Pool level detection method of visible-infrared dual-mode fusion improved YOLOv8

Abstract

The application discloses a molten pool liquid level detection method of visible-infrared dual-mode fusion improved YOLOv8, relates to the technical field of industrial visual detection and metallurgical process parameter measurement, and the visible light image and the infrared thermal image after the detection ruler is obtained, and after the registration processing of the two kinds of modal images is carried out, the dual-mode liquid level recognition model is input, the complementary information of the visible light texture feature and the infrared thermal radiation feature is utilized to recognize the slag layer boundary position and the molten pool layer boundary position on the detection ruler surface, the boundary length is determined according to the slag layer boundary position and the molten pool layer boundary position, the liquid level depth is cooperatively solved by combining the corresponding falling depth of the detection ruler end, and thus the molten pool reference liquid level depth is obtained.

Description

Technical Field

[0001] This invention relates to the field of industrial visual inspection and metallurgical process parameter measurement technology, and in particular to a molten pool level detection method based on the improved YOLOv8 with visible and infrared dual-mode fusion. Background Technology

[0002] Molten pool level is a crucial process parameter in metallurgical smelting, directly impacting smelting reaction efficiency, product quality, and production safety. To understand the internal state of the molten pool, production sites typically need to periodically measure the molten pool level depth for material feeding control, slag layer management, and process parameter adjustments. Existing methods for molten pool level detection mainly include mechanical gauge measurement and visual inspection. Mechanical gauge measurement involves inserting a gauge into the molten pool and obtaining level information based on the insertion depth. This method is widely used in high-temperature metallurgical applications due to its direct measurement and simple structure. However, after the gauge is removed from the molten pool, its surface is usually covered with slag and molten pool material. Because the slag layer and molten pool layer continuously cover the gauge surface, it is difficult to accurately determine the true boundary between them. This results in slag thickness errors in the molten pool level depth reflected by the gauge, affecting the accuracy of the level measurement results.

[0003] To avoid measurement errors caused by mechanical contact, some technical solutions employ machine vision to detect the molten pool level. These methods typically utilize edge information from visible light images to identify the liquid level boundary. However, metallurgical sites generally present complex conditions such as high-temperature radiation, smoke and dust obstruction, dust interference, and strong arc light. Visible light images are prone to problems like blurred boundaries, decreased contrast, and target occlusion, resulting in poor stability in liquid level boundary identification. While introducing infrared thermal imaging can obtain temperature distribution information of the target area and improve the perception of high-temperature regions, relying solely on infrared images still suffers from insufficient texture detail and limited boundary positioning accuracy, making it difficult to simultaneously achieve both identification accuracy and measurement precision. Therefore, accurately identifying the boundary between the slag layer and the molten pool layer on the surface of the measuring scale under high temperature, smoke, and strong light interference environments, and combining this with depth information obtained from mechanical measuring to eliminate measurement errors caused by slag adhesion, thereby achieving high-precision measurement of the molten pool level depth, has become a pressing technical problem to be solved in this field. Summary of the Invention

[0004] To address the technical problems existing in the background art, this invention proposes an improved YOLOv8 molten pool level detection method based on visible infrared dual-mode fusion.

[0005] The present invention proposes an improved YOLOv8 molten pool level detection method based on visible and infrared dual-mode fusion, comprising the following steps:

[0006] S1. Obtain image data of the inspection ruler after it is inserted into the molten pool and then removed;

[0007] The detection ruler image data includes a visible light image and an infrared thermal image corresponding to the same acquisition time;

[0008] The visible light image is used to characterize the surface texture information of the detection ruler;

[0009] The infrared thermal image is used to characterize the surface temperature distribution information of the detection ruler;

[0010] S2. Perform image enhancement processing on the visible light image and the infrared thermal image respectively to obtain a visible light enhanced image and an infrared enhanced image;

[0011] Based on the pre-established binocular heterogeneous vision calibration relationship, the infrared enhanced image is spatially mapped so that the infrared enhanced image and the visible light enhanced image are in the same image coordinate system;

[0012] Obtain dual-mode registration images corresponding to the same detection scale region;

[0013] S3. Input the dual-mode registration image into the pre-trained dual-mode liquid level recognition model;

[0014] The texture features in the visible light enhanced image and the thermal radiation features in the infrared enhanced image are extracted using the dual-mode liquid level recognition model.

