Image recognition based flotation froth monitoring system

Abstract

The present application relates to the field of image recognition and computer vision, and particularly relates to a flotation froth monitoring system based on image recognition. The system comprises an image sequence receiving end and a vision processor. The image sequence receiving end receives continuous flotation froth image frames, and the vision processor performs froth boundary enhancement and anti-reflection area suppression on the image frames, extracts froth unit candidate regions, froth wall intersection points, froth group density and surface flow direction, and constructs a froth topological evolution graph which is updated over time, taking the froth unit candidate regions as nodes and the boundary contact, inter-frame area overlap, motion adjacency and fusion splitting relationship as edges. According to the froth topological evolution graph, the present application performs before-and-after frame consistency correction on single-frame abnormal segmentation regions, and generates froth stability state, froth load state and abnormal trend labels, thereby reducing the interference of reflection, shadow, froth wall overlap and transient noise on the flotation froth state recognition result.

Description

Technical Field

[0001] This invention relates to the fields of image recognition and computer vision technology, and more specifically to a flotation foam monitoring system based on image recognition. Background Technology

[0002] Flotation foam images can reflect the surface morphology, size distribution, clarity of the foam walls, mineralization banding, and surface flow. Current flotation foam monitoring typically uses continuously acquired foam images as the data source. After brightness equalization, noise reduction, region cropping, and color space conversion, features such as grayscale mean, color components, texture contrast, foam area ratio, and edge density are extracted from the foam images. Then, rule-based discrimination or classification models are used to determine the foam's sparseness, thickness, stability, or abnormality. The above methods primarily process single-frame images, relying mainly on pixel statistics and local texture distribution within a single frame for identification. To reduce the impact of ambient lighting variations, conventional methods also perform grayscale normalization or highlight reduction to keep feature extraction within the same brightness range. For continuous images, a common practice is to repeat the same identification process frame by frame, then arrange the identification results of each frame over time to form a monitoring curve. Because the foam surface itself is in a state of continuous flow, breakage, and renewal, there is a clear morphological continuity between adjacent frames, which the above processing typically simplifies to numerical changes. This type of method can judge the overall appearance changes of foam images, but its state discrimination still relies on static visual features such as color, brightness and texture, and lacks expression of the continuous evolution relationship between foam units.

[0003] In foam region extraction, existing technologies generally employ threshold segmentation, edge detection, watershed segmentation, morphological closing operations, or neural network segmentation models to obtain foam contours, and then calculate foam morphological features based on contour area, perimeter, roundness, pore size distribution, and bubble wall width. Due to the presence of numerous reflective spots, bubble wall intersections, bubble overlaps, and localized collapse areas on the flotation foam surface, conventional segmentation often requires masking of bright areas, smoothing of fracture boundaries, and deletion of small noise areas. These processes still primarily rely on the continuity of edges, the prominence of brightness, and the closure of regions in the current image. When reflective spots cover bubble walls, mineralized bands are similar in color to bubble walls, or foam contours are occluded in localized areas, the segmentation results are prone to incorrect merging, incorrect segmentation, or missing boundaries of foam units. Existing processing typically repairs contours within the same frame, lacking a mechanism for verification using foam positions, adjacency relationships, and the continuation of intersection points between consecutive frames, and also failing to consider boundary changes of the same foam unit in consecutive images as a criterion for contour reliability. Since subsequent bubble diameter statistics, area ratio calculations, and state classification all rely on the contour results, single-frame segmentation errors will be propagated along the subsequent processing links, causing local visual disturbances to be amplified into the basis for judging changes in bubble state.

[0004] To utilize continuous image information, existing flotation foam monitoring methods also perform temporal smoothing on the average brightness, foam area ratio, texture change rate, or foam flow velocity of adjacent frames, and vote or accumulate the classification results of multiple consecutive frames. This type of method essentially performs temporal post-processing on the single-frame recognition results. Inter-frame correlations are usually only reflected in the rise and fall of numerical curves or overall movement trends, without treating specific foam units as trackable objects, nor describing the adjacency, aggregation, breakage, and splitting relationships between foam units. Changes in flotation foam state often manifest as local foam first coarsening, aggregation, or breakage, then affecting the connection structure and movement consistency of surrounding foam groups. If only the average features of the entire image or the classification results are used for voting, local real changes may be diluted by normal areas; if only local anomalies in a single frame are relied upon, instantaneous reflections, shadows, or image noise are easily included in the anomaly judgment. This type of temporal smoothing method does not establish a correspondence between foam units, foam wall intersections, and regional adjacency relationships, cannot identify the trajectory continuity of the same foam unit in continuous images, and is difficult to perform image-level backtracking verification of anomaly segmentation results. For large outlines formed after local foam aggregation, fractured outlines formed after rupture, and missing outlines formed after being covered by reflection, existing time smoothing can usually only record changes in characteristic values ​​and cannot determine whether these changes originate from the structural reorganization of the same foam group.

[0005] The main technical problem with existing technologies is that flotation foam identification methods based on single-frame image features or smooth single-frame results lack constraints on the topological relationships of foam units and their inter-frame evolution. This makes it difficult to distinguish between real foam state changes and single-frame visual disturbances in image recognition results. The reason is that foam stability, foam load changes, and foam layer disorder are not only manifested as color or area differences in a single frame, but are reflected in continuous events such as continuous movement of foam units, reconstruction of adjacency relationships, local merging, boundary breaks, and node disappearance. If the recognition process does not incorporate foam candidate regions, foam wall intersections, moving adjacencies, and fusion / splitting relationships into a unified temporal image recognition framework, then boundary loss caused by reflective occlusion, contour missegmentation caused by foam wall overlap, and feature mutations caused by local shadows may all be mistaken for changes in operating conditions. These misjudgments do not stem from the complexity of the classification model itself, but from the lack of continuous topological constraints on the image recognition object. This makes it difficult to separate local pixel anomalies, contour anomalies, and real foam evolution events in the recognition chain, thus causing a deviation between the flotation foam monitoring results and the real foam evolution state. The technical problem is that the image recognition process lacks continuous constraints oriented towards bubble evolution events, rather than simply lacking more classification samples or more complex static features. Summary of the Invention

[0006] The purpose of this invention is to provide an image recognition-based flotation foam monitoring system that can solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The image recognition-based flotation foam monitoring system includes an image sequence receiver for receiving continuous flotation foam image frames, and a vision processor connected to the image sequence receiver. The vision processor is used to perform foam boundary enhancement and reflective area suppression on each continuous flotation foam image frame, extracting foam unit candidate regions, foam wall intersection points, foam cluster density, and surface flow direction. Using foam unit candidate regions as nodes and boundary contact, inter-frame area overlap, motion adjacency, and fusion / splitting relationships as edges, a time-updated foam topology evolution map is constructed. Based on the foam topology evolution map, the system corrects the consistency between consecutive frames for abnormal segmentation regions in a single frame and outputs the flotation foam state identification result.

