Among these constraints, the boundary category mutual exclusion constraints are used to retain one boundary category affiliation from the multi-category competing relationships corresponding to shared boundary segments, the region inclusion constraints are used to restrict the coverage area of ​​a certain adjacency result from crossing the heterogeneous boundary that has formed a closed separation, and the relationship continuation constraints are used to prioritize maintaining the continuation of confirmed relationships between the current frame and the previous frame. After each round of updates, the category classification, conflict status, and association confidence results of each adjacency relationship are written back, and a new consistency check set is generated. When the category classification of each adjacency relationship remains consistent in two adjacent rounds of updates and the conflict relationships in the consistency check set no longer change, the current adjacency relationship set is determined as the confirmed adjacency relationship set. The output is the confirmed adjacency relationship set, which includes at least the relationship identifier, the corresponding stable reference area identifier, the boundary category, the coverage area, and the relationship retention status, and is read by subsequent directional boundary expansion calculations. If a conflict group still has multiple feasible classifications after two consecutive rounds of updates, the adjacency relationships with lower association cost value and the adjacent frame relationship continues to be valid are retained, and the remaining relationships are written into the elimination record. After confirming the formation of the adjacency relationship set, the edge computing device performs directional boundary expansion calculations on each slope stability reference candidate area based on the confirmed adjacency relationship to obtain candidate anomalous attachment zones. The input quantities are the confirmed adjacency relationship set and the corresponding slope stability reference candidate areas. The processing actions include: using the boundary segments in the confirmed adjacency relationship as the starting edge for expansion, determining the expansion direction by the normal direction of the boundary segments and the main extension direction of the candidate area, and generating expansion strips layer by layer along the expansion direction; after each layer of expansion strips is generated, the strip generation cost function is calculated. The strip generation cost function consists of the expansion width deviation, the cross-boundary leakage amount, and the apparent stability deviation within the strip. The expansion width deviation is the difference between the current expansion width and the statistical value of the effective adjacency width of the same type of confirmed adjacency relationship in the current batch. The statistical value comes from the statistical estimation results of the confirmed adjacency relationships in the current inspection batch; the cross-boundary leakage amount is the proportion of the area of ​​the current expansion strip crossing the heterogeneous boundary to the total area of ​​the strip. The apparent stability deviation within the band is taken as the joint change value of texture difference and brightness difference between continuous sampling units within the strip; then, the boundary direction expansion range is solved under the constraints of not crossing the heterogeneous boundary, covering the corresponding adjacent boundary constraints, and the expansion termination boundary constraints. The expansion termination boundary is taken as the last valid boundary before the end of continuous apparent stability maintenance; the output is the candidate anomalous attachment band and the corresponding band generation confidence result. The band generation confidence result is jointly determined by the convergence state of the band generation cost function, the expansion coverage integrity, and the boundary leakage state, and is written into the candidate band record table for subsequent multi-view consistency verification; if the cross-boundary leakage continues to increase during the expansion in a certain direction and a continuous apparent stability region cannot be formed, the expansion in that direction is terminated, and only the candidate anomalous attachment bands formed in other directions are retained. Furthermore, the edge computing device performs multi-view consistency verification and band filtering on candidate anomalous attachment bands to obtain a set of anomalous attachment bands that can be directly used for subsequent anomaly detection. The inputs are the candidate anomalous attachment bands in the current frame, the corresponding band generation confidence results, and the confirmed adjacency relationship set in adjacent frames. The processing actions include: firstly, based on the source boundary category and the corresponding stable reference region identifier of the candidate anomalous attachment band, searching for the same source confirmed adjacency relationship in adjacent frames, and establishing cross-frame corresponding bands accordingly; then calculating the band position drift, boundary category jump variable, and intra-band apparent fluctuation, where the band position drift is the projection offset of the center line of the candidate band in the current frame and the center line of the corresponding band in the adjacent frame in the stable reference region coordinate system, the boundary category jump variable is the number of changes in the source boundary category of the corresponding band in the current frame and the adjacent frame, and the intra-band apparent fluctuation is the joint fluctuation value of the continuous sampling strips in the candidate band in terms of texture encoding, brightness distribution, and edge direction. Subsequently, the confidence results of the generated band and the aforementioned consistency verification results are jointly written into the confidence correction model to correct the confidence results of the generated band. Band filtering is then performed based on the corrected results. Candidate bands with non-zero boundary category jump variables, continuously expanding band position drift, or no common maintenance area in adjacent frames are filtered out. The output is an abnormal attachment band set. Each record in the abnormal attachment band set includes at least the band identifier, source boundary category, boundary coordinates, band generation confidence result, inter-frame consistency record, and corresponding stable reference area identifier. These are written into the abnormal attachment band record table for subsequent extraction of suspected abnormal areas and reading of slope attachment position attribution. If there is no usable viewpoint in adjacent frames, the confidence results of the generated band in the current frame are retained and written into a single-frame valid marker. Consistency verification is then performed again in the next usable frame. Next, the edge computing device uses the abnormal attachment band set as a constraint region and performs texture perturbation detection and appearance continuity disruption identification within each abnormal attachment band to identify suspected abnormal regions. The input is the abnormal attachment band set and the corresponding local region of the inspection image. The processing actions include: first, segmenting the image segments within the band according to the boundary of the abnormal attachment band and establishing an in-band sampling grid; then, calculating the texture perturbation amount, boundary severance amount, brightness migration amount, and morphological extension amount for each sampling grid. Among them, the texture perturbation amount is the difference between the texture code of the current sampling grid and the texture code of the adjacent grid in the same band; the boundary severance amount is the interruption ratio of the original continuous boundary in the band in the current sampling grid; the brightness migration amount is the migration amount of the brightness distribution center in the band relative to the sampling grid at the same position in the adjacent frame; and the morphological extension amount is the newly added extension length of the abnormal outer edge along the band direction or the band edge direction. Subsequently, the aforementioned four types of quantities, along with the source boundary category of the current band, are input into the abnormal appearance identification unit. Regions meeting the conditions of texture abrupt change, boundary fragmentation, brightness migration anomaly, or morphological extension anomaly are aggregated into connected components. Regions located within the abnormal attachment band and whose connected component area reaches the statistical lower limit of the effective sampling area within the current band are identified as suspected abnormal regions. The output is a set of suspected abnormal regions. Each record in the suspected abnormal region set includes at least a region identifier, the identifier of the corresponding abnormal attachment band, the identifier of the corresponding stable reference region, an abnormal appearance type record, and a local image index. This record is written into the suspected abnormal region record table for subsequent slope attachment location attribution steps. If a region exhibits only a single brightness migration anomaly and this anomaly is consistent with the local reflection direction of the current frame, this region is not directly written into the suspected abnormal region set but is instead marked for verification in adjacent frames within the same band. To ensure that the aforementioned results can be directly invoked in subsequent steps, this specific implementation also organizes the slope stability reference candidate set, confirmed adjacency relationship set, abnormal attachment zone set, and suspected abnormal area set into a structured record chain of the same batch; the input is the output results of the aforementioned steps; the processing action is: to establish a hierarchical calling relationship according to the current inspection batch number, frame number, stability reference area identifier, abnormal attachment zone identifier, and suspected abnormal area identifier, and to uniformly write the source field, calculation field, status field, and reserved field of each intermediate result into the local result cache area of ​​the edge computing device; The source field records which frame, stable reference region, and boundary category the result comes from; the calculation field records intermediate quantities such as association cost, generated confidence result, location drift, and apparent fluctuation; the status field records statuses such as confirmed, pending review, removed, and valid for a single frame; and the reserved field is used for subsequent reading of slope attachment location attribution and attachment relationship consistency determination. The output is a structured record chain that can be called across steps and can be directly read by S2. When any intermediate result lacks a source field or a status field, it is not written to the local result cache, but instead falls back to the corresponding previous step for recalculation. Through the above specific implementation method, stable reference candidate regions with cross-frame preservation capability can be extracted from continuous inspection images. Then, anomaly attachment bands are generated and verified based on the adjacency relationship between the stable reference candidate regions and various boundaries. Finally, suspected anomaly regions with structural attribution basis are extracted within the anomaly attachment bands. This way, subsequent steps no longer deal with random anomalies in the full-frame image, but with anomaly objects with clear slope attachment basis and inter-frame correspondence basis. After this processing, the calling relationship between anomaly attachment bands, candidate adjacency relationships, band generation confidence results, and suspected anomaly regions is clear. The calculation process, solution process, judgment process, and screening process all have executable specifications, avoiding problems such as ambiguous terminology, non-unique output, and unclear step connection. In practical applications: When drones inspect the high slopes of hydropower stations, edge computing devices first remove areas obscured by the sky, water surface, and vegetation from continuously acquired inspection images, and then extract stable support component areas and stable bare rock texture areas. Next, candidate anomalous attachment zones are generated based on their adjacency with the boundaries of drainage traces and repair traces, and low-confidence zones affected by viewing angle changes are filtered out through adjacent frame consistency checks. Subsequently, within the remaining anomalous attachment zones, local areas where both boundary cleavage and morphological extension are established are identified. These local areas are written into a suspected anomalous area record table and cached together with the corresponding stable reference area and anomalous attachment zone record for subsequent reading of slope attachment location attribution, attachment relationship consistency determination, and real hazard marking.