[0015] Based on the fusion result of the texture features and the thermal radiation features, the surface attachments of the detection ruler are identified in layers.

[0016] The locations of the slag layer boundary and the molten pool layer boundary are obtained;

[0017] S4. Determine the boundary length between the slag layer boundary position and the molten pool layer boundary position based on the slag layer boundary position and the molten pool layer boundary position;

[0018] in,

[0019] The boundary length characterizes the actual boundary length between the slag layer attached to the surface of the detection ruler and the molten pool layer.

[0020] S5. Obtain the length reference established before the detection ruler is inserted into the molten pool;

[0021] The descent depth between the end of the detection ruler and the boundary of the molten pool layer is determined based on the length reference.

[0022] in,

[0023] The descent depth characterizes the actual immersion depth of the end of the detection ruler relative to the surface of the molten pool.

[0024] S6. Perform a collaborative calculation of liquid level depth based on the boundary length and the descent depth;

[0025] Obtain the reference liquid level depth of the molten pool;

[0026] in,

[0027] The reference depth of the molten pool is the actual depth value between the molten pool surface and the end of the measuring ruler.

[0028] Preferably, the image enhancement processing performed on the visible light image and the infrared thermal image in step S2 specifically includes:

[0029] The visible light image is subjected to Gaussian filtering to remove random noise;

[0030] Gamma correction was applied to the image after Gaussian filtering to enhance the contrast of the surface texture of the detection ruler.

[0031] The infrared thermal image is subjected to non-uniformity correction processing to eliminate differences in the response of the infrared detector;

[0032] The corrected infrared thermal image is subjected to contrast-limited adaptive histogram equalization to enhance the thermal feature information of the temperature boundary region.

[0033] Preferably, the binocular heterogeneous vision calibration relationship in step S2 is established in the following manner:

[0034] Acquire the corresponding calibration images of the calibration board in the visible light camera and the infrared camera;

[0035] Extract the corresponding calibrated feature points;

[0036] Establish the homography transformation matrix between the infrared image coordinate system and the visible light image coordinate system based on the corresponding calibrated feature points;

[0037] The infrared enhanced image is mapped to the coordinate system of the visible light enhanced image using the homography transformation matrix.

[0038] Preferably, the dual-mode liquid level recognition model in step S3 is an improved YOLOv8 model;

[0039] The improved YOLOv8 model includes:

[0040] Visible light feature extraction branch used to extract visible light image features;

[0041] Infrared feature extraction branch used to extract features from infrared thermal images;

[0042] Feature fusion network for fusing features from different modalities;

[0043] And a detection head network used to output target detection results.

[0044] Preferably, the feature fusion network employs an attention-weighted fusion method;

[0045] Channel stitching is performed between visible light feature maps and infrared feature maps of the same scale;

[0046] Generate corresponding attention weights based on the intensity of the feature response;

[0047] The attention weights are used to weight the concatenated features;

[0048] Obtain the fused feature map.

[0049] Preferably, the fused feature map is input into the detection head network after being aggregated by PAN-FPN multi-scale features;

[0050] The detection head network outputs slag layer boundary detection frames and molten pool layer boundary detection frames;

[0051] The positions of the slag layer boundary and the molten pool layer boundary are determined based on the slag layer boundary detection frame and the molten pool layer boundary detection frame.