[0008] Preferably, when the visual processor performs foam boundary enhancement and reflective area suppression on each consecutive flotation foam image frame, it first decomposes the image frame into a bubble wall brightness component, a mineralized color band component, and a local texture component. Then, it generates a reliable boundary region based on the continuity of the gray-level gradient around the bubble wall intersection point, performs boundary neighborhood preservation processing on the bright patches, excludes the bright isolated regions that are disconnected from the bubble wall intersection point from the initial contour, and merges the preserved reliable boundary region with the local texture component to form the initial contour of the foam unit candidate region.

[0009] Preferably, when the visual processor constructs the foam topology evolution graph, it establishes a node description for each foam unit candidate region, including contour segments, area sequences, centroid trajectories, adjacent foam numbers, and a set of intersection points; it performs contour overlap matching and motion direction matching on nodes in adjacent image frames, merging matching nodes into the same foam trajectory; when two or more foam trajectories correspond to the same candidate region in a later image frame, it records a merging edge; when a foam trajectory is separated into multiple candidate regions in a later image frame, it records a splitting edge; and it marks nodes that do not meet the matching conditions as nodes to be verified.

[0010] Preferably, when the visual processor performs frame-to-frame consistency correction, it retrieves the corresponding node trajectory in the preceding image frame and the continuing node trajectory in the following image frame for the single-frame abnormal segmentation region, and maps the boundary missing region, local reflective region and bubble wall overlapping region to the bubble topology evolution diagram respectively; when the node adjacency relationship of the abnormal region remains continuous in the preceding and following frames and the area change does not satisfy the fusion split relationship, the boundary of the abnormal region is completed with the contour fragments of the adjacent nodes; when the abnormal region does not form a continuous trajectory, the abnormal marker is retained and does not participate in the state discrimination.

[0011] Preferably, when the visual processor generates the boundary reliable region, it sets up multi-directional sampling lines around the bubble wall intersection point, extracts the brightness change position, texture direction consistent position and color band discontinuity position along each sampling line, and identifies the segments where the three types of positions overlap in space as candidate boundary segments; for candidate boundary segments covered by bright patches, it generates occlusion compensation segments according to the curvature continuation direction of adjacent uncovered segments and the color band distribution of the bubble regions on both sides, and combines the candidate boundary segments and occlusion compensation segments into a closed contour of the bubble unit candidate region.

[0012] Preferably, when the visual processor performs contour overlap matching and motion direction matching on nodes in adjacent image frames, it writes the centroid displacement direction, contour overlap range, changes in adjacent bubble sets, and the number of intersection points retained on the node into the candidate matching table. For bubble trajectories with multiple selectable nodes in the candidate matching table, the node with the most common adjacency relationship with the preceding adjacent bubble set and whose motion direction is consistent with the mainstream direction of the bubble group is selected first. For bubble trajectories lacking corresponding nodes, a temporary continuation node is generated based on the positions of adjacent nodes in the preceding and following frames.

[0013] Preferably, the visual processor is configured with event marking processing in the foam topology evolution graph. The event marking processing includes: marking a foam merging event when two or more foam trajectories merge and the outer contour of the merged node covers the end contour of each input trajectory; marking a foam bursting event when the node contour exhibits a combination of broken segments, disappearing intersection points, and reduced adjacency relationships in consecutive image frames; and marking a foam layer disorder event when the motion directions of multiple node trajectories in the region are dispersed and the adjacency relationship reconstruction frequency increases, and writing the event markers to the corresponding nodes and edges.

[0014] Preferably, the visual processor performs delayed verification on the abnormal markers that did not participate in state discrimination. The delayed verification includes searching for nodes in subsequent consecutive image frames that have spatial adjacency, motion continuity, or boundary morphology continuity with the abnormal markers. When a continuity node is found, the abnormal marker is transferred to the foam topology evolution graph and participates in consistency correction again according to the corresponding node trajectory. When no continuity node is found, the abnormal marker is classified into a fixed-position visual interference set, and the boundary weight of the highlighted area at the same position is reduced in the subsequent boundary enhancement processing.

[0015] Preferably, when the visual processor performs conflict verification on closed contours and temporary continuation nodes, it maps the overlapping closed contours within the same image frame, the temporary continuation nodes derived from the previous and next frames, and the confirmed bubble trajectories to a local topological subgraph. In the local topological subgraph, contour combinations that cannot exist simultaneously are filtered out according to the intersection point sharing relationship, the adjacent edge preservation relationship, and the contour coverage relationship. The abnormal markers corresponding to the filtered combinations are retained in the state pending verification, and the retained contour combinations are written back to the node description of the bubble topological evolution graph.

[0016] Preferably, when the vision processor generates the flotation foam state recognition result, it extracts foam aggregation events, foam bursting events, foam layer disorder events, and fixed-position visual interference sets from the foam topology evolution diagram; it writes the event combinations that occur consecutively in the same spatial area and are consistent with the direction of foam trajectory movement into the state candidate queue, removes the event combinations that only overlap with the fixed-position visual interference set from the state candidate queue, and generates foam stability state, foam load state, and abnormal trend labels according to the time order of event types in the state candidate queue.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention integrates foam boundary enhancement, reflective area suppression, foam unit candidate region extraction, and foam topology evolution map construction into a single image recognition process. Using foam units as nodes and boundary contact, inter-frame area overlap, motion adjacency, and fusion / splitting relationships as edges, it constrains and expresses the foam evolution process in continuously floating foam images. Abnormal segmentation regions appearing in a single frame are not directly used as state discrimination criteria, but are corrected for consistency through the corresponding node trajectories, adjacency relationships, and area changes in consecutive frames. Thus, local boundary defects caused by reflective occlusion can be filled in using adjacent node contour fragments, missegmentation of candidate regions caused by foam wall overlap can be verified through fusion / splitting relationships, and instantaneous anomalies that do not form continuous trajectories can be excluded from the state discrimination process, making the monitoring results more closely reflect the true state of continuous foam unit evolution.