[0018] S2. Perform reference correspondence identification on the stable reference set in adjacent frames, and map each suspected abnormal region to the corresponding slope attachment position according to the positional relationship between each suspected abnormal region and the corresponding stable reference region and the abnormal attachment zone, and output the attachment position attribution set of the suspected abnormal region. This specific implementation method is used to further resolve suspected anomaly regions from the image region level into slope attachment positions with entity attribution significance. This allows subsequent attachment relationship consistency determination to no longer be based on simple pixel proximity or region overlap, but rather on the relationship maintenance results of the same suspected anomaly region relative to the same stable reference region and the same anomalous attachment band. The basic principle is as follows: First, a cross-frame correspondence map is established using the stable reference set, anomalous attachment band set, and suspected anomaly region set in adjacent frames. The stable reference correspondence set in adjacent frames is solved by boundary morphology encoding, texture distribution encoding, and adjacent band category encoding of the stable reference region. Then, based on the stable reference correspondence results and anomalous attachment band constraints, the relative orientation, relative distance, boundary contact, and intra-band extension relationship of each suspected anomaly region in adjacent frames are jointly calculated to form a candidate set of slope attachment positions. Finally, the attribution consistency check resolves multiple candidate conflicts and outputs the attribution set of suspected anomaly region attachment positions. The slope attachment location here is not simply an image coordinate location, but rather the location of the slope entity defined by the relative relationships between the corresponding stable reference area, the corresponding anomalous attachment zone, and the suspected anomalous area; the implementation process includes the following steps: First, the edge computing device constructs a cross-frame correspondence map based on the stable reference set, anomalous attachment band set, and suspected anomalous region set in adjacent frames to complete the corresponding solution of stable reference regions, providing a unified reference basis for subsequent attachment position calculation. The input includes the stable reference set, anomalous attachment band set, and suspected anomalous region set in the current frame and adjacent frames. Each stable reference region in the stable reference set has a boundary morphology code, texture distribution code, and adjacent band category code. The boundary morphology code is used to characterize the polyline turning sequence, closure state, and main extension direction of the outer contour of the stable reference region. The texture distribution code is used to characterize the texture type, texture direction, and texture density inside the stable reference region and its boundary neighborhood. The adjacent band category code is used to characterize the anomalous attachment band category and its coverage direction that have been confirmed in the previous steps for the stable reference region. The processing steps include: using stable reference regions in the current frame and adjacent frames as reference nodes, and suspected abnormal regions as nodes to be assigned, constructing a cross-frame correspondence graph. Candidate correspondence edges are established between reference nodes in the graph based on boundary morphology encoding similarity, texture distribution encoding similarity, and adjacent band category encoding consistency. Then, a reference correspondence cost is calculated for each candidate correspondence edge. The reference correspondence cost is composed of boundary morphology difference, texture distribution difference, and adjacent band category difference. The boundary morphology difference is calculated based on the morphological offset of the boundary polyline sequence at a unified sampling point. The texture distribution difference is calculated based on the difference between the texture direction histogram and the texture density matrix. The adjacent band category difference is determined based on whether the adjacent band categories and coverage directions are consistent. Then, reference correspondence solution is performed based on the reference correspondence cost. Under the single correspondence constraint that the stable reference region of the same current frame corresponds to only one stable reference region of the adjacent frame and the stable reference region of the same adjacent frame is occupied by only one stable reference region of the current frame, the candidate correspondence with the lowest cost and consistent category is retained, the set of stable reference correspondences of adjacent frames is output, and the current frame reference identifier, adjacent frame reference identifier, corresponding cost and corresponding state of each correspondence are written into the stable reference correspondence record table for the next step of reading; if there is no candidate relationship that satisfies the single correspondence constraint in the adjacent frame for a certain stable reference region, a non-correspondence mark is written and the stable reference region is prohibited from being used as the basis for attachment position solution; if there are multiple candidate relationships with the same corresponding cost, the candidate relationships with consistent adjacent band category codes and small boundary morphology differences are retained first, and the remaining candidate relationships are written into the conflict record; Next, the edge computing device performs joint location attribution calculation on each suspected anomaly region based on the stable reference correspondence set of adjacent frames to form a candidate set of slope attachment positions in adjacent frames, providing a candidate basis for subsequent unique attribution output; the input includes the stable reference correspondence set of adjacent frames, the set of suspected anomaly regions, and the set of anomaly attachment zones; the processing actions include: reading the corresponding stable reference region and the corresponding anomaly attachment zone of each suspected anomaly region in its frame, and calculating the orientation relationship, distance relationship, and boundary contact relationship of the suspected anomaly region relative to the corresponding stable reference region, as well as the intra-zone positional relationship, zone edge adjacency relationship, and zone direction extension relationship relative to the corresponding anomaly attachment zone; Among them, the orientation relationship is determined by the azimuth angle interval between the center point of the suspected anomaly area and the center point of the stable reference area; the distance relationship is determined by the ratio of the Euclidean distance between the two center points to the equivalent width of the stable reference area; the boundary contact relationship is determined by whether there is a contact segment between the boundary of the suspected anomaly area and the boundary of the stable reference area and the proportion of the contact segment length; the intra-zone position relationship is calculated based on the longitudinal and lateral positions of the suspected anomaly area within the anomaly attachment zone; the zone edge adjacency relationship is determined by the shortest projected distance and contact state between the boundary of the suspected anomaly area and the boundary of the anomaly attachment zone; and the zone directional extension relationship is determined by the angle difference between the main extension direction of the suspected anomaly area and the main extension direction of the anomaly attachment zone and the proportion of the extension length. Subsequently, the six types of relationships of the same suspected anomaly region in the current frame and adjacent frames are input into the joint location attribution unit to generate a candidate set of slope attachment locations. The joint location attribution unit sorts the multiple candidate attachment locations of the same suspected anomaly region in adjacent frames according to the rules of relationship consistency priority, anomaly attachment band category consistency priority, and stable reference retention priority. Among them, relationship consistency priority is used to retain candidate locations with smaller changes in the six types of relationships, anomaly attachment band category consistency priority is used to retain candidate locations from the same anomaly attachment band category in the current frame and adjacent frames, and stable reference retention priority is used to retain candidate locations attached to the corresponding set of confirmed stable references. The output is a candidate set of slope attachment positions for each suspected anomaly region in adjacent frames. The candidate identifier, corresponding stable reference region identifier, corresponding anomaly attachment zone identifier, six types of relationship calculation results, and candidate ranking record of each candidate position are written into the attachment position candidate record table for the next step of reading. If a suspected anomaly region forms only one set of valid relationship combinations in its corresponding anomaly attachment zone, a single candidate position is directly generated. If a suspected anomaly region is adjacent to two anomaly attachment zone boundaries at the same time, two candidate positions are generated respectively and written into the multi-candidate mark, to be resolved by the next step of attribution consistency verification. If a suspected anomaly region lacks a corresponding stable reference region in adjacent frames, the joint position attribution solution is not performed, and the suspected anomaly region is written into the frame to be supplemented record table. Finally, the edge computing device performs a consistency check on the slope attachment position candidate set and the stable reference correspondence set of adjacent frames to resolve multiple candidate conflicts and output a unique attachment position attribution set for suspected abnormal regions. The input includes the slope attachment position candidate set and the stable reference correspondence set of adjacent frames for each suspected abnormal region. The processing actions include: performing attachment position coincidence check, attachment band category consistency check, and reference relationship maintenance check on all candidate positions of the same suspected abnormal region in adjacent frames one by one. The attachment position coincidence check is determined by the normalized position difference and intra-band position difference in the coordinate system of the stable reference region corresponding to the candidate position. The attachment band category consistency check is determined by whether the abnormal attachment band category corresponding to the candidate position in the current frame and the adjacent frame is consistent. The reference relationship maintenance check is determined by whether the stable reference region to which the candidate position is attached belongs to the same set of stable reference correspondences. Subsequently, candidate elimination is performed based on the above three verification results. When the same candidate simultaneously satisfies the conditions of overlapping attachment positions, consistent attachment zone types, and maintained reference relationships, the candidate is determined as the unique slope attachment position. When multiple candidates simultaneously satisfy the above three conditions, the candidate with the smaller normalized position difference and the smaller intra-zone position difference is retained, and the remaining candidates are written into the elimination record. When all candidates fail to satisfy the three conditions simultaneously, the candidate with consistent attachment zone types and maintained reference relationships is retained, and the attachment position overlap failure status is written into the attachment position aberration record, which serves as the basis for offset before subsequent attachment relationship consistency determination. The output is a set of attachment locations for suspected abnormal areas. Each record in the set includes at least the suspected abnormal area identifier, the corresponding stable reference area identifier, the corresponding abnormal attachment zone identifier, the slope attachment location identifier, the attribution status, and the verification result. These records are written to the attachment location attribution record table for S3 to read. If the same suspected abnormal area exists in the current frame but the adjacent frame lacks a corresponding suspected abnormal area, a single-frame attachment location pending confirmation mark is written, and it is not included in the attachment location attribution set of the current batch. If the stable reference correspondence is found to be invalid during the attribution consistency verification process, the stable reference correspondence record table is called back, and the reference correspondence solution is re-executed for the suspected abnormal area. Through the above specific implementation method, a stable reference correspondence between adjacent frames can be established first by using boundary morphology encoding, texture distribution encoding, and adjacent band category encoding of the stable reference region. Then, a candidate set of slope attachment positions is formed by combining the six types of relationships between the suspected abnormal region and the stable reference region and the abnormal attachment band. Finally, a unique attribution output is completed by the attachment position coincidence result, the attachment band category consistency result, and the reference relationship maintenance result. This transforms the suspected abnormal region from a local region at the image level into a slope attachment position with entity attribution significance. After this processing, the subsequent attachment relationship consistency determination no longer relies on simple region overlap or inter-frame proximity, but reads the structured attachment position attribution result for judgment, avoiding misattribution problems caused by changes in viewpoint, local occlusion, or changes in band edge contact. In practical applications: When a drone inspects the high slope of a hydropower station and continuously collects two adjacent frames of images, the edge computing device first reads the stable reference set, the abnormal attachment zone set, and the suspected abnormal area set output by the previous steps. Using the stable support component area and the stable bare rock texture area as stable reference nodes, it solves the stable reference corresponding set in adjacent frames. Then, for a suspected abnormal area located within a drainage-affected abnormal attachment zone, it calculates its orientation, distance, and boundary contact relationship relative to the corresponding stable support component area, as well as its intra-zone position, zone edge adjacency, and zone direction extension relationship relative to the corresponding drainage-affected abnormal attachment zone, forming a candidate set of slope attachment positions. Finally, through a consistency check, it retains the unique candidate position that has overlapping attachment positions, consistent attachment zone categories, and maintained reference relationships, and writes this candidate position into the suspected abnormal area attachment position attribution set for direct use in subsequent attachment relationship consistency determination.