[0052] Preferably, the method for determining the boundary length in step S4 includes:

[0053] Obtain the boundary pixel distance between the slag layer boundary position and the molten pool layer boundary position;

[0054] Based on the ratio between the pre-defined pixel length and the actual length;

[0055] Convert the boundary pixel distance into the actual boundary length.

[0056] Preferably, the method for determining the descent depth in step S5 includes:

[0057] Obtain the reference mark position corresponding to the original length of the measuring ruler;

[0058] Determine the pixel distance between the boundary of the molten pool layer and the end of the detection ruler;

[0059] Based on the pre-established conversion relationship between pixel length and actual length;

[0060] The actual descent depth corresponding to the position of the end of the detection ruler relative to the boundary of the molten pool layer is obtained.

[0061] Preferably, the collaborative calculation of liquid level depth in step S6 specifically includes:

[0062] The boundary length and the descent depth are superimposed for calculation;

[0063] Obtain the reference liquid level depth of the molten pool between the molten pool surface and the end of the measuring ruler.

[0064] Preferably, after obtaining the reference liquid level depth of the molten pool, the method further includes:

[0065] Obtain the reference liquid level depth of the molten pool at multiple consecutive sampling times;

[0066] Construct a time series of liquid level depths;

[0067] The liquid level depth time series was subjected to median filtering.

[0068] Obtain the smoothed melt pool level result.

[0069] The molten pool level detection method for YOLOv8 proposed in this invention, which integrates visible and infrared dual-mode fusion, has the following beneficial technical effects:

[0070] This invention acquires visible light and infrared thermal images of the probe after it is removed from the probe. These two modal images are then registered and input into a dual-mode liquid level recognition model. The complementary information of visible light texture features and infrared thermal radiation features is used to identify the slag layer boundary and molten pool boundary on the probe surface. The boundary length is determined based on these boundary positions, and the liquid level depth is calculated collaboratively using the drop depth corresponding to the probe tip, thus obtaining the reference liquid level depth of the molten pool. Because a dual-mode fusion recognition mechanism using visible light and infrared is employed, the texture details of the visible light image and the temperature distribution information of the infrared thermal image are fully utilized, improving the stability and accuracy of boundary recognition under complex working conditions. Furthermore, by independently characterizing and collaboratively correcting the length of the slag layer attached to the probe surface during the liquid level calculation process, the influence of slag layer attachment on the mechanical probe measurement results is effectively eliminated. This solves the technical problems of existing mechanical probe methods' difficulty in accurately distinguishing the slag layer and molten pool boundary, and the insufficient stability of existing visual inspection methods in environments with smoke, dust, and strong light interference, achieving high-precision measurement of the molten pool liquid level depth. Attached Figure Description

[0071] Figure 1 This is a flowchart of the method of the present invention;

[0072] Figure 2 This is a schematic diagram showing the slag layer, molten pool layer, and visual recognition boundary on the surface of the measuring ruler after it has been removed.

[0073] Figure 3 A schematic diagram illustrating the geometric relationship between the descent depth and the boundary length for jointly calculating the reference liquid level depth. Detailed Implementation

[0074] Embodiments of the present invention are described in detail below. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar symbols denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0075] like Figure 1-3 The visible-infrared dual-mode fusion improved YOLOv8 molten pool level detection method shown includes the following steps:

[0076] Step S1: Simultaneously and in real time acquire visible light images and infrared thermal images of the molten pool area using an infrared camera and a visible light camera fixed above the slag outlet platform;

[0077] Step S2: Preprocess the visible light image and the infrared thermal image respectively, and register the preprocessed infrared thermal image to the visible light image coordinate system to form a registered image pair;

[0078] Step S3: Input the registered image pair into the pre-trained liquid level recognition model to identify the boundary between the slag layer and the molten pool layer on the surface of the detection ruler after it is taken out of the molten pool, and obtain the slag layer boundary line and the molten pool layer boundary line.