[0018] 2. This invention generates a reliable boundary region by combining bubble wall intersections, brightness abrupt change locations, texture direction consistency locations, and color band discontinuity locations. It also synthesizes closed contours using occlusion compensation fragments, ensuring the continuous description of the bubble contour is preserved even when bright patches cover the bubble wall. Furthermore, it unifies the handling of multi-node matching, short-term node loss, and contour overlap through candidate matching tables, temporary continuation nodes, and local topological subgraph conflict checks. Finally, by distinguishing between bubble merging events, bubble bursting events, bubble layer disorder events, and fixed-position visual interference sets, the state candidate queue retains only event combinations consistent with the bubble trajectory movement direction, reducing interference from fixed highlights, blemishes, and local shadows on the bubble stability state, bubble load state, and abnormal trend labels. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the overall processing flow of the image recognition-based flotation foam monitoring system of the present invention. Figure 2 This is a flowchart illustrating the foam boundary enhancement and reflective area suppression process of the present invention. Figure 3 This is a flowchart of the foam topology evolution diagram construction and frame-to-frame consistency correction process of the present invention; Figure 4 This is a flowchart of the foam event marking, delay review, conflict verification and status output of the present invention. Detailed Implementation

[0020] refer to Figure 1 In one embodiment, the image recognition-based flotation foam monitoring system includes an image sequence receiver for receiving continuous flotation foam image frames, and a vision processor connected to the image sequence receiver. The image sequence receiver receives continuous image frames from the flotation foam monitoring field of view and forms an image sequence according to the acquisition time. The vision processor uses continuous flotation foam image frames as the processing object, and does not use the average color value or average brightness value of a single frame as a direct discrimination criterion. Instead, it extracts the foam boundary, foam wall intersection point, foam unit candidate region, and surface flow direction in each frame, and then organizes the corresponding foam candidate regions between adjacent frames into a foam topology evolution diagram with a time sequence relationship. In the topology evolution graph, nodes correspond to candidate regions of foam units, and edges correspond to boundary contact, inter-frame area overlap, motion adjacency, and fusion splitting relationships. Before outputting the floating foam state recognition results, the visual processor performs frame-to-frame consistency correction on abnormal segmentation regions corresponding to boundary loss, contour missegmentation, reflective occlusion, and foam wall overlap that appear in a single frame. This ensures that image events identified as foam state changes must be supported by node trajectories, adjacency relationships, or fusion splitting relationships in continuous images. In this embodiment, the above processing transforms the image recognition object from isolated pixel features into a topology evolution object of continuous foam units, providing a unified image recognition basis for subsequent judgment of foam stability state, foam load state, and abnormal trend labels.

[0021] In this embodiment, the image sequence receiver writes consecutive float bubble image frames into the frame buffer. The vision processor reads the current image frame and its preceding and following image frames from the frame buffer, establishes a pixel coordinate system and time index for the current image frame, and the current image frame is denoted as the [number missing]. Frame, the preceding image frame is denoted as the 1st frame. Frame, the subsequent image frame is denoted as the 1st frame. Frame, visual processor for the first When performing bubble boundary enhancement, the brightness abrupt changes, texture direction, and mineralization band distribution in the image are jointly analyzed. The boundary enhancement result does not directly replace the original image, but is used as a constraint layer for candidate contour generation. When suppressing reflective areas, the bright pixels are not simply deleted, but it is determined whether the bright area has a spatial continuity relationship with the bubble wall intersection point, the bubble wall texture direction, and the candidate boundary segment. If the bright area covers the bubble wall, its neighborhood boundary continuity relationship is retained. If the bright area is disconnected from the bubble wall intersection point and is fixed near the same image coordinate in consecutive frames, it is recorded as a fixed-position visual interference candidate. The visual processor obtains the bubble unit candidate area, bubble wall intersection point, bubble group density, and surface flow direction accordingly. The bubble group density is characterized by the number of closed candidate contours and their effective coverage area in a unit image area. The surface flow direction is determined by the centroid displacement direction of candidate nodes in adjacent frames and the offset direction of local texture over time. In this embodiment, image reception, boundary enhancement, reflective suppression, and inter-frame association are placed in the same processing link, so that single-frame anomalies will not skip the continuous image verification and directly enter the state recognition result.

[0022] refer to Figure 3 The visual processor for the first Each candidate region of a foam unit in a frame is described using a node description. The node description includes at least a contour fragment, an area sequence, a centroid trajectory, adjacent foam numbers, and a set of intersection points. The contour fragment represents the visible portion of the foam boundary in the current frame; the area sequence represents the area change of the same foam trajectory across multiple frames; the centroid trajectory represents the spatial displacement of the foam unit as it flows through the foam layer; adjacent foam numbers represent the shared bubble wall or mutual contact relationship between foam units; and the set of intersection points represents the topological anchor points formed by the intersection of multiple bubble walls. The visual processor performs contour overlap matching and motion analysis on candidate regions in adjacent image frames. Directional matching groups candidate regions that meet the matching relationship into the same bubble trajectory. When two or more bubble trajectories correspond to a candidate region in subsequent image frames, a clustering edge is recorded in the bubble topology evolution graph. When a bubble trajectory is separated into multiple candidate regions in subsequent image frames, a splitting edge is recorded. When a candidate region appears in the current frame but lacks a corresponding trajectory in the preceding and following frames, it is marked as a node to be verified. In this embodiment, the local morphological changes, adjacent relationship changes, and motion trend changes of bubble units are written into the same graph structure through nodes and edges, so that bubble recognition no longer depends on the isolated condition of whether the contour of a single frame is complete.

[0023] in, Indicates the first The first frame The boundary credibility of each bubble candidate region This indicates the value of the gray-level gradient continuity of the candidate boundary segment. This represents the connectivity integrity value at the intersection of bubble walls near the candidate boundary. This indicates the consistency value between the reflective area and the candidate boundary direction. This represents the value indicating the difference in the distribution of mineralized color bands on both sides of the candidate boundary. , , , The weights of the corresponding components are such that the sum is... ,when , , , , , , , At that time, the calculation yielded The result indicates that the candidate region has a high degree of boundary confidence but low reflectivity. The vision processor retains the candidate region in the subsequent topology matching and requires it to pass the trajectory verification of the previous and next frames. In this embodiment, the above calculation method is used to make boundary preservation, reflectivity suppression and topology verification have a unified numerical input.

[0024] Preferably, when the visual processor constructs the foam topology evolution graph, it establishes a candidate matching table for each candidate node. The candidate matching table includes the centroid displacement direction, contour overlap range, adjacent foam set changes, and the number of retained intersection points between the current node and its predecessor node. The centroid displacement direction is used to exclude erroneous matches that are opposite to the mainstream direction of the foam group and lack adjacent support. The contour overlap range is used to determine whether two candidate regions might belong to the continuous form of the same foam unit. The adjacent foam set changes are used to determine whether the surrounding structures of the foam unit maintain a continuity relationship. The number of retained intersection points is used to determine whether the boundary topological anchor points extend in adjacent frames. Continuing, when there are multiple selectable nodes in the candidate matching table, the visual processor prioritizes the node that maintains the most common adjacency relationship with the previous adjacent bubble set and whose movement direction is consistent with the mainstream direction of the bubble group. When a certain previous trajectory lacks a corresponding node in the current frame, the visual processor does not immediately determine that the bubble has burst. Instead, it generates a temporary continuation node based on the position of the adjacent nodes in the previous and next frames and writes the temporary continuation node into the bubble topology evolution graph for subsequent frame verification. This embodiment enables short-term node occlusion, reflective coverage, and local segmentation loss to be absorbed by continuous image relationships, avoiding the direct interpretation of short-term visual gaps as a sudden change in bubble state.

