High-temperature kiln overflow port stone AI intelligent detection method and storage medium

Abstract

The invention discloses a high-temperature kiln overflow port stone AI intelligent detection method and a storage medium. The method comprises the steps that a glass melt video is collected in real time; the method comprises the following steps: firstly, setting a parameter frame image, then acquiring a current video frame image from a glass melt video, performing difference operation on the current video frame image and the set parameter frame image, delimiting a block area according to a difference operation result, then filtering and judging the block area, and judging whether the block area contains stones or not. The method has the advantages that stone identification and alarm are carried out in the molten glass state, and the method has the technical advantage of accurate and reliable identification.

Description

Technical Field

[0001] The present invention relates to the field of glass production monitoring, and in particular to an AI intelligent detection method and storage medium for stones at the overflow outlet of a high-temperature kiln. Background Art

[0002] During the glass production process, calculi are solid impurities formed in the molten glass due to incomplete melting of raw materials, flaking of refractory materials, or the influx of external contaminants. Common types include siliceous calculi, aluminous calculi, and calculi caused by foreign contaminants. These impurities can significantly reduce the transparency, mechanical strength, and yield of the glass. Currently, glass factories primarily rely on manual inspections to detect and remove calculi: workers inspect the surface of the molten glass under high temperatures and manually pick out any impurities using hooks.

[0003] With the development of machine vision technology, computer vision inspection is gradually becoming a viable alternative to manual inspection. These systems typically consist of high-resolution industrial cameras, optical imaging modules, and image processing algorithms. However, these systems are currently primarily designed for post-molding inspection of glass products, rather than for real-time inspection of glass products in a high-temperature molten state.

[0004] Manual solutions suffer from low efficiency and high missed detection rates, and long hours can easily lead to misjudgments due to visual fatigue. Current computer vision solutions are mostly targeted at glass products after molding, rather than real-time detection in a high-temperature molten state. High-temperature environments also place higher demands on camera stability and image processing algorithms. First, the highly reflective nature of the hot melt leads to uneven image illumination, making it difficult for traditional threshold segmentation algorithms to accurately identify defective areas. Second, the dynamic flow of the melt places extremely high demands on real-time image processing, and existing models are prone to missed detections under complex textures. Finally, deep learning models rely on large amounts of labeled data for training, but in actual production, the morphology of stones is diverse and randomly distributed, resulting in insufficient model generalization. Summary of the Invention

[0005] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide an AI intelligent detection method and storage medium for stones at the overflow of a high-temperature furnace, which can identify and alarm stones in the state of molten glass, and has the technical advantages of accurate and reliable identification.

[0006] In order to achieve the above-mentioned purpose, the technical solution adopted by the present invention is: an AI intelligent detection method for stones at the overflow outlet of a high-temperature furnace, which collects a video of molten glass in real time; then obtains the current video frame image from the molten glass video, performs a difference operation on the current video frame image and the set parameter frame image, demarcates the block area according to the result of the difference operation, and then filters and judges the block area to determine whether the block area contains stones.

[0007] After filtering the block area to determine whether it contains stones, a voice alarm is issued through the alarm module. At the same time, the identified block area containing stones is marked and the marked video image data is stored.

[0008] Whether the block area is a stone is determined based on the movement characteristics of the stone in the glass melt.

[0009] After collecting the video data of the molten glass, the initialization process begins. During the initialization phase, a frame of image is extracted from the video data as a reference frame, and a stone detection area is created based on the reference frame. Creating the stone detection area includes defining a detection range in the video frame and identifying and detecting stones within the defined detection range.

[0010] The reference frame is set during initialization and updated according to preset conditions during the detection process.

[0011] The preconditions for reference frame update include:

[0012] The reference frame is updated regularly according to a set time period or the pixel information of the detection area of the glass melt is monitored in real time. The reference frame is updated when the pixel change reaches the set condition.