[0019] S3. Perform a consistency determination of attachment relationship for the attachment location set of suspected abnormal areas. When the slope attachment location, abnormal attachment zone category and attachment relationship of the same suspected abnormal area are consistent in adjacent frames, output the attachment retention result; otherwise, output the attachment offset result. This specific implementation is used to further determine whether a suspected anomaly area maintains the same attachment relationship between adjacent frames after the slope attachment position has been assigned in the suspected anomaly area, thereby providing stable input for subsequent risk performance condition determination. Its basic principle is as follows: First, based on the attachment position assignment set of the suspected anomaly area, the slope attachment position, anomaly attachment band category, and attachment relationship of the same suspected anomaly area in adjacent frames are organized into attachment determination units, and position correspondence verification, band category correspondence verification, and relationship type correspondence verification are performed respectively to form attachment correspondence results. Then, conflict merging and main conflict identification are performed on the position changes, category switching, and relationship transfers in the attachment correspondence results to solve the attachment consistency state of the same suspected anomaly area in adjacent frames. Finally, the result assignment is performed based on the attachment consistency state, and the attachment maintenance result or attachment offset result is output. Here, the attachment relationship refers to the spatial relationship type exhibited by the suspected anomaly area relative to the corresponding anomaly attachment band and the corresponding stable reference area, including at least intra-band maintenance, band edge adjacency, cross-boundary contact, and band extension. This implementation process includes the following steps: First, the edge computing device constructs an attachment determination unit based on the attachment location attribution set of suspected abnormal areas, and performs three types of corresponding verification on each attachment determination unit to convert the attachment status of the same suspected abnormal area in adjacent frames into comparable verification results. The input includes the attachment location attribution set of suspected abnormal areas, the corresponding stable reference area identifier, the abnormal attachment band identifier, and the attachment relationship record. Each attachment location attribution record includes at least the suspected abnormal area identifier, the slope attachment location identifier of the current frame and the adjacent frames, the abnormal attachment band category, the attachment relationship type, and the attribution status. The processing action includes: pairing the attachment location attribution records of the current frame and the adjacent frames that belong to the same suspected abnormal area according to the suspected abnormal area identifier to form an attachment determination unit. Subsequently, position correspondence verification, band category correspondence verification, and relationship type correspondence verification are performed sequentially for each attachment determination unit. Position correspondence verification is determined by reading the slope attachment position identifiers of the current frame and adjacent frames and comparing their normalized position difference, intra-band position difference, and boundary contact position difference in the same stable reference area coordinate system. If all three differences are within the position maintenance interval formed by the statistics of the current inspection batch, the position correspondence verification is successful. The position maintenance interval comes from the position fluctuation statistics of the confirmed stable attachment areas in the current inspection batch. Band category correspondence verification is determined by comparing whether the abnormal attachment band categories corresponding to the attachment positions of the current frame and adjacent frames are consistent. Relationship type correspondence verification is determined by comparing whether the attachment relationship types in the current frame and adjacent frames belong to the same relationship type group. The relationship type group is predefined by the implementation rules as intra-band maintenance group, band edge adjacency group, cross-boundary contact group, and band extension group. The output is the attachment corresponding result set. Each record in the attachment corresponding result set includes at least the suspected abnormal area identifier, the attachment judgment unit identifier, the position corresponding verification result, the category corresponding verification result, the relationship type corresponding verification result, and the verification status. These are written into the attachment corresponding record table for the next step of reading. When a suspected abnormal area lacks an attachment position attribution record in an adjacent frame, an attachment judgment unit is not constructed. Instead, a missing frame marker is written and reserved as an attachment offset warning object for subsequent frames to supplement and verify. Next, the edge computing device performs relationship conflict resolution based on the attachment correspondence result set to distinguish whether the change in the attachment status of the same suspected abnormal area is caused by a single local disturbance or has formed an attachment relationship migration, thereby outputting an attachment consistency judgment result; the input includes the attachment correspondence result set and the slope attachment location identifier, abnormal attachment zone category and attachment relationship type record read when constructing the attachment judgment unit; the processing actions include: firstly mapping the cases of location correspondence failure, zone category correspondence failure and relationship type correspondence failure to slope attachment location change, abnormal attachment zone category switch and attachment relationship transfer, respectively, and organizing them into conflict judgment groups according to the same suspected abnormal area; Subsequently, for each conflict determination group, the position deviation direction, band category preservation status, and relationship transfer path are calculated. The position deviation direction is determined based on the displacement vector of the current frame attachment position relative to the adjacent frame attachment position in the stable reference region coordinate system, and is recorded according to four types of directions: towards the in-band, towards the band edge, cross-band outward movement, and along the band. The band category preservation status is determined based on whether the abnormal attachment band category of the current frame and the adjacent frame is consistent. The relationship transfer path is determined based on the order of change of the current frame attachment relationship type and the adjacent frame attachment relationship type among the relationship type groups. For example, the change from in-band preservation to band edge adjacency and from band edge adjacency to cross-boundary contact are recorded separately as different relationship transfer paths. Subsequently, the edge computing device performs conflict merging on multiple conflicts in the same conflict determination group. The conflict merging rules are as follows: priority is given to retaining conflict items that are consistent with the direction of position deviation and have continuous relationship transfer paths, and conflict items that contradict the state of the belt category and do not form a continuous transfer path are removed. After the conflict merging is completed, the main conflict is identified. The main conflict is determined according to the order of its impact on the stability of the attachment relationship. First, the abnormal attachment belt category switch is compared, then the attachment relationship transfer is compared, and finally the slope attachment position change is compared. When the previous type of conflict is established, the next type of conflict is only recorded as an accompanying conflict and no longer dominates the determination on its own. The output is the attachment consistency judgment result corresponding to each suspected abnormal area. The judgment result includes at least the suspected abnormal area identifier, the main conflict type, the accompanying conflict type, the location deviation direction, the relationship transfer path, and the consistency status. It is written into the attachment consistency judgment record table for the next step of reading. If there are multiple conflict items in the same suspected abnormal area and the unique main conflict cannot be determined after merging the conflicts, the conflict item with the category-maintained status is retained, and the remaining conflict items are written into the attachment conflict pending review record. Finally, the edge computing device assigns the results of each suspected abnormal region based on the attachment consistency determination results to form the attachment retention results and attachment offset results directly read in the subsequent risk performance condition determination. The input includes the attachment consistency determination results, the attachment corresponding result set, and the corresponding attachment position assignment record of the suspected abnormal region. The processing actions include: reading the consistency status, main conflict type, and accompanying conflict type in the attachment consistency determination results of each suspected abnormal region, and assigning the results according to the unified assignment rule. The unified assignment rule is: when the slope attachment position, abnormal attachment band category, and attachment relationship of the same suspected abnormal region are consistent in adjacent frames, it is determined to be an attachment retention result. When the categories of abnormal attachment zones are inconsistent, it is directly determined as an attachment shift result; when the categories of abnormal attachment zones are consistent but the attachment relationships are inconsistent, it is determined whether it is a relationship migration based on the main conflict identification result. If the main conflict is an attachment relationship transfer, it is determined as an attachment shift result; when the categories of abnormal attachment zones are consistent and the attachment relationships are consistent, but the attachment positions on the slope are inconsistent, the position deviation direction is read. If the position deviation direction is still within the same abnormal attachment zone and does not cross the zone edge, a position disturbance retention mark is written and it is retained as an attachment retention result. If the position deviation direction points out of the zone or crosses the zone edge, it is determined as an attachment shift result. The output includes attachment retention results and attachment offset results. The attachment retention results include at least the suspected abnormal area identifier, retention status, corresponding slope attachment location identifier, and retention type. The attachment offset results include at least the suspected abnormal area identifier, offset status, offset direction, and main conflict type. Both types of results are written to the attachment result record table for S4 to read. If the attachment consistency determination result is found to be missing the main conflict type or location deviation direction field during the assignment process, the attachment consistency determination record table is called back to re-execute the relationship conflict resolution, and the final assignment result is not directly output. Through the above specific implementation method, the slope attachment position, abnormal attachment zone category, and attachment relationship of the same suspected abnormal area in adjacent frames can be organized into attachment determination units. Then, through three types of correspondence verification, conflict merging, and main conflict identification, an attachment consistency determination result with a unique attribution caliber is formed. Finally, the attachment retention result and attachment offset result are output. After this processing, attachment retention and attachment offset no longer depend on a single position change or a single category change, but are determined based on the joint result of three types of information: position, category, and relationship. This avoids misjudging local position disturbances as attachment offsets and also avoids misjudging real relationship migrations as attachment retention. In practical applications: When a suspected anomaly is located within the same drainage-affected anomaly attachment zone in both the current frame and adjacent frames, and its attachment relationship extends along the zone, while its slope attachment position only shifts within a small range within the same zone, the edge computing device will record the position disturbance in the position correspondence verification, record the validity in the zone category correspondence verification and relationship type correspondence verification, and determine it as an attachment retention result when attributing the results; when another suspected anomaly is located within the repair boundary-type anomaly attachment zone in the current frame and transitions to the support adjacency-type anomaly attachment zone in adjacent frames, and the attachment relationship changes from zone edge adjacency to cross-boundary contact, the edge computing device will identify the anomaly attachment zone category as the main conflict and determine the suspected anomaly as an attachment offset result, for direct differentiation and processing when determining subsequent risk performance conditions.