[0025] refer to Figure 2In a preferred embodiment, the visual processor employs component decomposition to enhance foam boundaries and suppress reflective areas. The image frame is decomposed into a bubble wall brightness component, a mineralized color band component, and a local texture component. The bubble wall brightness component describes the grayscale abrupt changes between the bubble wall and the interior of the foam. The mineralized color band component describes the color band distribution formed by materials carried on the foam surface. The local texture component describes bubble wall wrinkles, foam surface graininess, and edge direction continuity. The visual processor sets up multi-directional sampling lines around the bubble wall intersection points and extracts the brightness abrupt change locations, texture direction consistency locations, and color band discontinuities along each sampling line. When the three types of locations coincide in space or are within the extensible range of the same continuous curve, the corresponding segment is identified as a candidate boundary segment. When a bright patch covers a candidate boundary segment, the visual processor generates an occlusion compensation segment along the curvature continuation direction of the adjacent uncovered segment. It also uses the distribution of mineralized color bands in the foam regions on both sides to determine whether the compensation segment is located between two foam units. Finally, the candidate boundary segment and the occlusion compensation segment are combined to form a closed contour of the foam unit candidate region. This embodiment discloses the specific processing method for contour generation in reflective scenes through the combined constraints of brightness, texture, color bands and intersection points.

[0026] Table 1 shows the correspondence between different image components and processing results during the generation of foam candidate contours. The component names in the table are used to explain how the visual processor distinguishes bubble walls, reflections, and mineralization bands in the same image frame. The processing content is used to explain the constraint method before the data enters the closed contour. The output content is used to explain the form in which the corresponding components are written into the foam unit candidate region.

[0027] Table 1. Correspondence between different image components and processing results during foam candidate contour generation. ; In this embodiment, the candidate boundary segments, color band boundary constraints, texture direction constraints, and occlusion compensation segments in Table 1 are not output as independent state results, but rather enter the closed contour synthesis process together. When synthesizing the closed contour, the visual processor checks whether the candidate segments can form non-self-intersecting foam unit boundaries. If the segment given by the luminance component is inconsistent with the texture direction constraint, the closure is temporarily suspended and candidate contours for the same region in adjacent frames are retrieved. If the mineralized color band component shows that the regions on both sides belong to the same continuous bubble surface, the weight of the candidate segment as a bubble wall boundary is reduced. If the compensation segment of the reflective coverage area intersects with the node trajectory in the previous and next frames, the compensation segment is included in the closed contour and the compensation source is recorded. After the closed contour is formed, the visual processor generates a corresponding set of contour segments and a set of intersection points for each foam unit candidate region. The set is written into the node description of the foam topology evolution graph. This embodiment, through component association and cross-frame verification, ensures that reflective areas are not simply deleted or directly mistaken for foam boundaries, thereby reducing the formation of abnormal segmentation regions from the source of image recognition.

[0028] Furthermore, when the visual processor calculates the curvature continuation of candidate boundary segments, it extracts the endpoint positions and endpoint directions of adjacent uncovered segments, restricts the compensation curve within the highlighted coverage area to the local neighborhood between the two endpoints, and the directional change of the compensation curve needs to remain continuous with the tangential direction of the uncovered segments at both ends. The visual processor simultaneously detects the color band distribution on both sides of the compensation curve. If the difference in color bands on both sides remains within the same direction range as the difference in color bands on both sides of adjacent uncovered segments, then the compensation curve is used as an occlusion compensation segment. If the compensation curve crosses multiple bubble wall intersection points or has an inexplicable overlap with other confirmed closed contours, then it is retained as a segment to be verified, pending subsequent local topological subgraph conflict verification. This embodiment provides implementable constraints for boundary compensation in highlighted coverage scenarios, so that when the bubble wall is covered by reflective light, a verifiable contour can still be obtained in the continuous topological structure.

[0029] in, Indicates the first The first frame The node and the first The first frame The matching score of each node. This indicates the degree of overlap between the projected contours of two nodes. This indicates the degree of continuity in the area changes between two nodes. Indicates the degree of preservation of adjacent foam clusters. This indicates the degree of consistency between the direction of the centroid displacement and the mainstream direction of the foam swarm. , , , The weights of the corresponding components are such that the sum is... ,when , , , , , , , At that time, the calculation yielded If another candidate node has a higher degree of contour overlap but a significantly lower degree of adjacency set preservation, the visual processor will continue to compare the topological constraint components instead of determining the matching object solely based on the overlap area. This embodiment discloses the quantitative determination entry point for node trajectory continuation through this formula.

[0030] Preferably, when the visual processor performs frame-to-frame consistency correction on abnormal segmentation regions in a single frame, for boundary missing regions, local reflective regions, and bubble wall overlapping regions appearing in the current frame, it searches for the corresponding node trajectories in the preceding image frame and the continuing node trajectories in the following image frame in the bubble topology evolution graph. If the abnormal region is located near the contour gap of the same bubble trajectory, and the adjacency relationship of the corresponding nodes in the preceding and following frames remains continuous, the visual processor selects a segment from the contour segment of the adjacent node that is consistent with the curvature continuity of the gap endpoint to complete the boundary of the abnormal region. If the abnormal region covers the contact position of two or more bubble trajectories, and a common candidate region appears in the following frame, the visual processor includes the abnormal region in the clustering event candidate. If the abnormal region lacks motion continuity in both the preceding and following frames and is fixed at a similar position in the image coordinates, the abnormal marker is retained and it does not participate in the current state judgment. This embodiment makes the frame-to-frame consistency correction have a clear graph structure retrieval path and boundary write-back rules, solving the problem of the propagation of existing single-frame segmentation errors along subsequent state judgments.

[0031] refer to Figure 4 In a preferred embodiment, the visual processor is configured with event labeling processing in the foam topology evolution graph. The event labeling processing does not replace the foam evolution process with single-frame classification labels, but generates foam merging events, foam bursting events, and foam layer disorder events based on node trajectories and edge relationships. When two or more foam trajectories merge and the outer contour of the merged node covers the end contour of each input trajectory, the visual processor labels the foam merging event and writes the merging edge between the input trajectory participating in the merging and the merged node. When the node contour shows a combination of broken segments, disappearing intersection points, and reduced adjacency relationships in consecutive image frames, the visual processor labels the foam bursting event and writes the node position where the broken segment is located, the record of the disappearance of the intersection point, and the record of the reduced adjacency relationship into the corresponding node. When the movement directions of multiple node trajectories in the region are dispersed and the frequency of adjacency relationship reconstruction increases, the visual processor labels the foam layer disorder event and writes the event label into the nodes and edges in the region. In this embodiment, the foam state is no longer directly given by static features, but is formed by the combination of traceable topological events to form state candidate criteria.