[0013] The delineation of block areas includes:

[0014] The difference between the pixels in the detection area of the current video frame and the pixels in the detection area of the reference frame is calculated, and the pixels whose differences are greater than the threshold and the same are merged to form block areas. The block areas are used as candidate areas for the existence of stones.

[0015] Filtering and judging the block area includes:

[0016] The size of the block area is judged, and the block area whose size does not meet the preset conditions is filtered out; multi-frame matching tracking is performed on the filtered block area to identify the motion trajectory status of the block area in different video frames, and whether the motion trajectory meets the stone motion trajectory is judged according to the motion trajectory status. If it meets the requirements, it is judged as a stone, otherwise the block area is eliminated; and the new frame of video is re-detected.

[0017] The falling speed of the block area is calculated based on the motion trajectory of the block area identified in sequence, and when the block area is in uniform motion, it is determined to be a stone.

[0018] A computer storage medium stores a computer program, which, when executed by a processor, implements the AI intelligent detection method for high-temperature kiln overflow stones.

[0019] The advantages of the present invention are: accurate and reliable identification and alarm for stones in the state of molten glass. The advantages of the above scheme include:

[0020] 1. Real-time and targeted detection: Unlike existing computer vision solutions, which primarily focus on post-molding glass products, this invention focuses on real-time stone detection in molten glass at high temperatures. This enables timely detection of stones at critical stages of glass production, preventing them from further impacting glass quality during subsequent processing, effectively improving quality control efficiency in glass production.

[0021] 2. Adapt to complex high-temperature environments: In order to solve the problem of uneven image illumination caused by strong reflection of high-temperature molten liquid, the present invention divides the detection area by creating a pixel-level mask, effectively shielding interference factors, and combining reasonable image processing algorithms to improve the ability to identify stones under uneven illumination. For scenes with dynamically flowing molten liquid, the detection algorithm of the present invention can process images in real time, reduce missed detections caused by dynamic changes in the molten liquid, and has higher detection accuracy under complex textures than existing models. In addition, the present invention does not need to rely on a large amount of labeled data to train deep learning models. Through multi-dimensional algorithm design (such as recognition, matching, filtering algorithms, etc.), it can adapt to the situation where stones are diverse in shape and randomly distributed in actual production, effectively solving the problem of insufficient model generalization ability.

[0022] 3. Reduced labor costs and false positive rates: This approach changes the low efficiency and high missed detection rate of traditional manual stone inspections. Manual inspections are not only inefficient but also prone to misjudgments due to visual fatigue after long hours. This invention utilizes an automated detection system, significantly improving detection efficiency. Furthermore, through multi-dimensional filtering algorithms and intelligent decision-making algorithms, it effectively reduces the false positive rate and minimizes production losses caused by these misjudgments.

[0023] 4. Flexible and Convenient Operation: The system features flexible configuration capabilities, allowing users to configure parameters and adjust inspection areas based on diverse production scenarios and requirements through a simple workflow. For example, dynamic geometric correction of inspection frame vertices can be achieved through drag-and-drop operations in the GUI, quickly and easily adapting to diverse inspection needs. Furthermore, the system's simple and intuitive interface design makes it easy for operators to master and use, reducing training costs and operational complexity.

[0024] 5. Intelligent Alarm and Data Storage: The alarm mechanism utilizes a graded response and multimodal alarm method, enabling appropriate responses based on the number of stones and ongoing detection, prompting operators to take appropriate measures. The system also automatically saves videos of stone appearances, facilitating subsequent tracing and analysis of the production process, helping to further optimize production processes and improve product quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The following is a brief description of the contents and symbols in the drawings of the present invention:

[0026] Figure 1 Schematic diagram of the detection method of the present invention.

[0027] Figure 2 Schematic diagram of the linked list data structure and matching algorithm of the present invention;

[0028] Figure 3 This is a schematic diagram of the comparison of rectangular frame matching in the previous and next two frames of images;

[0029] Figure 4 This is a schematic diagram of the division of the glass melt detection area according to the present invention. DETAILED DESCRIPTION

[0030] The specific implementation of the present invention will be further explained in detail below by describing the best embodiment with reference to the accompanying drawings.