[0020] S4. For the suspected abnormal areas corresponding to the attachment retention results, perform risk performance condition judgment according to the corresponding abnormal attachment zone category. When the abnormal performance continuously meets the risk performance conditions of the corresponding abnormal attachment zone, output the real hidden danger mark; otherwise, output the imaging pseudo-anomaly mark. This specific implementation is used to further determine whether a suspected abnormal area continuously exhibits a risk appearance consistent with the attachment band category within the corresponding abnormal attachment band after the suspected abnormal area has been identified as being in an attachment-maintaining state. This further distinguishes the attachment-maintaining results into genuine hidden dangers and imaging pseudo-anomalies. Its basic principle is as follows: First, the suspected abnormal area, abnormal attachment band category, and adjacent frame inspection images corresponding to the attachment-maintaining results are read. Then, directional extension, boundary adhesion, apparent continuous change, and intra-band occupancy change are extracted around the abnormal attachment band category to form a risk performance feature set. Finally, based on the risk performance constraints corresponding to different abnormal attachment band categories, the aforementioned features are applied... The process involves: validating conditions within each category to form a risk performance determination group and outputting the risk performance establishment result; finally, assigning a label based on the risk performance establishment result; labels that remain valid in adjacent frames are judged as genuine hidden danger labels, otherwise they are judged as imaging pseudo-anomaly labels; the risk performance constraints here refer to the pre-defined feature combination establishment rules for different anomaly attachment band categories, which are derived from the category determination rules jointly determined by the inspection business rule constraints and historical sample statistical results; the directional extension, boundary adhesion, apparent continuous change, and band occupancy change are all calculated based on the same suspected anomaly area and corresponding anomaly attachment band in adjacent frames; this implementation process includes the following steps: First, the edge computing device extracts risk performance features matching the abnormal attachment band category around the suspected abnormal region corresponding to the attachment retention result, so as to establish the subsequent true / false determination on the basis of continuous apparent changes. The input includes the suspected abnormal region corresponding to the attachment retention result, the abnormal attachment band category, the corresponding abnormal attachment band boundary record, and the inspection image of adjacent frames. Each attachment retention record includes at least the suspected abnormal region identifier, the corresponding abnormal attachment band identifier, the retention type, and the correspondence with adjacent frames. The processing actions include: reading local image segments in the current frame and adjacent frames according to the suspected abnormal region identifier, and unifying the position reference under the corresponding abnormal attachment band coordinate system; then, extracting four types of performance quantities according to the abnormal attachment band category. Among them, the directional extension quantity is calculated by calculating the main extension direction of the suspected abnormal region. The difference in angle between the suspected anomalous region and the main extension direction of the anomalous attachment band, and the proportion of the newly added coverage length of the suspected anomalous region along the main extension direction to the original coverage length in adjacent frames; the boundary fitting amount is obtained by calculating the proportion of the contact segment length between the boundary of the suspected anomalous region and the boundary of the anomalous attachment band to the total boundary length of the suspected anomalous region, and the change in this proportion in adjacent frames; the apparent continuous change amount is obtained by comparing the continuous change results of the texture coding sequence, brightness distribution sequence and edge direction sequence inside the suspected anomalous region in adjacent frames, and the continuous change results are determined by comparing the difference between two adjacent frames with the statistical interval of normal fluctuations within the same band in the current batch; the change in in-band occupancy is obtained by calculating the changes in the projection length, projection width and center position proportion of the suspected anomalous region within the anomalous attachment band in adjacent frames; Next, the four types of performance indicators and the abnormal attachment band categories are written together into the risk performance feature record to form the corresponding risk performance feature set, and written into the risk feature record table for the next step of conditional verification within the category; if a local image segment in an adjacent frame is affected by occlusion or strong reflection and cannot extract complete features, the remaining extractable features are retained and a feature missing mark is written; if only one of the four types of features is missing, it is allowed to proceed to the next step of verification; if two or more types are missing, the suspected abnormal area is written into the frame to be supplemented record and is not included in the current round of risk performance judgment; Next, the edge computing device performs in-category condition verification based on the risk performance feature set and the risk performance constraints corresponding to the abnormal attachment zone category to form a result indicating that the risk performance is valid. The input includes the risk performance feature set and the risk performance constraint library corresponding to the abnormal attachment zone category. The risk performance constraint library stores directional extension continuity constraints, boundary fit continuity constraints, apparent continuous change consistency constraints, and zone occupancy retention constraints according to the category. The constraint library is generated by the statistical results of historical inspection samples and inspection business rule constraints and is pre-written into the edge computing device. The processing actions include: firstly, selecting the corresponding risk performance constraint group according to the abnormal attachment zone category. For example, for drainage-affected abnormal attachment zones, directional extension continuity constraints and zone occupancy retention constraints are preferentially called; for support-adjacent abnormal attachment zones, boundary fit continuity constraints and apparent continuous change consistency constraints are preferentially called; for repair trace abnormal attachment zones, boundary fit continuity constraints and directional extension continuity constraints are preferentially called; and for slope apparent boundary abnormal attachment zones, apparent continuous change consistency constraints and zone occupancy retention constraints are preferentially called. Subsequently, in-category conditional checks were performed on each suspected abnormal region. Specifically, the directional extension continuity check was determined by whether the difference in the angle between the main extension directions in adjacent frames fell within the allowed directional fluctuation range of the current category and whether the proportion of newly added coverage length remained in the same direction. The allowed directional fluctuation range was derived from the statistical estimation results of confirmed attachment-maintaining regions within the same type of abnormal attachment band in the current batch. The boundary fit continuity check was determined by whether the proportion of boundary contact segments maintained a similar continuous change state in adjacent frames. The apparent continuity consistency check was determined by whether the difference in texture coding sequence, brightness distribution sequence, and edge direction sequence all fell within the allowed apparent change range of the current category. The allowed apparent change range was derived from the statistical results of historical samples of the same category. The in-band occupancy retention check was determined by whether the proportion of projection length, projection width, and center position remained in the same direction or unchanged in adjacent frames. Subsequently, the four types of verification results are grouped into risk performance judgment groups according to their categories. The validity of the judgment group is then determined based on the requirements of the current category. The validity rule is constrained by business rules, requiring that all mandatory items be valid and at least one optional item be valid. The output is the risk performance validity result corresponding to each suspected abnormal area. The result includes at least the suspected abnormal area identifier, the abnormal attachment category, the risk performance judgment group, the validity status of each verification item, and the overall validity status. This result is written into the risk performance result record table for the next step of marking and attribution. If a suspected abnormal area is missing a feature item that is a mandatory item for the current category, the overall validity status is directly recorded as invalid. If only an optional item is missing, the judgment is performed based on the remaining available items. Finally, the edge computing device performs labeling based on the risk performance results to form real hazard labels and imaging pseudo-anomaly labels. The input includes the risk performance results, the correspondence between suspected anomaly areas in adjacent frames, and the attachment and retention result records. The processing actions include: reading the risk performance judgment groups in the current frame and adjacent frames according to the suspected anomaly area identifier, and comparing their overall establishment status and the status of each component; then executing the labeling rules. When the risk performance judgment group of the same suspected anomaly area is established as a whole in adjacent frames, and at least one of the mandatory items in the current frame and adjacent frames remains continuously established, it is determined as a real hazard label. When the risk performance judgment group is not established as a whole, or although it is established in the current frame but not in adjacent frames, and the non-established item involves a mandatory item of the current category, it is determined as an imaging pseudo-anomaly label. When both the current frame and adjacent frames are established, but only the optional items in the components change, the real hazard label is retained and written into the component fluctuation record. Afterwards, the real hazard markers or imaging pseudo-anomaly markers, along with the corresponding suspected anomaly area identifiers, anomaly attachment band categories, risk performance judgment groups, and judgment criteria, are written into the marking result record table for S5 to read. If the missing adjacent frames make it impossible to compare risk performance judgment groups, the current frame's status is retained and a single-frame pending confirmation marker is written, without directly outputting the final marker. If the correspondence between the suspected anomaly areas in the current frame and adjacent frames fails, the attachment retention result record table is read back, and the inter-frame correspondence for the suspected anomaly area is re-verified before marking is assigned. Through the above specific implementation method, risk performance features matching the abnormal attachment band category can be extracted from the suspected abnormal area corresponding to the attachment retention result. Then, the in-category condition verification is performed according to the categorized risk performance constraints. Finally, the real hazard mark or imaging pseudo-anomaly mark is output, thereby further converging the attachment retention result into a true / false judgment result that can be directly used for hazard identification output. After this processing, the real hazard mark no longer depends on whether a single abnormal appearance appears, but on the joint retention result of directional extension, boundary fitting, continuous appearance change and in-band occupancy retention in adjacent frames. The imaging pseudo-anomaly mark is no longer directly triggered by a single frame anomaly, but is determined by the failure of the risk performance judgment group or the continuous interruption. Therefore, it can reduce misjudgment caused by local illumination changes, shadow occlusion and texture disturbance. In practical applications: When a suspected abnormal area is located within the drainage impact anomaly attachment zone, the edge computing device prioritizes reading its directional extension and intra-zone occupancy change, and combines them with the boundary adhesion and apparent continuous change to form a risk performance feature set. If the suspected abnormal area extends along the drainage direction in adjacent frames, and its proportion of projection length within the zone continuously increases, while the apparent continuous change falls within the allowable change range for drainage impact, then the risk performance judgment group remains valid, and the system outputs a genuine hazard marker. If another suspected abnormal area shows brightness changes in a single frame, but the directional extension is interrupted in adjacent frames, the boundary adhesion is discontinuous, and the intra-zone occupancy change has no common maintenance range, then the risk performance judgment group is invalid, and the system outputs it as an imaging pseudo-anomaly marker for subsequent result generation steps to distinguish and process.