[0079] Step S4: Based on the identified slag layer boundary line and molten pool layer boundary line, and in conjunction with the pre-calibrated original length benchmark of the detection ruler, calculate the descent depth of the end of the detection ruler relative to the benchmark mark;

[0080] Step S5: Based on the pixel distance between the identified slag layer boundary line and the molten pool layer boundary line, calculate the actual boundary length between the slag layer and the molten pool layer on the detection ruler, and perform a joint calculation with the descent depth to obtain the reference liquid level depth of the molten pool.

[0081] Step S2 specifically includes: performing Gaussian filtering to denoise the visible light image and Gamma correction to enhance contrast; performing non-uniformity correction on the infrared thermal image and using contrast-limited adaptive histogram equalization to enhance edges; the registration is based on the homography matrix obtained by offline calibration, mapping the infrared thermal image pixel by pixel to the coordinates of the visible light image.

[0082] The liquid level recognition model described in step S3 is an improved YOLOv8 network. This network has a dual-stream backbone structure, including parallel visible light feature extraction branches and infrared feature extraction branches, which are used to extract visible light image features and infrared thermal image features, respectively. In the neck network, visible light feature maps and infrared feature maps of the same scale are concatenated in the channel dimension and weighted and fused through an attention mechanism. Then, multi-scale feature aggregation is performed through a PAN-FPN structure. Finally, the detection head outputs the target detection boxes of the slag layer boundary line and the molten pool layer boundary line.

[0083] The method for calculating the descent depth of the detection ruler end in step S4 is as follows: obtain the pixel length of the original length of the detection ruler in the image as a reference mark; compare the identified molten pool layer boundary line position with the reference mark to obtain the pixel distance from the molten pool layer boundary line to the end of the detection ruler; convert the pixel distance into the actual physical length according to the pre-calibrated pixel-to-actual distance conversion coefficient, which is the descent depth of the detection ruler end.

[0084] The specific calculation method for obtaining the reference liquid level depth of the molten pool in step S5 is as follows: add the boundary length between the slag layer boundary line and the molten pool layer boundary line to the descent depth of the end of the detection ruler to obtain the reference liquid level depth from the molten pool surface to the bottom of the detection ruler.

[0085] Step S5 is followed by: performing time-series median filtering smoothing on the calculated molten pool reference liquid level depth, and then outputting it to the host computer or control system via an industrial communication protocol.

[0086] This application has the following:

[0087] Eliminating slag adhesion error: Traditional mechanical measuring scales rely solely on the scale's descent distance, failing to distinguish the boundary between the slag layer adhering to the scale surface and the molten pool layer. This invention uses a visual recognition model to accurately identify the boundary between the slag layer and the molten pool layer on the removed measuring scale surface, quantifies the slag layer length separately, and calculates it collaboratively from the overall measurement, thus eliminating measurement deviations caused by slag adhesion.

[0088] Leveraging the advantages of dual-mode vision: By fusing visible light and infrared thermal images, visible light provides clear texture and boundary details, while infrared thermal images highlight the thermal boundary between the slag layer and the molten pool layer due to temperature differences. The dual-stream backbone network extracts specific features separately before attention fusion, achieving stable and accurate boundary recognition even under harsh conditions such as smoke, dust, and arc light.

[0089] Visual guidance and mechanical measurement complement each other: visual recognition is responsible for solving the problem of "inaccurate boundary measurement", while mechanical measuring rulers are responsible for providing "absolute depth reference". The two work together to locate and calculate, which not only retains the direct reliability of mechanical measurement, but also obtains the ability of visual recognition to analyze complex surface conditions, realizing the leap from a single method to multi-source fusion.

[0090] A non-contact-dominated, contact-assisted safety solution: In this invention, the visual camera system is kept at a safe distance above the molten pool throughout the process, and only the detection ruler is briefly immersed in the molten pool. The overall system has a small heat load, long equipment life, and low maintenance cost.