[0031] This solution aims to accurately and rapidly identify stones in molten glass and generate an alarm. The algorithm primarily processes stone detection and alarm processing from a series of video frames. The core idea is to perform a difference calculation between the current video frame and a reference frame to identify clustered areas. The algorithm then determines whether the stone is a stone based on its uniform descent. The detailed algorithm flow is as follows.

[0032] like Figure 1 As shown, an AI intelligent detection method for stones at the overflow of a high-temperature furnace includes real-time acquisition of a video of molten glass; then obtaining a current video frame image from the video of the molten glass, performing a difference operation between the current video frame image and a set parameter frame image, demarcating a block area according to the result of the difference operation, and then filtering and judging the block area to determine whether the block area contains stones.

[0033] After filtering the block area to determine whether it contains stones, a voice alarm is issued through the alarm module. At the same time, the identified block area containing stones is marked and the marked video image data is stored.

[0034] After collecting the video data of the molten glass, the initialization process begins. During the initialization phase, a frame of image is extracted from the video data as a reference frame, and a stone detection area is created based on the reference frame. Creating the stone detection area includes defining a detection range in the video frame and identifying and detecting stones within the defined detection range.

[0035] The reference frame is set during initialization and updated during the detection process according to preset conditions. The preset conditions for reference frame update include:

[0036] The reference frame is updated regularly according to a set time period or the pixel information of the detection area of the glass melt is monitored in real time. The reference frame is updated when the pixel change reaches the set condition.

[0037] The delineation of block areas includes:

[0038] The difference between the pixels in the detection area of the current video frame and the pixels in the detection area of the reference frame is calculated. Pixels with the same difference greater than a threshold are merged to form a block region. The block region is used as a candidate for the presence of stones. The movement characteristics of the stones in the molten glass are used to determine whether the block region contains stones.

[0039] Filtering and judging the block area includes:

[0040] The size of the block regions is determined, and those that do not meet the preset size requirements are filtered out. Multi-frame matching tracking is performed on the filtered block regions to identify their motion trajectories across different video frames. Based on the motion trajectory, whether the motion trajectory meets the stone motion trajectory is determined. If so, the block region is identified as a stone; otherwise, the block region is removed. The next video frame is retested. The falling velocity of the block region is calculated based on the motion trajectories of the block regions identified over time. When the block region is in uniform motion, it is identified as a stone.

[0041] This embodiment also provides a computer storage medium, which stores a computer program. When the computer program is executed by a processor, the AI intelligent detection method for stones at the overflow outlet of a high-temperature furnace in this embodiment is implemented, thereby realizing the detection, identification and alarm of stones in the molten glass.

[0042] In this embodiment, at the beginning of the detection, a frame of video frame image is first obtained from the collected real-time video data as a reference frame, and then in chronological order, frames are extracted from the video at intervals or periods to obtain real-time video frames, and the real-time video frame is subtracted from the reference frame to obtain the difference of each pixel. The points with pixel differences greater than the set threshold are regarded as suspected stone points, and the area composed of pixel points with the same difference and greater than the set threshold is called a block area. The block area may or may not be a stone. The next step is to determine whether the block area is a stone.

[0043] Since the block area needs to meet conditions such as size and motion trajectory if it is a stone, the difference between two or more frames of real-time video frame images and the reference image can be used to collect block areas in multiple groups of real-time video frames arranged in chronological order. According to the block areas in the video frames in chronological order, the position change of the block area in the previous and next video frame images can be calculated, and whether it is a stone can be judged based on the position change. When it falls at a uniform speed along the falling direction through the position of the block area, it means that this block area is a stone, otherwise it is judged not to be a stone. In order to reduce or avoid false alarms, the block area is filtered, and the filtering criteria include the size of the block area, the movement pattern, etc., so as to achieve more accurate stone identification.