[0021] S5. For suspected abnormal areas with real hazard markers, generate the safety hazard identification results of the high slope of the hydropower station according to the corresponding slope attachment location, abnormal attachment zone category and risk performance condition judgment results, and output them to the edge computing device. This specific implementation method is used to further group suspected abnormal areas that have been identified as real hidden dangers into outputtable high slope hidden danger identification units, and generate high slope safety hidden danger identification results for hydropower stations based on these units. The basic process is as follows: First, read the suspected abnormal areas, slope attachment locations, abnormal attachment zone categories, and risk performance condition judgment results corresponding to the real hidden danger markers. Then, perform hidden danger merging calculations according to area merging relationships, category merging relationships, and result merging relationships to form a set of high slope hidden danger identification units. Next, generate structured safety hidden danger identification results based on the constituent fields of each high slope hidden danger identification unit, and write them to an edge computing device for subsequent display, early warning, and feedback. Here, a high slope hidden danger identification unit refers to a hidden danger merging result unit jointly defined by the same slope attachment location, the same abnormal attachment zone category, and the same risk performance condition judgment result. This implementation process includes the following steps: First, the edge computing device performs hazard merging calculation on suspected abnormal areas marked with real hazard markers to eliminate duplicate outputs of the same hazard in adjacent areas or adjacent frames. The input includes the suspected abnormal area corresponding to the real hazard marker, the corresponding slope attachment location, the abnormal attachment zone category, and the risk performance condition judgment result. The processing actions include: reading the corresponding slope attachment location identifier, abnormal attachment zone category, and risk performance condition judgment result according to the suspected abnormal area identifier, and constructing a hazard merging candidate table; then, the first round of merging is performed according to the area merging relationship of the same slope attachment location. The area merging relationship is determined by comparing the area overlap ratio, boundary continuity relationship, and center position offset of suspected abnormal areas in the same slope attachment location coordinate system. The allowable range of the area overlap ratio and center position offset comes from the statistical estimation results of similar real hazard areas in the current inspection batch. A second round of merging is performed according to the category merging relationship of the same abnormal attachment zone, merging records from the same abnormal attachment zone category that have completed area merging; then, a third round of merging is performed according to the result merging relationship of the same risk performance condition judgment result, merging records with consistent risk performance condition judgment results into the same hazard merging result; after completing the three rounds of merging, a set of high slope hazard identification units is output, and the unit identifier, the set of suspected abnormal areas, the slope attachment location identifier, the abnormal attachment zone category, the risk performance condition judgment result, and the merging status of each high slope hazard identification unit are written into the hazard identification unit record table for the next step of reading; if a suspected abnormal area meets the area merging conditions of two slope attachment locations at the same time, the merging relationship with the higher area overlap ratio and the smaller center position offset is retained, and the remaining merging relationships are written into the conflict record; if a suspected abnormal area lacks a risk performance condition judgment result, it will not participate in the third round of merging and will be marked as pending completion; Next, the edge computing device generates the safety hazard identification results of the high slope of the hydropower station based on the high slope hazard identification unit set, so as to form a unified output result. The input includes the high slope hazard identification unit set and its corresponding slope attachment location, abnormal attachment zone category and risk performance condition judgment result. The processing actions include: reading the constituent fields of the high slope hazard identification unit one by one, and generating the safety hazard identification result according to the predefined output field template. The output field template includes at least the hazard result identifier, the corresponding slope attachment location, the corresponding abnormal attachment zone category, the risk performance condition judgment result, the range of the suspected abnormal area, and the output status. Subsequently, the output fields corresponding to each high slope hazard identification unit are written into the result buffer, and an integrity check is performed on all safety hazard identification results within the same inspection batch. The integrity check is used to check whether each result simultaneously has the slope attachment location field, abnormal attachment zone category field, and risk performance condition judgment result field. When the integrity check is successful, the hydropower station high slope safety hazard identification result is generated and output to the edge computing device. When the integrity check is unsuccessful, the hazard identification unit record table is read back, and the high slope hazard identification units with missing fields are supplemented with field verification, and the corresponding results are not directly output. The output is the hydropower station high slope safety hazard identification result, and the result is written into the local result cache area of ​​the edge computing device for subsequent early warning display, task feedback, and historical comparison. Through the above specific implementation methods, suspected abnormal areas with real hazard markers can be uniformly merged according to the slope attachment location, abnormal attachment zone category and risk performance condition judgment results to form a high slope hazard identification unit. Then, based on this unit, a structured hydropower station high slope safety hazard identification result is output, thereby avoiding the same hazard being output repeatedly and avoiding different hazards being incorrectly merged. In practical applications: When multiple suspected abnormal areas with real hazard markers exist at the same slope attachment location, and these areas all originate from the same drainage-affected abnormal attachment zone and correspond to the same risk performance condition judgment result, the edge computing device first merges them into the same high slope hazard identification unit, and then generates a safety hazard identification result that includes the slope attachment location, the drainage-affected abnormal attachment zone category, and the corresponding risk performance condition judgment result, and writes it into the edge computing device for subsequent display and feedback.