[0091] In an optional embodiment:

[0092] This embodiment uses copper smelting furnace liquid level detection as an application scenario. The system hardware includes: a camera bracket mounted above the slag outlet platform, with an infrared camera and a visible light camera fixed on the bracket, their optical axes parallel and perpendicularly aligned with the observation station after the detection ruler is removed. The mechanical detection ruler is driven by a servo motor and can be precisely inserted into or lifted out of the molten pool in a vertical direction. The surface of the detection ruler is sandblasted to enhance the adhesion and differentiation between the slag layer and the molten pool layer.

[0093] Step S1: Image Acquisition

[0094] The infrared and visible light cameras are synchronously triggered by an external signal. When the inspection ruler is fully removed from the molten pool and rests at the observation station, it simultaneously acquires visible light and infrared thermal images in real time. The acquisition frame rate is 30 frames per second, and the resolution is 1920×1080 pixels.

[0095] Step S2: Image preprocessing and registration

[0096] The visible light image is first filtered and denoised using a 3×3 Gaussian kernel, and then gamma correction with γ=1.2 is used to stretch the details in the dark areas, making the slag layer texture on the surface of the detection ruler clearer. The infrared thermal image first undergoes non-uniformity compensation based on two-point correction to eliminate fixed pattern noise caused by inconsistent detector pixel responses, and then contrast-limited adaptive histogram equalization (CLAHE) is used to enhance the thermal boundary contrast between the slag layer and the molten pool layer.

[0097] Step S3: Liquid Level Recognition Model Inference

[0098] Registered visible light image and infrared thermal images (In this embodiment, H=W=640) Input a pre-trained improved YOLOv8 liquid level recognition model. The pre-trained improved YOLOv8 liquid level recognition model is used as input. This model has a dual-flow backbone network and includes structurally parallel visible light feature extraction branches. and infrared feature extraction branch Both are based on the CSPDarknet architecture.

[0099] Visible light feature map of layer k (k=3,4,5) and infrared feature map Extract them respectively using the following formulas:

[0100]

[0101]

[0102] In the neck network, dual-mode feature maps of the same scale are concatenated along the channel dimension and then weighted and fused through a squeeze-excitation attention module.

[0103]

[0104] Visible light and infrared feature maps of the same scale are concatenated along the channel dimension. The SE attention module generates channel weights, highlighting feature channels that contribute significantly to boundary discrimination and suppressing noisy channels. The weighted fused feature map then enters the PAN-FPN structure for top-down and bottom-up multi-scale feature aggregation, enhancing the detection capability of boundaries at different scales. In the top-down path, high-level semantic features are upsampled and fused with low-level features.

[0105]

[0106] The detection head outputs two categories of target detection boxes: "slag layer boundary line" and "molten pool layer boundary line." Each detection box includes center coordinates, size, and confidence score. During model training, 5000 sets of dual-mode images were collected under different slag layer thicknesses, different adhesion morphologies, and different lighting and dust conditions. The ground truth boxes for the two boundary lines were manually labeled, and training was performed using Mosaic data augmentation and the CIoU loss function.

[0107] Step S4: Calculate the descent depth

[0108] Obtain the pixel length of the pre-calibrated original length of the detection ruler in the image as a reference marker. (The pixel length of the original length of the detection ruler in the image). The center ordinate of the detection box for the melt pool layer boundary output by the model is used as the location of the melt pool layer boundary. (The center ordinate of the detection box for the molten pool layer boundary line output by the improved YOLOv8 liquid level recognition model), the ordinate of the end of the detection ruler in the image is... (The vertical coordinate of the end of the detector ruler in the image). Then the pixel distance from the molten pool boundary line to the end of the detector ruler is...

[0109]

[0110] Using the pixel-to-actual distance conversion coefficient k (unit: mm / pixel) obtained from on-site calibration, the pixel distance is converted into the actual physical length: =k× This refers to the descent depth of the end of the measuring ruler. The calibration coefficient k is calculated by placing a standard gauge block of known length at the observation station of the measuring ruler and measuring its pixel length in the image.