[0044] In this solution, the detection method will continue to read video frames from the video stream after it is started. At the beginning of the algorithm startup, a frame in the video will be selected as a reference frame, and some parameters such as the length and width of the image will be initialized using this frame. One of the key steps in initialization is to create a detection area. The detection area only covers the surface part of the molten glass in the video, and there is no need to process all pixels. In addition, in order to eliminate interference factors within the selected area. It is necessary to divide the molten glass detection area to avoid detection errors and false alarms caused by exceeding the detection area. Therefore, the purpose of creating a detection area is to generate a pixel-level mask to indicate the specific processing method for each pixel.

[0045] Over time, some factors in the video, such as lighting, will gradually change. Failure to update the reference frame in a timely manner will affect the accuracy of the detection algorithm. Therefore, the function of the reference frame update module is to regularly update the reference frame. It is important to note that in this example, the reference frame update module is located in the subsequent steps of stone detection, but this location is not fixed and can be set before or after stone detection. The conditions for reference frame update are set, and once the reference frame update conditions are met, the reference frame is immediately updated.

[0046] In the recognition module, a difference image is generated by performing a difference calculation on the corresponding pixels between the current frame and the reference frame. If a stone is present in the current frame, the difference in the difference image will be relatively large at the stone's location. Pixels with large differences are then merged into blocks and represented by rectangular boxes. This process is the recognition process. However, due to factors such as heat waves causing jitter in the video content and ripples on the surface of the molten glass in high-temperature environments, false positives are very likely to occur.

[0047] Figure 1The search and matching process in the link is to match the rectangle of each frame with each node in the linked list. If the match is successful, the node will update the recorded data (including timestamp, distance, rectangle, etc.); if no match is found after traversing all nodes, a new node will be created based on the rectangle, as shown in the diagram below. Figure 2 shown.

[0048] To analyze the velocity characteristics of stones, we need information about their positions across different frames. If multiple rectangular boxes are detected in two adjacent frames, we need to know which boxes point to the same stone so that the positions of the same stone across different frames can be recorded together. The process of determining which box in a given frame belongs to which stone is called search and match. In this solution, stones are recorded as linked list nodes, with each node representing a stone, and each node contains the location information for each stone.

[0049] like Figure 3 As shown, the rectangular boxes represent possible stone locations. Each rectangular box is labeled with a serial number, with different serial numbers representing different stones. Because stones rarely move between adjacent frames, the positions of the detected rectangular boxes are roughly the same. Therefore, stones 11 and 12 should be the same stone in the first and second frames, and stones 21 and 22 should be the same stone in the first and second frames. Stones 31 and 32 do not match, likely indicating a false detection.

[0050] How search and matching works: Search and matching specifically refers to querying nodes one by one along the linked list for the rectangle of the current frame. If a match is found, the data is added to the end of the node and the latest position is updated (for the next frame to match). If a match is not found, it means that the rectangle may be a new stone, so a new node is added to the linked list to represent the stone. Stones are recorded and stored in the form of a linked list, and then the search and matching algorithm is used to match the positions of the rectangles of two adjacent frames. A successful match indicates that the stone position of the rectangle in the latter frame is the same as that of the rectangle in the previous frame. Then, by identifying the position change of the same rectangle in the previous and subsequent frames, the speed change of the rectangle can be calculated and studied.

[0051] To further enhance the false alarm prevention function, three filtering algorithms have been added to the algorithm. In practice, large stones are generally darker in color, while small stones vary in color, but generally, lighter stones are not very large. Therefore, filter 1 limits the size of the block based on the difference; the block is graded based on the pixel difference:

[0052] Strength 1: m_thr1 ≤ m_PicYdiffblock <m_thr2

[0053] Strength 2: m_thr2 ≤ m_PicYdiffblock <m_thr3

[0054] Intensity 3: m_PicYdiffblock ≥ m_thr3

[0055] In this example, m_thr1 = 15, m_thr2 = 30, m_thr3 = 75

[0056] Limiting by size means limiting by strength:

[0057] ●Intensity 1: width≤32, height≤32

[0058] Strength 2: width ≤ 100

[0059] Intensity 3: None

[0060] When a certain amount of matching data accumulates, it enters filter 2 to determine whether it meets the characteristics of stone drop. Filter 3 is used to eliminate nodes with low matching rates. The matching rate is the number of times a corresponding rectangle matches the node since its establishment, divided by the total number of detections the node has undergone (detecting one frame is considered a detection).