[0022] Furthermore, the present invention also includes an intelligent identification system for safety hazards on high slopes of hydropower stations based on inspection images. The system includes an anomaly extraction module, an attachment mapping module, a judgment module, a discrimination module, and a result generation module. The anomaly extraction module is used to acquire inspection images continuously collected by edge computing devices during the inspection of high slopes in hydropower stations. It identifies stable reference areas on the slope in each inspection image, extracts abnormal attachment zones and suspected abnormal areas located within the abnormal attachment zones, and outputs a set of suspected anomalies, a set of stable reference zones, and a set of abnormal attachment zones in adjacent frames. The attachment mapping module is used to perform reference correspondence identification on the stable reference set in adjacent frames, and based on the positional relationship between each suspected abnormal region and the corresponding stable reference region and the abnormal attachment zone, it maps each suspected abnormal region to the corresponding slope attachment position and outputs the attachment position attribution set of the suspected abnormal region. The determination module is used to perform a consistency determination of the attachment relationship of the attachment position set of suspected abnormal areas. When the slope attachment position, abnormal attachment band category and attachment relationship of the same suspected abnormal area are consistent in adjacent frames, the attachment retention result is output; otherwise, the attachment offset result is output. The discrimination module is used to determine the risk performance conditions of suspected abnormal areas corresponding to the adhesion retention results based on their corresponding abnormal adhesion band categories. When the abnormal performance continuously meets the risk performance conditions of the corresponding abnormal adhesion band, a real hazard mark is output; otherwise, an imaging pseudo-anomaly mark is output. The results generation module is used to generate safety hazard identification results for high slopes of hydropower stations for suspected abnormal areas with real hazard markers, based on the corresponding slope attachment location, abnormal attachment zone category and risk performance conditions, and output the results to the edge computing device.