[0111] Step S5: Collaboratively calculate the reference liquid level depth

[0112] Based on the slag layer boundary detection box and the molten pool layer boundary detection box output by the model, the ordinate of the slag layer boundary center is extracted respectively. and the ordinate of the center of the molten pool layer boundary The pixel distance between the two is .

[0113] The actual boundary length between the slag layer and the molten pool layer on the detection ruler is calculated by using the pixel-to-actual distance conversion coefficient k. =k× .

[0114] Molten pool reference liquid level depth This is the sum of the depth of the probe tip and the length of the slag layer boundary. This value represents the total depth from the bottom of the probe tip to the surface of the molten pool.

[0115] Step S5 Post-processing

[0116] Calculated from 20 consecutive frames Time-series median filtering is performed to eliminate jumps caused by false detections in a single frame. The filtered liquid level depth value is then transmitted in real time to the host DCS control system via the OPC-UA industrial communication protocol.

[0117] For clarification, "acquisition" in this application refers to obtaining the required content or data using existing technical means.

[0118] Furthermore, any content not described in detail in this specification is existing technology known to those skilled in the art.

[0119] In the embodiments provided by this invention, it should be understood that the disclosed system or method can be implemented in other ways. For example, the embodiments of the invention described above are merely illustrative; for instance, the division of modules is only a logical functional division, and there may be other division methods in actual implementation.

[0120] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0121] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or in the form of hardware plus software functional modules.

[0122] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the basic characteristics of the present invention.

[0123] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for detecting molten pool level in YOLOv8 using visible and infrared dual-mode fusion, characterized in that, Includes the following steps: S1. Obtain image data of the inspection ruler after it is inserted into the molten pool and then removed; The image data of the detection ruler includes a visible light image and an infrared thermal image corresponding to the same acquisition time; The visible light image is used to characterize the surface texture information of the detection ruler; The infrared thermal image is used to characterize the surface temperature distribution information of the detection ruler; S2. Perform image enhancement processing on the visible light image and the infrared thermal image respectively to obtain a visible light enhanced image and an infrared enhanced image; Based on the pre-established binocular heterogeneous vision calibration relationship, the infrared enhanced image is spatially mapped so that the infrared enhanced image and the visible light enhanced image are in the same image coordinate system; Obtain dual-mode registration images corresponding to the same detection scale region; S3. Input the dual-mode registered image into the pre-trained dual-mode liquid level recognition model; The texture features in the visible light enhanced image and the thermal radiation features in the infrared enhanced image are extracted using the dual-mode liquid level recognition model. Based on the fusion result of the texture features and the thermal radiation features, the surface attachments of the detection ruler are identified in layers. The locations of the slag layer boundary and the molten pool layer boundary are obtained; S4. Determine the boundary length between the slag layer boundary position and the molten pool layer boundary position based on the slag layer boundary position and the molten pool layer boundary position; in, The boundary length characterizes the actual boundary length between the slag layer attached to the surface of the detection ruler and the molten pool layer. S5. Obtain the length reference established before the detection ruler is inserted into the molten pool; The descent depth between the end of the detection ruler and the boundary of the molten pool layer is determined based on the length reference. in, The descent depth characterizes the actual immersion depth of the end of the detection ruler relative to the surface of the molten pool. S6. Perform a collaborative calculation of liquid level depth based on the boundary length and the descent depth; Obtain the reference liquid level depth of the molten pool; in, The reference depth of the molten pool is the actual depth value between the molten pool surface and the end of the measuring ruler.