[0061] Based on search matching, suppose that starting from a certain frame, a rectangle becomes a new node because it does not match the previous node. Then the algorithm processes x frames, and i frames have corresponding rectangles matching the node, so the matching rate is i / x.

[0062] Figure 1 The decision module in will determine whether it is a stone based on factors such as the number of times the node succeeds in filtering 2, and decide whether to trigger the alarm mechanism.

[0063] The alarm mechanism means that when the decision-making algorithm determines that a stone is present, the system will make a voice announcement and mark the stone. Finally, the system will save the video of the stone locally.

[0064] 2. Detection area creation algorithm

[0065] When creating the detection area, the edge of the kiln will shake, which can easily cause false alarms. At the same time, there are 1-2 "lines" on the surface of many molten glass liquids, and the shaking of these "lines" will also cause false alarms. In view of the fact that stones usually fall from top to bottom, in order to effectively avoid false alarms, the original detection area is further divided and some constraints are added: First, the area from 0 to AreaBound1 in the detection area (here the upper boundary is recorded as 0, the lower boundary is recorded as 1, and then divided by percentage, AreaBound1 and AreaBound2 range from 0 to 1, indicating the percentage of division) will not trigger an alarm. Second, the rectangular box that appears in the area from AreaBound2 to 1 in the detection area will not become a newly added node. For example Figure 4 As shown in FIG. 1 , a schematic diagram of the division of the glass melt detection area is shown.

[0066] The entire molten glass detection area can be simply represented as a rectangular area (the specific shape can be changed according to actual production requirements, and the actual shape shall prevail). The upper boundary of the detection area is 0, the lower boundary is 1, and AreaBound1 and AreaBound2 are numbers from 0 to 1, which are used to subdivide the detection area. The molten glass detection area can be divided into three parts by two straight lines (AreaBound1 and 2):

[0067] Upper Bound 0 to AreaBound 1: When selecting the detection area, the upper boundary of the kiln is often included. However, due to video jitter, this boundary may cause false alarms. Therefore, even if the conditions are met, stones in this area will not be alarmed until the stone reaches AreaBound 1-AreaBound 2. This is to observe for a while to prevent false alarms.

[0068] AreaBound1-AreaBound2: This area is the normal alarm area.

[0069] AreaBound2-1: Because stones typically move from top to bottom, they won't first appear in this area. The ideal processing interval is AreaBound1-AreaBound2. Therefore, rectangles appearing in this area that don't match existing linked list nodes are not considered new targets. This strategy effectively avoids false positives in the lower area due to misjudgment, improving the accuracy and reliability of stone detection.

[0070] The images in the video frame are distinguished and detected in the following five situations:

[0071] 1) Undetectable area;

[0072] 2) The area from 0 to AreaBound1 in general;

[0073] 3) The area from AreaBound1 to AreaBound2 in general;

[0074] 4) The area from AreaBound2 to 1 in general;

[0075] 5) Detection areas for special situations.

[0076] Use the identifiers Idx1, Idx2, Idx3, Idx4, and Idx5 to distinguish these situations in turn. In this example, binary bits are used for marking, where Idx1 = 126 (binary representation is 01111110), Idx2 = 123 (binary representation is 0111 1011), Idx3 = 57 (binary representation is 00111001), Idx4 = 0, and Idx5 = 49 (binary representation is 00110001). The advantage of doing this is that it is convenient to merge conditions. For example, to distinguish between detection areas and non-detection areas, you only need to determine whether the 6th bit (from left to right) is 0. Different alarms and reminders can be achieved by demarcating different detection areas.