[0023] Working principle: This solution first allows the edge computing device to perform a step-by-step screening of continuous inspection images: First, it identifies reference areas on the slope that consistently exist in adjacent images. Then, it generates attachment zones around these reference areas and the boundaries of drainage marks, support components, repair marks, etc., where anomalies are likely to occur. Next, it extracts suspected anomaly areas only within these attachment zones. Then, the system determines whether these suspected anomalies consistently correspond to the same slope location in the previous and subsequent images, whether they still belong to the same attachment zone, and whether they maintain the same attachment relationship. If these relationships remain stable, it continues to determine whether the anomaly behavior is consistently valid. Only those consistently valid are identified as real hazards; otherwise, they are considered imaging pseudo-anomalies. Finally, real hazards at the same location are merged and the results are output. For example, when a drone inspects a high slope of a hydropower station, the footage may show cracks, wet spots, or signs of falling debris, or it may just show shadows, reflections, or swaying vegetation. This solution does not just look at whether a single frame looks like an anomaly, but first sees whether the anomaly consistently appears at the same location on the slope, and then sees whether it consistently changes along the boundaries of drainage marks, repairs, or support. If multiple frames match and the change pattern matches the actual manifestation of a potential hazard, it is output as a high slope safety hazard. If it only appears briefly in a single frame, is in an unstable location, or does not match the previous frames, it is not output as a hazard.

[0024] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent identification of safety hazards on high slopes of hydropower stations based on inspection images, characterized in that, include: S1. Acquire inspection images continuously collected by the edge computing device during the inspection of the high slope of the hydropower station, identify the slope stability reference area in each inspection image, extract the abnormal attachment zone and the suspected abnormal area located in the abnormal attachment zone, and output the suspected abnormal set, stable reference set and abnormal attachment zone set of adjacent frames. S2. Perform reference correspondence identification on the stable reference set in adjacent frames, and map each suspected abnormal region to the corresponding slope attachment position according to the positional relationship between each suspected abnormal region and the corresponding stable reference region and the abnormal attachment zone, and output the attachment position attribution set of the suspected abnormal region. S3. Perform a consistency determination of attachment relationship for the attachment location set of suspected abnormal areas. When the slope attachment location, abnormal attachment zone category and attachment relationship of the same suspected abnormal area are consistent in adjacent frames, output the attachment retention result; otherwise, output the attachment offset result. S4. For the suspected abnormal areas corresponding to the attachment retention results, perform risk performance condition judgment according to the corresponding abnormal attachment zone category. When the abnormal performance continuously meets the risk performance conditions of the corresponding abnormal attachment zone, output the real hidden danger mark; otherwise, output the imaging pseudo-anomaly mark. S5. For suspected abnormal areas marked with real hidden dangers, generate the safety hazard identification results of the high slope of the hydropower station according to the corresponding slope attachment location, abnormal attachment zone category and risk performance condition judgment results, and output them to the edge computing device.

2. The intelligent identification method for safety hazards of high slopes in hydropower stations based on inspection images as described in claim 1, characterized in that: S1 includes: S1-1. Obtain the slope foreground area in each inspection image, perform structural boundary extraction, and identify the fixed structure edge, stable bare rock texture area and stable support component area based on boundary continuity, texture repetition stability and cross-frame consistency retention, and output the slope stability reference candidate set. S1-2. Based on the candidate set of slope stability references, perform neighborhood relationship aggregation and boundary expansion calculation. Based on the adjacency relationship between each candidate region of slope stability references and the slope surface boundary, drainage trace boundary, support component boundary and repair trace boundary, generate the corresponding abnormal attachment zone and output the abnormal attachment zone set. S1-3. Using the set of abnormal attachment bands as a constraint region, perform in-band texture disturbance detection and appearance continuity destruction identification, and identify regions that meet the conditions of texture abrupt change, boundary breakage, brightness migration abnormality or morphological extension abnormality as suspected abnormal regions, and output the set of suspected abnormal regions located in each abnormal attachment band.

3. The intelligent identification method for safety hazards of high slopes in hydropower stations based on inspection images as described in claim 2, characterized in that: The process of outputting the abnormal attachment band set in S1-2 also includes: S1-21. Obtain the slope stability reference candidate set and the slope apparent boundary, drainage trace boundary, support component boundary and repair trace boundary. Construct an adjacency candidate graph with each slope stability reference candidate region and each boundary segment as nodes. The association cost function is constructed with the distance deviation between nodes, direction deviation, texture continuity deviation and occlusion violation. Perform association solution under the boundary single-class ownership constraint, local connectivity constraint and adjacent frame preservation constraint, and output the initial adjacency relationship set and the corresponding association confidence results. S1-22. Based on the initial adjacency set, perform neighborhood relationship aggregation and conflict resolution. Construct a consistency verification set from shared boundary segments, overlapping coverage areas, and multi-class competing adjacency relationships. Perform confidence updates based on boundary category mutual exclusion constraints, region inclusion constraints, and adjacent frame relationship continuation constraints. When the category affiliation of each adjacency relationship remains consistent in two adjacent confidence updates and the conflict relationships in the consistency verification set no longer change, output the confirmed adjacency set.