2. The molten pool level detection method of the visible infrared dual-mode fusion improved YOLOv8 according to claim 1, characterized in that, Step S2, which involves performing image enhancement processing on the visible light image and the infrared thermal image respectively, specifically includes: The visible light image is subjected to Gaussian filtering to remove random noise; Gamma correction was applied to the image after Gaussian filtering to enhance the contrast of the surface texture of the detection ruler. The infrared thermal image is subjected to non-uniformity correction processing to eliminate differences in the response of the infrared detector; The corrected infrared thermal image is subjected to contrast-limited adaptive histogram equalization to enhance the thermal feature information of the temperature boundary region.

3. The molten pool level detection method of the visible infrared dual-mode fusion improved YOLOv8 according to claim 1, characterized in that, The binocular heterogeneous vision calibration relationship described in step S2 is established in the following way: Acquire the corresponding calibration images of the calibration board in the visible light camera and the infrared camera; Extract the corresponding calibrated feature points; Establish the homography transformation matrix between the infrared image coordinate system and the visible light image coordinate system based on the corresponding calibrated feature points; The infrared enhanced image is mapped to the coordinate system of the visible light enhanced image using the homography transformation matrix.

4. The molten pool level detection method of the visible infrared dual-mode fusion improved YOLOv8 according to claim 1, characterized in that, The dual-mode liquid level recognition model mentioned in step S3 is an improved YOLOv8 model; The improved YOLOv8 model includes: Visible light feature extraction branch used to extract visible light image features; Infrared feature extraction branch used to extract features from infrared thermal images; Feature fusion network for fusing features from different modalities; And a detection head network used to output target detection results.

5. The molten pool level detection method of the visible infrared dual-mode fusion improved YOLOv8 according to claim 4, characterized in that, The feature fusion network employs an attention-weighted fusion method. Channel stitching is performed between visible light feature maps and infrared feature maps of the same scale; Generate corresponding attention weights based on the intensity of the feature response; The attention weights are used to weight the concatenated features; Obtain the fused feature map.

6. The molten pool level detection method of the visible infrared dual-mode fusion improved YOLOv8 according to claim 5, characterized in that, The fused feature map is then input into the detection head network after being aggregated by PAN-FPN multi-scale features. The detection head network outputs slag layer boundary detection frames and molten pool layer boundary detection frames; The positions of the slag layer boundary and the molten pool layer boundary are determined based on the slag layer boundary detection frame and the molten pool layer boundary detection frame.

7. The molten pool level detection method of the visible infrared dual-mode fusion improved YOLOv8 according to claim 1, characterized in that, The method for determining the boundary length in step S4 includes: Obtain the boundary pixel distance between the slag layer boundary position and the molten pool layer boundary position; Based on the ratio between the pre-defined pixel length and the actual length; Convert the boundary pixel distance into the actual boundary length.

8. The molten pool level detection method of the visible infrared dual-mode fusion improved YOLOv8 according to claim 1, characterized in that, The method for determining the descent depth in step S5 includes: Obtain the reference mark position corresponding to the original length of the measuring ruler; Determine the pixel distance between the boundary of the molten pool layer and the end of the detection ruler; Based on the pre-established conversion relationship between pixel length and actual length; The actual descent depth corresponding to the position of the end of the detection ruler relative to the boundary of the molten pool layer is obtained.

9. The molten pool level detection method of the visible infrared dual-mode fusion improved YOLOv8 according to claim 1, characterized in that, The collaborative calculation of liquid level depth in step S6 specifically includes: The boundary length and the descent depth are superimposed for calculation; Obtain the reference liquid level depth of the molten pool between the molten pool surface and the end of the measuring ruler.

10. The molten pool level detection method of the visible infrared dual-mode fusion improved YOLOv8 according to claim 1, characterized in that, After obtaining the reference liquid level depth of the molten pool, the process further includes: Obtain the reference liquid level depth of the molten pool at multiple consecutive sampling times; Construct a time series of liquid level depths; The liquid level depth time series was subjected to median filtering. Obtain the smoothed molten pool level result.

Images