[0032] Furthermore, the visual processor performs delayed verification on the abnormal markers that did not participate in state discrimination. The delayed verification process searches for nodes in subsequent consecutive image frames that have spatial adjacency, motion continuity, or boundary morphology continuity with the abnormal markers. If a node appears in a subsequent frame that is adjacent to the position of the abnormal marker and whose motion direction is consistent with the mainstream direction of the foam group, the visual processor transfers the abnormal marker to the foam topology evolution graph and re-participates in consistency correction according to the corresponding node trajectory. If no continuing node is found in a subsequent frame and the position of the abnormal marker in the image coordinates remains fixed, the visual processor classifies the abnormal marker into a fixed-position visual interference set and reduces the boundary weight of the bright area at the same position in the subsequent boundary enhancement processing. The fixed-position visual interference set records the coordinate range, continuous frame segment, brightness distribution, and detachment relationship with the foam trajectory of the interference area. The subsequent state discrimination removes the event combination that overlaps with the fixed-position visual interference set from the state candidate queue. This embodiment avoids immediately judging transient occlusion as noise by delayed verification and also avoids continuously including fixed highlights or blemishes in the foam event.

[0033] in, Indicates the first The event intensity of each event fragment. Indicates the start frame of the event segment. Indicates the end frame of the event segment. Indicates the number of frames the event lasts. Indicates the first The topological change values ​​of event-related nodes in the frame. Indicates the first The morphological changes of the event-related contours within the frame. Indicates the first The value of the change in the motion direction of the event-related nodes in the frame, when a certain merging event changes from the first... Frame continues until the Frames, and the three values ​​in the three frames are respectively , , hour, This value is written into the state candidate queue to describe the intensity of the event fragment in consecutive frames. In this embodiment, the topological change, morphological change and motion change are merged into the same event record through this calculation.

[0034] In this embodiment, when the visual processor generates the flotation foam state recognition result, it extracts foam aggregation events, foam bursting events, foam layer disorder events, and fixed-position visual interference sets from the foam topology evolution diagram. A state candidate queue is established according to the chronological order of events occurring within the same spatial region. The state candidate queue only accepts event combinations that are consistent with the foam trajectory movement direction and have continuous node trajectory support. If an event combination only overlaps with the fixed-position visual interference set, or if its position is fixed across multiple frames and detached from the foam group movement, it is removed from the state candidate queue. If foam aggregation and foam bursting events occur consecutively in the same spatial region, and the movement directions of adjacent nodes are dispersed, the visual processor considers it as a candidate for changes in foam stability state. If the distribution of mineralized color bands accompanies the movement of the foam trajectory and changes synchronously with the node area sequence, the visual processor considers it as a candidate for changes in foam load state. If events gradually appear in multiple regions according to the foam flow direction, the visual processor generates an abnormal trend label. This embodiment ensures that the final recognition result is constrained by event type, chronological order, trajectory direction, and interference set, avoiding directly writing single-frame pixel anomalies into the state result.

[0035] Preferably, when the visual processor performs conflict verification on closed contours and temporary continuation nodes, it maps the overlapping closed contours within the same image frame, the temporary continuation nodes derived from the preceding and following frames, and the confirmed bubble trajectory to a local topological subgraph. The nodes of the local topological subgraph include the closed contours to be verified, the temporary continuation nodes, and the confirmed nodes. The edges of the local topological subgraph include the intersection point sharing relationship, the adjacent edge preservation relationship, and the contour coverage relationship. The visual processor checks whether different contour combinations can simultaneously satisfy the intersection point sharing, adjacent edge preservation, and contour coverage constraints in the local topological subgraph. If two closed contours generate mutually exclusive connections at the same bubble wall intersection point, or if a temporary continuation node completely covers the confirmed trajectory but lacks support from adjacent frames, then contour combinations that cannot coexist are filtered out. The anomaly markers corresponding to the filtered combinations are retained in the state pending verification, and the retained contour combinations are written back to the node description of the bubble topology evolution graph. In this embodiment, the generation of closed contours, trajectory continuation, and anomaly delay verification are connected into a unified verification process through the local topological subgraph.

[0036] Table 2 shows the data structure of nodes, edges, and event records in the foam topology evolution graph. The fields in the table are used to explain how the visual processor saves foam units, topological relationships, and state candidate criteria in continuous images. The field content is not limited to a specific storage format and can be implemented using arrays, table entries, graph database records, or memory objects.

[0037] Table 2. Data structure of nodes, edges, and event records in the foam topology evolution graph. ; In this embodiment, the node descriptions in Table 2 are the basic data of the foam topology evolution diagram, the edge descriptions are used to express the spatial and temporal relationships between nodes, the anomaly markers are used to distinguish image regions that cannot be explained temporarily, and the event records are used to enter the state candidate queue. The visual processor updates the data objects shown in Table 2 after processing each frame of image. When a node is matched with a continuation node in a new frame, the area sequence and centroid trajectory are added. When a node generates a new boundary contact or fusion / splitting relationship with an adjacent node, the edge description is rewritten or added. When an anomaly marker finds a continuation node in the delay check, its record is transferred from the anomaly marker to the node description or event record. When an anomaly marker is fixed in multiple frames and does not move with the foam trajectory, its record enters the fixed-position visual interference set. This embodiment makes the foam image recognition process traceable and backtrackable by continuously updating the data structure.

[0038] in, In the local topological subgraph, the first... Conflict loss of a combination of contours, This indicates that the intersection point shares conflicting values. This indicates that adjacent edges maintain conflicting values. Indicates the value of contour coverage conflict. , , The weights of the corresponding conflicting terms are such that the sum is... When the intersection points of a certain contour combination share a conflict Adjacent edges maintain conflict Outline coverage conflict ,and , , At that time, the calculation yielded If the other combination for The visual processor selects combinations with lower conflict loss and writes them back to the foam topology evolution graph, while retaining the anomaly markers corresponding to combinations with higher conflict loss for later review. This embodiment discloses an implementable determination method for local topology verification through conflict loss.

[0039] In a preferred embodiment, when the visual processor calculates the mainstream direction of the foam swarm, it does not use a single motion vector for the entire image. Instead, it divides the image into several local regions that are updated with the movement of the foam surface. Within each local region, nodes with relatively complete continuous trajectories are selected as directional references. The centroid displacement direction, texture displacement direction, and adjacency migration direction of the reference node are synthesized into a local mainstream direction. If the displacement direction of a node to be matched is inconsistent with the local mainstream direction, but its adjacent foam set is well preserved, the visual processor retains it as a suspicious match and waits for subsequent frame verification. If the displacement direction of a node to be matched is inconsistent with the local mainstream direction and its adjacency set lacks a connection relationship, it is downgraded to a node to be verified. The local mainstream direction is also used to determine whether event combinations move with the foam trajectory. Event combinations entering the state candidate queue need to appear continuously in space along the local mainstream direction or its interpretable offset direction. In this embodiment, the local mainstream direction connects the foam flow characteristics with topology matching, event generation, and interference elimination.