4. The intelligent identification method for safety hazards of high slopes in hydropower stations based on inspection images according to claim 3, characterized in that: The process of outputting the abnormal attachment band set in S1-2 also includes: S1-23. Perform directional boundary expansion calculations on each slope stability reference candidate area according to the confirmed adjacency relationship set. Construct the zone generation cost function with the expansion width deviation, cross-boundary leakage amount and apparent stability deviation within the zone. Solve the boundary direction expansion range without crossing heterogeneous boundary constraints, covering the corresponding adjacent boundary constraints and expansion termination boundary constraints, and output the candidate abnormal attachment zone and the corresponding zone generation confidence results. S1-24. Perform multi-view consistency verification on the candidate abnormal attachment bands and the confirmed adjacency relationship set in the adjacent frames. Perform joint verification on the band position drift, boundary category jump variable and band apparent fluctuation. Perform confidence correction and band filtering based on the band generation confidence result and consistency verification result, and output the abnormal attachment band set.

5. The intelligent identification method for safety hazards of high slopes in hydropower stations based on inspection images according to claim 4, characterized in that: S2 includes: S2-1. Based on the stable reference set, abnormal attachment band set, and suspected abnormal region set in adjacent frames, construct a cross-frame correspondence map with stable reference regions as reference nodes and suspected abnormal regions as nodes to be assigned. Then, perform reference correspondence solving based on the boundary morphology encoding, texture distribution encoding, and adjacent band category encoding of each stable reference region, and output the stable reference correspondence set of adjacent frames. S2-2. Based on the stable reference set of adjacent frames, perform joint position attribution calculation on the orientation, distance and boundary contact relationship of each suspected anomaly region relative to the corresponding stable reference region, as well as the intra-band position relationship, band edge adjacency relationship and band extension relationship relative to the corresponding anomaly attachment band, and output the candidate set of slope attachment positions of each suspected anomaly region in adjacent frames. S2-3. Perform a consistency check between the candidate set of slope attachment positions of each suspected abnormal region in adjacent frames and the corresponding set of stable references in adjacent frames. Based on the overlap of attachment positions of the same suspected abnormal region in adjacent frames, the consistency of attachment zone categories, and the retention of reference relationships, determine the corresponding slope attachment positions and output the set of attachment positions of suspected abnormal regions.

6. The intelligent identification method for safety hazards of high slopes in hydropower stations based on inspection images as described in claim 5, characterized in that: S3 includes: S3-1. Based on the attachment location attribution set of suspected abnormal areas, construct attachment determination units according to the slope attachment location, abnormal attachment band category and attachment relationship of the same suspected abnormal area in adjacent frames, and perform position correspondence verification, band category correspondence verification and relationship type correspondence verification on each attachment determination unit, and output the attachment correspondence result set. S3-2. Based on the attachment corresponding result set, perform relationship conflict resolution, and form a conflict judgment group by changing the slope attachment position, switching of abnormal attachment zone categories, and transferring attachment relationships. Then, perform conflict merging and main conflict identification based on the position deviation direction, zone category maintenance status, and relationship transfer path in each conflict judgment group, and output the attachment consistency judgment result corresponding to each suspected abnormal area. S3-3. Assign the results of each suspected abnormal region according to the attachment consistency judgment results. When the slope attachment position, abnormal attachment band category and attachment relationship of the same suspected abnormal region are consistent in adjacent frames, it is determined to be the attachment retention result; otherwise, it is determined to be the attachment offset result.

7. The intelligent identification method for safety hazards of high slopes in hydropower stations based on inspection images as described in claim 6, characterized in that: S4 includes: S4-1. Obtain the suspected abnormal regions, abnormal attachment band categories, and adjacent frame inspection images corresponding to the attachment retention results. Extract the directional extension, boundary fitting, apparent continuous change, and intra-band occupancy change of each suspected abnormal region in adjacent frames according to the abnormal attachment band category, and construct the corresponding risk performance feature set.

8. The intelligent identification method for safety hazards of high slopes in hydropower stations based on inspection images according to claim 7, characterized in that: S4 also includes: S4-2. Based on the risk performance feature set, perform in-category condition verification on each suspected abnormal region according to the risk performance constraints corresponding to each abnormal attachment zone category, and form a risk performance judgment group by directional extension continuity, boundary fitting continuity, apparent continuous change consistency and in-band occupancy retention, and output the risk performance results corresponding to each suspected abnormal region. S4-3. Based on the risk performance results corresponding to each suspected abnormal area, perform the labeling and attribution. When the risk performance judgment group of the same suspected abnormal area remains valid in adjacent frames, output the real hidden danger label; otherwise, output the imaging pseudo-anomaly label.

9. The intelligent identification method for safety hazards of high slopes in hydropower stations based on inspection images according to claim 8, characterized in that: S5 includes: S5-1. Take the suspected abnormal areas with real hazard markers, the corresponding slope attachment locations, abnormal attachment zone categories and risk performance condition judgment results, and perform hazard merging calculation according to the area merging relationship of the same slope attachment location, the category merging relationship of the same abnormal attachment zone category and the result merging relationship of the same risk performance condition judgment results, and output the high slope hazard identification unit set. S5-2. Based on the high slope hazard identification unit set, generate the hydropower station high slope safety hazard identification results according to the slope attachment location, abnormal attachment zone category and risk performance condition judgment results corresponding to each high slope hazard identification unit, and output the hydropower station high slope safety hazard identification results to the edge computing device.

10. A smart identification system for safety hazards on high slopes of hydropower stations based on inspection images, used to implement the smart identification method for safety hazards on high slopes of hydropower stations based on inspection images as described in any one of claims 1-9, the system comprising an anomaly extraction module, an attachment mapping module, a judgment module, a discrimination module, and a result generation module, characterized in that: The anomaly extraction module is used to acquire inspection images continuously collected by edge computing devices during the inspection of high slopes in hydropower stations. It identifies stable reference areas on the slope in each inspection image, extracts abnormal attachment zones and suspected abnormal areas located within the abnormal attachment zones, and outputs a set of suspected anomalies, a set of stable reference zones, and a set of abnormal attachment zones in adjacent frames. The attachment mapping module is used to perform reference correspondence identification on the stable reference set in adjacent frames, and based on the positional relationship between each suspected abnormal region and the corresponding stable reference region and the abnormal attachment zone, it maps each suspected abnormal region to the corresponding slope attachment position and outputs the attachment position attribution set of the suspected abnormal region. The determination module is used to perform a consistency determination of the attachment relationship of the attachment position set of suspected abnormal areas. When the slope attachment position, abnormal attachment band category and attachment relationship of the same suspected abnormal area are consistent in adjacent frames, the attachment retention result is output; otherwise, the attachment offset result is output. The discrimination module is used to determine the risk performance conditions of suspected abnormal areas corresponding to the adhesion retention results based on their corresponding abnormal adhesion band categories. When the abnormal performance continuously meets the risk performance conditions of the corresponding abnormal adhesion band, a real hazard mark is output; otherwise, an imaging pseudo-anomaly mark is output. The results generation module is used to generate safety hazard identification results for high slopes of hydropower stations for suspected abnormal areas with real hazard markers, based on the corresponding slope attachment location, abnormal attachment zone category and risk performance conditions, and output the results to the edge computing device.

Images