[0077] The steps to create the mask m_EdgePicY are as follows:

[0078] 1) Initialization operation: first set the mask m_EdgePicY to the default value 0 (i.e. Idx4);

[0079] 2) To prevent the detection frame from being affected by edge jitter, the detection area is appropriately retracted. In this example, the area is retracted by 200 pixels on each side.

[0080] 3) Set the pixel values that are not in the detection area to 126 (i.e., Idx1);

[0081] 4) Detect the position of the "line" and then create a mask band above and below these points, that is, write its pixel value to 126 (i.e. Idx1).

[0082] 5) For the red-framed area (as the glass melt), write 123 (i.e., Idx2) in its upper half (i.e., the area from 0 to AreaBound1 in general), and write 57 (i.e., Idx3) in its lower half (i.e., the area from AreaBound1 to AreaBound2 in general);

[0083] 6) For the blue frame area (border, special scene detection area, etc.), within the detection area, if the pixel color value is not lower than the threshold (the threshold in this example is 100), write 49 (i.e., Idx5); otherwise, write 126 (i.e., Idx1).

[0084] 3. Reference frame update mechanism

[0085] The trigger condition is to meet one of the following conditions:

[0086] 1) Each segment is triggered once;

[0087] 2) There are large-area pixel changes in the glass melt detection area;

[0088] Count the number of black pixels in the detection area of the reference image (i.e., the number of pixel values less than the threshold, which is 110 in this example) Cnt1, and the number of black pixels in the detection area of the current frame Cnt2. Then determine whether the absolute value of Cnt1-Cnt2 is greater than the threshold, which is 16*208*(video height / 1080). If so, it is considered that a large-area pixel change has occurred; otherwise, it is determined that no large-area pixel change has occurred.

[0089] Update method:

[0090] Condition 1) corresponds to delayed update;

[0091] Condition 2) corresponds to immediate update.

[0092] In this example, the reference frame is updated every 5 minutes. To prevent sudden updates from affecting detection, a delayed update is used. In this example, the delay is 60 frames, that is, one frame is acquired into the buffer every 5 minutes and replaced with the reference frame after a delay of 60 frames.

[0093] 4. Stone Detection Algorithm

[0094] (1) Recognition algorithm

[0095] Core logic: Extract potential stone areas through frame difference analysis and block merging technology.

[0096] step:

[0097] 1) Pixel difference calculation

[0098] Traverse the detection area of the current frame and the reference frame and calculate the Y channel (brightness) pixel difference: m_PicYdiff[y][x] = max(0,refY[y][x]-currentY[y][x])

[0099] 2) Block-level difference analysis

[0100] Divide the image into blocksize × blocksize blocks (2 × 2 in this example) and calculate the block difference sum: m_PicYdiffblock. Because pixel-by-pixel processing is too time-consuming, the image is downsampled when calculating the difference. The block difference sum is the sum of the differences of adjacent blocksize × blocksize pixels. In other words, the sum of the differences of four or more pixels is used as a block to obtain a new difference image. This way, the resulting difference image is smaller and easier to process.

[0101] 3) Block merging: Merge adjacent blocks with the same intensity to generate candidate regions.

[0102] (2) Matching algorithm

[0103] Core logic: Dynamic tracking technology based on a bidirectional linked list to achieve multi-frame matching, as shown in Figure 5.

[0104] step:

[0105] 1) Linked list traversal: traverse the historical detection records from the end of the linked list (m_end_rect), such as Figure 2 shown

[0106] 2) Matching conditions

[0107] Position matching: The bottom and left and right deviations of the current block and the reference block are less than the threshold, which means that the following conditions must be met at the same time:

[0108] √ref_bottom-5≤cur_bottom

[0109] √cur_bottom-ref_bottom<16

[0110] √cur_left≥ref_left-16

[0111] √cur_right≤ref_right+16

[0112] √ Two rectangular boxes intersect

[0113] in,

[0114] ref_bottom, the bottom coordinate of the rectangular box in the node;

[0115] ref_left, the coordinates of the left border of the rectangular box in the node;

[0116] ref_right, the coordinates of the right border of the rectangular box in the node;

[0117] cur_bottom, the bottom coordinate of the current rectangle;

[0118] cur_left, the coordinates of the left border of the current rectangle;

[0119] cur_right, the coordinates of the right border of the current rectangle.