[0040] Furthermore, when the vision processor performs image recognition of the foam load state, it uses the migration relationship of the mineralized color band component with the node trajectory as the criterion. If the mineralized color band only appears at a fixed pixel position and does not move with the centroid trajectory of the node, it is recorded as an image background or local interference candidate. If the mineralized color band moves along the same foam trajectory and maintains a continuous distribution between adjacent nodes, it is written into the state supplementary field of the corresponding node. When generating the foam load state, the vision processor reads this supplementary field and, together with the foam merging event, the foam bursting event, and the foam layer disorder event, forms a state candidate. If a color band change is accompanied by a change in the foam node area sequence but lacks topological event support, it is retained as a load candidate to be observed. If a color band change is accompanied by the expansion of the foam outline after merging or the disappearance of the node after bursting, it is written into the foam load state recognition result. This embodiment discloses a limited path from the mineralized color band to the state result, so that the color change no longer participates in the judgment independently of the foam unit trajectory.

[0041] In a preferred embodiment, when the vision processor determines a bubble bursting event, it needs to simultaneously satisfy the combined conditions of increased contour breakage segments, disappearance of bubble wall intersection points, and decreased adjacency relationships in consecutive image frames. The increase in contour breakage segments is given by the breakage boundary in the closed contour that cannot be explained by the occlusion compensation segment. The disappearance of bubble wall intersection points is given by the difference between the intersection point set of the preceding node and the intersection point set of the current node. The decrease in adjacency relationships is given by the decrease in the adjacent bubble numbers of the same node trajectory in adjacent frames. If only one of the above conditions occurs, the vision processor retains the relevant node as an anomaly marker or a node to be verified. If the above conditions form a combined change on the same node trajectory in consecutive frames, the bubble bursting event is marked and written into the event record. This embodiment avoids directly interpreting a single breakage boundary, short-term intersection point occlusion, or local adjacency loss as a bursting event by combining multiple conditions.

[0042] In a preferred embodiment, when the vision processor determines a bubble merging event, it searches whether two or more preceding bubble trajectories in the same local area correspond to the same closed contour in the current frame. If the outer contour of the merged node can cover the end contour of each input trajectory, and there is a traceable union relationship between the merged node and the adjacent bubble set of the input trajectory, then a merging edge is written in the bubble topology evolution graph. If two preceding trajectories are only temporarily connected in the current frame due to reflective occlusion, but are separated again in subsequent frames and the set of intersection points does not change continuously, then the vision processor rewrites the connection as an abnormal marker caused by reflective occlusion. If two preceding trajectories are connected in the current frame and maintain the same centroid trajectory and the same adjacent set in subsequent frames, then it is confirmed as a bubble merging event. This embodiment limits the merging event by outer contour coverage, adjacent set union relationship and subsequent trajectory maintenance, so that short-term erroneous connections of bubble walls will not be directly written into the state candidate queue.

[0043] in, Indicates the first The state candidate score for each spatial region. This represents the average intensity of events within the spatial region that align with the direction of movement of the bubble trajectory. This indicates the continuous values ​​of the event type formed in chronological order. This indicates the degree of overlap between the spatial region and the set of visual disturbances at a fixed location. , , For the corresponding component weights, and ,when , , , , , At that time, the calculation yielded If the intensity of the event is similar in another region but If the degree of interference overlap is low, the score of the region will be lowered. When generating state recognition results, the vision processor will give priority to the region candidates with lower degree of interference overlap and higher event continuity. This embodiment discloses the numerical screening method of the state candidate queue through this formula.

[0044] Furthermore, when the visual processor generates the foam stability state, it reads the temporal order of foam aggregation events, foam bursting events, and foam layer disorder events in the same spatial region within the state candidate queue. If the event combination appears continuously along the local mainstream direction and the visual interference set at the fixed position has a low degree of overlap, the event combination is written into the foam stability state recognition result. If the event combination overlaps with the interference set or lacks node trajectory support, it is removed from the candidate queue. When the visual processor generates anomaly trend labels, it reads the order of changes in state candidate scores over time in multiple spatial regions. If event records in adjacent spatial regions appear recursively along the direction of foam group movement, anomaly trend labels are generated. If event records appear randomly in non-adjacent spatial regions and lack a common trajectory direction, they are retained as local anomaly records. In this embodiment, the state recognition result is derived from continuous events and regional propagation relationships, rather than from instantaneous changes in the average features of the entire image.

[0045] In a preferred embodiment, when the vision processor maintains the fixed-position visual interference set, it records the image coordinate range, continuous frame segment, brightness distribution, and separation relationship with the bubble trajectory for each interference candidate. When a bright area within the same coordinate range remains fixed across multiple frames and does not move with the centroid trajectory of the bubble node, the vision processor writes the bright area into the fixed-position visual interference set. During subsequent boundary enhancement, the boundary weight of the bright area at the same position is reduced. If the area exhibits a movement relationship consistent with the bubble trajectory or forms a continuous intersection point relationship with the node contour in subsequent frames, the vision processor removes its fixed-position interference mark and re-enters the delayed verification process. This embodiment enables fixed reflections, blemishes, and local shadows to be recorded independently in the graph structure, avoiding long-term fixed interference repeatedly affecting the bubble contour and state discrimination.

[0046] In a preferred embodiment, when the visual processor re-incorporates the anomaly marker into the foam topology evolution graph, it reads the spatial location, boundary morphology, and subsequent frame continuation nodes of the anomaly marker. If there is a spatial adjacency between the subsequent continuation node and the anomaly marker, a candidate connection edge is established. If the motion direction of the continuation node is consistent with the local mainstream direction, the credibility of the candidate connection edge is increased. If the contour segment of the continuation node can close with the boundary missing area in the anomaly marker, the anomaly marker is transformed into part of the node description. If the above conditions cannot be met simultaneously, the anomaly marker continues to remain in the pending review state or is classified into the fixed-position visual interference set. This embodiment controls the classification of anomaly markers through candidate connection edges, motion direction, and contour closure, so that delayed review can not only restore the real foam area blocked by reflection, but also eliminate image noise unrelated to foam evolution.

[0047] In this embodiment, the vision processor retains a traceable record for each state recognition result. The traceable record includes the node trajectories involved in the recognition, event records, records of visual interference removed from fixed positions, and records of consistency correction between consecutive frames. When the state recognition result is a change in foam stability, the traceable record includes foam aggregation events, foam bursting events, or foam layer disorder events that caused the state. When the state recognition result is a change in foam load, the traceable record includes the node trajectories and area sequences corresponding to the mineralization color band components. When the state recognition result is an abnormal trend label, the traceable record includes the temporal order of events occurring in adjacent spatial regions and the local mainstream direction. The vision processor can replay the state source based on the traceable record and check the corrected single-frame abnormal segmentation region. This embodiment uses traceable records to keep the image recognition result consistent with the foam topology evolution process, which facilitates subsequent verification of monitoring results and algorithm iteration.