[0120] Strength matching: The difference between the current block strength and the historical block strength is ≤ 1 level, that is, strength 3 cannot match strength 1;

[0121] Size matching: aspect ratio difference ≤ 3 times.

[0122] 3) Matching results

[0123] Matching success: update the reference block coordinates, count (count++), record the bottom position dstVector, and time ptsVector;

[0124] Matching failed: Create a new node for the current block, such as Figure 2 shown.

[0125] Core logic of filtering algorithm: multi-dimensional conditional filtering to suppress false positives.

[0126] 1) Filter 1: Constrain the size of the block according to the intensity:

[0127] Strength 1: width≤32, height≤32;

[0128] Strength 2: width ≤ 100;

[0129] Intensity 3: None;

[0130] 2) Filter 2: Eliminate nodes that do not conform to the movement pattern of stones

[0131] When the recorded bottom position dstVector and time ptsVector reach a certain number num_data, motion analysis is started.

[0132] Preliminary data check: Calculate the movement speed of the node and eliminate nodes with a speed greater than the upper threshold or less than the lower threshold.

[0133] Motion analysis: Determines the speed of movement at different times. If it is uniform, it is considered a stone. If it meets the requirements, it is considered a stone and the alarm mechanism is activated.

[0134] Alarm mechanism: hierarchical response, multi-modal alarm.

[0135] Single stone alarm: Green frame mark (stone location) + directional voice (such as "There is a stone on the north board of the first line")

[0136] Multiple stone alarm: Voice prompt (such as "There is abnormal stone on the north plate of the first line")

[0137] Continuous alarm: When stones are detected continuously for more than 80 frames, the voice is repeated every 80 frames

[0138] Stop condition: Automatically stops when there is no detection for 150 consecutive frames, and the alarm light turns gray.

[0139] The stone detection method in this embodiment has the following technical advantages:

[0140] 1. System Architecture and Functional Configuration: The unique system interface design includes a link input area, parameter configuration area, and real-time monitoring area, each with clear and collaborative functions. The parameter configuration area integrates multi-dimensional detection strategy configuration modules, such as the line identification module, the direction identification module, the dynamic detection domain configuration component, and the debug log management unit. This system architecture and functional configuration is one of the key protection points of this application.

[0141] 2. Detection area creation algorithm: A method of creating pixel-level masks by distinguishing five different areas (non-detection areas, areas of different ranges in general cases, and detection areas in special cases) through binary bit markings, as well as specific steps for establishing masks (initialization, shrinkage processing, setting pixel values in non-detection areas, edge detection and shielding, red and blue frame area processing, etc.). This algorithm effectively eliminates interference and accurately delineates the detection area, which is the core innovation and protection point of this application.

[0142] 3. Reference frame update mechanism: A dynamic reference frame update mechanism with dual trigger conditions, including time interval trigger (such as 60-frame update every 5 minutes) and blue frame pixel change trigger (immediate update).

[0143] 4. Stone detection algorithm:

[0144] Identification algorithm: Based on frame difference analysis and block merging technology, the method extracts potential stone areas through steps such as pixel difference calculation, block difference analysis, intensity grading, edge shielding and block merging.

[0145] Matching algorithm: Based on the dynamic tracking technology of the bidirectional linked list, the algorithm traverses the linked list, sets strict matching conditions (position matching, strength matching, size matching) and performs corresponding processing based on the matching results.

[0146] Filtering algorithm: Multi-dimensional conditional filtering algorithm, including removing larger blocks, removing non-compliant nodes according to motion rules (by setting analysis parameter thresholds at different intensities, performing preliminary data inspection, motion analysis and result processing, etc.), and removing nodes with low matching rates, etc., effectively suppressing false alarms.

[0147] Decision-making algorithm: This method uses a combination of factors such as the number of matches, the number of successful motion analysis results, and the distance of descent to determine whether a stone is present. The combination and specific implementation of these algorithms are the key technologies and protection focus of this application.

[0148] 5. Alarm mechanism: A graded response and multi-modal alarm mechanism, including single stone alarm (green box mark + directional voice), multiple stone alarm (voice prompt), continuous alarm (repeated voice when stones are detected continuously for more than a certain number of frames) and stop condition (automatically stops when no detection is made for a certain number of consecutive frames, and the alarm light turns gray). This comprehensive and intelligent alarm mechanism is one of the important features and protection points of this application.

[0149] Obviously, the specific implementation of the present invention is not limited to the above-mentioned methods. As long as various non-substantial improvements are made using the method concept and technical solution of the present invention, they are all within the scope of protection of the present invention.

Claims

1. An AI intelligent detection method for high-temperature kiln overflow stones, characterized by: The glass melt video is collected in real time; the current video frame image is then obtained from the glass melt video, and a difference operation is performed between the current video frame image and the set parameter frame image. The block area is delineated according to the result of the difference operation, and then the block area is filtered and judged to determine whether the block area contains stones.

2. The AI intelligent detection method for high-temperature kiln overflow stones according to claim 1, characterized in that: After filtering the block area to determine whether it contains stones, a voice alarm is issued through the alarm module. At the same time, the identified block area containing stones is marked and the marked video image data is stored.

3. The AI intelligent detection method for high-temperature kiln overflow stones according to claim 1, characterized in that: Whether the block area is a stone is determined based on the movement characteristics of the stone in the glass melt.

4. The AI intelligent detection method for high-temperature kiln overflow stones according to claim 1, characterized in that: After collecting the video data of the molten glass, the initialization process begins. During the initialization phase, a frame of image is extracted from the video data as a reference frame, and a stone detection area is created based on the reference frame. Creating the stone detection area includes defining a detection range in the video frame and identifying and detecting stones within the defined detection range.

5. The AI intelligent detection method for high-temperature kiln overflow stones according to claim 1 or 4, characterized in that: The reference frame is set during initialization and updated according to preset conditions during the detection process.

6. The AI intelligent detection method for high-temperature kiln overflow stones according to claim 5, characterized in that: The preconditions for reference frame update include: The reference frame is updated regularly according to a set time period or the pixel information of the detection area of the glass melt is monitored in real time. The reference frame is updated when the pixel change reaches the set condition.

7. The AI intelligent detection method for high-temperature kiln overflow stones according to claim 1, characterized in that: The delineation of block areas includes: The difference between the pixels in the detection area of the current video frame and the pixels in the detection area of the reference frame is calculated, and the pixels whose differences are greater than the threshold and the same are merged to form block areas. The block areas are used as candidate areas for the existence of stones.

8. The AI intelligent detection method for high-temperature kiln overflow stones according to any one of claims 1 to 7, characterized in that: Filtering and judging the block area includes: The size of the block area is judged, and the block area whose size does not meet the preset conditions is filtered out; multi-frame matching tracking is performed on the filtered block area to identify the motion trajectory status of the block area in different video frames, and whether the motion trajectory meets the stone motion trajectory is judged according to the motion trajectory status. If it meets the requirements, it is judged as a stone, otherwise the block area is eliminated; and the new frame of video is re-detected.

9. The AI intelligent detection method for high-temperature kiln overflow stones according to claim 8, characterized in that: The falling speed of the block area is calculated based on the motion trajectory of the block area identified in sequence, and when the block area is in uniform motion, it is determined to be a stone.

10. A computer storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by the processor, the AI intelligent detection method for high-temperature kiln overflow stones as described in any one of claims 1 to 9 is implemented.

Images