[0048] In a preferred embodiment, the visual processor uses a sliding time window to update the bubble topology evolution graph. The sliding time window stores node descriptions, edge descriptions, anomaly markers, and event records for the current frame, several previous frames, and several subsequent frames. After the current frame is processed, the window moves forward and releases data that has exceeded the window range and has been verified. If an anomaly marker cannot find a continuing node before the window ends, the visual processor classifies it into a fixed-position visual interference set or archives it as an unexplained local anomaly based on its fixed-position characteristics and trajectory separation relationship. If a temporary continuing node is continued by a real node in the later part of the window, the temporary continuing node is rewritten as part of the trajectory of the formal node. If the temporary continuing node cannot be continued and coincides with the interference set, the temporary continuing node is deleted and the original trajectory of the covered node is restored. This embodiment controls the time range of consistency correction between consecutive frames through a sliding time window, avoids unbounded storage of historical images, and retains the ability to process short-term occlusion and delayed verification.

[0049] In a preferred embodiment, the vision processor retains the hierarchical relationship between the original image frame, the reliable boundary region, the closed contour, the bubble topology evolution graph, and the state candidate queue when performing image recognition. The original image frame provides the pixel basis, the reliable boundary region provides candidate boundary constraints, the closed contour provides candidate bubble unit regions, the bubble topology evolution graph provides node trajectories and event records, and the state candidate queue provides input for the bubble state recognition results. If the upper-level processing finds a state candidate that is inconsistent with the continuous topology relationship, the vision processor backtracks down to the node description and the reliable boundary region for correction. If the lower-level processing generates multiple feasible closed contours, the vision processor calls the local topology subgraph conflict check to select the contour combination that can coexist. This embodiment forms a closed loop in the image processing link through hierarchical backtracking, reducing the accumulation of single-layer processing errors in subsequent links.

[0050] In this embodiment, the working process of the image recognition-based flotation foam monitoring system is manifested as a continuous processing flow of receiving continuous image frames, enhancing foam boundaries, suppressing reflective areas, generating candidate contours, establishing node descriptions, matching adjacent frames, updating the topology evolution graph, correcting anomaly segmentation consistency, marking events, delay verification, checking local topological conflicts, and generating state recognition results. Each processing step revolves around the same core object, namely the foam unit and its topological evolution relationship. The process does not rely on adding new acquisition objects unrelated to the flotation foam image, nor does it replace continuous evolution verification with a fixed threshold. The vision processor, through the collaborative processing of image components, node trajectories, adjacency relationships, fusion and splitting relationships, and interference sets, enables real foam aggregation, bursting, and disorder events to be distinguished from single-frame visual disturbances such as reflections, shadows, and stains. The advantage of this embodiment is that it forms a complete technical link from foam image pixels to foam state labels, and makes the foam state recognition results have continuous topological basis.

[0051] In a preferred embodiment, the aforementioned processing steps can be deployed as an image sequence receiving program, a boundary reliable region generation program, a candidate contour generation program, a foam topology evolution graph maintenance program, an anomaly label verification program, and a state candidate queue generation program. Multiple programs run in the same vision processor according to data dependencies. The image sequence receiving program outputs continuously floating foam image frames, the boundary reliable region generation program reads the image frames and outputs candidate boundary segments and occlusion compensation segments, the candidate contour generation program synthesizes foam unit candidate regions, the foam topology evolution graph maintenance program establishes node descriptions and edge descriptions, the anomaly label verification program processes boundary missing, local reflection, and foam wall overlapping areas, and the state candidate queue generation program reads event records and generates foam stability states, foam load states, and anomaly trend labels. The data objects passed between programs are consistent with the node descriptions, edge descriptions, anomaly labels, and event records shown in Table 2. The advantage of this embodiment is that the system can completely realize continuous image recognition and topology evolution constraint processing in a computer program manner, while maintaining clear reference relationships between data objects.

[0052] In this embodiment, if local reflection of the foam layer occurs in consecutive floating foam image frames and the reflected area covers the intersection point of the foam walls, the vision processor first generates candidate boundary segments and occlusion compensation segments in the current frame, and then searches for the node trajectory of the same area in the previous and next frames in the foam topology evolution diagram. If the adjacent foam numbers of the nodes in the previous and next frames remain continuous and the centroid trajectory moves along the local mainstream direction, the vision processor writes the occlusion compensation segment into the closed contour of the current frame. If the reflected area is fixed at the same image coordinates in subsequent frames and does not move with the node trajectory, the reflected area is transferred to the fixed position visual interference set. If the reflected area disappears in subsequent frames and does not form a node continuation, the abnormal marker does not participate in the state candidate queue. The advantage of this embodiment is that it can retrospectively correct the boundary loss caused by reflection, and at the same time avoids long-term interference of fixed high brightness with the foam state recognition result.

[0053] In this embodiment, if multiple foam unit contours are temporarily connected in consecutive floating foam image frames, the vision processor does not directly determine it as a foam merging event. Instead, it reads the area sequence, intersection point set, and adjacent foam set of each node before and after the connection. If the connected regions separate in subsequent frames and each node recovers its original centroid trajectory, the temporary connection is recorded as an anomaly marker corresponding to bubble wall overlap or reflective coverage. If the connected regions remain the same closed contour in subsequent frames, and the outer contour of the merged node covers the end contour of each input trajectory, and the adjacent foam set shows a traceable union of the adjacent relationships of the input trajectories, the connection is recorded as a foam merging event and written into the state candidate queue. The advantage of this embodiment is that it distinguishes between real merging and contour misconnection, so that the fusion and splitting relationship of foam units can accurately reflect the structural reorganization in the continuous image.

[0054] In this embodiment, if a local bubble wall breakage occurs in a continuous flotation bubble image frame, the vision processor reads the set of intersection points of the node in the previous frame, the broken fragment in the current frame, and the changes in the adjacency relationship in the subsequent frame. Only when the broken fragment cannot be explained by occlusion compensation, the intersection point disappears in the continuous frame, and the adjacent bubble number decreases, is the node recorded as a bubble bursting event. If only a single frame of broken fragment occurs but the intersection points and adjacency relationships in the preceding and following frames remain continuous, the boundary of the current frame is completed with the contour fragment of the adjacent node and the abnormality mark is retained for verification. If the broken area is fixed in the image coordinates and does not move with the bubble trajectory, it is classified into the fixed position visual interference set. The advantage of this embodiment is that by combining the determination of contour, intersection point, and adjacency relationship, the bubble bursting event has continuous image evidence and reduces the introduction of single-frame noise.

[0055] In this embodiment, the flotation foam state identification results output by the system include foam stability state, foam load state, and abnormal trend label. The foam stability state is derived from the continuous combination of foam agglomeration events, foam bursting events, and foam layer disorder events within the same spatial region. The foam load state is derived from the movement of mineralized color band components along node trajectories and their association with area sequences and event records. The abnormal trend label is derived from the propagation order of event combinations in adjacent spatial regions along the local mainstream direction. When outputting the results, the vision processor simultaneously saves the state candidate scores, the set of excluded fixed-position visual interferences, and the nodes and edges involved in the result generation. The advantage of this embodiment is that each type of output corresponds to a traceable data object in the foam topology evolution diagram, avoiding the generation of state labels independently of the image recognition link.

Claims

1. A flotation foam monitoring system based on image recognition, characterized in that, It includes an image sequence receiver for receiving continuous flotation foam image frames, and a vision processor connected to the image sequence receiver; The vision processor is used to perform foam boundary enhancement and reflective area suppression on each consecutive flotation foam image frame, and extract the foam unit candidate region, foam wall intersection point, foam group density and surface flow direction. When the vision processor performs foam boundary enhancement and reflective area suppression on each consecutive flotation foam image frame, it first decomposes the image frame into foam wall brightness component, mineralization color band component and local texture component, and then generates a reliable boundary region based on the continuity of gray level gradient around the foam wall intersection point. It performs boundary neighborhood preservation processing on bright patches, excludes bright isolated regions that are disconnected from the foam wall intersection point from the initial contour, and merges the preserved reliable boundary region with the local texture component to form the initial contour of the foam unit candidate region. Using candidate regions of foam units as nodes and boundary contact, inter-frame area overlap, motion adjacency, and fusion splitting relationships as edges, a foam topology evolution graph that is updated over time is constructed. When constructing the foam topology evolution graph, the visual processor establishes a node description for each candidate region of foam units, which includes contour segments, area sequences, centroid trajectories, adjacent foam numbers, and a set of intersection points. Perform contour overlap matching and motion direction matching on nodes in adjacent image frames, and merge the matched nodes into the same bubble trajectory; When two or more bubble trajectories correspond to the same candidate region in the next image frame, the merging edge is recorded. When a bubble trajectory is separated into multiple candidate regions in the next image frame, the splitting edge is recorded. Nodes that do not meet the matching conditions are marked as nodes to be verified. Based on the foam topology evolution diagram, the consistency of the abnormal segmentation region in a single frame is corrected between consecutive frames, and the flotation foam state identification result is output. When the visual processor performs frame-to-frame consistency correction, it retrieves the corresponding node trajectory in the preceding image frame and the continuing node trajectory in the following image frame for the abnormal segmentation region of a single frame, and maps the boundary missing region, local reflective region and bubble wall overlapping region to the bubble topology evolution diagram respectively. When the adjacency relationship of nodes in the abnormal region remains continuous in the previous and next frames and the area change does not satisfy the fusion and split relationship, the boundary of the abnormal region is completed by the contour fragments of the adjacent nodes. When an abnormal region does not form a continuous trajectory, the abnormal marker is retained but not used for state determination.

2. The image recognition-based flotation foam monitoring system according to claim 1, characterized in that, When the visual processor generates the boundary reliable region, it sets up multi-directional sampling lines around the bubble wall intersection point, extracts the brightness change position, texture direction consistent position and color band discontinuity position along each sampling line, and identifies the segments with spatial overlap of the three types of positions as candidate boundary segments. For candidate boundary segments covered by bright patches, occlusion compensation segments are generated based on the curvature continuation direction of adjacent uncovered segments and the color band distribution of foam regions on both sides. The candidate boundary segments and occlusion compensation segments are then combined to form the closed contour of the foam unit candidate region.

3. The image recognition-based flotation foam monitoring system according to claim 2, characterized in that, When the visual processor performs contour overlap matching and motion direction matching on nodes in adjacent image frames, it writes the centroid displacement direction, contour overlap range, adjacent bubble set change and intersection point retention of the node into the candidate matching table. For a bubble trajectory with multiple selectable nodes in the candidate matching table, the node with the most common adjacency relationship with the preceding adjacent bubble set and whose movement direction is consistent with the mainstream direction of the bubble group is selected first. For bubble trajectories lacking corresponding nodes, temporary continuation nodes are generated based on the positions of adjacent nodes in the preceding and following frames.

4. The image recognition-based flotation foam monitoring system according to claim 3, characterized in that, The visual processor configures event tagging processing in the foam topology evolution graph, and the event tagging processing includes: A bubble merging event is marked when two or more bubble trajectories merge and the outer contour of the merged node covers the end contour of each input trajectory. A bubble burst event is marked when a node profile exhibits a combination of broken segments, disappearance of intersections, and reduced adjacency relationships in consecutive image frames. When the movement directions of multiple node trajectories within the region are dispersed and the frequency of adjacency relationship reconstruction increases, the foam layer disorder event is marked, and the event marker is written to the corresponding node and edge.

5. The image recognition-based flotation foam monitoring system according to claim 4, characterized in that, The visual processor performs delayed verification on abnormal markers that are not involved in state discrimination. The delayed verification includes searching for nodes in subsequent consecutive image frames that have spatial adjacency, motion continuity, or boundary morphology continuity with the abnormal markers. When a continuing node is found, the anomaly marker is transferred to the foam topology evolution graph and participates in the consistency correction again according to the corresponding node trajectory; When no continuation node is found, the abnormal marker is classified into the fixed-position visual interference set, and the boundary weight of the highlighted area at the same position is reduced in the subsequent boundary enhancement process.

6. The image recognition-based flotation foam monitoring system according to claim 5, characterized in that, When the visual processor performs conflict verification on closed contours and temporary continuation nodes, it maps the overlapping closed contours within the same image frame, the temporary continuation nodes derived from the previous and next frames, and the confirmed bubble trajectories to a local topological subgraph. In the local topological subgraph, contour combinations that cannot exist simultaneously are filtered out according to the intersection point sharing relationship, the adjacent edge preservation relationship and the contour coverage relationship. The abnormal markers corresponding to the filtered combinations are retained for review, and the retained contour combinations are written back to the node description of the foam topology evolution graph.

7. The image recognition-based flotation foam monitoring system according to claim 6, characterized in that, When the vision processor generates the flotation foam state recognition result, it extracts foam aggregation events, foam bursting events, foam layer disorder events, and fixed-position visual interference sets from the foam topology evolution diagram. Event combinations that occur consecutively within the same spatial region and are consistent with the direction of foam trajectory movement are written into the state candidate queue. Event combinations that only overlap with the set of visual disturbances at fixed locations are removed from the state candidate queue. Based on the time order of event types in the state candidate queue, foam stability status, foam load status, and abnormal trend labels are generated.

Images